Introduction to Reason Codes

2018-01-17

A few questions always asked by business leaders after seeing the results of highly accurate machine learning models are as follows
- Are machine learning models interpretable and transparent?
- How can the results of the model be used to develop a business strategy?
- Can the predictions from the model be used to explain to the regulators why something was rejected or accepted based on model prediction?

DataRobot does provide many diagnostics like partial dependence, feature impact, reason codes to answer the above questions and using those diagnostics predictions can be converted to prescriptions for the business. In this vignette we would be covering reason codes. Partial dependence has been covered in detail in the companion vignette “Interpreting Predictive Models Using Partial Dependence Plots”.

Introduction

The DataRobot modeling engine is a commercial product that supports the rapid development and evaluation of a large number of different predictive models from a single data source. The open-source R package datarobot allows users of the DataRobot modeling engine to interact with it from R, creating new modeling projects, examining model characteristics, and generating predictions from any of these models for a specified dataset. This vignette illustrates how to interact with DataRobot using datarobot package, build models, make prediction using a model and then use reason codes to explain why a model is predicting high or low. Reason codes can be used to answer the questions mentioned earlier.

Let’s load datarobot and other useful packages

library(datarobot)
library(httr)
library(knitr)
library(data.table)

Connecting to DataRobot

To access the DataRobot modeling engine, it is necessary to establish an authenticated connection, which can be done in one of two ways. In both cases, the necessary information is an endpoint - the URL address of the specific DataRobot server being used - and a token, a previously validated access token.

token is unique for each DataRobot modeling engine account and can be accessed using the DataRobot webapp in the account profile section.

endpoint depends on DataRobot modeling engine installation (cloud-based, on-prem…) you are using. Contact your DataRobot admin for endpoint to use. The endpoint for DataRobot cloud accounts is https://app.datarobot.com/api/v2

The first access method uses a YAML configuration file with these two elements - labeled token and endpoint - located at $HOME/.config/datarobot/drconfig.yaml. If this file exists when the datarobot package is loaded, a connection to the DataRobot modeling engine is automatically established. It is also possible to establish a connection using this YAML file via the ConnectToDataRobot function, by specifying the configPath parameter. The second method of establishing a connection to the DataRobot modeling engine is to call the function ConnectToDataRobot with the endpoint and token parameters. endpoint <- "https://<YOUR ENDPOINT HERE>/api/v2" apiToken <- "<YOUR API TOKEN HERE>" ConnectToDataRobot(endpoint = endpoint, token = apiToken) Data We would be using a sample dataset related to credit scoring open sourced by LendingClub (https://www.lendingclub.com/). Below is the information related to the variables.  Id Min. : 1 1st Qu.: 2501 Median : 5000 Mean : 5000 3rd Qu.: 7500 Max. :10000 NA is_bad Min. :0.0000 1st Qu.:0.0000 Median :0.0000 Mean :0.1295 3rd Qu.:0.0000 Max. :1.0000 NA emp_title Length:10000 Class :character Mode :character NA NA NA NA emp_length Length:10000 Class :character Mode :character NA NA NA NA home_ownership Length:10000 Class :character Mode :character NA NA NA NA annual_inc Min. : 2000 1st Qu.: 40000 Median : 58000 Mean : 68203 3rd Qu.: 82000 Max. :900000 NA’s :1 verification_status Length:10000 Class :character Mode :character NA NA NA NA pymnt_plan Length:10000 Class :character Mode :character NA NA NA NA Notes Length:10000 Class :character Mode :character NA NA NA NA purpose_cat Length:10000 Class :character Mode :character NA NA NA NA purpose Length:10000 Class :character Mode :character NA NA NA NA zip_code Length:10000 Class :character Mode :character NA NA NA NA addr_state Length:10000 Class :character Mode :character NA NA NA NA debt_to_income Min. : 0.00 1st Qu.: 8.16 Median :13.41 Mean :13.34 3rd Qu.:18.69 Max. :29.99 NA delinq_2yrs Min. : 0.0000 1st Qu.: 0.0000 Median : 0.0000 Mean : 0.1482 3rd Qu.: 0.0000 Max. :11.0000 NA’s :5 earliest_cr_line Length:10000 Class :character Mode :character NA NA NA NA inq_last_6mths Min. : 0.000 1st Qu.: 0.000 Median : 1.000 Mean : 1.067 3rd Qu.: 2.000 Max. :25.000 NA’s :5 mths_since_last_delinq Min. : 0.00 1st Qu.: 18.00 Median : 34.00 Mean : 35.89 3rd Qu.: 53.00 Max. :120.00 NA’s :6316 mths_since_last_record Min. : 0.00 1st Qu.: 0.00 Median : 86.00 Mean : 61.65 3rd Qu.:101.00 Max. :119.00 NA’s :9160 open_acc Min. : 1.000 1st Qu.: 6.000 Median : 9.000 Mean : 9.335 3rd Qu.:12.000 Max. :39.000 NA’s :5 pub_rec Min. :0.00000 1st Qu.:0.00000 Median :0.00000 Mean :0.06013 3rd Qu.:0.00000 Max. :3.00000 NA’s :5 revol_bal Min. : 0 1st Qu.: 3524 Median : 8646 Mean : 14271 3rd Qu.: 16952 Max. :1207359 NA revol_util Min. : 0.00 1st Qu.: 25.00 Median : 48.70 Mean : 48.45 3rd Qu.: 71.80 Max. :100.60 NA’s :26 total_acc Min. : 1.00 1st Qu.:13.00 Median :20.00 Mean :22.01 3rd Qu.:29.00 Max. :90.00 NA’s :5 initial_list_status Length:10000 Class :character Mode :character NA NA NA NA collections_12_mths_ex_med Min. :0 1st Qu.:0 Median :0 Mean :0 3rd Qu.:0 Max. :0 NA’s :32 mths_since_last_major_derog Min. :1.000 1st Qu.:1.000 Median :2.000 Mean :2.002 3rd Qu.:3.000 Max. :3.000 NA policy_code Length:10000 Class :character Mode :character NA NA NA NA Divide data into train and test and setup the project Let’s divide our data in train and test. We can use train data to create a datarobot project using SetupProject function and test data to make predictions and generate reason codes. Detailed explanation about creating projects was described in the vignette , “Introduction to the DataRobot R Package.” The specific sequence used here was: target <- "is_bad" projectName <- "Credit Scoring" numWorkers <- 10 set.seed(1111) split <- sample(nrow(Lending), round(0.9 * nrow(Lending)), replace = FALSE) train <- Lending[split,] test <- Lending[-split,] project <- SetupProject(dataSource = train, projectName = projectName) SetTarget(project = project, target = target) Once the modeling process has completed, the ListModels function returns an S3 object of class “listOfModels” that characterizes all of the models in a specified DataRobot project. Calling this function before the modeling process is complete causes a partial result to be returned, with a warning; to avoid this problem, the WaitForAutopilot function is used before calling ListModels: # increase the number of workers used by this project UpdateProject(project = project$projectId,
workerCount = numWorkers)
WaitForAutopilot(project, verbosity = 1, timeout = 999999)

