PASS Summit 2018 Evaluation Ratings & Comments

Following up on Brent Ozar’s post on the topic, I figured I’d post my own ratings (mostly because they’re not awful!).  This was my first PASS Summit at which I was a speaker, so I don’t have a good comp for scores except what other speakers publish.  I had the privilege of giving three presentations at PASS Summit this year, and I’m grateful for everyone who decided to sit in on these rather than some other talk.  All numeric responses are on a 1-5 scale with 5 being the best.

Applying Forensic Accounting Techniques Using SQL and R

This was a talk that I’ve given a few times and even have an extra-long director’s cut version available.  I had 71 attendees and 14 responses.

Eval Question
Avg Rating
Rate the value of the session content.
How useful and relevant is the session content to your job/career?
How well did the session’s track, audience, title, abstract, and level align with what was presented?
Rate the speaker’s knowledge of the subject matter.
Rate the overall presentation and delivery of the session content.
Rate the balance of educational content versus that of sales, marketing, and promotional subject matter.

The gist of the talk is, here are techniques that forensic accountants use to find fraud; you can use them to learn more about your data.  I fell flat in making that connection, as the low “useful” score shows.  That’s particularly bad because I think this is probably the most “immediately useful” talk that I did.

Event Logistics Comments

  • Room was very cold
  • very cold rooms, very aggressive air-con
  • The stage was squeaky and made banging noises when the speaker was trying to present. Not their fault! The stage just didn’t seem very stable. Also the room had a really unpleasant smell.
  • Everything was great!

The squeak was something I noticed before the talk.  I thought about staying in place to avoid the squeak, but this is a talk where I want to gesticulate to emphasize points—like moving from one side of the stage to the other to represent steps in a process.  My hope was that the squeak wouldn’t be too noticeable but the microphone may have picked it up.

Speaker Comments

  • I was just curious about the topic, but the speaker inspired me with many smaller, but very feasible tips and tricks of how to look at a data! Thank You!
  • The Jupyter notebooks were awesome. I felt the speaker really knew their stuff.  But the downsides were that the analysis methods discussed weren’t really shown to us, or were so far out of context I didn’t quite see how to use them or how they related to the demos. Multiple data sets were used and maybe just focusing all the methods on one of them may have worked better? I just felt overall it was a really interesting topic with a lot of work done but it just didn’t come together for me. Sorry.
  • I like understanding how fraud got uncovered by looking at data. Thanks
  • Very interesting session. Speak made the subject very interesting. I’ve picked up a few ideas I can use in my job.
  • Some examples of discovery of fraud would have been more effective.
  • One of my favorite sessions at PASS. Thank you for making the jupyter notebook available for download.
  • Great speaker/content, would attend again.

The second comment is exactly the kind of comment I want.  My ego loves the rest of the comments, but #2 makes me want to tear this talk apart and rebuild it better.  The biggest problem that I have with the talk is that my case study involved actual fraud, but none of the data sets I have really show fraud.  I’m thinking of rebuilding this talk using just one data set where I seed in fraudulent activities and expose it with various techniques.  Ideally, I’d get a copy of the case study’s data, but I never found it anywhere.  Maybe I could do a FOIA request or figure out some local government contact.

Getting Started with Apache Spark

My second session was Friday morning.  I had 100 somewhat-awake attendees but only 5 responses, so take the ratings with a grain of salt.

Eval Question
Avg Rating
Rate the value of the session content.
How useful and relevant is the session content to your job/career?
How well did the session’s track, audience, title, abstract, and level align with what was presented?
Rate the speaker’s knowledge of the subject matter.
Rate the overall presentation and delivery of the session content.
Rate the balance of educational content versus that of sales, marketing, and promotional subject matter.

This is a talk that I created specifically for PASS Summit.  I’m happy that it turned out well, considering that there was a good chance of complete demo failure:  my Portable Hadoop Cluster was finicky that morning and wanted to connect to the Internet to grab updates before it would let me run anything.  Then I had to restart the Apache Zeppelin service mid-talk to run any notebooks, but once that restarted successfully, the PHC ran like a champ.

Event Logistics Comments

  • Good

Speaker Comments

  • Session was 100, but I would say 200-300
  • Great presentation!

Getting rating levels right is always tricky.  In this case, I chose 100 rather than 200 because I spent the first 30+ minutes going through the history of Hadoop & Spark and a fair amount of the remaining time looking at Spark SQL.  But I did have a stretch where I get into RDD functions and most T-SQL developers will be unfamiliar with map, reduce, aggregate, and other functions.  So that’s a fair point—calling it a 200 level talk doesn’t bother me.

Cleaning is Half the Battle:  Launching a Data Science Project

This was my last PASS Summit talk, which I presented in the last session slot on Friday.  I had 31 attendees and 7 responses.

Eval Question
Avg Rating
Rate the value of the session content.
How useful and relevant is the session content to your job/career?
How well did the session’s track, audience, title, abstract, and level align with what was presented?
Rate the speaker’s knowledge of the subject matter.
Rate the overall presentation and delivery of the session content.
Rate the balance of educational content versus that of sales, marketing, and promotional subject matter.

Again, small sample size bias applies.

Event Logistics Comments

  • Good
  • Great

Speaker Comments

  • You have a lot content, leading to a rushed talk. Also, your jokes have potential, if u slowed down and sold them better
  • Good presentation. I expected more on getting the project off the ground, but enjoyed the info.
  • Funny and informative–I truly enjoyed your presentation!
  • Great
  • Kevin knows how to present, especially for getting stuck w/ the last session of the last day. He brought a lot of energy to the room. Content was on key too, he helped me understand more about handling data that we don’t want to model.

The slowing down comment is on point.  This is a 90-minute talk by its nature.  I did drop some content (like skipping slides on data cleansing and analysis and just showing the demos) so that I could spend a little more time on the neural network portion of the show, but I had to push to keep on time and technically went over by a minute or two.  I was okay with the overage because it was the final session, so I wasn’t going to block anybody.

Synthesis and Commentary

The ratings numbers are something to take with several grains of salt:  26 ratings over 3 sessions isn’t nearly a large enough sample to know for sure how these turned out.  But here are my thoughts from the speaker’s podium.

  • I speak fast.  I know it and embrace it—I know of the (generally good) advice that you want to go so slow that it feels painful, but I’ll never be that person.  In the Ben Stein — Billy Mays continuum, I’d rather err on the Oxyclean side.
  • I need to cut down on material.  In the first and last talks, they could both be better with less.  The problem with cutting material in the data science process talk is that I’d like to cover the whole process with a realistic-enough example and that takes time.  So this is something I’ll have to work on.
  • I might need to think of a different title for my data science process talk.  I explicitly call out that it’s about launching a data science project, but as I was sitting in a different session, I overheard a couple of people mention the talk and one person said something along the lines of not being interested because he’s already seen data cleansing talks.  The title is a bit jokey and has a punchline in the middle of the session, so I like it, but maybe something as simple as swapping the order of the segments to “Launching a Data Science Project:  Cleaning is Half the Battle” would be enough.
  • Using a timer was a really good idea.  I normally don’t use timers and instead go by feel at SQL Saturdays and user group meetings, and that leads to me sometimes running short on time.  I tend to practice talks with a timer to get an idea of how long they should last, but rarely re-time myself later, so they tend to shift in length as I do them.  Having the timer right in front of me helped keep me on track.
  • For the Spark talk, I think when I create my normal slide deck, I’m going to include the RDD (or “Spark 1.0”) examples as inline code segments and walk through them more carefully.  For an example of what I mean, I have a section in my Classification with Naive Bayes talk where I walk through classification of text based on word usage.  Normally, I’d make mention of the topic and go to a notebook where I walk through the code.  But that might have been a little jarring for people brand new to Spark.
  • I tend to have a paucity of images in talks, making up for it by drawing a lot on the screen.  I personally like the effect because action and animation keep people interested and it’s a lot easier for me to do that by drawing than by creating the animations myself…  It does come with the downside of making the slides a bit more turgid and making it harder for people to review the slides later as they lose some of that useful information.  As I’ve moved presentations to GitPitch I’ve focused on adding interesting but not too obtrusive backgrounds in the hopes that this helps.  Still, some of the stuff that I regularly draw should probably show up as images.

