Random Forest Regression
- Ananya Bhattacharya
Introduction
Spark plugins that trains and predicts label data based on Random Forest Regression Algorithm.
Use-case
User wants to predict if the flight will be delayed or not based on some features of airline data:
Label → delayed and not delayed - delayed if 1.0 and 0.0 otherwise
Features → {dayOfMonth, weekday, scheduledDepTime, scheduledArrTime, carrier, elapsedTime, originId, destId}
User Stories
User should be able to train the data.
User should be able to provide the list of columns(features) to use for training.
User should be able to provide the list of columns(features) to be used for prediction.
User should be able to provide the column to be used as prediction field while training/prediction.
User should be able to specify the maximum depth of the decision tree.
User should be able to specify maximum number of bins used for splitting features.
User should be able to specify the cardinality for categorical features.
User should be able to specify random seed for bootstrapping and choosing feature subsets.
User should be able to specify number of features to consider for splits at each node. Supported: "auto", "all", "sqrt", "log2", "onethird".
User should be able to provide the file set name to save the training model.
User should be able to provide the path of the file set.
Example
Following is a simple example showing how Random Forest Trainer and Regression would work to predict if the flight will be delayed or not.
For each flight, we have the following information:
Delayed | Day of Week | Carrier | TailNum | FlightNum | OriginId | DestinationId | Day of Month | Distance | Arrival Time | Departure Time |
---|---|---|---|---|---|---|---|---|---|---|
1.0 | 4 | 1.0 | N787AA | 21 | 101 | 111 | 1 | 2475 | 1230 | 855 |
0.0 | 6 | 2.0 | N457ER | 34 | 105 | 203 | 1 | 1589 | 1530 | 1700 |
The Random Forest Trainer will train the data based on some features, for example : {dayOfMonth, weekday, scheduledDepTime, scheduledArrTime, carrier, elapsedTime, originId, destId.
The label for the first and second rows will be set to 1.0 and 0.0(delayed column value).
Trainer will save the model in a fileSet, which will be used later for predicting the delayed value using regression.
Conditions
- Fields to be used for training and prediction using Regression(features) should be of simple type : int, double, float, long.
- Fields to be used for training and prediction using Regression(features) should not be of type NULL.
Design
Random Forest Trainer:
Properties:
- fileSetName : The name of the FileSet to save the model to.
- path : Path of the FileSet to save the model to.
- featuresToInclude: A comma-separated sequence of fields to use for training. If empty, all fields except label field will be used for training. Features to be used, must be of type: int, double, float, long. Both "featuresToInclude" and "featuresToExclude" fields cannot be specified.
- featuresToExclude: A comma-separated sequence of fields to be excluded when training. If empty, all the fields except label field will be usefor training. Both "featuresToInclude" and "featuresToExclude" fields cannot be specified.
- cardinalityMapping: Mapping of the feature to the cardinality of that feature; required for categorical features.
- predictionField : The field from which to get the prediction. It must be of type double.
- maxDepth : Maximum depth of the tree.E.g., depth 0 means 1 leaf node; depth 1 means 1 internal node + 2 leaf nodes. Default is 10.
- maxBins : Maximum number of bins used for splitting when discretizing continuous features. Random Forest requires maxBins to be at least as large as the number of values in each categorical feature. Default is 100.
- seed : Random seed for bootstrapping and choosing feature subsets.
featureSubsetStrategy
: Number of features to consider for splits at each node. Supported: "auto", "all", "sqrt", "log2", "onethird". If "auto" is set, this parameter is set based on numTrees: if numTrees == 1, set to "all"; if numTrees > 1 (forest) set to "onethird".
Input Json Format
{ "name": "RandomForestTrainer", "type": "sparksink", "properties": { "fileSetName": "random-forest-model", "path": "RandomForest", "featuresToInclude": "dofM,dofW,scheduleDepTime,scheduledArrTime,carrier, elapsedTime,originId,destId", "labelField": "delayed", "maxBins": "100", "maxDepth": "9", "seed": "12345", "featureSubsetStrategy": "auto" } }
Random Forest Predictor:
Properties:
- fileSetName : The name of the FileSet to load the model from.
- path : Path of the FileSet to load the model from.
featuresToInclude: A comma-separated sequence of fields to be used for Decision Tree Regression. If empty, all
fields except the prediction field will be used for prediction. Features to be used, must be from one of the following
type: int, long, float or double. Both featuresToInclude and featuresToExclude fields cannot be specified.featuresToExclude: A comma-separated sequence of fields to be excluded for prediction. If empty, all fields except
the prediction field will be used for prediction. Both featuresToInclude and featuresToExclude fields cannot be
specified.predictionField : The field on which to set the prediction. It will be of type double.
{ "name": "RandomForestRegression", "type": "sparkcompute", "properties": { "fileSetName": "random-forest-model", "path": "RandomForest", "features": "dofM,dofW,scheduleDepTime,scheduledArrTime,carrier, elapsedTime,originId,destId", "predictionField": "delayed" } }
Table of Contents
Checklist
- User stories documented
- User stories reviewed
- Design documented
- Design reviewed
- Feature merged
- Examples and guides
- Integration tests
- Documentation for feature
- Short video demonstrating the feature