Research On Recommendation Algorithm For Sparse Data

Recommendation system is an effective information filtering technology,which can filter out products,news,videos,music and other items in line with users’ preferences from massive information.Recommendation system is widely used in every field of life,so it is of great practical significance to improve the performance of recommendation system.Collaborative filtering algorithm is the main algorithm in recommendation system.It can infer the user’s preference by mining the interactive data between users and items.However,collaborative filtering algorithms face challenges when facing sparse data.On the one hand,the supervisory signal in the recommendation data is weak,and each user only interacts with a small number of items in the item library,resulting in a lack of positive samples and high-quality negative samples.On the other hand,there are long tail distribution characteristics and noise data interference in the recommended data.In order to achieve more effective recommendation in the recommendation scenario with sparse data,this study made improvements on the basis of the graph collaborative filtering algorithm.The main work is as follows:1.Aiming at the problem that the existing negative sampling strategy is insufficient to distinguish positive and negative samples,a collaborative filtering algorithm based on improved Softmax is proposed.Firstly,based on the collaborative filtering idea of "birds of a feather flock together and people flock together",the graph convolutional network is used to generate the embedded representation of target nodes and their similar nodes.By combining the characteristic information of the groups to which nodes belong,the positive sample data sparsity is alleviated.Secondly,the non-sampling training method based on the improved Softmax is proposed.The non-sampling strategy is used to expand the number of negative samples,and the loss function of the improved Softmax is used to adaptively mine difficult negative samples,so as to realize effective discrimination between positive and negative samples in the sparse data and achieve more accurate recommendation effect.According to the experiments on four public data sets,the algorithm has stronger adaptability to sparse data sets,and the recommendation performance has been significantly improved.2.To solve the problem of long tail effect and noise interference in interactive data,a recommendation algorithm of contrast enhanced graph neural network is proposed.A lightweight graph convolutional network was first used to encode high-level collaboration information between users and items into the embedded representation,and an improved Softmax loss function was applied in the training batch to construct supervised learning tasks to better distinguish between positive and negative samples.Secondly,self-supervision signals are generated by edge clipping to alleviate data sparsity,and comparative learning tasks are constructed between different item representations to learn the consistency and uniqueness of item representations,increase the recommendation opportunities of long-tail items,reduce the popularity bias of popular items and noise data interference,and realize more robust and personalized recommendation.Compared with the advanced algorithm SimGCL,in the Alibaba-iFashion dataset,the recall rate and NDCG of the proposed algorithm increased by 6.1% and 5.1% respectively,which further proved the effectiveness of the algorithm.3.In order to solve the sparsity problem of user and item interaction,a knowledge graph recommendation algorithm based on alignment and dispersion is proposed,which uses knowledge graph data to assist recommendation.The algorithm first uses graph attention networks to model knowledge graphs to generate embedded representations of items.Next,the graph convolutional network is used to model the user-item interaction graph,and the knowledge graph information and interaction graph information are fused into the final embedded representation to alleviate the data sparsity.Finally,the embedded representations are aligned and dispersed on the hypersphere to distinguish between positive and negative samples and mitigate popular bias,improving the accuracy and personalization of recommendations.Experiments on Yelp 2018 and Amazon-Book data sets show that compared with the advanced algorithm KGCL,the recall rate of the proposed algorithm increases by 2.5% and 4.3% respectively,and NDCG increases by 3.7% and 6.8% respectively,which proves the effectiveness of the proposed algorithm.In summary,the recommendation algorithm proposed in this paper is optimized from the directions of positive and negative sample differentiation,contrastive learning and knowledge graph,etc.,and achieves more accurate and personalized recommendation in the recommendation scenario with sparse data,which has wide application value.