So it’s not perfect, but I didn’t have people hounding me with pitchforks and torches after any of the sessions.  I have some specific areas of focus and intend to take a closer look at most of my talks to improve them.


What Comes After Go-Live?

This is part eight of a series on launching a data science project.

At this point in the data science process, we’ve launched a product into production.  Now it’s time to kick back and hibernate for two months, right?  Yeah, about that…

Just because you’ve got your project in production doesn’t mean you’re done.  First of all, it’s important to keep checking the efficacy of your models.  Shift happens, where a model might have been good at one point in time but becomes progressively worse as circumstances change.  Some models are fairly stable, where they can last for years without significant modification; others have unstable underlying trends, to the point that you might need to retrain such a model continuously.  You might also find out that your training and testing data was not truly indicative of real-world data, especially that the real world is a lot messier than what you trained against.

The best way to guard against unbeknownst model shift is to take new production data and retrain the model.  This works best if you can keep track of your model’s predictions versus actual outcomes; that way, you can tell the actual efficacy of the model, figuring out how frequently and by how much your model was wrong.

Depending upon your choice of algorithm, you might be able to update the existing model with this new information in real time.  Models like neural networks and online passive-aggressive algorithms allow for continuous training, and when you’ve created a process which automatically feeds learned data back into your continuously-training model, you now have true machine learning. Other algorithms, however, require you to retrain from scratch.  That’s not a show-stopper by any means, particularly if your underlying trends are fairly stable.

Regardless of model selection, efficacy, and whether you get to call what you’ve done machine learning, you will want to confer with your stakeholders and ensure that your model actually fits their needs; as I mentioned before, you can have the world’s best regression, but if the people with the sacks of cash want a recommendation engine, you’re not getting the goods.  But that doesn’t mean you should try to solve all the problems all at once; instead, you want to start with a Minimum Viable Product (MVP) and gauge interest.  You’ve developed a model which solves the single most pressing need, and from there, you can make incremental improvements.  This could include relaxing some of the assumptions you made during initial model development, making more accurate predictions, improving the speed of your service, adding new functionality, or even using this as an intermediate engine to derive some other result.

Using our data platform survey results, assuming the key business personnel were fine with the core idea, some of the specific things we could do to improve our product would be:

  • Make the model more accurate.  Our MAE was about $19-20K, and reducing that error makes our model more useful for others.  One way to do this would be to survey more people.  What we have is a nice starting point, but there are too many gaps to go much deeper than a national level.
  • Introduce intra-regional cost of living.  We all know that $100K in Manhattan, NY and $100K in Manhattan, KS are quite different.  We would want to take into account cost of living, assuming we have enough data points to do this.
  • Use this as part of a product helping employers find the market rate for a new data professional, where we’d ask questions about the job location, relative skill levels, etc. and gin up a reasonable market estimate.

There are plenty of other things we could do over time to add value to our model, but I think that’s a nice stopping point.

What’s Old Is New Again

Once we get to this phase, the iterative nature of this process becomes clear.

The Team Data Science Project Lifecycle (Source)

On the micro level, we bounce around within and between steps in the process.  On the macro level, we iterate through this process over and over again as we develop and refine our models.  There’s a definite end game (mostly when the sacks of cash empty), but how long that takes and how many times you cycle through the process will depend upon how accurate and how useful your models are.

In wrapping up this series, if you want to learn more, check out my Links and Further Information on the topic.

Deploying A Model: The Microservice Approach

This is part seven of a series on launching a data science project.

Up to this point, we’ve worked out a model which answers important business questions.  Now our job is to get that model someplace where people can make good use of it.  That’s what today’s post is all about:  deploying a functional model.

Back in the day (by which I mean, say, a decade ago), one team would build a solution using an analytics language like R, SAS, Matlab, or whatever, but you’d almost never take that solution directly to production.  These were analytical Domain-Specific Languages with a set of assumptions that could work well for a single practitioner but wouldn’t scale to a broad solution.  For example, R had historically made use of a single CPU core and was full of memory leaks.  Those didn’t bother analysts too much because desktops tended to be single-core and you could always reboot the machine or restart R.  But that doesn’t work so well for a server—you need something more robust.

So instead of using the analytics DSL directly in production, you’d use it indirectly.  You’d use R (or SAS or whatever) to figure out the right algorithm and determine weights and construction and toss those values over the wall to an implementation team, which would rewrite your model in some other language like C.  The implementation team didn’t need to understand all of the intricacies of the problem, but did need to have enough practical statistics knowledge to understand what the researchers meant and translate their code to fast, efficient C (or C++ or Java or whatever).  In this post, we’ll look at a few changes that have led to a shift in deployment strategy, and then cover what this shift means for practitioners.

Production-Quality Languages

The first shift is the improvement in languages.  There are good libraries for Java, C#, and other “production” languages, so that’s a positive.  But that’s not one of the two positives I want to focus on today.  The first positive is the general improvement in analytical DSLs like R.  We’ve gone from R being not so great when running a business to being production-quality (although not without its foibles) over the past several years.  Revolution Analytics (now owned by Microsoft) played a nice-sized role in that, focusing on building a stable, production-ready environment with multi-core support.  The same goes for RStudio, another organization which has focused on making R more useful in the enterprise.

The other big positive is the introduction of Python as a key language for data science.  With libraries like NumPy, scikit-learn, and Pandas, you can build quality models.  And with Cython, a data scientist can compile those models down to C to make them much faster.  I think the general acceptance of Python in this space has helped spur on developers around other languages (whether open-source like R or closed-source commercial languages like SAS) to get better.

The Era Of The Microservice

The other big shift is a shift away from single, large services which try to solve all of the problems.  Instead, we’ve entered the era of the microservice:  a small service dedicated to providing a single answer to a single problem.  A microservice architecture lets us build smaller applications geared toward solving the domain problem rather than trying to solve the integration problem.  Although you can definitely configure other forms of interoperation, most microservices typically are exposed via web calls and that’s the scenario I’ll discuss today.  The biggest benefit to setting up a microservice this way is that I can write my service in R, you can call it from your Python service, and then some .NET service could call yours, and nobody cares about the particular languages used because they all speak over a common, known protocol.

One concern here is that you don’t want to waste your analysts time learning how to build web services, and that’s where data science workbenches and deployment tools like DeployR come into play.  These make it easier to deploy scalable predictive services, allowing practitioners to build their R scripts, push them to a service, and let that service host the models and turn function calls into API calls automatically.

But if you already have application development skills on your team, you can make use of other patterns.  Let me give two examples of patterns that my team has used to solve specific problems.

Machine Learning Services

The first pattern involves using SQL Server Machine Learning Services as the core engine.  We built a C# Web API which calls ML Services, passing in details on what we want to do (e.g., generate predictions for a specific set of inputs given an already-existing model).  A SQL Server stored procedure accepts the inputs and calls ML Services, which farms out the request to a service which understands how to execute R code.  The service returns results, which we interpret as a SQL Server result set, and we can pass that result set back up to C#, creating a return object for our users.

In this case, SQL Server is doing a lot of the heavy lifting, and that works well for a team with significant SQL Server experience.  This also works well if the input data lives on the same SQL Server instance, reducing data transit time.

APIs Everywhere

The second pattern that I’ll cover is a bit more complex.  We start once again with a C# Web API service.  On the opposite end, we’re using Keras in Python to make predictions against trained neural network models.  To link the two together, we have a couple more layers:  first, a Flask API (and Gunicorn as the production implementation).  Then, we stand nginx in front of it to handle load balancing.  The C# API makes requests to nginx, which feeds the request to Gunicorn, which runs the Keras code, returning results back up the chain.

