Analysis Of User Preference Modeling With Probabilistic Generative Models

With the exponential increasing of data,recommendation system is nowadays ubiquitous in various domains because it has ability to provide users with the information of potential interest.By analyzing the user's historical behavior information,we can effectively mine the user preference representation to achieve the recommendation on different situations and different tasks.Research on personalized recommendation not only has important guiding significance for users to obtain effective information,but also has important business value.Efficient recommendation can effectively enhance the market value of online services,and even has important significance for social development and national security.In addition,the research on the related theories and methods of recommendation system combines different branches of different disciplines,plays a promoting role in the development of different disciplines,and is also a typical representative of the combination of industry,university and research,which is conducive to promoting the interaction between industry and scientific research,and reflects the comprehensive advantages.While recommendation systems are so important,they also face many challenges.Although the number of users and items on the Internet platform is tens of millions or even billions,for a single user,their interaction behavior often only accounts for a very rare part of the potential interaction of large user items,which leads to serious data sparsity problems.In addition,even if users only interact with a small number of items,user preferences are often diverse due to different categories of items.Most of the existing recommendation methods cannot effectively portray users' diverse interests and preferences.Finally,the explanatory nature of the recommendation system is lacking in existing methods that focus only on the accuracy of the recommendation.However,explainable results can not only enhance user trust,but also provide direction for model designers to improve.Therefore,it is still the focus of the existing recommendation system research to effectively address the challenges of data sparsity while modeling a variety of user preference patterns and to achieve interpretable results.In this work,we focus on the sparsity of data,diversity of preferences and explanatory problems of models in recommendation data and models.We study the generative recommendation models from four aspects: main preference modeling,social preference modeling,diverse preference modeling and explainable preference modeling.The main contributions are as follows:1.Main preference modeling: The task of the generative recommendation model focuses on learning user preference representation,but the highly sparse user-item behavior information poses a challenge to user preference learning.In this work,we propose a deep generative ranking model DGR based on the characteristics of user behavior data.Based on Wasserstein auto-encoder,DGR combines the flexibility of the generative model with the strong feature representation ability of the neural network to effectively model sparse user behavior data.Specifically,by defining the Beta-Bernoulli generative process for sparse data,user preference modeling is achieved by combining the pointwise ranking mechanism for behavior data generation with the pair-wise ranking mechanism for the relationship between partial ordering of behavior data.Point-wise ranking mechanism is dedicated to reconstructing user preference behavior data,while the pair-wise ranking mechanism models preference by considering the positive and negative feedback information.In addition,the generalized error bound of the overall generated model is analyzed from the theoretical level to ensure the validity of the model.The experimental results on four real-world datasets show that the proposed DGR model can model sparse user behavior data well and effectively improve the prediction accuracy on cold-start user.2.Social preference modeling: The challenge of sparse user behavior information can be mitigated by effectively introducing auxiliary information.As a unique user information,social networks effectively portray the relevance between different users.In this work,we presents a probabilistic matrix factorization model for social recommendation,In SRMF,which combines indirect social relations,to effectively achieve a unified learning of users' social preferences and behavioral preferences.Specifically,we propose a PoissonBernoulli generative process to model the internal relationship of simple binary social relationships,in which the potential social strength among users is effectively portrayed by the discrete nature of Poisson distribution.The social relation generative process can be effectively integrated with the user preference data generative process.For efficient optimization,we develop a parallel graph vertex programming algorithm for efficiently handling large scale social recommendation data.The experimental results have shown that In SRMF has ability to mine the proper indirect social relations and improve the recommendation performance compared with the testing methods in the literature,especially on the users with few social neighbors,Near-cold-start Users,Pure-cold-start Users and Long-tail Items.3.Diverse preference modeling: While the sparseness of user interaction is exacerbated by the increase in large-scale user and item sizes,it also brings diversity of user preferences,especially when the item categories are rich.For user preference diversity modeling,we introduce adaptive learning mechanism to construct the optimal representation of different user groups from the perspective of traditional Bayesian generative models and deep generative models.Global preference modeling is introduced to mining user intrinsic diverse preference.Specifically,we propose an adaptive local matrix approximation recommendation model,ALo MA,for rating prediction task.User preference diversity is modeled by effectively combining adaptive local submatrix learning,optimal rank learning of submodels,and missing non at random mechanisms of sparse data.We propose a scalable inference algorithm for Gibbs sampling to infer the model efficiently.At the same time,in order to combine the ability of deep neural network to express the non-linear features,we propose a deep generative recommendation model,DGLGM,which combines global and local representations.By introducing a non-parameterized Dirichlet prior,user population adaptive partitioning can be effectively achieved.In order to alleviate the model error caused by the variational inference optimization algorithm,we introduce a local variational optimization strategy to effectively raise the lower bound of model evidence.Our models ensure the validity from both the theoretical level and the real practical level on datasets.4.Explainable preference modeling: The improvement of the accuracy in recommendation does not necessarily lead to the improvement of user satisfaction.The model can explain not only to make users more convinced of the recommendation results,but also to make the model designer understand the internal mechanism of the model and optimize the model accurately.In view of the existing traditional generative recommendation models and deep generative recommendation models,we propose an explainable generative mechanism oriented to data level and feature level.Specifically,in view of the low interpretability of the probabilistic matrix factorization model,we propose an influence-based recommendation model,In2 Rec,to measure the importance of users' historical behavior so that users' predictions can be interpreted and traced.In addition,for sparse discretized rating data,a discretized Gaussian distribution is proposed to effectively describe the fitting process of different rating values.By combining the missing not at random mechanism,the sparse data generative process is described.In2 Rec can be efficiently solved by the proposed iterative conditional mode algorithm.At the same time,for In DGRM,which is an explainable model with deep generative model,we propose a factor sparse mapping mechanism to effectively correlate the relationship between different hidden representation factors and item groups.In addition,a non-parameterized prior for user group partitioning not only effectively models the diversity of user preferences,but also facilitates the learning of disentangled representations.We have validated the model performance in terms of accuracy and explainability in real-world datasets.