results <- as.data.frame(ListModels(project))
kable(head(results), longtable = TRUE, booktabs = TRUE, row.names = TRUE)
modelType expandedModel modelId blueprintId featurelistName featurelistId samplePct validationMetric
1 Advanced AVG Blender Advanced AVG Blender 58d9501cf2ff94714e225839 300fb24a331bd7e8313023faf0db0a4d Informative Features 58d94dddfbfa573bd3aa3295 63.9929 0.34306
2 ENET Blender ENET Blender 58d9501cf2ff94714e22583b 9df9e8c1439604171b9a025faf88aa5a Informative Features 58d94dddfbfa573bd3aa3295 63.9929 0.34310
3 AVG Blender AVG Blender 58d9501bf2ff94714e225837 e3aadd9331abf90ea27fa28dcbd72fc7 Informative Features 58d94dddfbfa573bd3aa3295 63.9929 0.34570
4 ENET Blender ENET Blender 58d9501cf2ff94714e22583d 6ed326ca6e31c0ce27aa7729b256e814 Informative Features 58d94dddfbfa573bd3aa3295 63.9929 0.34572
5 eXtreme Gradient Boosted Trees Classifier with Early Stopping eXtreme Gradient Boosted Trees Classifier with Early Stopping::Tree-based Algorithm Preprocessing v21 58d94ec5f2ff947071225847 8565246f48026082fac7fb3153ed8aa0 Informative Features 58d94dddfbfa573bd3aa3295 63.9929 0.34636
6 eXtreme Gradient Boosted Trees Classifier with Early Stopping eXtreme Gradient Boosted Trees Classifier with Early Stopping::Tree-based Algorithm Preprocessing v1 58d94ec5f2ff94707122584b c2cf08be8a3d2c7d91bbb299ac00fe83 Informative Features 58d94dddfbfa573bd3aa3295 63.9929 0.34636

Generating Model Predictions

The generation of model predictions is a three-step process:
2. Create a predict job using RequestPredictionsForDataset function, which returns the predictJobId.
3. Pass the predictJobId to GetPredictions along with the projectId for the DataRobot project containing the model. The result returned by this function is a vector of predicted responses; in the case of binary classification projects, the optional type parameter may be used to request a vector of probabilities instead of binary responses; refer to the help files for details.