So why have the C# service if we’ve already got nginx running?  That way I can cache prediction results (under the assumption that those results aren’t likely to change much given the same inputs) and integrate easily with the C#-heavy codebase in our environment.


If you don’t need to run something as part of an automated system, another deployment option is to use notebooks like JupyterZeppelin, or knitr.  These notebooks tend to work with a variety of languages and offer you the ability to integrate formatted text (often through Markdown), code, and images in the same document.  This makes them great for pedagogical purposes and for reviewing your work six months later, when you’ve forgotten all about it.

Using a Jupyter notebook to review Benford’s Law.

Interactive Visualization Products

Another good way of getting your data into users’ hands is Shiny, a package which lets you use Javascript libraries like D3 to visualize your data.  Again, this is not the type of technology you’d use to integrate with other services, but if you have information that you want to share directly with end users, it’s a great choice.


Over the course of this post, I’ve looked at a few different ways of getting model results and data into the hands of end users, whether via other services (like using the microservice deployment model) or directly (using notebooks or interactive applications).  For most scenarios, I think that we’re beyond the days of needing to have an implementation team rewrite models for production, and whether you’re using R or Python, there are good direct-to-production routes available.

The Basics Of Data Modeling

This is part five of a series on launching a data science project.

At this point, we have done some analysis and cleanup on a data set.  It might not be perfect, but it’s time for us to move on to the next step in the data science process:  modeling.

Modeling has five major steps, and we’ll look at each step in turn.  Remember that, like the rest of the process, I may talk about “steps” but these are iterative and you’ll bounce back and forth between them.

Feature Engineering

Feature engineering involves creating relevant features from raw data.  A few examples of feature engineering include:

  • Creating indicator flags, such as IsMinimumAge: Age >= 21, or IsManager: NumberOfEmployeesManaged > 0.  These are designed to help you slice observations and simplify model logic, particularly if you’re building something like a decision tree.
  • Calculations, such as ClickThroughRate = Clicks / Impressions.  Note that this definition doesn’t imply multicollinearity, though, as ClickThroughRate isn’t linearly related to either Clicks or Impressions.
  • Geocoding latitude and longitude from a street address.
  • Aggregating data.  That could be aggregation by day, by week, by hour, by 36-hour period, whatever.
  • Text processing:  turning words into arbitrary numbers for numeric analysis.  Common techniques for this include TF-IDF and word2vec.

Feature Selection

Once we’ve engineered interesting features, we want to use feature selection to winnow down the available set, removing redundant, unnecessary, or highly correlated features.  There are a few reasons that we want to perform feature selection:

  1. If one explanatory variable can predict another, we have multicollinearity, which can make it harder to give credit to the appropriate variable.
  2. Feature selection makes it easier for a human to understand the model by removing irrelevant or redundant features.
  3. We can perform more efficient training with fewer variables.
  4. We reduce the risk of an irrelevant or redundant feature causing spurious correlation.

For my favorite example of spurious correlation:

The only question here is, which causes which?

Model Training

Now that we have some data and a clue of what we’re going to feed into an algorithm, it’s time to step up our training regimen.  First up, we’re going to take some percentage of our total data and designate it for training and validation, leaving the remainder for evaluation (aka, test).  There are no hard rules on percentages, but a typical reserve rate is about 70-80% for training/validation and 20-30% for test.  We ideally want to select the data randomly but also include the relevant spreads and distributions of observations by pertinent variables in our training set; fortunately, there are tools available which can help us do just this, and we’ll look at them in a bit.

First up, though, I want to cover the four major branches of algorithms.

Supervised Learning

The vast majority of problems are supervised learning problems.  The idea behind a supervised learning problem is that we have some set of known answers (labels).  We then train a model to map input data to those labels in order to have the model predict the correct answer for unlabeled records.

Going back to the first post in this series, I pointed out that you have to listen to the questions people ask.  Here’s where that pays off:  the type of algorithm we want to choose depends in part on the nature of those questions.  Major supervised learning classes and their pertinent driving questions include:

  • Regression — How many / how much?
  • Classification — Which?
  • Recommendation — What next?

For example, in our salary survey, we have about 3000 labeled records:  3000(ish) cases where we know the salary in USD based on what people have reported.  My goal is to train a model which can then take some new person’s inputs and spit out a reasonable salary prediction.  Because my question is “How much money should we expect a data professional will make?” we will solve this using regression techniques.

Unsupervised Learning

With unsupervised learning, we do not know the answers beforehand, so we’re trying to derive answers within the data.  Typically, we’ll use unsupervised learning to gain more insight about the data set, which can hopefully give us some labels we can use to convert this into a relevant supervised learning problem.  The top forms of unsupervised learning include:

  • Clustering — How can we segment?
  • Dimensionality reduction — What of this data is useful?

Typically your business users won’t know or care about dimensionality reduction (that is, techniques like Principal Component Analysis) but we as analysts can use dimensionality reduction to narrow down on useful features.

Self-Supervised Learning

Wait, isn’t self-supervised learning just a subset of supervised learning?  Sure, but it’s pretty useful to look at on its own.  Here, we use heuristics to guesstimate labels and train the model based on those guesstimates.  For example, let’s say that we want to train a neural network or Markov chain generator to read the works of Shakespeare and generate beautiful prose for us.  The way the recursive model would work is to take what words have already been written and then predict the most likely next word or punctuation character.

We don’t have “labeled” data within the works of Shakespeare, though; instead, our training data’s “label” is the next word in the play or sonnet.  So we train our model based on the chains of words, treating the problem as interdependent rather than a bunch of independent words just hanging around.

Reinforcement Learning

Reinforcement learning is where we train an agent to observe its environment and use those environmental clues to make a decision.  For example, there’s a really cool video from SethBling about MariFlow:

The idea, if you don’t want to watch the video, is that he trained a recurrent neural network based on hours and hours of his Mario Kart gameplay.  The neural network has no clue what a Mario Kart is, but the screen elements below show how it represents the field of play and state of the game, and uses those inputs to determine which action to take next.

“No, mom, I’m playing this game strictly for research purposes!”

Choose An Algorithm

Once you understand the nature of the problem, you can choose the form of your destructor algorithm.  There are often several potential algorithms which can solve your problem, so you will want to try different algorithms and compare.  There are a few major trade-offs between algorithms, so each one will have some better or worse combination of the following features:

  • Accuracy and susceptibility to overfitting
  • Training time
  • Ability for a human to be able to understand the result
  • Number of hyperparameters
  • Number of features allowed.  For example, a model like ARIMA doesn’t give you many features—it’s just the label behavior over time.

Microsoft has a nice algorithm cheat sheet that I recommend checking out:

It is, of course, not comprehensive, but it does set you in the right direction.  For example, we already know that we want to predict values, and so we’re going into the Regression box in the bottom-left.  From there, we can see some of the trade-offs between different algorithms.  If we use linear regression, we get fast training, but the downside is that if our dependent variable is not a linear function of the independent variables, then we won’t end up with a good result.

By contrast, a neural network regression tends to be fairly accurate, but can take a long time to finish or require expensive hardware to finish in any reasonable time.

Once you have an algorithm, features, and labels (if this is a supervised learning problem), you can train the model.  Training a model involves solving a system of equations, minimizing a loss function.  For example, here is an example of a plot with a linear regression thrown in:

This plot might look familiar if you’re read my ggplot2 series.

In this chart, I have a straight line which represents the best fitting line for our data points, where best fit is defined as the line which minimizes the sum of the squares of errors (i.e., the sum of the square of the distance between the dot and our line).  Computers are great at this kind of math, so as long as we set up the problem the right way and tell the computer what we want it to do, it can give us back an answer.

But we’ve got to make sure it’s a good answer.  That’s where the next section helps.

Validate The Model

Instead of using up all of our data for training, we typically want to perform some level of validation within our training data set to ensure that we are on the right track and are not overfitting our model.  Overfitting happens when a model latches onto the particulars of a data set, leaving it at risk of not being able to generalize to new data.  The easiest way to tell if you are overfitting is to test your model against unseen data.  If there is a big dropoff in model accuracy between the training and testing phases, you are likely overfitting.

