Semantic Explanation for Deep Neural Networks Using Feature Interactions

Given the promising results obtained by deep-learning techniques in multimedia analysis, explainability of predictions made networks has become important practical applications. We present a method to generate semantic and quantitative explanations that are easily interpretable humans. The previous work obtain such focused on contributions each feature, taking their sum be prediction result for target variable; lack discriminative power due this simple additive formulation led low explanatory performance. Our considers not only individual features but also interactions, more detailed interpretation decisions networks. algorithm is based factorization machine, calculates factor vectors feature. conducted experiments multiple datasets with different models validate our method, achieving higher performance than work. show including interactions generates makes them richer able convey information. examples produced visual format verify they plausible.