Research On Personalized Recommendation Algorithm Based On Multimodal Data Source

With the rapid development of the Internet,online products and contents are exploding,providing users with more convenient shopping experiences.Meanwhile,overwhelming choices,such as clothes,movies,music,news,and books induce the information overload problem.To deal with this problem,recommendation system and search engine emerge.Compared to the search engine,recommendation system is more active to provide users with personalized services and products,creating business benefits.Benefit from the rapid development of mobile internet and continuous breakthrough in computing power,the production and transmission of unstructured data(picture,video,audio,text,etc.)has become more convenient,online contents and products can be presented in a more diverse manner.On the one hand,the number of products is increasing rapidly in the recommendation system.Under the large-scale product set,the traditional recall strategy will face the test of algorithm performance and resource consumption,how to improve the recall performance in limited time is very important for the implementation of the recommendation system.On the other hand,a lot of content information has not been used by traditional recommendation algorithms,leading to a lower utilization rate of information,also most of the applications of multimodal data are weighted addition or splicing in other fields of artificial intelligence,which means the common underlying information between multiple modalities is neglected to some extent.To solve these problems,this paper proposes a new recall algorithm and reorder algorithm to realize the personalized recommendation system based on multimodal data sources.Firstly,this paper implements the product recall algorithm under sparse data scenario based on deep learning,which mainly includes: 1)conducting data analysis and preprocessing based on the desensitized real data set of the operator platform to make it more suitable for this algorithm;2)achieving matrix restoration and completion under sparse data scenarios based on denoising stacking Auto Encoder(AE).The main idea is to learn and construct nonlinear implicit variable models based on partial observation data.The results show that the performance of denoising stacking Auto Encoder is better than that of the traditional collaborative filtering and Restricted Boltzmann Machine(RBM),which enables the recall task to be conducted accurately and efficiently,so as to reduce the complexity of the subsequent reordering phase.Next,this paper proposes an end-to-end refined ordering model based on deep learning and attention mechanism termed as Interest-Related Item Similarity model(Multimodal IRIS)to provide recommendations based on multimodal data source.Specifically,the Multimodal IRIS consists of three modules,i.e.,multimodal feature learning module,the Interest-Related Network(IRN)and item similarity recommendation module.The multimodal feature learning module adds knowledge sharing unit among different modalities.Then IRN learn the interest relevance between target item and different historical items respectively.At last,the multimodal data feature learning,IRN and item similarity recommendation modules are unified into an integrated system to achieve performance enhancements and to accommodate the addition or absence of different modal data.Extensive experiments on real-world datasets show that,by dealing with the multimodal data which people may pay more attention to when selecting items,the proposed Multimodal IRIS not only improves the information utilization,but also significantly improves accuracy and interpretability on top-N recommendation task over the mainstream methods.Finally,combined with the product recall algorithm based on deep learning proposed in this paper and the Interest-Related Item Similarity model based on multimodal data,we designed and implemented a multi-functional recommendation system,including similarity recommendation,association recommendation and personalized recommendation based on the most widely used public data set.The experimental results show that the personalized system has good performance and can be well adapted to the recommended scenario.It can also be applied to film,clothing,books,catering and other online content personalized recommendation.