Here’s one technique for validation:  let’s say that we reserved 70% of our data for training.  Of the 70%, we might want to slice off 10% for validation, leaving 60% for actual training.  We feed the 60% of the data to our algorithm, generating a model.  Then we predict the outcomes for our validation data set and see how close we were to reality, and how far off the accuracy rates are for our validation set versus our training set.

Another technique is called cross-validation.  Cross-validation is a technique where we slice and dice the training data, training our model with different subsets of the total data.  The purpose here is to find a model which is fairly robust to the particulars of a subset of training data, thereby reducing the risk of overfitting.  Let’s say that we cross-validate with 4 slices.  In the first step, we train with the first 3/4 of the data, and then validate with the final 1/4.  In the second step, we train with slices 1, 2, and 4 and validate against slice 3.  In the third step, we train with 1, 3, and 4 and validate against slice 2.  Finally, we train with 2, 3, and 4 and validate against slice 1.  We’re looking to build up a model which is good at dealing with each of these scenarios, not just a model which is great at one of the four but terrible at the other three.

Often times, we won’t get everything perfect on the first try.  That’s when we move on to the next step.

Tune The Model

Most models have hyperparameters.  For example, a neural network has a few hyperparameters, including the number of training epochs, the number of layers, the density of each layer, and dropout rates.  For another example, random forests have hyperparameters like the maximum size of each decision tree and the total number of decision trees in the forest.

We tune our model’s hyperparameters using the validation data set.  With cross-validation, we’re hoping that our tuning will not accidentally lead us down the road to spurious correlation, but we have something a bit better than hope:  we have secret data.

Evaluate The Model

Model evaluation happens when we send new, never before seen data to the model.  Remember that 20-30% that we reserved early on?  This is where we use it.

Now, we want to be careful and make sure not to let any information leak into the training data.  That means that we want to split this data out before normalizing or aggregating the training data set, and then we want to apply those same rules to the test data set.  Otherwise, if we normalize the full data set and then split into training and test, a smart model can surreptitiously  learn things about the test data set’s distribution and could train toward that, leading to overfitting our model to the test data and leaving it less suited for the real world.

Another option, particularly useful for unlabeled or self-learning examples, is to build a fitness function to evaluate the model.  Genetic algorithms (for a refresher, check out my series) are a common tool for this.  For example, MarI/O uses a genetic algorithm to train a neural network how to play Super Mario World.

He’s no Evander Holyfield, but Mario’s worth a genetic algorithm too.


Just like with data processing, I’m going to split this into two parts.  Today, we’ve looked at some of the theory behind modeling.  Next time around, we’re going to implement a regression model to try to predict salaries.

Data Processing: An Example

This is part four of a series on launching a data science project.

An Example Of Data Processing

Last time around, I spent a lot of time talking about data acquisition, data cleansing, and basic data analysis.  Today, we’re going to walk through a little bit of it with the data professional salary survey.

First, let’s install some packages:

if(!require(tidyverse)) {
  install.packages("tidyverse", repos = "")

if(!require(XLConnect)) {
  install.packages("XLConnect", repos = "")

if(!require(caret)) {
  install.packages("caret", repos = "")

if(!require(recipes)) {
  install.packages("recipes", repos = "")

if(!require(data.table)) {
  install.packages("data.table", repos = "")

if(!require(devtools)) {
  install.packages("devtools", repos = "")

if(!require(keras)) {
  install_keras(method = "auto", conda = "auto", tensorflow = "default", extra_packages = NULL)

The tidyverse package is a series of incredibly useful libraries in R, and I can’t think of doing a data science project in R without it. The XLConnectpackage lets me read an Excel workbook easily and grab the salary data without much hassle. The caret library provides some helpful tooling for working with data, including splitting out test versus training data, like we’ll do below. The recipes package will be useful for normalizing data later, and we will use data.table to get a glimpse at some of our uneven data. We need the devtools package to install keras from GitHub. Keras is a deep learning library which implements several neural network libraries, including TensorFlow, which we will use later in this series. We need to install TensorFlow on our machine. Because this is a small data set, and because I want this to run on machines without powerful GPUs, I am using the CPU-based version of TensorFlow. Performance should still be adequate for our purposes.

Once we have the required packages loaded, we will then load the Excel workbook. I have verified the Excel worksheet and data region are correct, so we can grab the survey from the current directory and load it into salary_data.

wb <- XLConnect::loadWorkbook("2018_Data_Professional_Salary_Survey_Responses.xlsx")
salary_data <- XLConnect::readWorksheet(wb, sheet = "Salary Survey", region = "A4:Z6015")

We can use the glimpse function inside the tidyverse to get a quick idea of what our salary_data dataframe looks like. In total, we have 6011 observations of 26 variables, but this covers two survey years: 2017 and 2018. Looking at the variable names, we can see that there are some which don’t matter very much (like Timestamp, which is when the user filled out the form; and Counter, which is just a 1 for each record.

Observations: 6,011
Variables: 26
$ Survey.Year                <dbl> 2017, 2017, 2017, 2017, 2017, 2017, 2017...
$ Timestamp                  <dttm> 2017-01-05 05:10:20, 2017-01-05 05:26:2...
$ SalaryUSD                  <chr> "200000", "61515", "95000", "56000", "35...
$ Country                    <chr> "United States", "United Kingdom", "Germ...
$ PostalCode                 <chr> "Not Asked", "Not Asked", "Not Asked", "...
$ PrimaryDatabase            <chr> "Microsoft SQL Server", "Microsoft SQL S...
$ YearsWithThisDatabase      <dbl> 10, 15, 5, 6, 10, 15, 16, 4, 3, 8, 4, 22...
$ OtherDatabases             <chr> "MySQL/MariaDB", "Oracle, PostgreSQL", "...
$ EmploymentStatus           <chr> "Full time employee", "Full time employe...
$ JobTitle                   <chr> "DBA", "DBA", "Other", "DBA", "DBA", "DB...
$ ManageStaff                <chr> "No", "No", "Yes", "No", "No", "No", "No...
$ YearsWithThisTypeOfJob     <dbl> 5, 3, 25, 2, 10, 15, 11, 1, 2, 10, 4, 8,...
$ OtherPeopleOnYourTeam      <chr> "2", "1", "2", "None", "None", "None", "...
$ DatabaseServers            <dbl> 350, 40, 100, 500, 30, 101, 20, 25, 3, 5...
$ Education                  <chr> "Masters", "None (no degree completed)",...
$ EducationIsComputerRelated <chr> "No", "N/A", "Yes", "No", "Yes", "No", "...
$ Certifications             <chr> "Yes, and they're currently valid", "No,...
$ HoursWorkedPerWeek         <dbl> 45, 35, 45, 40, 40, 35, 40, 36, 40, 45, ...
$ TelecommuteDaysPerWeek     <chr> "1", "2", "None, or less than 1 day per ...
$ EmploymentSector           <chr> "Private business", "Private business", ...
$ LookingForAnotherJob       <chr> "Yes, but only passively (just curious)"...
$ CareerPlansThisYear        <chr> "Not Asked", "Not Asked", "Not Asked", "...
$ Gender                     <chr> "Not Asked", "Not Asked", "Not Asked", "...
$ OtherJobDuties             <chr> "Not Asked", "Not Asked", "Not Asked", "...
$ KindsOfTasksPerformed      <chr> "Not Asked", "Not Asked", "Not Asked", "...
$ Counter                    <dbl> 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1...

