Employee turnvover (attrition) is a major cost to an organization, and predicting turnover is at the forefront of needs of Human Resources (HR) in many organizations. Until now the mainstream approach has been to use logistic regression or survival curves to model employee attrition. However, with advancements in machine learning (ML), we can now get both better predictive performance and better explanations of what critical features are linked to employee attrition. In this post, we’ll use two cutting edge techniques. First, we’ll use the
h2o package’s new FREE automatic machine learning algorithm,
h2o.automl(), to develop a predictive model that is in the same ballpark as commercial products in terms of ML accuracy. Then we’ll use the new
lime package that enables breakdown of complex, black-box machine learning models into variable importance plots. We can’t stress how excited we are to share this post because it’s a much needed step towards machine learning in business applications!!! Enjoy.
Employee Attrition: A Major Problem
Bill Gates was once quoted as saying,
“You take away our top 20 employees and we [Microsoft] become a mediocre company”.
His statement cuts to the core of a major problem: employee attrition. An organization is only as good as its employees, and these people are the true source of its competitive advantage.
Organizations face huge costs resulting from employee turnover. Some costs are tangible such as training expenses and the time it takes from when an employee starts to when they become a productive member. However, the most important costs are intangible. Consider what’s lost when a productive employee quits: new product ideas, great project management, or customer relationships.
With advances in machine learning and data science, its possible to not only predict employee attrition but to understand the key variables that influence turnover. We’ll take a look at two cutting edge techniques:
Machine Learning with
h2opackage: This function takes automated machine learning to the next level by testing a number of advanced algorithms such as random forests, ensemble methods, and deep learning along with more traditional algorithms such as logistic regression. The main takeaway is that we can now easily achieve predictive performance that is in the same ball park (and in some cases even better than) commercial algorithms and ML/AI software.
Feature Importance with the
limepackage: The problem with advanced machine learning algorithms such as deep learning is that it’s near impossible to understand the algorithm because of its complexity. This has all changed with the
limepackage. The major advancement with
limeis that, by recursively analyzing the models locally, it can extract feature importance that repeats globally. What this means to us is that
limehas opened the door to understanding the ML models regardless of complexity. Now the best (and typically very complex) models can also be investigated and potentially understood as to what variables or features make the model tick.
Employee Attrition: Machine Learning Analysis
With these new automated ML tools combined with tools to uncover critical variables, we now have capabilities for both extreme predictive accuracy and understandability, which was previously impossible! We’ll investigate an HR Analytic example of employee attrition that was evaluated by IBM Watson.
IBM Watson (Where we got the data)
The example comes from IBM Watson Analytics website. You can download the data and read the analysis here:
To summarize, the article makes a usage case for IBM Watson as an automated ML platform. The article shows that using Watson, the analyst was able to detect features that led to increased probability of attrition.
Automated Machine Learning (What we did with the data)
In this example we’ll show how we can use the combination of H2O for developing a complex model with high predictive accuracy on unseen data and then how we can use LIME to understand important features related to employee attrition.
Load the following packages.
Download the data here. You can load the data using
read_excel(), pointing the path to your local file.
Let’s check out the raw data. It’s 1470 rows (observations) by 35 columns (features). The “Attrition” column is our target. We’ll use all other columns as features to our model.
