Recommended Method For Low Exposure Items Based On Bias Calibration

The bias calibration recommendation method aims to alleviate the impact of exposure bias on the recommendation task,and through the definition and analysis of bias,the corresponding bias calibration strategy is formulated,and the rich and fine-grained item list is accurately recommended for the user,which is not only conducive to increasing user stickiness,but also crucial for the iterative optimization and innovative development of the recommendation platform.Starting from the reality and hot issues facing the industry,this thesis aims to calibrate the exposure bias in the recommendation system and improve the recommendation performance,and mainly studies the bias calibration based on graph representation learning and the bias calibration based on deep neural networks around the bias problem in the recall link of the recommendation system.The main contributions and innovations of this thesis include the following two aspects:(1)Aiming at the problem that the traditional collaborative filter recommendation algorithm tends to recommend high-exposure items and causes the bias amplification,a new bias calibration method that combines multimodal deepwalk and bias calibration factor(Mm DW-BC)is proposed.The algorithm first introduces the multimodal attribute features of the item as auxiliary information,and alleviates the problems of sparse interaction and cold start of low-exposure items by constructing a richer multimodal item graph with edges and nodes.Based on this,the improved Deep Walk algorithm proposes a new node transfer probability,which can learn more characteristics of uninterlaced items.To prevent exposure bias from being further enhanced,a bias calibration factor is designed to correct the value of the item’s contribution to the recommended score.Finally,the proposed model is applied to the Amazon and ML-1M datasets,and the experimental results fully demonstrate the necessity and effectiveness of this method to improve the recommended accuracy by showing the modeling exposure bias.(2)Aiming at the problem that the existing serialization recommendation system cannot well capture the dynamic evolution in user interaction,so that user decisions are highly susceptible to exposure bias interference,the multi-head self-attention sequence perception recommendation method oriented to interest drift is innovatively proposed.In order to effectively capture the dependencies between long-term and short-term user behaviors,LSTM deep network understanding and modeling of dynamic updates of user profiles are applied.Since the attention mechanism of focusing on the user’s single interest can no longer meet the user’s multi-faceted and fine-grained preferences,this thesis introduces a multi-head self-attention mechanism to model long-term interest,thereby assisting short-term decision-making.The model is applied to datasets with different sparsity,and experimental results show that this method reduces the drift of interest caused by random operations of users on the basis of accurately capturing sequence dependencies.