Our first data cleansing activity will be to filter our data to include just 2018 results, which gives us a sample size of 3,113 participants. There are also results for 2017, but they asked a different set of questions and we don’t want to complicate the analysis or strip out the new 2018 questions.

survey_2018 <- filter(salary_data, Survey.Year == 2018)
nrow(survey_2018) # << 3113 records returned

Looking at the survey, there are some interesting data points that we want:

  • SalaryUSD (our label, that is, what we are going to try to predict)
  • Country
  • YearsWithThisDatabase
  • EmploymentStatus
  • JobTitle
  • ManageStaff
  • YearsWithThisTypeOfJob
  • OtherPeopleOnYourTeam
  • DatabaseServers
  • Education
  • EducationIsComputerRelated
  • Certifications
  • HoursWorkedPerWeek
  • TelecommuteDaysPerWeek
  • EmploymentSector
  • LookingForAnotherJob
  • CareerPlansThisYear
  • Gender

For each of these variables, we want to see the range of options and perform any necessary cleanup. The first thing I’d look at is the cardinality of each variable, followed by a detailed anlaysis of the smaller ones.

PrimaryDatabase is another variable which looks interesting, but it skews so heavily toward SQL Server that there’s more noise than signal to it. Because there are so many platforms with 10 or fewer entries and about 92% of entrants selected SQL Server, we’ll throw it out.

rapply(survey_2018, function(x) { length(unique(x)) })

Survey.Year – 1
Timestamp – 3112
SalaryUSD – 865
Country – 73
PostalCode – 1947
[… continue for a while]

  1. ‘United States’
  2. ‘Australia’
  3. ‘Spain’
  4. ‘United Kingdom’
    [… continue for a while]


  1. ‘Full time employee’
  2. ‘Full time employee of a consulting/contracting company’
  3. ‘Independent consultant, contractor, freelancer, or company owner’
  4. ‘Part time’

We can use the setDT function on data.table to see just how many records we have for each level of a particular factor. For example, we can see the different entries for PrimaryDatabase and EmploymentSector below. Both of these are troublesome for our modeling because they both have a number of levels with 1-2 entries. This makes it likely that we will fail to collect a relevant record in our training data set, and that will mess up our model later. To rectify this, I am going to remove PrimaryDatabase as a feature and remove the two students from our sample.

data.table::setDT(survey_2018)[, .N, keyby=PrimaryDatabase]

To the three MongoDB users: you have my sympathy.

data.table::setDT(survey_2018)[, .N, keyby=EmploymentSector]

In a way, aren’t we all students? No. Only two of us are.

Most of these columns came from dropdown lists, so they’re already fairly clean. But there are some exceptions to the rule. They are:

  • SalaryUSD
  • YearsWithThisDatabase
  • YearsWithThisTypeOfJob
  • DatabaseServers
  • HoursWorkedPerWeek
  • Gender

All of these were text fields, and whenever a user gets to enter text, you can assume that something will go wrong. For example:

survey_2018 %>%
  distinct(YearsWithThisDatabase) %>%
  arrange(desc(YearsWithThisDatabase)) %>%

Some are older than they seem.

Someone with 53,716 years working with their primary database of choice? That’s commitment! You can also see a couple of people who clearly put in the year they started rather than the number of years working with it, and someone who maybe meant 10 years? But who knows, people type in weird stuff.

Anyhow, let’s see how much that person with at least 10 thousand years of experience makes:

survey_2018 %>%
  filter(YearsWithThisDatabase > 10000)

Experience doesn’t pay after the first century or two.

That’s pretty sad, considering their millennia of work experience. $95-98K isn’t even that great a number.

Looking at years of experience with their current job roles, people tend to be more reasonable:

survey_2018 %>%
  distinct(YearsWithThisTypeOfJob) %>%
  arrange(desc(YearsWithThisTypeOfJob)) %>%

Next up, we want to look at the number of database servers owned. 500,000+ database servers is a bit excessive. Frankly, I’m suspicious about any numbers greater than 5000, but because I can’t prove it otherwise, I’ll leave them be.

survey_2018 %>%
  distinct(DatabaseServers) %>%
  arrange(desc(DatabaseServers)) %>%

survey_2018 %>%
  filter(DatabaseServers >= 5000) %>%

500K servers is a lot of servers.

The first entry looks like bogus data: a $650K salary, a matching postal code, and 500K database servers, primarily in RDS? Nope, I don’t buy it.

The rest don’t really look out of place, except that I think they put in the number of databases and not servers. For these entrants, I’ll change the number of servers to the median to avoid distorting things.

Now let’s look at hours per week:

survey_2018 %>%
  distinct(HoursWorkedPerWeek) %>%
  arrange(desc(HoursWorkedPerWeek)) %>%

One of these numbers is not like the others.  The rest of them are just bad.

To the person who works 200 hours per week: find a new job. Your ability to pack more than 7*24 hours of work into 7 days is too good to waste on a job making just $120K per year.

survey_2018 %>%
  filter(HoursWorkedPerWeek >= 168) %>%

What would I do with an extra day and a half per week? Sleep approximately an extra day and a half per week.

As far as Gender goes, there are only three with enough records to be significant: Male, Female, and Prefer not to say. We’ll take Male and Female and bundle the rest under “Other” to get a small but not entirely insignificant set there.

survey_2018 %>%
  group_by(Gender) %>%
  summarize(n = n())

To the one Reptilian in the survey, I see you and I will join forces with Rowdy Roddy Piper to prevent you from taking over our government.

survey_2018 %>%
  group_by(Country) %>%
  summarize(n = n()) %>%
  filter(n >= 20)

Probably the most surprising country on this list is The Netherlands.  India is a close second, but for the opposite reason.

There are only fifteen countries with at least 20 data points and just eight with at least 30. This means that we won’t get a great amount of information from cross-country comparisons outside of the sample. Frankly, I might want to limit this to just the US, UK, Canada, and Australia, as the rest are marginal, but for this survey analysis, I’ll keep the other eleven.

Building Our Cleaned-Up Data Set

Now that we’ve performed some basic analysis, we will clean up the data set. I’m doing most of the cleanup in a single operation, but I do have some comment notes here, particularly around the oddities with SalaryUSD. The SalaryUSD column has a few problems:

  • Some people put in pennies, which aren’t really that important at the level we’re discussing. I want to strip them out.
  • Some people put in delimiters like commas or decimal points (which act as commas in countries like Germany). I want to strip them out, particularly because the decimal point might interfere with my analysis, turning 100.000 to $100 instead of $100K.
  • Some people included the dollar sign, so remove that, as well as any spaces.

It’s not a perfect regex, but it did seem to fix the problems in this data set at least.

valid_countries <- survey_2018 %>%
                    group_by(Country) %>%
                    summarize(n = n()) %>%
                    filter(n >= 20)

# Data cleanup
survey_2018 <- salary_data %>%
  filter(Survey.Year == 2018) %>%
  filter(HoursWorkedPerWeek < 200) %>%
  # There were only two students in the survey, so we will exclude them here.
  filter(EmploymentSector != "Student") %>%
  inner_join(valid_countries, by="Country") %>%
    SalaryUSD = stringr::str_replace_all(SalaryUSD, "\\$", "") %>%
      stringr::str_replace_all(., ",", "") %>%
      stringr::str_replace_all(., " ", "") %>%
      # Some people put in pennies.  Let's remove anything with a decimal point and then two numbers.
      stringr::str_replace_all(., stringr::regex("\\.[0-9]{2}$"), "") %>%
      # Now any decimal points remaining are formatting characters.
      stringr::str_replace_all(., "\\.", "") %>%
    # Some people have entered bad values here, so set them to the median.
    YearsWithThisDatabase = case_when(
      (YearsWithThisDatabase > 32) ~ median(YearsWithThisDatabase),
      TRUE ~ YearsWithThisDatabase
    # Some people apparently entered number of databases rather than number of servers.
    DatabaseServers = case_when(
      (DatabaseServers >= 5000) ~ median(DatabaseServers),
      TRUE ~ DatabaseServers
    EmploymentStatus = as.factor(EmploymentStatus),
    JobTitle = as.factor(JobTitle),
    ManageStaff = as.factor(ManageStaff),
    OtherPeopleOnYourTeam = as.factor(OtherPeopleOnYourTeam),
    Education = as.factor(Education),
    EducationIsComputerRelated = as.factor(EducationIsComputerRelated),
    Certifications = as.factor(Certifications),
    TelecommuteDaysPerWeek = as.factor(TelecommuteDaysPerWeek),
    EmploymentSector = as.factor(EmploymentSector),
    LookingForAnotherJob = as.factor(LookingForAnotherJob),
    CareerPlansThisYear = as.factor(CareerPlansThisYear),
    Gender = as.factor(case_when(
      (Gender == "Male") ~ "Male",
      (Gender == "Female") ~ "Female",
      TRUE ~ "Other"

