| Traditional news recommendation algorithms have been unable to meet user's browsing needs,especially in today's artificial intelligence era,how to improve user's browsing experience has become a research hotspot,so personalized news recommendation algorithms came into being.Compared with the traditional news recommendation algorithm,it considers more dimensional factors,which improves the interpretability of recommended content.Some algorithms introduce an external knowledge graph,which improves the diversity of recommendations,but there are the following two factors make it difficult to personalize the news recommendation algorithm.One is the characteristics of the news itself,such as its large number and wide coverage,fast update and high timeliness;the second is that it is not easy to unify user portraits when constructing user portraits.The user's behavior and interests need to construct different user portraits.Although the current mainstream personalized news recommendation algorithm has achieved good recommendation results,most of the algorithms ignore the picture information in the news,and these pictures have potential value.Because the information conveyed to the user by the picture is more intuitive and more likely to affect the user's reading interest,this paper proposes a model for reconstructing news tags and proposes an adaptive tag algorithm based on these new tags.The detailed research content is summarized as follows:(1)This paper proposes a model of reconstructing news tags(MRNT)by combining pictures and text in news.Considering that the number of original tags of news is small and rarely contains news picture information,and considering that these tags are usually written by humans,it can only represent some people's views and understanding of the news.This paper first introduces the open source image recognizer and word vector,followed by the basic knowledge of neural network and Word2 Vec model,and finally uses the original news tag as a reference to remove the low correlation features extracted from news pictures and text,and uses the remaining features as new tags for news.These new tags are the basis of the tag correlation graph and AT algorithm proposed later.(2)This paper proposes an adaptive tag algorithm(AT)that focuses on the shortterm interests of users.Due to the characteristics of news,the short-term interest of users is easily affected by news content.If the user's interest model is not well constructed,it will reduce the user's browsing experience.Therefore,this paper proposes an AT algorithm for how to construct the user's short-term interest model.The algorithm uses the user's feedback to filter out the tags that the user is currently most interested in.In order to filter out the tags that the user is most interested in,this paper scores each tag in the user's current set of interest,first calculates the specific score of a tag in the user's browsing history,then calculates the popularity score of the tag,and then calculates the weight score of the tag in the candidate news,followed by calculating the depth score of the tag.Finally,the scores of the first four are added up to be the final score of the tag,and then the tag with the highest score is the tag that is most interesting to the user.(3)Experimental results and analysis.Based on the performance evaluation criteria of news recommendation(F1,AUC,MRR),the method proposed in this paper is verified.When the recommended items are small,it has better recommendation results compared with other baselines. |
PDF Full Text Request