Back2Back Project

In this project, we aim to build a tool allowing us to easily launch back to back testing for machine learning algorithms. We work on three different machine learning libraries : Weka, Renjin and SparkML.

How to launch our project ?

After cloning the project, the application can be launched with the launch.java class (in the default package) :

git clone https://github.com/AlkaSaliss/Back2Back_Project.git

Our application requires to instantiate a Data object with all the parameters required to open the file and apply a model on it. We then need to instantiate a DecisionTree or RandomForest object. The idea is that with those two objects, we can run a decision tree/random forest on the same dataset with the 3 considered machine learning libraries. However, we can't guarantee the exact same model configuration for each library, because they require different parameters which don't necessarly match.

• Parameters about the data:
  • path : path of the file (String)
  • header : does the file have header or not ? (true or false in String)
  • targetName : name of the target Y, the variable to explain (String)
  • classif : what is the model's goal ? a classification or regression ? (Is the target categorical or continuous ?) (boolean : true if classification, false if regression)
  • numClasses : if the target is categorial, the user have to provide its number of modality (int)
  • hasRowNames : does the file have row names or id ? (should the first column be removed ?) (true or false in String)
  • sep, dec : what are the separator and the decimal in the file ? (String)
  • catNames : have the dataset some categorical features ? If so, the user have to fill a list with the names of the categorical variables. (ArrayList)
• Parameters for a DecisionTree:
  • maxDepth : maximum depth of the tree (needed for SparkML) (int)
  • minSplit : minimum number of instances required for splitting a node (int)
  • cp : complexity criteria (double)
  • gini : criteria used for splitting (boolean true for gini/classifcation, false for variance/regression)
  • minPerLeaf : minimum number of instances required in each final leaf (int)
  • maxBins : number of bins used when discretizing continuous features (int)
• Parameters for a RandomForest:
  • ntrees : number of trees to aggregate (*int)
  • mtry : percentage of features to select for each tree (double)
  • sampsize : percentage of individuals to select for each tree (double)
  • maxDepth : maximum depth for each tree (int)
  • maxBins : number of bins used when discretizing continuous features (int)
  • seed : seed to fix the randomness (int)
  • featureSubsetStrategy : type of random forest algorithm in SparkML (String, "auto" by default)
  • gini : split criteria (gini for classifcation problem, variance for regression) (boolean true for gini/classification, false for variance/regression)

We have chosen in our launch.java to compare the libraries on a decision tree for a regression task. The dataset used has been generated randomly for the purpose of the application. We wanted a dataset with a continuous target and with both continuous and categorical features.

Besides, our application also handle classification problem: we can apply a decision tree or a random forest on the iris dataset (located in data/iris.csv).

Authors

• ABOUBACAR ALKA Mahamadou Salissou
• FORMAL Thibault
• GIRARD Naomi
• KEITA Seydou