Research And Application Of Machine Learning For Recommender Systems

The recommender system is an effective solution for handling information overload and has been widely used in various fields in recent years.However,large-scale users and items slow down the online recommendation stage of the recommendation system,caus-ing inefficient bottlenecks for recommendations.The fast retrieval of hash technology has become an effective solution to solve the online recommendation efficiency bottle-neck.Currently,there are two types of hash-based recommendation algorithms.The first one is a two-stage quantization based hash algorithm;the second is a learning-based hash algorithm.The drawback of the former is that the quantization-based hashing method oversimplifies the discrete optimization problem,resulting in a large amount of informa-tion loss,which greatly affects the recommendation accuracy;and the shortage of the latter is that the discrete optimization model formulated for the recommendation system is inconsistent with the ultimate goal of the recommender system,besides,updating hash codes with the discrete coordinate method is expensive.In order to solve the above issues,this thesis conducts an indepth research based on the above two hash algorithms,the main contributions are summarized as follows:Firstly,to solve the problem of low recommendation accuracy caused by information loss in the quantization-based hash algorithm.This Thesis proposes a new quantization-based framework:Quantization-based Hashing?QBH?,which consists of similarity quan-tization and norm quantization.By this fined quantization method,the information loss is greatly reduced.This thesis propose two different preference prediction model based on hash codes from QBH:inner-product preserving hashing?QBH₁?and constraint-free preference preserving hashing?QBH₂?.Motivated by the preference denoted by the inner product in the matrix factorization,this thesis design an inner product preserving prefer-ence model based on hash codes from similarity quantization and norm quantization.Be-sides,by learning real-valued vectors from the matrix factorization model without norm constraints and then quantize the learned real-valued vectors into binary vectors via sim-ilarity quantization and norm quantization,this thesis propose the constraint-free prefer-ence preserving hashing.Finally,for the above two QBH methods,this thesis design a model to learn the optimal bits of norm quantization,which aims to reduce information loss caused by quantization stage,and hence improve the recommendation accuracy.Secondly,to furtherly reduce the information loss of the hashing algorithm and im-prove the accuracy of the recommender system,this thesis proposes a learning-based hashing framework:Discrete personalized ranking?DPR?,to learn hash codes by directly solving a discrete ranking-based objective based on implicit feedback.The ranking-based objective is consistent with the ultimate goal of the recommender system–providing a specific user with a personalized items'ranking list,thus DPR provides a more accurate recommendation.Besides,to get compact and informative hash codes,DPR adds balance and uncorrelation constraints on binary codes.Then,DPR transforms the original dis-crete problem into a series of mixed integer programming sub-problems by the alternating optimization algorithm,and learn the binary codes of users and items by the discrete coor-dinate descent algorithm.It's worth noting that DPR designs an objective consistent with the target of recommendation and then directly solves the discrete optimization problem,which greatly reduces the information loss,and thus DPR can achieve better performance compared with quantization-based hashing and learning-based hashing with rating-based objectives.Then,to solve the problem of data sparse and cold start in the recommender system,this thesis puts forward two learning-based hashing hybrid recommendation frameworks:Discrete Deep Learning?DDL?and Discrete Pairwise Hashing?DPH?.Specifically,DDL first pretrain items'deep representations from the content information,and then integrates the learned deep representations into a collaborative filtering objective,and thus the pro-posed DDL and DPH can overcome the problem of data sparse and cold start.By adding discrete constraints to users and items,and using the alternating optimization algorithm to finetune the deep network,and finally learning hash codes of users and items,DDL and DPH are efficient recommendation systems,which effectively solve the sparse data problem and the cold start issue.Finally,this thesis desgins a new hash learning algorithm based on integer program-ming:discrete ranking-based matrix factorization?DRMF?.The previous learning-based hashing frameworks are often solved by the discrete coordinate descent algorithm,which is a bitwise iterative optimization.Due to the non-convex and discontinuity of the dis-crete optimization problem,the coordinate descent algorithm is easy to fall into a locally optimal solution.DRMF designs a new discrete optimization algorithm that updates hash codes as a whole vector instead of one bit in each step.The new discrete optimization algorithm learns hash codes by solving a series of binary quadratic programming?BQP?subproblems,which is called block-wise method that can be used to avoid locally optimal solutions.Besides,self-paced learning are added into the learning procedure to learn bet-ter initialization solution.The self-paced learning furtherly avoids the algorithm falling into local optimal solution,and thus significantly improve the recommendation accuracy.