|41||Yes||Travel_Rarely||1102||Sales||1||2||Life Sciences||1||1||2||Female||94||3||2||Sales Executive||4||Single||5993||19479||8||Y||Yes||11||3||1||80||0||8||0||1||6||4||0||5|
|49||No||Travel_Frequently||279||Research & Development||8||1||Life Sciences||1||2||3||Male||61||2||2||Research Scientist||2||Married||5130||24907||1||Y||No||23||4||4||80||1||10||3||3||10||7||1||7|
|37||Yes||Travel_Rarely||1373||Research & Development||2||2||Other||1||4||4||Male||92||2||1||Laboratory Technician||3||Single||2090||2396||6||Y||Yes||15||3||2||80||0||7||3||3||0||0||0||0|
|33||No||Travel_Frequently||1392||Research & Development||3||4||Life Sciences||1||5||4||Female||56||3||1||Research Scientist||3||Married||2909||23159||1||Y||Yes||11||3||3||80||0||8||3||3||8||7||3||0|
|27||No||Travel_Rarely||591||Research & Development||2||1||Medical||1||7||1||Male||40||3||1||Laboratory Technician||2||Married||3468||16632||9||Y||No||12||3||4||80||1||6||3||3||2||2||2||2|
|32||No||Travel_Frequently||1005||Research & Development||2||2||Life Sciences||1||8||4||Male||79||3||1||Laboratory Technician||4||Single||3068||11864||0||Y||No||13||3||3||80||0||8||2||2||7||7||3||6|
|59||No||Travel_Rarely||1324||Research & Development||3||3||Medical||1||10||3||Female||81||4||1||Laboratory Technician||1||Married||2670||9964||4||Y||Yes||20||4||1||80||3||12||3||2||1||0||0||0|
|30||No||Travel_Rarely||1358||Research & Development||24||1||Life Sciences||1||11||4||Male||67||3||1||Laboratory Technician||3||Divorced||2693||13335||1||Y||No||22||4||2||80||1||1||2||3||1||0||0||0|
|38||No||Travel_Frequently||216||Research & Development||23||3||Life Sciences||1||12||4||Male||44||2||3||Manufacturing Director||3||Single||9526||8787||0||Y||No||21||4||2||80||0||10||2||3||9||7||1||8|
|36||No||Travel_Rarely||1299||Research & Development||27||3||Medical||1||13||3||Male||94||3||2||Healthcare Representative||3||Married||5237||16577||6||Y||No||13||3||2||80||2||17||3||2||7||7||7||7|
The only pre-processing we’ll do in this example is change all character data types to factors. This is needed for H2O. We could make a number of other numeric data that is actually categorical factors, but this tends to increase modeling time and can have little improvement on model performance.
Let’s take a
glimpse at the processed dataset. We can see all of the columns. Note our target (“Attrition”) is the first column.
Modeling Employee Attrition
We are going to use the
h2o.automl() function from the H2O platform to model employee attrition.
First, we need to initialize the Java Virtual Machine (JVM) that H2O uses locally.
Next, we change our data to an
h2o object that the package can interpret. We also split the data into training, validation, and test sets. Our preference is to use 70%, 15%, 15%, respectively.
Now we are ready to model. We’ll set the target and feature names. The target is what we aim to predict (in our case “Attrition”). The features (every other column) are what we will use to model the prediction.
Now the fun begins. We run the
h2o.automl() setting the arguments it needs to run models against. For more information, see the h2o.automl documentation.
x = x: The names of our feature columns.
y = y: The name of our target column.
training_frame = train_h2o: Our training set consisting of 70% of the data.
leaderboard_frame = valid_h2o: Our validation set consisting of 15% of the data. H2O uses this to ensure the model does not overfit the data.
max_runtime_secs = 30: We supply this to speed up H2O’s modeling. The algorithm has a large number of complex models so we want to keep things moving at the expense of some accuracy.
All of the models are stored the
automl_models_h2o object. However, we are only concerned with the leader, which is the best model in terms of accuracy on the validation set. We’ll extract it from the models object.
Now we are ready to predict on our test set, which is unseen from during our modeling process. This is the true test of performance. We use the
h2o.predict() function to make predictions.
Now we can evaluate our leader model. We’ll reformat the test set an add the predictions as column so we have the actual and prediction columns side-by-side.
We can use the
table() function to quickly get a confusion table of the results. We see that the leader model wasn’t perfect, but it did a decent job identifying employees that are likely to quit. For perspective, a logistic regression would not perform nearly this well.
We can review from a percentage standpoint. We had 7% Type I Error (predicted to quit, but actually stayed) and 5% Type II Error (predicted to stay, but actually quit).
And the final performance from a percentage-standpoint is about 88% accuracy.
