Modelplotr v1.0 now on CRAN: Visualize the Business Value of your Predictive Models
Explaining the business value of your predictive models to your business colleagues is a challenging task. Using Modelplotr, an R package, you can easily create stunning visualizations that clearly communicate the business value of your models.
In this blog, we explain most valuable evaluation plots to assess the business value of a predictive model. Since these visualisations are not included in most popular model building packages or modules in R and Python, we show how you can easily create these plots for your own predictive models with our modelplotr r package and our modelplotpy python module (Prefer python? Read all about modelplotpy here!). This will help you to explain your model's business value in layman's terms to non-techies.
‘...And as we can see clearly on this ROC plot, the sensitivity of the model at the value of 0.2 on one minus the specificity is quite high! Right?…’.
If your fellow business colleagues didn’t already wander away during your presentation about your fantastic predictive model, it will definitely push them over the edge when you start talking like this. Why? Because the ROC curve is not easy to quickly explain and also difficult to translate into answers on the business questions your spectators have. And these business questions were the reason you’ve built a model in the first place!
What business questions? We build models for all kinds of supervised classification problems. Such as predictive models to select the best records in a dataset, which can be customers, leads, items, events... For instance: You want to know which of your active customers have the highest probability to churn; you need to select those prospects that are most likely to respond to an offer; you have to identify transactions that have a high risk to be fraudulent. During your presentation, your audience is therefore mainly focused on answering questions like Does your model enable us to our target audience? How much better are we, using your model? What will the expected response on our campaign be? What is the financial impact of using your model?
During our model building efforts, we should already be focused on verifying how well the model performs. Often, we do so by training the model parameters on a selection or subset of records and test the performance on a holdout set or external validation set. We look at a set of performance measures like the ROC curve and the AUC value. These plots and statistics are very helpful to check during model building and optimization whether your model is under- or overfitting and what set of parameters performs best on test data. However, these statistics are not that valuable in assessing the business value the model you developed.
One reason that the ROC curve is not that useful in explaining the business value of your model, is because it’s quite hard to explain the interpretation of ‘area under the curve’, ‘specificity’ or ‘sensitivity’ to business people. Another important reason that these statistics and plots are useless in your business meetings is that they don’t help in determining how to apply your predictive model: What percentage of records should we select based on the model? Should we select only the best 10% of cases? Or should we stop at 30%? Or go on until we have selected 70%?... This is something you want to decide together with your business colleague to best match the business plans and campaign targets they have to meet. The four plots - the cumulative gains, cumulative lift, response and cumulative response - and three financial plots - costs & revenues, profit and return on investment - we are about to introduce are in our view the best ones for that cause.
Before we start exploring the plots, let's install our modelplotr package. Since it is available on CRAN it can easily be installed. All source code and latest developments are also available on github.
Now we're ready to use modelplotr. In the package vignette (also available here) we go into much more detail on our modelplot package in r and all its functionalities. Here, we'll focus on using modelplotr with a business example.
Example: Predictive models from caret, mlr, h2o and keras on the Bank Marketing Data Set
Let's get down to business! Our example is based on a publicly available dataset, called the Bank Marketing Data Set. It is one of the most popular datasets which is made available on the UCI Machine Learning Repository. The data set comes from a Portuguese bank and deals with a frequently-posed marketing question: whether a customer did or did not acquire a term deposit, a financial product. There are 4 datasets available and the bank-additional-full.csv is the one we use. It contains the information of 41.188 customers and 21 columns of information.
To illustrate how to use modelplotr, let's say that we work as a data scientist for this bank and our marketing colleagues have asked us to help to select the customers that are most likely to respond to a term deposit offer. For that purpose, we will develop a predictive model and create the plots to discuss the results with our marketing colleagues. Since we want to show you how to build the plots, not how to build a perfect model, we'll only use six of the available columns as features in our example. Here’s a short description on the data we use:
- y: has the client subscribed a term deposit?
