Personalized Recommendation System Based On Category Tag Or User Sequence Behaviors

With the increasing number of registered users,the significance of the recommendation system for e-commerce,social network and video sites is becoming more and more important.The recommendation system not only helps users find a large number of alternative products in a short time when facing a large amount of commodities,but also helps the internet platforms to improve the user experience and maximize the total number of active users.However,the growth of registered users of many internet platforms has tended to be gradual,and the impact of different marketing strategies on the internet has gradually disappeared as the user's online time has grown quietly.While simple similarity recommendation and collaborative filtering algorithm can no longer meet the needs of users and the expectations of the internet platforms,it is important to deeply dig out the personalized recommendation of user information,which means meticulous and considering services and deeper interpretation of user value are the future direction of development of internet platforms.Modeling and understanding the interaction between users and items and the relationship between the projects is the core task of the recommendation.Texts,images,geographies,social relations,etc.which are integrated into the recommendation algorithm to generate personalized recommendation items,are often used to generate corresponding features.In addition,the Category labels have distinctive features,and the user's sequential behavior is more distinctive and personalized.The user-specific and product-related information are important auxiliary features for generating personalized recommendations.In recommendation system,Category labels are information carriers for recommending products of the same Category or making a recommendation by clustering of Category tags.However,categories are the labels of not only products,but also users.The user searches for the product according to his favorite Category tags.Actually,the fact that the Category architecture is an intermediary that collects the preferences of the user and the attribute of the product,is ignored by many researchers.By exploiting the integral entirety of “user-Category-item”,a large number of users' purchase histories will result in different collocations of the Category labels,and each user will search for a Category label related to the opinionated Category label set and then select the corresponding merchandise.Predicting user behaviors is always the ultimate goal of the recommendation system.Which clothing will be bought by users next,where he will go next,and which movie to watch following.The sequence of user behavior is the user's decision to switch between commodities,for example,based on the existing perfume,the user would like to match a small dress according to his own needs.This is the formation of the conversion of the user's buying psychology between commodities and the Markov Chain is a common model for the transition probability matrix between modeling commodities.In order to model user behavior sequence,the Markov Chain probabilistic transfer matrix between commodities is decomposed,which means decomposing the first-order Markov chain to modeling user's future behavior only depends on the previous behavior.And decomposing higher-order Markov chains is the user's future behavior from the front multiple consumer behaviors are jointly determined.However,most users have similar behaviors because their following psychological or similar preferences,and many of these behaviors have obvious relationships with each other and they are tied together for sale.Therefore,many of them are obviously potential bonds.For example,it is obvious that beer and diapers are related to bundling,but the binding relationship between perfumes and gowns is not obvious,while they often appear together in the same user's consumption records.From the perspective of Category labels and user behavior sequences,we explore the deeper level of personalization features on the basis of existing research.This article contributes as follows:1.In order to supplement the user's personalized information and mitigate the data sparsity of the recommendation system at the same time,we mine the circumstance that categories fill in the gap between users and items,propose the LICL model and rebuild the “user-Category-item” relationship in two steps:1)We propose a Category label weighting strategy for the hierarchical structure of Category labels,and build a weight matrix of "user-Category" and "Category-item",respectively,and decompose the above two matrices to obtain user-Category,Category,and product-Category characteristics vector.The inner product of user-Category and item-Category feature vectors is the user's score on the Category of the product.2)The score of the Category and the score of the user together form the rate prediction,which represents Category factor and the user factor together determine the recommended alternative commodity.In addition,the Bayesian a priori mitigation data of user preferences and product attributes subject to the user-Category and commodity-Category feature vectors are sparse,and Gibbs sampling is used to optimize the model parameters.In addition,user preferences and product attributes are subject to Bayesian priors of user-Category and item-Category feature vectors to mitigate data sparseness.And the BPR model is integrated into the LLIC to achieve recommended sorting accuracy and reduce the rating forecasting ambiguity which Gibbs sampling is used for optimization to automate model parameters.2.The strength of the latent binding relationship between commodities is separated by the behavior of the two commodities at the time of the user's buying behavior: the closer the two commodities are spaced or the two apples are directly adjacent to the purchasing order,the stronger the binding relationship is;The longer the interval between each product's purchase behavior is,the weaker the binding relationship is.Exploration of potential binding relationships is based on high-order behavior sequences.We propose that the LBS-Ranking model by introducing potential binding relationship feature vectors and integrate commodity review semantic feature extracted by neural network convolutional layer to rich model information.We train the inner of semantic features and potential binding relationship features as the commodity binding intensity that is affixed to the higher order Markov chain transition probability of the user behavior sequence as the weight of the product transfer relationship to predict the future behavior of the user.The LICL model and the LBS-Ranking model have been subjected to conventional evaluation criteria on a real Amazon dataset.The experimental results prove that our model far exceeds the current benchmark algorithm,especially on the problem of data sparsity.