A neighborhood-based method for mining and fusing positive and negative false samples

Qingwei Pan, Tiansheng Zheng, Shuqing Li, Zhiwang Zhang, Jiuchuan Jiang

Abstract: sIn the face of prevalent data bias and data sparsity problems, recommender systems still face issues with limited effectiveness and low accuracy. In this paper, a false sample mining and fusion approach based on neighbor information is proposed. Firstly, this paper proposes a method for identifying False Positive Samples (FPS) to address the issue of bias. Specifically, this paper combines neighbor information to mine True Negative Samples (TNS) from the data. Then FPS can be recognized since recommendation models can learn reverse user preferences from TNS. For these FPS, this paper presents three schemes to correct explicit ratings of TPS so that bias problem can be mitigated. Secondly, this paper proposes a method for mining False Negative Samples (FNS) to address the issue of sparsity. By establishing a correlation between positive samples of users and neighbors of negative samples, FNS can be recognized from negative samples. Since FNS can represent user preferences, incorporating them into the data can alleviate the issue of sparsity. Finally, in order to fully utilize FPS and FNS, this paper proposes an outer multi-round fusion method and an inner multi-round fusion method, which can address both data bias and data sparsity issues simultaneously. Through extensive experiments on real datasets, this paper verifies that the fusion method of FPS and FNS is indeed effective in alleviating issues of bias and sparsity. The method not only has high accuracy, but also has strong stability and robustness. In addition, the experiments suggest that addressing bias should take priority over dealing with sparsity when tackling both issues simultaneously.