A Study On Recommender System Based On Deep Learning

Just as information retrieval system, recommendation system was proposed to solve the problems brought by massive amount of data in the internet era, now it has become a new research area. The main aspects of researching on recommendation system include two parts:platform research and algorithm research, of course, the core part is algorithm and strategy researching. Taking use of other area such as artificial intelligence and machine learning, recommendation research developed rapidly; at the same time it also promote other research areas.In this article, we aim to study old and new recommendation algorithms, especially study the newest algorithm from machine learning area:deep learning. We also aim at applying this new deep structure of representation to recommendation system so that we could make recommendation system more humanity and intelligent.In order to achieve these goals, we did several works in many ways. We start our research from basic methods and old recommendation system:content based recommendation system and collaborative-filter based recommendation system, and we analysis the advantages and shortcomings of those systems. Then we state a very typical deep learning structure:Deep Belief Network, which is made up of several Restricted Boltzmann Machines. We use strict mathematic way to state and analysis this model and explain their physical meaning and their energy function from the view of thermodynamics.In the next, we will explain how a two layer Restricted Boltzmann Machine works when it applied to collaborative recommendation. Inspired by the application of the DBN, we proposed a new model which combine Deep Networks with traditional collaborative methods to put it used in recommendation system. Meanwhile, we try to maximum the model’s likelihood function using a new way called Contrastive Divergence, which run limited steps of Gibbs sampling to approximate the joint probability distribution of the model when it reaches its stable state, and we will derive the variational bound of the likelihood function to prove that this deep learning method is rational for training the deep network structure.At last, we complete such a brilliant algorithm and conduct a group of experiments on several datasets using different algorithms and models. We’ll see the deep network structure has shown its fabulous ability in feature detecting and feature learning, and has better property of resisting noise and effectiveness from the result of experiments.