| With the emergence of the Internet,in the era of information explosion,it is difficult for people to find the information they want.In recent years,recommender systems have become a widely concerned research area in academia and industry,capturing users’ feedback to understand their intentions and recommend items that may be of interest to them.Methods based on deep learning use deep neural networks to convert user interests into embedding vectors,and calculate the similarity with item embedding vectors in vector space to complete the recommendation task.Recently,the user’s multi-interest driven methods have shown better performance.These methods usually learn multiple embedding vectors from users’historical interaction sequences to capture users’different interests.However,existing methods only extract the interests of respective subsequences independently,without considering the global correlation of the entire interaction sequence.This may fail to capture users’ inherent preference for generalization of potential interests and inevitably results in recommended items that are too similar to historical behavior.This paper aims to propose novel methods from the two perspectives of modeling scale and supervision signal mining,so that the two complement each other and promote the development of recommendation tasks.The main work of this article involves:1.The existing multi-interest driven methods are mainly based on subsequences of user interaction sequences,which independently extract each interest and ignore the global correlation of the sequence,which may not accurately capture the inherent preferences of users.This study proposes a new Dual Scale Interest Extractor(DSIE)method,which can not only model user interests at the local scale,but also consider user inherent preferences and fine-grained multiple interests at the global scale.Specifically,DSIE explicitly models the inherent preferences of users at the global scale by focusing on their complete interaction sequences;At the local scale,a multi-interest extraction layer was designed to obtain finegrained user interests through intention prototype clustering.This article further develops an interest aggregation module that actively integrates multiple interests of users based on their inherent preferences to predict their current intentions.Therefore,DSIE can not only more accurately estimate the user’s current intention,but also generate a more diverse recommendation list beyond the product categories that the user has recently interacted with.The experimental results show that compared with existing methods,DSIE has improved both recommendation accuracy and novelty.This article also conducted in-depth exploration of the performance of DSIE through ablation experiments and parameter sensitivity experiments.2.In response to the challenge of insufficient supervised training of labeled data when modeling user inherent preferences,this study proposes a self-supervised learning framework.This framework is built on top of the dual scale interest modeling framework,incorporating self-supervised signals into the inherent preference modeling at the global scale.Due to the effective differentiation of multiple user interests in dual scale modeling,it can effectively solve the problem faced by some current selfsupervised learning work,which is to force different interests to be constrained at similar positions in the encoding space.Specifically,selfsupervised learning achieves the training of interest embedding vectors by designing a similarity comparison task,and proposes three data augmentation strategies based on the characteristics of recommendation tasks.The experimental results indicate that the dual scale interest modeling method combined with self-supervised learning can further improve the recommendation performance.In addition,this study also used the t-SNE algorithm to reduce the dimensionality of the learned embedding vectors and visualized them to more intuitively demonstrate the effectiveness of self-supervised learning. |