- duration: last contact duration, in seconds (numeric)
- campaign: number of contacts performed during this campaign and for this client
- pdays: number of days that passed by after the client was last contacted from a previous campaign
- previous: number of contacts performed before this campaign and for this client (numeric)
- euribor3m: Euribor 3 month rate
Let's load the data and have a quick look at it:
On this data, we've applied some predictive modeling techniques. To show modelplotr can be used for any kind of model, built with numerous packages, we've created some models with the caret package, the mlr package, the h2o package and the keras package. These four are very popular R packages to build models with many predictive modeling techniques, such as logistic regression, random forests, XGboost, SVM, neural nets and many, many others. When you've built your model with one of these packages, using modelplotr is super simple.
mlr, caret, h2o and keras provide numerous different algorithms for binary classification. It should be noted, that to use our modelplotr package, you don't have to use one of these packages to build your models. It's just a tiny bit more work to use modelplotr for models created differently, using R or even outside or R! More on this in the vignette (see: vignette(modelplotr) or go here. For now, to have a few models to evaluate with our plots, we do take advantage of mlr's/caret's/h2o's/keras' greatness.
Ok, we’ve generated some predictive models. Now, let’s prepare the data for plotting! We will focus on explaining to our marketing colleagues how good our predictive model works and how it can help them select customers for their term deposit campaign.
What just happened? In the prepare_scores_and_ntiles function, we've scored the customers in the train dataset and test dataset with their probability to acquire a term deposit, according to the predictive models we've just built with caret, mlr, h2o and keras. Aside from the datasets and model objects, we had to specify the name of the target variable and for our convenience, we gave our datasets and models some useful labels. These labels are used in the plots. And we could specify how many ntiles we want - bear with us for a few seconds, we'll get into the details of ntiles in a while. Let's have a look at some lines from the dataframe scores_and_ntiles we just created:
The function created a dataframe, with per record in each dataset, per model the actual target value (y), the predicted probability per target class ('no' and 'term.deposit') and the bucketing in ntiles (1 to 100, we'll explain what they mean shortly).
Now that we have scored all models we want on all datasets we want for the target we want, we can take the last step before plotting. In this second step, we specify the scope of the analysis. For this, we use the plotting_scope function and use the dataframe as our prepared_input. There are some other parameters you might want to tweak. An important parameter is scope, enabling you to set the analysis perspective you want to see in the plots. You can use modelplotr to evaluate your model(s) from several perspectives:
- Interpret just one model (the default)
- Compare the model performance across different datasets
- Compare the performance across different models
- Compare the performance across different target classes
We don't set the "scope" parameter now, therefore the default - no comparison - is chosen. We will keep it simple and evaluate - from a business perspective - how well a selected model will perform in a selected dataset for one target class. We can use other parameters to focus on a specific model (let's say the gradient boosted trees model) and on a specific dataset (let's pick the test data). When not specified, modelplotr will set a value for you. The default value for the target class is the smallest target category, in our case term.deposit ; since we want to focus on customers that do take term deposits, this default is perfect so we don't specify it explicitly here:
As you can see in the output above, modelplotr prints some guiding information about the options chosen and unused options to the console. We have generated an aggregated dataframe, with the scope and selections we want and many metrics used for our plots:
(plot_input has 101 rows and 25 columns)
We're ready to start plotting now!
Let’s introduce the Gains, Lift and (cumulative) Response plots.
Before we throw more code and output at you, let’s get you familiar with the plots we so strongly advocate to use to assess a predictive model’s business value. Although each plot sheds light on the business value of your model from a different angle, they all use the same data:
- Predicted probability for the target class
- Equally sized groups based on this predicted probability, named ntiles
- Actual number of observed target class observations in these groups (ntiles)
Regarding the ntiles: It’s common practice to split the data to score into 10 equally large groups and call these groups deciles. Observations that belong to the top-10% with highest model probability in a set, are in decile 1 of that set; the next group of 10% with high model probability are decile 2 and finally the 10% observations with the lowest model probability on the target class belong to decile 10.
