Interpreting Black Box Deep Learning Models Using Data Mining Techniques

Modern machine learning models are used in a wide variety of real-world applications. However in order for them to be fully accepted in certain fields, the reasoning behind their predictions must be easily comprehensible. Models for whom the explanations behind their predictions cannot be comprehended are called "black boxes" due to their opaque nature. Opposite to them are "white boxes" whose predictions are more explainable. This work explores the idea of interpreting black boxes using a combination of data mining techniques and approximation using white boxes. The following white boxes were augmented using a data mining technique: 1)Bayesian Rule Lists 2)RIPPER algorithm 3)OneR algorithm and this white box was created using a data mining technique: 4)Lazy Support based Classifier the black boxes interpreted were 1)Feedforward neural network 2)Support-vector machine The following results obtained were the following: 1)The black boxes were approximated by the white boxes thereby fulfilling the surrogacy requirement of the interpretability method. 2)A existence of common kernel of prominent features were found for the white boxes. 3)The existence of similar rules were found between whiteboxes. The following conclusions are made: 1) This method of interpretability is feasible. 2) The existence of a common kernel shows that there exist some prominent features that contribute significantly to the predictions made by the black boxes, thereby helping to explain its predictions better. 3)Similar rules found for the whiteboxes based on the common kernel further proves that these prominent features are learned across all models.