Joint Representation Learning With Heterogeneous Data For Personalized Recommendation

With the development of technology,both information producers and consumers are facing considerable challenges in the era of information overload,and recommender system plays a crucial role in guiding users to explore their actual needs.Matrix factorization based collaborative filtering(CF)algorithm is the most representative model in recommender system,whose basic idea is to make use of users' historical interactions.However,the performance of CF is easily affected by data sparsity issues.Since traditional retail and lifestyle services transfer to online platforms rapidly,application platforms in many fields have accumulated rich reviews and images,which can be used as the auxiliary data,providing new methods to handle the problem at the data level.On the other hand,model agnostic meta-learning(MAML)has been proposed in recent years,the paradigm focuses on adapting to new tasks with sparse labeled data at the model optimization level,which can be extended to new users and items with few interactions.Typically,existing works incorporate reviews to alleviate the data sparsity and cold-start issue to some extent.However,these studies have the following problems:(1)Existing efforts model the representation of users and items from static dimensions,neglecting the time signals and behavior patterns hidden in the user-item interaction history,which may fail to capture users' instant interests and items' temporal attributes.(2)These works follow the paradigm of splicing user reviews into a total document and generating context features by CNN or RNN.While the stitched documents have problems such as incoherent semantics and confusion of context information.(3)There is no effective framework that unifies recent behavior sequences and reviews.Therefore,we propose a novel unified model with dynamic-static features(UMDSF).Specifically,the proposed model extracts both temporal sequence and review features by two parallel feature extractors based on self-attention and a multi-head attention mechanism.Subsequently,an adaptive fusion module is utilized to combine the fine-grained representations for the downstream recommendation tasks.Extensive experiments on four real-world datasets demonstrate the superiority of U-MDSF and additional ablation studies verify the effectiveness of the components designed in the proposed model.Traditional recommendation algorithms work well with large-scale feedback information,but it is hard for them to approach cold-start issues.On one hand,Existing works have alleviated the problem at the data level by utilizing auxiliary information(reviews and images)as the features of users and items to some extent.On the other hand,recent meta-learning paradigm provides new ideas for mitigating the cold-start issues from the perspective of the model level.Given its ability to rapidly adapt to new tasks with scarce labeled data,or in the context of cold-start recommendation,new users and items with very few interactions.Combining the work of data level and model optimization level,we propose the meta-learning tower network based on heterogeneous data representation and meta-learning optimization(MLTN).Specifically,through the mutual attention mechanism between the latent factor of the text and the image,MLTN extracts the high quality part from the original data and eliminates the noise information.Moreover,model agnostic meta-learning(MAML)paradigm is integrated into MLTN's parameters optimization process,so that the model has the ability to rapidly adapt to new tasks with scarce labeled data.Extensive experiments on industrial datasets and public datasets show that MLTN achieves significantly better results in various cold-start scenarios compared to the state-of-the-art methods.