*Notice that modelplotr does support that you specify the number of equally sized groups with the parameter ntiles. Hence, ntiles=100 results in 100 equally sized groups with in the first group the 1% with the highest model probability and in group 100 the 1% with the lowest model probability. These groups are often referred to as percentiles; modelplotr will also label them as such. Any value between 4 and 100 can be specified for ntiles.
Each of the plots in modelplotr places the ntiles on the x axis and another measure on the y axis. The ntiles are plotted from left to right so the observations with the highest model probability are on the left side of the plot. This results in plots like this:
Now that it’s clear what is on the horizontal axis of each of the plots, we can go into more detail on the metrics for each plot on the vertical axis. For each plot, we’ll start with a brief explanation what insight you gain with the plot from a business perspective. After that, we apply it to our banking data and show some neat features of modelplotr to help you explain the value of your wonderful predictive models to others.
1. Cumulative gains plot
The cumulative gains plot - often named ‘gains plot’ - helps you answer the question:
When we apply the model and select the best X ntiles, what % of the actual target class observations can we expect to target?
Hence, the cumulative gains plot visualises the percentage of the target class members you have selected if you would decide to select up until ntile X. This is a very important business question, because in most cases, you want to use a predictive model to target a subset of observations - customers, prospects, cases,... - instead of targeting all cases. And since we won't build perfect models all the time, we will miss some potential. And that's perfectly all right, because if we are not willing to accept that, we should not use a model in the first place. Or build only perfect models, that scores all actual target class members with a 100% probability and all the cases that do not belong to the target class with a 0% probability. However, if you’re such a wizard, you don’t need these plots any way or you should have a careful look at your model - maybe you’re cheating?....
So, we'll have to accept we will lose some. What percentage of the actual target class members you do select with your model at a given ntile, that’s what the cumulative gains plot tells you. The plot comes with two reference lines to tell you how good/bad your model is doing: The random model line and the wizard model line. The random model line tells you what proportion of the actual target class you would expect to select when no model is used at all. This vertical line runs from the origin (with 0% of cases, you can only have 0% of the actual target class members) to the upper right corner (with 100% of the cases, you have 100% of the target class members). It’s the rock bottom of how your model can perform; are you close to this, then your model is not much better than a coin flip. The wizard model is the upper bound of what your model can do. It starts in the origin and rises as steep as possible towards 100%. Say, exactly 10% of all cases belong to the target category. This means that this line goes steep up from the origin to the value of decile 1 (or percentile 10 in case ntiles=100) and cumulative gains of 100% and remains there for all other ntiles as it is a cumulative measure. Your model will always move between these two reference lines - closer to a wizard is always better - and looks like this:
Back to our business example. How many of the term deposit buyers can we select with the top-20% of our predictive models? Let's find out! To generate the cumulate gains plot, we can simply call the function plot_cumgains():
We don't need to specify any other parameters than the data to use, plot_input, which is generated with the plotting_scope() function we ran previously. There are several other parameters available to customize the plot, though. If we want to emphasize the model performance at a specific ntile, we can add both highlighting to the plot and add some explanatory text below the plot. Both are optional, though:
Our highlight_ntile parameter adds some guiding elements to the plot at ntile 20 as well as a text box below the graph with the interpretation of the plot at ntile 20 in words. This interpretation is also printed to the console. Our simple model - only 6 predictors were used - seems to do a nice job selecting the customers interested in buying term deposits. When we select 20% with the highest probability according to gradient boosted trees, this selection holds 87% of all term deposit cases in test data. With a perfect model, we would have selected 100%, since less than 20% of all customers in the test set buy term deposits. A random pick would only hold 20% of the customers with term deposits. How much better than random do we do? This brings us to plot number two!