We have a very good model that is capable of making very accurate predictions on unseen data, but what can it tell us about what causes attrition? Let’s find out using LIME.
lime package implements LIME in R. One thing to note is that it’s not setup out-of-the-box to work with
h2o. The good news is with a few functions we can get everything working properly. We’ll need to make two custom functions:
model_type: Used to tell
limewhat type of model we are dealing with. It could be classification, regression, survival, etc.
predict_model: Used to allow
limeto perform predictions that its algorithm can interpret.
The first thing we need to do is identify the class of our model leader object. We do this with the
Next we create our
model_type function. It’s only input is
x the h2o model. The function simply returns “classification”, which tells LIME we are classifying.
Now we can create our
predict_model function. The trick here is to realize that it’s inputs must be
x a model,
newdata a dataframe object (this is important), and
type which is not used but can be use to switch the output type. The output is also a little tricky because it must be in the format of probabilities by classification (this is important; shown next). Internally we just call the
Run this next script to show you what the output looks like and to test our
predict_model function. See how it’s the probabilities by classification. It must be in this form for model_type = “classification”.
Now the fun part, we create an explainer using the
lime() function. Just pass the training data set without the “Attribution column”. The form must be a data frame, which is OK since our
predict_model function will switch it to an
h2o object. Set
model = automl_leader our leader model, and
bin_continuous = FALSE. We could tell the algorithm to bin continuous variables, but this may not make sense for categorical numeric data that we didn’t change to factors.
Now we run the
explain() function, which returns our
explanation. This can take a minute to run so we limit it to just the first ten rows of the test data set. We set
n_labels = 1 because we care about explaining a single class. Setting
n_features = 4 returns the top four features that are critical to each case. Finally, setting
kernel_width = 0.5 allows us to increase the “model_r2” value by shrinking the localized evaluation.
Feature Importance Visualization
The payoff for the work we put in using LIME is this feature importance plot. This allows us to visualize each of the ten cases (observations) from the test data. The top four features for each case are shown. Note that they are not the same for each case. The green bars mean that the feature supports the model conclusion, and the red bars contradict. We’ll focus in on Cases with Label = Yes, which are predicted to have attrition. We can see a common theme with Case 3 and Case 7: Training Time, Job Role, and Over Time are among the top factors influencing attrition. These are only two cases, but they can be used to potentially generalize to the larger population as we will see next.
What Features Are Linked To Employee Attrition?
Now we turn to our three critical features from the LIME Feature Importance Plot:
- Training Time
- Job Role
- Over Time
We’ll subset this data and visualize to detect trends.
From the violin plot, the employees that stay tend to have a large peaks at two and three trainings per year whereas the employees that leave tend to have a large peak at two trainings per year. This suggests that employees with more trainings may be less likely to leave.
The plot below shows a very interesting relationship: a very high proportion of employees that turnover are working over time. The opposite is true for employees that stay.
Several job roles are experiencing more turnover. Sales reps have the highest turnover at about 40% followed by Lab Technician, Human Resources, Sales Executive, and Research Scientist. It may be worthwhile to investigate what localized issues could be creating the high turnover among these groups within the organization.
There’s a lot to take away from this article. We showed how you can use predictive analytics to develop sophisticated models that very accurately detect employees that are at risk of turnover. The autoML algorithm from H2O.ai worked well for classifying attrition with an accuracy around 87% on unseen / unmodeled data. We then used LIME to breakdown the complex ensemble model returned from H2O into critical features that are related to attrition. Overall, this is a really useful example where we can see how machine learning and data science can be used in business applications.
Our mission is to enable ANY organization to have access to data science. We have a full suite of data science services to supercharge your financial and business performance. How do we do it? Using our network of data science consultants, we pull together the right team to get custom projects done on time, within budget, and of the highest quality. Find out more about Business Science or contact us!
Grow with us! We are seeking top-tier data scientists. Let us know if you are interested in joining our network of data scientist consultants. If you have expertise in Marketing Analytics, Data Science for Business, Financial Analytics, or Data Science in general, we’d love to talk. Contact us!
- @bizScienc is on twitter!
- Check us out on Facebook page!
- Check us out on LinkedIn!
- Sign up for our insights blog to stay updated!
- If you like our software, star our GitHub packages!