Now we can pare out variables we don’t need. Some of these, like postal code, are interesting but we just don’t have enough data for it to make sense. Others, like Kinds of Tasks Performed or Other Job Duties, have too many varieties for us to make much sense with a first pass. They might be interesting in a subsequent analysis, though.

survey_2018 <- survey_2018 %>%
  # One person had a salary of zero.  That's just not right.
  filter(SalaryUSD > 0) %>%
  select(-Counter, -KindsOfTasksPerformed, -OtherJobDuties, -OtherDatabases, -Timestamp, -Survey.Year, 
         -PostalCode, -n, -PrimaryDatabase)

Now that we have our salary data fixed, we can finally look at outliers. I’d consider a salary of $500K a year to be a bit weird for this field. It’s not impossible, but I am a little suspicious. I am very suspicious of the part-timer making $1.375 million, the federal employee making $1 million, or the New Zealander making $630K at a non-profit.

I’m kind of taking a risk by removing these, but they’re big enough outliers that they can have a real impact on our analysis if they’re bad data.

survey_2018 %>%
  filter(SalaryUSD > 500000) %>%

I think I’d be willing to accept $1.4 million a year to be a manager of none.

On the other side, there are 12 people who say they earned less than $5K a year. Those also seem wrong. Some of them look like dollars per hour, and maybe some are monthly salary. I’m going to strip those out.

survey_2018 %>%
  filter(SalaryUSD < 5000) %>%

For just over a dollar a week, you can hire a data architect.

survey_2018 <- filter(survey_2018, SalaryUSD >= 5000 & SalaryUSD <= 500000)

Data Analysis

We did some of the data analysis up above. We can do additional visualization and correlation studies. For example, let’s look at a quick distribution of salaries after our cleanup work:

   Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
   5000   70000   92000   95186  115000  486000

We can also build a histogram pretty easily using the ggplot2 library. This shows the big clump of database professionals earning beween $70K and $115K per year. This salary distribution does skew right a bit, as you can see.

ggplot(data = survey_2018, mapping = aes(x = SalaryUSD)) +
  geom_histogram() +
  theme_minimal() +
  scale_x_log10(label = scales::dollar)

Not including that guy making $58 a year.

We can also break this down to look by primary job title, though I’ll limit to a couple of summaries instead of showing a full picture.

survey_2018 %>% filter(JobTitle == "Data Scientist") %>% select(SalaryUSD) %>% summary(.)
 Min.   : 45000  
 1st Qu.: 76250  
 Median :111000  
 Mean   :102000  
 3rd Qu.:122000  
 Max.   :160000
survey_2018 %>% filter(JobTitle == "Developer: App code (C#, JS, etc)") %>% select(SalaryUSD) %>% summary(.)
 Min.   : 22000  
 1st Qu.: 60000  
 Median : 84000  
 Mean   : 84341  
 3rd Qu.:105000  
 Max.   :194000
survey_2018 %>% filter(JobTitle == "Developer: T-SQL") %>% select(SalaryUSD) %>% summary(.)
 Min.   : 12000  
 1st Qu.: 66000  
 Median : 87000  
 Mean   : 88026  
 3rd Qu.:110000  
 Max.   :300000

This fit pretty well to my biases, although the max Data Scientist salary seems rather low.


This is only a tiny sample of what I’d want to do with a real data set, but it gives you an idea of the kinds of things we look at and the kinds of things we need to fix before a data set becomes useful.

In the next post, we will get started with the wide world of modeling.

Data Processing: The Other 90%

This is part three of a series on launching a data science project.

The Three Steps Of Data Processing

Data processing is made up of a few different activities:  data gathering, data cleansing, and data analysis.  Most estimates are that data scientists spend about 80% of their time in data processing (particularly in data cleansing).  If anything, I consider this an underestimation:  based on my experiences, I would probably put that number closer to 90%.

As with everything else, these separate concepts tend to jumble together and you’ll bounce back and forth between them as you learn about new resources, determine the viability of sources, and try to link data sources together.  But for the purposes of elucidation, we’ll treat them independently.  So let’s start with data gathering.

Data Gathering

At first, you might not know exactly what data you need, but as you flesh out your models and gain a better understanding of the problem, you will go back to the well several times and probably end up getting data from a variety of sourses.  These sources can include, but are not limited to, the following:

  • Internal proprietary data
  • Open data sources (often governmental or academic, but can also be private sources like our data professional survey)
  • Paid APIs or data sources from third parties
  • Data from surveys that you commission

Data Gathering In Our Example

For our example, we will stick to using just the data professional survey.  But if we wanted to expand our analysis out a bit, there are some places we could take it.  For example, we could use the Penn World Table and calculate Purchasing Power Parity GDP per capita to normalize salaries across countries.  This could help us find if there were countries in which database professionals were relatively higher-paid compared to the norm, and is pretty similar to work I did last year.

We could get a geocoding data set to visualize results on a map, or get a data set which has cost of living by ZIP Code.  This would help us normalize salaries within the United States, as long as the ZIP Code data in our set is good enough (spoilers:  it’s not).  We could also use census information to build out more data by ZIP Code or greater metropolitan area.  Finally, we could use data from other surveys to add more information to the sample.  Our sample is pretty small, at just 6000 entries across two years, so supplementing this data could help us a lot.

Data Cleansing

Once we have relevant (or at least potentially relevant) data sets, we probably will want to join them together to gain insight from the mashup of these data sets.  It is rare, however, to have all of your data sets match up exactly on the first try.  Instead, we’re going to have to work with those data sets to get the puzzle pieces to fit together.

Soft Corinthian Data With A Beautiful Grain

For example, we probably need to change the grain of some of our data sets.  Suppose have a data set which is at the daily grain and another which is at the monthly grain.  We can connect these two together, but we have to make a decision.  Do we take the daily grain and aggregate it up to the monthly level?  Do we take the monthly level and create some kind of daily allocation rule?  Do we figure out the month to which the date belongs and include month-level aggregations for each of those daily values?  Each of these is a viable strategy, but has different implications on what you can do with the data.

Next up, you might have to define join criteria.  There won’t always be an obvious natural join key like a date.  You might have a data set from one source which contains product titles and manufacturer codes, and another data set which contains UPCs and gussied up titles.  In this case, you might be able to map manufacturer codes to UPCs using a third data set, or maybe you can write some regular expressions to get the product titles to match up, but there’s work involved.

We may also need to reshape the data.  As mentioned above with grains, we might need to aggregate or disaggregate data to fit a join.  We may also need to perform operations like pivoting or unpivoting data (or gathering and spreading using the tidyverse parlance).

Dealing With Mislabels, Mismatches, And Just Plain Wrong Data

This next category of data cleansing has to do with specific values.  I want to look at three particular sub-categories:  mislabeled data, mismatched data, and incorrect data.

Mislabeled data happens when the label is incorrect.  In a data science problem, the label is the thing that we are trying to explain or predict.  For example, in our data set, we want to predict SalaryUSD based on various inputs.  If somebody earns $50,000 per year but accidentally types 500000 instead of 50000, it can potentially affect our analysis.  If you can fix the label, this data becomes useful again, but if you cannot, it increases the error, which means we have a marginally lower capability for accurate prediction.

