The Spiral Of Silence And Its Application In Recommender System

Recently,the recommender system is a very popular research topic.The rating data handled by the recommender system is very sparse.Most of the recommender system models assume that the ratings are randomly missing,but this does not reflect the actual situation very well.This often results in biases in the learning of the recommender system model and reduces its recommended effect.In this paper,we consider the user's rating miss mechanism,empirically test the user's rating behavior,and then correct the deviation of the recommender system,according to the characteristics of the user rating behavior obtained,use the MNAR model framework to design a new recommender system model to improve the recommender system effect.First,we testify the spiral of silence theory in largescale cross domain real recommender systems.Our study exhibits an spiral process for a silent minority,people whose opinions are not supported by the majority(a.k.a the minority)are less likely to give ratings than majority opinion holders;the possibility of a majority opinion holder to rate is intensifying as the majority opinion becomes more dominant,while the possibility of a minority opinion holder to rate is shrinking.To verify the spiral of silence,we present methods for different scenarios where a user's willingness to rate is explicitly provided or absent.Furthermore,we study the factors which contribute to the formation of the spiral of silence,i.e.a user's assessment of opinion climate and the characteristics of a hardcore person.In addition,we utilize our empirical discoveries to guide the building of models.We proposed novel models which assume that the probability a user willing to express opinion is dependent on the rating and how the rating is divergent from the perceived opinion climate,we model the formation of community,opinion leaders,hardcore personality and item popularity to enhance recommendation performance.Finally,we test the model on the real recommender system dataset and verified the validity of model based on comparisons with other recommender system models.