Research On Joint Learning Method For Sequential Recommendation

In recent years,data on the Internet has shown an explosive growth trend,which makes it difficult for users to quickly find the information they really need.Recommendation System plays a pivotal role and brings huge commercial value.Due to the natural sequential characteristics of user interactions,Sequential Recommendation(SR),as an important branch of recommendation system,has become a research hot-spot in the fields of information retrieval and natural language processing.SR mines the user’s preferences behind large-scale sequences and aims to obtain the user and item representations so as to recommend the next item to the user.One of the key challenges of SR is how to obtain the accurate user and item representations.Recent works most use the next item prediction as a supervision signal to train the recommendation model.Under real application scenarios,the interaction sequences often have huge sparsity due to the huge scale of items.At the same time,user interaction sequences may be driven by complex preferences.Only using these sparse interactions and the single supervision signal introduces noise to model learning.Joint learning can model multiple tasks simultaneously and can improve the performance of each task.Using joint learning can introduce auxiliary data reasonably and improve existing training methods,which can break through the bottleneck of current tasks.Based on the requirements of real recommendation scenarios,this paper studies the joint learning method for SR and focuses on the following three works:(1)Aiming at the problem of sparse data in the single domain,this paper conducts research on joint learning for shared-account cross-domain sequential recommendations.It proposes a novel cross-domain shared-account sequential recommendation task,and a parallel information sharing network,which can recurrently extract and share useful information between two domains.It performs joint learning on SR tasks on both domains.It constructs a new dataset HVIDEO from real-world smart TV watching logs.Experimental results show that the model can improve the recommendation performance in different domains.(2)Aiming at the problem of inadequate interpretability under the cross-domain scenario,this paper conducts research on joint learning for mixed information flow for cross-domain sequential recommendations.It proposes a mixed information flow network.It introduces a knowledge graph between cross-domain items,which is used to help transfer the user preference on the mixed information flow composed of behavioral flow and knowledge flow.It performs joint learning on SR tasks on both domains.Experimental results show that it improves recommendation performance in different domains.(3)Aiming at the problem that the user preferences cannot be mined by the simple supervision signal of next item prediction,this paper conducts research on joint learning for preference editing based sequential recommendations.It designs a novel preference editing mechanism to mine the supervision signal between sequence pairs.It consists of the preference discrimination and preference recombination,which is used to discriminate the common and unique preferences between sequences,and recombine the sequence representations to introduce new supervision signals.It performs joint learning on preference editing and SR tasks.It conducts experiment on real-world e-commerce datasets.Experimental results show that the model significantly outperforms state-of-the-art sequential recommendation methods.