The repository consists of the implementation of neural collaborative filtering based on the course term paper.
The aim is to build recommendation systems for explicit feedback systems where the task is to recommend items to users based on a rating matrix which is a matrix where cell (i,j) corresponds to the rating user ’i’ gave to item ’j’. Based on these ratings, the system tries to predict unfilled ratings in the matrix i.e. predicts intelligently what rating value an item might get by a user. The traditional collaborative filtering technique for recommender systems depends upon the similarity values between two given user or item vectors in the rating matrix. But the technique in general suffers from some major problems such as sparseness of user profiles and scalability. [1] Deep neural networks have achieved success in various fields such as image classification, segmentation, speech recognition and and many other domains. The idea is to replace the inner product in the matrix factorization with a deep neural network. [2] The technique of neural collaborative filtering [2] has been used in systems based on implicit feedback which classifies a given item by all the users as either might be interested in or not interested in, the task in explicit feedback systems is to have techniques for getting the the exact ratings from 1 to 5 that user might give to an item. We implement two different approaches to use Neural Networks with collaborative filtering (implemented in NeuMF.ipynb) and matrix factorization and concatenation (implemented in embed_concat.ipynb) where we concatenate user and item embeddings and then develop a model over them.