As a specific example, the following code sequence identifies the model with the best performance, extracts it as bestModel, and generates predictions for it from the test dataframe we created earlier:

allModels <- ListModels(project)
modelFrame <- as.data.frame(allModels)
metric <- modelFrame$validationMetric bestIndex <- which.min(metric) bestModel <- allModels[[bestIndex]] dataset <- UploadPredictionDataset(project, test, maxWait = 1200) bestPredictJobId <- RequestPredictionsForDataset(project, bestModel$modelId, dataset$id) bestPredictions <- GetPredictions(project, bestPredictJobId, type="probability") testPredictions <- data.frame(original = test$is_bad, prediction = bestPredictions)
kable(head(testPredictions), longtable = TRUE, booktabs = TRUE, row.names = TRUE)
original prediction
1 0 0.1882636
2 0 0.0890647
3 0 0.0888415
4 0 0.2131941
5 0 0.0966023
6 0 0.0956107

Calculate Feature Impact

We need to generate Feature Impact for the model before we can get reason codes using that model. Feature Impact, which is available for all model types, works by altering input data and observing the effect on a model’s score. It is an on-demand feature, meaning that you must initiate a calculation to see the results. Once you have had DataRobot compute the feature impact for a model, that information is saved with the project (you do not need to recalculate feature impact each time you re-open the project or each time you request reason codes in new data).

Feature Impact for a given column measures how much worse a model’s error score would be if DataRobot made predictions after randomly shuffling that column (while leaving other columns unchanged). This technique is sometimes called Permutation Importance.

featureImpactJobId <- RequestFeatureImpact(bestModel)
featureImpact <- GetFeatureImpactForJobId(project, featureImpactJobId, maxWait = 1200)
#Print top 10 features
kable(featureImpact[1:10,], longtable = TRUE, booktabs = TRUE, row.names = TRUE)
featureName impactNormalized impactUnnormalized
1 purpose_cat 1.0000000 0.0684926
2 total_acc 0.6393432 0.0437903
3 Notes 0.3740867 0.0256222
4 open_acc 0.2490012 0.0170547
5 annual_inc 0.1822035 0.0124796
6 revol_util 0.1383699 0.0094773
7 inq_last_6mths 0.1377131 0.0094323
8 emp_title 0.1308886 0.0089649
9 purpose 0.0708564 0.0048531
10 emp_length 0.0614025 0.0042056

Calculate Reason Codes

For each prediction, DataRobot provides an ordered list of reasons; the number of reasons is based on the setting. Each reason is a feature from the dataset and its corresponding value, accompanied by a qualitative indicator of the reason’s strength—strong (+++), medium (++), or weak (+) positive or negative (-) influence.

There are three main inputs you can set for DataRobot to use when computing reason codes
1. maxCodes: the Number of reasons for each predictions. Default is 3.
2. thresholdLow: Probability threshold below which DataRobot should calculate reason codes.
3. thresholdHigh: Probability threshold above which DataRobot should calculate reason codes.

# Calculate reason codes
reasonCodeJobID <- RequestReasonCodesInitialization(bestModel)
reasonCodeJobIDInitialization <- GetReasonCodesInitializationFromJobId(project,reasonCodeJobID)
# Calculate reason codes for our dataset
reasonCodeRequest <- RequestReasonCodes(bestModel, dataset$id, maxCodes = 3, thresholdLow = 0.25, thresholdHigh = 0.75) # Get the reason codes we calculated reasonCodeRequestMetaData <- GetReasonCodesMetadataFromJobId(project, reasonCodeRequest, maxWait = 1800) reasonCodeMetadata <- GetReasonCodesMetadata(project, reasonCodeRequestMetaData$id)
reasonCodeAsDataFrame <- GetAllReasonCodesRowsAsDataFrame(project, reasonCodeRequestMetaData$id) reasonCodeAsDataFrame$rowId <- NULL
#subset top 3 and bottom 3 predictions
reasonCodeAsDataFrameTopBottom <- rbind(reasonCodeAsDataFrame[order(reasonCodeAsDataFrame$class1Probability),][1:3,], reasonCodeAsDataFrame[order(reasonCodeAsDataFrame$class2Probability),][1:3,])
kable(head(reasonCodeAsDataFrameTopBottom), longtable = TRUE, booktabs = TRUE, row.names = TRUE)
prediction class1Label class1Probability class2Label class2Probability reason1FeatureName reason1FeatureValue reason1QualitativeStrength reason1Strength reason1Label reason2FeatureName reason2FeatureValue reason2QualitativeStrength reason2Strength reason2Label reason3FeatureName reason3FeatureValue reason3QualitativeStrength reason3Strength reason3Label
590 0 1 0.0152559 0 0.9847441 inq_last_6mths 5 -1.1721930 1 purpose CDT CARD RAISING RATE FR/17 TO 27 % -0.4290407 1 total_acc 38 -0.3091894 1
823 0 1 0.0170943 0 0.9829057 inq_last_6mths 5 -1.2399844 1 annual_inc 230000 -0.6646366 1 total_acc 40 -0.4963727 1
596 0 1 0.0196080 0 0.9803920 inq_last_6mths 6 -0.9582067 1 total_acc 28 -0.3250731 1 annual_inc 110000 -0.1900297 1
600 1 1 0.9692447 0 0.0307553 purpose_cat credit card small business +++ 5.1677089 1 inq_last_6mths 2 + 0.2373998 1 revol_util 84.1 + 0.2321250 1
414 1 1 0.9651950 0 0.0348050 purpose_cat major purchase small business +++ 4.9394082 1 Notes ++ 0.3794873 1 annual_inc 38400 + 0.1755247 1
540 1 1 0.9576436 0 0.0423564 purpose_cat small business small business +++ 4.5879754 1 open_acc 5 -0.8902392 1 earliest_cr_line (Year) 2005 - -0.3356800 1

