Yao Ming, Panpan Xu, Huamin Qu, and Liu Ren
A Brief Introduction
The interpretability of machine learning models is becoming increasingly important for crucial decision-making scenarios. It is especially challenging for deep learning models which consist of massive parameters and complicated architectures.
ProSeNet (Prototype Sequence Network) is a sequence model that is interpretable while retaining the accuracy of sequence neural networks like RNN and LSTM.
ProSeNet is interpretable in the sense that its predictions are produced under a case-based reasoning framework, which naturally generates explanations by comparing the input to the typical cases. For each input sequence, the model computes its similarity with the prototype sequences that we learned form the training data, and computes the final prediction by consulting the most similar prototypes. For example, the prediction and explanation of a sentiment classifier for text based on ProSeNet would be something like:
Here the numbers (0.69, 0.30) indicates the similarity between the inputs and the prototypes. You may find such a framework is similar to k-nearest neighbor models. And yes, the idea of the model originates in classical k-nearest neighbors and metrics learning!
The architecture of the model is illustrated as in the upmost figure. It uses an LSTM encoder r to map sequences to a fixed embedding space, and learns a set of k prototype vectors that are used as a basis for inference. The embedding of the sequence is compared with the prototype vectors and produces k similarity scores. Then a fully connected layer f is used to produce the final output. Here the weight of the layer f assigns the relation between the prototypes and the final classes.
However, the model is still not interpretable, cuz the prototypes are vectors in the embedding space! Thus, we use a projection technique to replace the prototype vectors by its closest embedding vector during training, which associates each prototype vector with a "real" readable sequence. For more details, please check our paper listed below.
To appear in Proceedings of KDD 19. [preprint]
The code of ProSeNet will be hosted on github at here.
We are actively working on the code review to meet legal and copyright requirements of Bosch. The code will be release once the review is finished.
The major idea of the paper, and most of the experiments are done during Yao's internship at Bosch Research North America.