Study On Small Sample Problem In Review-aware Recommendation System

In the era of big data,recommendation system is an effective tool to help people alleviate information overload.As effective supplementary information,review text is widely used to improve the performance of recommendation systems.However,the existing review text based recommendation system(review-aware recommendation system)only focuses on how to learn more feature information from the review text,and ignores the extremely serious small sample problem.Existing recommendation systems are mostly using only the observed user-item interaction information(labeled sample)to train the model,then a large number of unlabeled samples,and user interactions on the number of items is far less than the total number of items,the model is difficult to achieve ideal performance.How to effectively solve small sample problem is the fundamental problem to improve the performance of review perception recommendation.In view of this,semisupervised learning paradigm and self-supervised learning paradigm are introduced in this paper to alleviate small sample problem in the review-aware recommendation system by using the widely existing unlabeled data in the recommendation system and effectively mining the rich information contained in the original data.The innovative achievements made in this paper include the following two parts:A semi-supervised review-aware recommendation algorithm Co FM is proposed.In this method,semi-supervised learning is introduced into the review-aware recommendation system,and small sample problem can be alleviated by using a large number of unlabeled samples reasonably.Co FM uses two factorization machines(FM)as base learners.In particular,the two base learners in this paper are divided into user side and item side and use different review text information respectively to increase the difference between the base learners.In the course of training,the labeled samples were divided into training set and validation set.Then use two set of training samples to initialize the two base learners,each learners to forecast the unlabeled samples and generate pseudo label samples,at the same time,integrating with the result of prediction confidence authentication,select small samples of the pseudo scalar of credibility high sample to join another learning training sets and used to update the model parameters;The process is repeated until convergence occurs.In this paper,extensive experiments on three public data sets are carried out and compared with many review-aware recommendation algorithms,which fully verify the superiority of the proposed algorithm.A self-supervised review-aware recommendation algorithm SRMA is proposed.In this work,self-supervised learning paradigm is considered to fully mine the inherent correlation information of the data itself,so as to enhance the overall performance of the model.In this paper,the design is based on the two-tower model.The two-tower model is generally composed of two sub-networks on the user side and the item side.The two networks have the same structure but are independent from each other and only fuse in the final prediction.Taking the user side as an example,this paper sets up two parallel networks with the same structure but different inputs on the user side to construct different views on the user side.Then the self-supervision task is designed to make the similarity of the same user’s review embedded representation in different views larger,and the similarity of different users smaller,so as to enhance the network’s representation learning ability.At the same time,due to the lack of information exchange between the two towers in the traditional two-tower model,this paper introduces the mutual attention module as a bridge for information exchange between the user side and the item side to help the model more accurately discover high-information review text and improve the information extraction ability of the model.In this paper,a large number of experiments are carried out on three real amazon data sets to prove the effectiveness of the proposed algorithm.