Research On Hybrid Latent Factor Model For Recommendation Based On Deep Learning Techniques

Nowadays,with the development of intelligent device renewal,people has gradually changed their lifestyles and some basic habits,and the Internet has become the most convenient and fast way for people to obtain information in their daily life.However,while the Internet brings convenience to people’s lives,the"information explosion"also follows.Recommendation system and Search engine,as two different information filtering techniques,can effectively alleviate the problems caused by information overloading to the network.Unlike search engines,the recommendation system is an implicit and active personalized information filtering technology that does not require a user’s display trigger（e.g.,active provision of retrieval keywords）.It takes users’historical data（such as users’rating scores and browsing-clicking records）as proof to reduce the volume of items in which users may be interested.For example,the recommendation system can help content providers（portals,self-media,merchants）to better locate potential customers,thus it can bring rich economic benefits to merchants;at the same time,the recommendation system can also help users to find target goods more accurately,save users’information filtering time and enhance users’satisfaction with content and loyalty to content providers.Statistics show that 80%of Netflix’s viewing records come from the recommendation system of the Netflix website^[1],and60%of the video hits on YouTube come from the YouTube home page recommendation^[2].In addition,the application of the recommendation system is not limited to the application of video websites,and we can sense the recommendation system around us everywhere.The recommendation system can recommend for people goods,music,restaurants,jobs,financial products and so on.It can be seen that the recommendation system has become an indispensable part of people’s life.Therefore,it is very meaningful and necessary to deeply analyze the core algorithms of the recommendation system.According to different recommendation strategies and practical application scenarios,the commonly used recommendation systems can be divided into three categories:collaborative filtering recommendation,content-based recommendation system,and hybrid recommendation system.Collaborative filtering recommendation aims to utilize the relationship between items,users,and a user and an item to make recommendations;The content-based recommendation system is based on the content which the user has shown interested in to make recommendations;Hybrid recommendation system can combine these two methods in four strategies to make recommendations.From the effect of the model,the hybrid recommendation system is better than the collaborative recommendation system,and the collaborative recommendation system is better than the content-based recommendation system.Among the collaborative filtering recommendation systems,Latent Factor Model is a kind of recommendation model with the best accuracy and is applied in many scenarios.As its name implies,“latent"is the core of the model.The model can map users and items into a shared latent vector space based on the rating matrix between users and items.The latent factor vector can be used to represent users and items respectively.Then,we can model the rating based on the implicit feature vectors.Because of its effectiveness,many recommendation systems are based on the idea of"latent factor".Although the latent factor model has many advantages mentioned above,with the increasing volume of data,the user-item rating matrix becomes more and more sparse,and the traditional latent factor model using rating matrix only is greatly affected by the sparsity.By deeply analyzing the internal mechanism of the traditional latent factor model,we find that there are still many aspects in which we can improve the latent factor model:1)The user’s real preference is diverse,while the traditional latent factor model constructs a fixed user’s implicit preference vector representation,which does not take into account the"personality diversity"feature of the user’s preference;2)In addition to the rating matrix,there are too many side-information related with items,such as text descriptions,images and etc.However,the traditional latent factor model does not take these side-information into account;3)There are complex relationships between users and items,but the traditional latent factor model only uses simple inner product to model the interaction between user preference vector and item feature vector when predicting users’rating.Aiming to improve the latent factor model from the above three aspects,we focus on how to build models integrating the user-item rating matrix with side-information.We propose three novel hybrid latent factor models.First,in order to solve the lack of"personality diversity"in user preference,we propose an adaptive deep latent factor model（Adaptive Deep Latent Factor Model--ADLFM）.ADLFM model combines the text description of the object with the rating matrix between the user items and uses the convolution neural networks to extract the global high-order semantic features to construct the user preference vector and the item feature vector.At the same time,the attention mechanism is used to construct an"adaptive"user preference representation,which can dynamically construct the user preference vector according to a specific candidate item,focusing on the preference related to the candidate item,so that the user preference representation can be more accurate and the rating prediction can be more accurate than the way traditional latent factor model has done.Next,a novel latent factor model（Deep Latent Factor Model with Fusion of Multi-modal Features--FLFM）based on multimodal feature fusion is proposed.Inspired by the cross-modal retrieval in the field of information retrieval,we design a novel multi-modal feature fusion module for items,which can effectively fuse the text description information of items with the image information of items,and learn more appropriate feature vectors of items,thus improving the accuracy of score prediction;Finally,aiming at the problem that the traditional latent factor model is too simple to model the relationships between users and items,we propose a deep hybrid latent factor model（Deep Latent Factor Model with Hierarchical Similarity Measure--DLFM-HSM）based on a hierarchical metric.The DLFM-HSM model replaces the user preference vector in the traditional latent factor model with a user preference matrix so that user preferences can be expressed more comprehensively.By hierarchical modeling users’ratings,DLFM-HSM can effectively distinguish the local preferences of users for different characteristics of items and enlarge these meaningful local preferences to accurately depict the relationship between users and items and improve the accuracy of rating prediction.Moreover,the hierarchical similarity measure method associated with hierarchical modeling does not contain any hyperparameters and therefore does not require an additional parameter training process.A large number of experimental results show that the hybrid latent factor models we propose in this paper can effectively alleviate the problems caused by the sparse data and improve the accuracy of rating prediction.