From the example above, you could answer “Why did the model give one of the customers a 97% probability of defaulting?” Top reason explains that purpose_cat of loan was “credit card small business”" and we can also see in above example that whenever model is predicting high probability of default, purpose_cat is related to small business.

Some notes on reasons:
- If the data points are very similar, the reasons can list the same rounded up values.
- It is possible to have a reason state of MISSING if a “missing value” was important in making the prediction.
- Typically, the top reasons for a prediction have the same direction as the outcome, but it’s possible that with interaction effects or correlations among variables a reason could, for instance, have a strong positive impact on a negative prediction.

In some projects – such as insurance projects – the prediction adjusted by exposure is more useful to look at than just raw prediction. For example, the raw prediction (e.g. claim counts) is divided by exposure (e.g. time) in the project with exposure column. The adjusted prediction provides insights with regard to the predicted claim counts per unit of time. To include that information, set excludeAdjustedPredictions to False in correspondent method calls.

reasonCodeAsDataFrameWithExposure <- GetAllReasonCodesRowsAsDataFrame(project, reasonCodeRequestMetaData\$id, excludeAdjustedPredictions = FALSE)
kable(head(reasonCodeAsDataFrameWithExposure), longtable = TRUE, booktabs = TRUE, row.names = TRUE)
rowId prediction class1Label class1Probability class2Label class2Probability reason1FeatureName reason1FeatureValue reason1QualitativeStrength reason1Strength reason1Label reason2FeatureName reason2FeatureValue reason2QualitativeStrength reason2Strength reason2Label reason3FeatureName reason3FeatureValue reason3QualitativeStrength reason3Strength reason3Label
1 0 0 1 0.0673742 0 0.9326258 revol_util 37.1 -0.2352395 1 earliest_cr_line (Month) 9 ++ 0.1299938 1 mths_since_last_delinq 16 ++ 0.1047859 1
2 1 0 1 0.1144156 0 0.8855844 total_acc 27 -0.5951351 1 inq_last_6mths 2 ++ 0.2421135 1 annual_inc 50000 -0.1259684 1
3 2 0 1 0.0946424 0 0.9053576 total_acc 23 -0.5863631 1 annual_inc 40000 ++ 0.2572561 1 purpose FICO score 762 want’s to buy a new car -0.2285383 1
4 3 0 1 0.0532636 0 0.9467364 total_acc 24 -0.5805456 1 annual_inc 112000 -0.3023279 1 revol_util 39.5 -0.2645794 1
5 4 0 1 0.1360759 0 0.8639241 total_acc 19 -0.3870679 1 annual_inc 30000 ++ 0.1768992 1 revol_util 62.1 ++ 0.1716356 1
6 5 0 1 0.0907338 0 0.9092662 total_acc 24 -0.3785583 1 inq_last_6mths 3 ++ 0.2634575 1 revol_util 3.1 -0.1964933 1

Summary

This note has described the Reason Codes which are useful for understanding why model is predicting high or low predictions for a specific case. DataRobot also provides qualitative stregth of each reason. Reason Codes can be used in developing good business strategy by taking prescriptions based on the reasons which are responsible for high or low predictions. They are also useful in explaining the actions taken based on the model predictions to regulatory or compliance department within an organiation.