Hybrid Recommendation Systems Via Exploiting Temporal Dynamics?Texts And Item Correlation For Cold Start Items

The traditional recommendation system(RS)can learn the user's potential personalized preferences and item's potential attribute characteristics through the historical rating records between users and items,so that it can help users to target content quickly and accurately.However,for the new items which have just entered the market,there is no or lack of sufficient historical rating record information,the traditional recommendation system often suffers from the classic cold start problem,which has a serious impact on the market introduction and promotion of newcome items.Therefore,it is of great significance to improve the cold start recommendation of items and mine the attributes of new products to implement personalized recommendations.The item cold start recommendation model needs to use the additional auxiliary information to mine the attributes of items.This thesis focuses on feature modeling of cold-start items through the description text information and item correlation.At the same time,we further bring in temporal dynamics that can be inferred from time-stamped historical scoring records and design scientific and effective item cold-start recommendations model.This thesis starts with the above three aspects,takes the accuracy of the rating prediction of the recommendation system as the evaluation metric,and makes an in-depth exploration of the item cold start recommendation.The main innovations and research results are as follows:(1)Statistical data shows that the scores of users and items all exhibit strong temporal dynamic characteristics,and the rating scores are closely related to the time status of users and items at the particular time.In order to give the most likely prediction score for an item at a particular recommendation moment,this thesis brings in the temporal dynamics into the item cold start recommendation model.Based on the different temporal dynamic characteristics of the user and the items,we model them with different granularity.(2)The items in the multimedia resource service platforms are usually have related item description texts.Such text information is a brief description of the item contents and reflects the content characteristics of the items.In this thesis,we select the convolutional neural network to extract the feature of item description texts,and inject them into the item cold start recommendation models.The extracted text features help to understand the content of the texts and characterize the features of cold start item,thus the models can overcome the limitations of the traditional recommendation system when dealing with cold start items.(3)For incomplete cold start items with a small number of rating scores,this thesis further explores the correlation between items.Based on basic word embedding model which exploits word co-occurrence frequency to build correlation matrix,we proposed an adjusted correlation computing method for our specific dataset and secondly,relate the standard matrix factorization model and the correlation matrix by the shared item latent factor.(4)This thesis proposes two new item cold start recommendation models TmTe-CCS and TmTeCo-ICS for different cold start scenarios.These two models incorporate features such as temporal dynamics,texts,and item correlation,which can effectively break through the limitations of item cold start on recommendation performance.This thesis evaluates the two proposed models on a real data set.The experimental results show that the two models proposed in this thesis can perform more effective item cold start recommendation than the existing excellent models.