Research On Discrete Recommendation Based On Gumbel-Softmax

The recommender system,as an important tool for solving information overload,has been continuously developed and widely used in various fields.However,the amount of information is increasing,making the recommendation speed gradually slow,and it is difficult to meet the speed requirements of the current information age.Therefore,the development of recommender systems has fallen into a bottleneck.As a technology that supports fast retrieval,hash recommendation has become an effective solution to this problem,and the technology is being gradually developed and improved.There are two types of hash-based recommendation algorithms,one is hash algorithm two-stage quantization based,and the other is learning-based hash algorithm.Both algorithms have drawbacks.The former needs to be divided into two steps,each step is approximated,and a lot of information is lost,so that the recommended results are not accurate enough.The latter is that the designed recommendation model is inconsistent with the real target,and cause the recommendation result to be far from the target.Based on this,this thesis deeply studies the discrete recommendation model,and proposes the model Gumbel-Softmax based discrete collaborative filtering model(GDCF).The framework is a learning-based hash recommendation which can avoid losses due to two-stage learning.According to the specific recommendation problem,a corresponding objective function can be designed to match the objective,so that the recommendation result and the target result is matched.Consistent.Based on collaborative filtering,this thesis integrates neural network to do binary recommendation algorithm research.The main contributions are as follows:1.In order to solve the problem of gradient update that cannot do discrete values in neural networks,it is proposed to use Gumbel-Softmax trick to build the recommendation system framework GDCF.The framework supports learning the interactive relationship between the users and the target items in the Hamming space,and obtains the corresponding binary representation.In order to solve the discrete optimization problem,the Gumbel distribution sampling is used to fit the Bernoulli distribution to relax the constraints,while use Softmax to solve the gradient update problem,and finally implement the task of recommendation.2.The thesis mainly combines the neural network research recommendation model on the basis of GDCF,and puts forward two models,GDCF-N and GDCF-G,respectively.The GDCF-N model combines GDCF and Neural Collaborative Filtering(NCF)to construct a learning-based hybrid hash recommendation method.The GDCF-G model combines GDCF and the Generated Adversarial Network(GAN)to construct a hash recommendation method for generating an adversarial network.In the design comparison experiment on the classic public data set,the GDCF-N model and the GDCF-G model are compared with the classic hash algorithm.The experiment shows that the model can provide better recommendation services.