Profile Construction For Personalized News Recommendation

With the rapid development and popularization of Internet technology and smartphones,the Internet has become the most important and broadest way for people to obtain news articles.However,the amount of news articles published on the Internet every day is far beyond the processing capability of people.Hence,how to help users discover news articles with timeliness,novelty and usefulness becomes an urgent problem to online news sites.Personalized news recommendation is an effective solution.News profiles and user profiles are two important supports in the field of personalized news recommendation.Since it is hard to collect information of users' attributes due to privacy protection,a user profile is usually treated as a set of multiple news profiles.Therefore,how to construct news profiles with high quality becomes a key and difficult point in the field.Many previous studies employed keywords to build news profiles and seldom took the context semantics into account.However,the context semantics are an indispensable element when expressing the gist,emotional tendency and implicit information of news articles.Although some studies tried to overcome this shortage by employing ontologies,constructing or maintaining an ontology requires a long cycle,which violates the timeliness and novelty principles of personalized news recommendation.Since keyphrases can preserve the context semantics,and crowdsourcing can complete a time-consuming task within a short period,this thesis focuses on the problem of constructing profiles for personalized news recommendation by combining keyphrases and crowdsourcing.Our research contributions are as follows.(1)Keyphrases extraction from news articles using sequential pattern mining and entropy.A keyphrase is an ordered list of words that includes context semantics.Based on a common sense that a word does not repeatedly appear in an effective keyphrase,this thesis proposes a novel keyphrase extraction approach KeyRank,which consists two components:KCSP and PF-H.KCSP is a sequential-pattern-mining-based algorithm for searching keyphrase candidates,and PF-H evaluates three inherent properties(meaningfulness,uncertainty and uselessness)of keyphrase candidates by calculating their entropy.Experiments show that KeyRank performs very well in keyphrase extraction.(2)Keyphrase extraction and quality improvement by crowdsourcing.Since the news articles recommended by personalized news recommendation are usually short and include emerging entities with low frequencies,extracting keyphrases from such news articles requires strong ability of understanding the context.Although machine-based approaches can obtain such an ability with the help of knowledge bases,it is still not enough because of the low accuracy.Experts can achieve higher accuracy but requires a longer period.Therefore,this thesis employs crowdsourcing to complete the two tasks of"directly extracting keyphrases" and "improving the quality of existing keyphrases".Experiments show that crowdsourcing performs well in both tasks.(3)Crowdsourcing-based influence relevance assessment of news articles.Although the explicit content expressed by a news article does not match users' preferences,the implicit information delivered by the news article may have a great impact on the regions where users live and the industries that users are involved in.Recommending such news articles to users can help them make decisions in their daily lives.Since implicit information delivered by a news article cannot be obtained literally,and normally the influenced regions and industries are not explicitly mentioned,machine-based approaches cannot effectively assess the influence relevance of news articles.This thesis employs crowdsourcing to assess the influence relevance of news articles.Experiments show that crowdsourcing performs well in assessing influence relevance of news articles.