Mismatched data happens when we join together data from sources which should not have been joined together.  Let’s go back to the product title and UPC/MFC example.  As we fuss with the data to try to join together these two data sets, we might accidentally write a rule which joins a product + UPC to the wrong product + MFC.  We might be able to notice this with careful observation, but if we let it through, then we will once again thwart reality and introduce some additional error into our analysis.  We could also end up with the opposite problem, where we have missed connections and potentially drop useful data out of our sample.

Finally, I’m calling incorrect data where something other than the label is wrong.  For example, in the data professional salary survey, there’s a person who works 200 hours per week.  While I admire this person’s dedication and ability to create 1.25 extra days per week that the rest of us don’t experience, I think that person should have held out for more than just $95K/year.  I mean, if I had the ability to generate spare days, I’d want way more than that.

Missing Data

People don’t always fill out the entirety of every form, and when you make fields optional, people are liable not to fill them in.  The flipside to this is that if you make optional fields required, people might input junk data just to get the form to submit, so simply marking fields as required isn’t the answer.

When we are missing data, we have a few options available to us.  First of all, if we’re looking at a small percentage of the total, we might just get rid of those records with important, missing data.  Once we start removing more than 2-3% of the data set for this reason, though, I’m liable to get a bit edgy about removing more.  If we’re missing information from non-vital columns like middle initial or address line 3, I’d just remove those features altogether and save the hassle.

Otherwise, if we can’t remove the feature and don’t want to remove the row, we can use statistical techniques to fill in the blanks.  We might substitute missing values with a dummy value, such as “Unknown” or a reasonable default like 0 (assuming 0 is a reasonable default in this case).  We might substitute with an “average” value like mean, median, or mode.  The default here is probably median, but there are times when you want to use mean or mode.  We could also do a quick regression and replace our missing value with the regressed value, assuming that there’s some correlation (positive or negative) between the filled-in features and our missing feature.

These statistical techniques can help fill in gaps, but I wouldn’t use them for more than a few percent of the data; if you’re seeing 15+ percent of the feature values missing, this is probably a compromised feature and you don’t want to use it in analysis.  Also, by applying these statistical techniques, we are knowingly introducing noise, making our models that much less accurate.  That is, we are saying that the value is X, but it’s probably somewhere in the range { X-y, X+y} and we are giving undue precision based on our own made-up calculations in the hopes that y is small enough that it won’t matter too much.  If this hope turns out to be the case, the benefit to salvaging incomplete records could overcome the cost of adding more noise to our analysis.

Inconsistent Data

As I mentioned earlier in the series, one of my data adages is that independent systems tend to end up with inconsistent data.  Sometimes people forget to fill out both systems the same way.  Sometimes there are typos in one data source but not the other (or the typos are different in each).  Sometimes there are subtle differences in data sets which lead to differing results.  Sometimes one data source is newer than the other, or at least one particular record is newer than what’s in the other system.

Regardless of the reason, we need to be able to handle these differences.  One way to handle differences is to make one data source canonical and trust it over the other when there comes time for a discrepancy.  If there’s an authoritative source, it makes sense to behave this way, and your business team might be able to explain which sources (if any) are authoritative.

More commonly, you will have to institute one or more rules to handle data discrepancies.  For example, if both records have a timestamp, pick the later timestamp…although wall clocks lie.  Another example of a rule might be, if the discrepancy is over a numeric value, go with the lower number as it is more likely to be correct.  These rules will be specific to your data sources and systems, and will often require quite a bit of sussing out.

Misshapen Data

If you happen to have data stored in a textual format, you can end up with misshapen data.  For example, you might have XML or JSON data where some of the records are in an invalid format.  Or maybe you have newlines in the middle of a record.  Maybe you have too many delimiters in that CSV, or maybe not enough.  Maybe the file starts as a CSV and then moves to pipe-delimited in the middle because the source system appended data over time using different output mechanisms.

Regular expressions can help out here sometimes.  You might also be able to get satisfaction from the owner of the source data and have them fix the problem.  Sadly, it’s just as likely that your team becomes the owner of the data and gets the pleasure of figuring it out.  If you store data in relational databases and adhere to first normal form (or better), data shape is already enforced, so you can pat yourself on the back for this one and move on to the next problem.

Data Shaping

Speaking of data shapes, we often need to perform some kinds of data shaping activities to turn our raw data into something useful for modeling.  Examples of data shaping include:

  • Vectorizing words.  That is, turning words in a document into numbers for analysis.
  • Categorizing data.  That is, turning strings (or maybe numeric values) into factors.
  • Normalizing values.  That is, transforming numeric data to have a mean of 0 and a standard deviation of 1.  The benefit of normalization is that you can talk about two features which have radically different scales in the context of the same model.  Let’s say that we want to measure a person’s desire to purchase a car based on its gas mileage and its sticker price.  Gas mileage typically ranges from about 15-50, but sticker price can range from $9000-90,000 (or more).  Both the sheer numbers and the wide range of sticker price will overwhelm gas mileage in many models, so the way we get around this is to normalize both features so that each has a mean of 0 and standard deviation of 1.  That way, one feature’s effect cannot swamp another’s.
  • Binning data.  That is, converting continuous values to discrete.  For example, we might have the age of each person in our data set, but we’re more interested in age cohorts, like 18-25 or 30-34.

Data Analysis

As part of building a data dictionary, I’ve already brought up a couple simple data analysis concepts.  Here, I’ll go into a few, focusing on techniques common in Exploratory Data Analysis (EDA).  The goal with EDA is to summarize and visualize the data in a way that furthers your understanding of the data.

Cardinality Analysis

The first simple technique you can use is finding the cardinality of variables.  For example, in the salary survey, there were four options available for Employment Status.

Not included: use multiverse versions of self to do work, thereby freeing up time to lounge on a beach in Cabo.

Cardinality tells us a few things.  For a text field like employment status, it indicates that we probably got the data from a proscribed list—often a dropdown list or set of radio buttons—and we probably want to treat this as categorical data, as people choose one to the exclusion of others.

For free-form entry, like number of hours worked, looking at the cardinality lets us see how many different values people chose, and if you compare the number selected versus the range from minimum to maximum, you can get a fair idea of how packed this set is.

Summarization And Visualization

Next up, we can run a five-number summary of a feature.

The 5-Number Summary: buy five numbers, get one free.

This summary, which is easy to perform in R, gives us five pertinent values:  the minimum, maximum, 25th percentile, 75th percentile, and 50th percentile (i.e., median).  R also gives you the mean, which isn’t technically part of the five-number summary but is helpful, as the difference between median and mean clues you into how skewed the distribution of your sample data is.

You can also visualize these five-number summaries with a box plot.

A box plot.  Buy five boxes, get one free?

This type of plot will show you min, max, 25th, 50th, and 75th percentiles, where 25-50-75 make up the box and min and max are the furthest points in the plot.  This plot also shows you “expected” values (those within 1.5 * [75th percentile – 25th percentile], shifted to the next data point in each direction), as well as outliers—those which fall outside the interquartile range.  Box plots are great at visualizing differences by category.

When looking at the distribution of data within a category, you can easily build a histogram.

A histogram using base-10 logged values because if you can’t do exponential math in your head, get out of my classroom.

This distribution visualization lets us see the skew—in this case, our plot skews left, meaning that the left-hand side of the median is more filled out than the right-hand side.  This is fairly common when looking at income data, so I wasn’t surprised to see this setup when I created the above graph using the 2017 survey data.

Correlation Checking

Another part of data analysis involves testing for correlation between explanatory variables.  You want correlation between your independent (explanatory) variables and your dependent variable—that’s what data analysis is all about, after all—but having tightly related explanatory variables can lead to analytical problems down the road.  This correlation is called multicollinearity and is one of the bugbears of linear regression analysis.

As an example of this, suppose that we want to regress plant growth rates using two variables:  amount of rainfall in inches and amount of precipitation in centimeters.  There is collinearity between rainfall in inches and precipitation in inches:  precipitation is rainfall + snowfall, so we can easily get from one to the other and their movements over time will be linked.  If we use both of these variables to try to model plant growth, it becomes harder for the algorithm we choose to differentiate between the two variables and assign them the correct rate.

