Research On Cross-domain Sequential Recommendation Based On Adversarial Learning

Compared with the traditional recommendation system represented by the collaborative filtering algorithm,the sequential recommendation system has gradually become a research hotspot of the current recommendation system by virtue of its advantage in effectively capturing the time-series correlation between the user and the item interaction.Among them,the cross-domain sequential recommendation algorithm introduces auxiliary information from other domains through knowledge transfer,which can effectively deal with the data sparsity issues and cold-start problems that are common in recommendation systems,and has received extensive attention in recent years.However,most of the existing cross-domain sequential recommendation methods are only based on overlapped user sample spaces for training,while ignoring the objective fact that overlapped users account for a relatively small proportion in the real scene.This limitation leads to a significant distribution difference between the training space and the inference space in practical applications,which deteriorates the final recommendation performance.In addition,these methods can not effectively recommend non-overlapped users,and has high requirements for the sample quality and quantity of overlapped users in actual application scenarios,which greatly limits the scope of application of the model.In order to solve the above problems,we propose a Cross-Domain Sequential Recommendation based on Adversarial Learning(CDSR-AL).The model is composed of a generative adversarial network based on an auto-encoder and a feature extractor based on a self-attention model.The former learns the long-term preferences of users across domains,and the latter aims to model users' instant interests.Specifically,the generative adversarial network first utilizes an auto-encoder to learn user representations based on the interaction sequences of the source domain and the target domain,and then reduces the distribution differences between the user representations in these two domains through adversarial learning.By effectively migrating to the target domain,the model obtains better feature representation and generalization capabilities.In addition,considering the impact of the user's dynamic instant interest on the recommendation,the model also applies a self-attention based feature extractor to model the user's recent interaction sequence,which assigns different weights to items based on their importance.Finally,a sequential recommendation model is established based on the user's long-term preferences and instant interests to provide users with more comprehensive and accurate next item recommendations.To evaluate the performance of the proposed model,we conduct extensive experiments based on four datasets published by Amazon.The results show that the proposed model performs better than the state-of-the-art sequential recommendation models in terms of Normalized Discounted Cumulative Gain and Hit Rate indicators,which verifies the effectiveness of cross-domain knowledge transfer and user instant interest extraction.In addition,we have designed a sequence recommendation framework and a sequence recommendation system for video recommendation.The system has now been applied to actual engineering projects,allowing workers to more conveniently and effectively provide better quality recommendations to video platform users.