CSVLoader.loadDataSource() cannot load data without objective variable

problem

CSVLoader.loadDataSource() cannot load data which does not have objective (response) variable. For example:

// Load Titanic test data downloaded from Kaggle (https://www.kaggle.com/c/titanic)
ListDataSource dataource = csvLoader.loadDataSource(Paths.get("test.csv"),"Survived");
List<Prediction> predicts = model.predict(dataource);

this code results:

Exception in thread "main" java.lang.IllegalArgumentException: Response Survived not found in file file:/home/tamura/git/tribuo-examples/titanic/test_removed.csv
    at org.tribuo.data.csv.CSVLoader.validateResponseNames(CSVLoader.java:286)
    at org.tribuo.data.csv.CSVLoader.innerLoadFromCSV(CSVLoader.java:244)
    at org.tribuo.data.csv.CSVLoader.loadDataSource(CSVLoader.java:238)
    at org.tribuo.data.csv.CSVLoader.loadDataSource(CSVLoader.java:209)
    at org.tribuo.data.csv.CSVLoader.loadDataSource(CSVLoader.java:184)
    at org.tribuo.data.csv.CSVLoader.loadDataSource(CSVLoader.java:138)
    at TitanicSurvivalClassifier.main(TitanicSurvivalClassifier.java:74)

It’s a little inconvenient.

Solution

Add methods to load data without response variable name.

[Craigacp]

The CSVLoader is designed for purely numerical data, not the mixture of numerical and text data that comprises the Kaggle Titanic dataset, so this will probably error out later on anyway. For mixed numerical and text data you should use CSVDataSource and construct a RowProcessor which allows full control of how the features are extracted from the columnar data. CSVDataSource has a boolean outputRequired flag which if set to false will generate Examples which contain the unknown output for that output type (e.g. Label.UNKNOWN for classification tasks) if the output doesn’t exist in the csv file.

However you’re right that there should be a way to load a numerical CSV which doesn’t have a response. We’ll fix that.

[k-tamura]

Thank you! I will firstly implement a FeatureProcessor.