The cor() function in R gives us correlation between two variables.

Diamond dimension correlations.

In this case, depth and table have a mild, negative correlation.  By contrast, the x and y dimensions are highly correlated.  If I were performing a linear regression, I would have no qualms about including both depth and table, but would think twice about including both x and y.

Conclusion (For Now)

In the next post, I will perform some data cleansing & analysis work on the survey, as this post is long enough as-is.  As mentioned, this is by far the longest phase of any data science project; what I’ve touched on in this post is just a brief smattering of the things you should expect to try when working through a project.

Building Business Understanding

This is part two of a series on launching a data science project.

How Is Babby Data Science Project Formed?

Behind each data science project, there is (hopefully) someone higher up on the business side who wants it done.  This person might have been the visionary behind this project or might simply be the sponsor who drives it because of the project’s potential value.  Nevertheless, the data science team needs to seek out and gather as much information about that champion’s vision as possible.  In a perfect scenario, this is the person handing out sacks of cash to you, and you want those sacks of cash because they buy pretty hardware and let you bring in smart people to work with (or for) you.  You might even have several people interested in your project; if so, you’ll want to build a composite vision, one which hopefully includes all of the members’ visions.  Just keep in mind that sometimes you can’t combine everybody’s dreams and get a coherent outcome, so you’ll need to drive the champion(s) toward clarity.  Forthwith are a few clues to help.

Learn The Domain

The first clue is figuring out the domain.  This is asking questions like what the company does, what other companies in the industry do, what kind of jargon people use, etc.  The more you know, the better prepared you are to understand the mind(s) of your champion(s).  But even if you’ve been at that company for a long time and have a detailed understanding of the business, you still want to interview your champion(s) and ask questions which expose the ideal outcome.

Stop, Collaborate, and Listen

The second clue is simple:  listen.  When interviewing people, listen for the following types of questions:

  • How much / how many?
  • Which category does this belong to?
  • How can we segment this data?
  • Is this weird?
  • Which option should I choose?

Each of these is a different type of problem with its own set of statistical techniques and rules.  For example, the “Is this weird?” question relates to anomaly detection:  finding outliers or “weird” results.  Figuring out which of these types of questions is most important to your champion is crucial to delivering the right product.  You can build the absolute best regression ever, but if the person was expecting a recommendation engine, you’re going to disappoint.

As you listen to these types of questions, your goal is to nail down a specific problem with a specific answer.  You want to narrow down the scope to something that your team can achieve, ideally something with a built-in measure for success.  For example, here are a few specific problems that we could go solve:

  • Find a model which predicts quarterly sales to within 5% no later than 30 days into the quarter.
  • Given a title and description for a product, tell me a listing category which Amazon will, with at least 90% confidence, consider valid for this product.
  • Determine the top three factors which most affect the number of years the first owner holds onto our mid-range sedan.

With a specific problem in mind, you can look for relevant data.  Of course, you’ll probably need to modify the scope of this problem over time as you gather new information, but this gives you a starting point for success.  Also, don’t expect something as clear-cut as the above early on; instead, people will hem and haw, not quite sure what they really want.  You can take a fuzzy goal into data acquisition, but as you acquire data, you will want to work with the champion to focus down to a targeted and valuable problem.

Dig For Data

Once you have an interesting question, or even the bones of a question, start looking for data.  Your champion likely has a decent understanding of what is possible given your data, so the first place to look is in-house data sources like databases, Excel, flat files, data accessible through internal APIs, and even reports (generated reports like PDFs or even printed-out copies).  Your champion will hopefully be able to point you in the right direction and tell you where some of this data is located—especially if it’s hidden in paper format or in a spreadmart somewhere—but you’re going to be doing a lot of legwork between here and the data processing phase, and you’ll likely bounce back and forth between the two phases a number of times.

As you gather data resources, you will probably want to build a data dictionary for your data sources.  A great data dictionary will include things like:

  • The data type of each attribute:  numeric, string, categorical, binary.
  • The data format:  CSV, SQL Server table, Hive table, JSON file, etc.
  • The size of the data and number of records.
  • The enumeration of valid categorical values.
  • Other domain rules (if known).

I’d love to say that I always do this…but that’d be a lie.  Still, as the saying goes, hypocrisy is the tribute that vice pays to virtue.

Learn Your Outputs

While you’re looking for data and focusing in on the critical problem, you also need to figure out the endgame for your product.  Will there be a different engineering team (or multiple teams?) expecting to call a microservice API and get results back?  Will you get a set of files each day and dump the results into a warehouse?  What’s the acceptable latency?

The Engineering team should help solve this technical problem, although your champion should have insight here depending upon how the business side will need to use your results.  If the business side is already getting files in once a day, they may be fine with your process running overnight and having results in a system by 8 AM, when the analysts start showing up at work.  By contrast, you may have a fraud detection system which needs to make a decision in milliseconds.  These two systems will have radically different sets of requirements, even if the output looks the same to the business side.

Going Through An Example

My motivating example, as mentioned yesterday, is data professional salaries—figuring out how to get more money does, after all, motivate me!

Let’s suppose we work for Data Platform Specialists, a company dedicated to providing DBAs and other data platform professionals with valuable market knowledge.  We have come into possession of a survey of data professionals and want to build insights that we can share with our client base.

If you haven’t seen the survey yet, I recommend checking it out.

You can tell that this is a quality survey because it’s in blue, which is the most serious of survey result colors.

Once we have the salary data, we want to start building a data dictionary and see what the shape of the data looks like.  We’d get information on the total number of rows, note that this is stored in Excel on a single worksheet, and then make some notes on the columns.  For example, a number of these features are categorical:  for example, TelecommuteDaysPerWeek has six options, ranging from “less than 1” to “5 or more.”  By contrast, hours worked per week is an integer, which ranges from 5 to 200 (umm…).

There are quite a few columns, most of which originally came from dropdowns rather than users typing the data in.  This is good, because users are typically the biggest enemy of clean data.  But even in this example, we can see some interesting results:  for example, about halfway through the image, you can see “111,000” in the SalaryUSD column.  It turns out that this was a string field rather than numeric.  If you simply try to turn it into numeric, it will fix “111,000” but it’d turn a German’s entry of “111.000” from $110,000 to $111 if you’re in the US.  But I’m getting a bit ahead of myself here…

Where I want to go first is, what are the interesting types of questions we can ask given our vast wealth of domain knowledge and this compendium of valuable pricing insight?  (Too obsequious?  Maybe.)

  • How much money does a DBA make?  Similarly, how much does a data scientist make, or a developer who specializes in writing T-SQL?
  • Which category of DBA (e.g., junior, mid-level, senior) does a particular type of work?
  • How can we segment the DBAs in our survey?
  • Suppose I work 80 hours per week.  Compared to my peers, is this many hours per week weird?
  • Which option should I choose as a career path?  DBA?  Data scientist?  BI specialist?

Talking this through with our champion and other stakeholders, we can talk through some of the information we’ve already gathered, prime them toward asking focused questions, and narrow down to our question of interest:

How much money should we expect a data professional will make?

Well, that’s…broad.  But this early on in the process, that’s probably a reasonable question.  We might not want to put hard boundaries on ranges yet, but as we go along, we can narrow the question down further to something like, “Predict, within $5K USD, how much we should expect a data professional to make depending upon country of residence, education level, years of experience, and favorite Teenage Mutant Ninja Turtle (original series only).”

The end product that we want to support is a small website which allows people to build profiles and then takes that profile information and gets an estimate of how much they could make in different roles.  Our job is to build a microservice API which returns a dollar amount based on the inputs we define as important.


Today’s post was all about getting into the business users’ heads.  These are the people handing us big sacks of cash, after all, so we have to give them something nice in return.  In the next post, we’ll go a lot deeper into data processing and ask the question, if data platform professionals are the gatekeepers between good data and bad, why are we so bad at filling out forms?