Research On Context-aware Mobile User News Preferences Acquisition And Recommendation Algorithms

With the popularity of mobile devices and the quick growth of the mo-bile Web, more and more users access to news via their mobile devices. Given this status quo, it is the main research task of context-aware mobile news rec-ommendation, how to acquire news preferences and recommend news to im-prove rec,ommendation performance and mobile user satisfaction according to the context awareness of mobile news. Although conventional news preference acquisition and recommendation for web users have achieved great success in both academia and industry, differently, mobile users' news preferences and mobile news recommendation are usually affected by various, personalized and dynamic contexts; it is difficult to apply the web news preference acquisition and recommendation algorithms to the context-aware mobile news preference acquisition and recommendation directly. Therefore, this thesis focuses on how to acquire context-aware news preferences of mobile users for news recom-mendation according to various contexts (e.g., geographical context, temporal context and social context) in mobile environment. The main contributions of this thesis are as follows.1) We propose three algorithms for location-based mobile users' news preferences acquisition and mobile news recommendation. Users can now browse news wherever they want, so their news preferences are usually strong-ly correlated with their geographical contexts. The location-based mobile news recommendations can mainly be divided into physical distance-based and ge-ographical topic-based approaches. As for geographical topic-based location-aware news recommendation, ELSA (Explicit Localized Semantic Analysis)is the state-of-the-art geographical topic model: it has been reported to out-perform many other topic models, e.g., ESA (Explicit Semantic Analysis) and LDA (Latent Dirichlet Allocation). However, the Wikipedia-based topic space in ELSA suffers from the problems of high dimensionality, sparsity, and re-dundancy, which greatly degrade the recommendation performance of ELSA.Therefore, to overcome these problems, we propose three novel geographical topic feature models, CLSA, ALSA, and DLSA, which integrate clustering, au-toencoders, and recommendation-oriented deep neural networks, respectively,with ELSA to extract dense, abstract, low dimensional, and effective topic fea-tures from the Wikipedia-based topic space for the representation of news and locations. Experimental results show that our three algorithms all greatly out-perform the state-of-the-art geographical topic model, ELSA, in location-based mobile news recommendation in terms of both the recommendation effective-ness and efficiency; while DLSA with recommendation-oriented deep neural networks achieves the most significant improvements. Specifically, the three algorithms can also remedy the "cold-start" problem by uncovering users' la-tent news preferences at new locations.2) We propose two algorithms for location-aware mobile users' personal-ized news preferences acquisition and personalized mobile news recommenda-tion. Mobile users' news preferences are usually related to their geographical contexts. Consequently, many research efforts have been put on location-aware mobile news recommendation, which recommends to users news happening n-earest to them. Nevertheless, in a real-world context, mobile users' news pref-erences are not only related to their locations, but also strongly related to their personal interests. Therefore, we propose a hybrid algorithm called location-aware personalized news recommendation with explicit semantic analysis (LP-ESA), which recommends news using both the users' personal interests and their geographical contexts. However, the Wikipedia-based topic space in LP-ES A suffers from the problems of high dimensionality, sparsity, and redundan-cy, which greatly degrade the performance of LP-ESA. To address these prob-lems, we further propose a novel algorithm called LP-DSA (Location-aware Personalized News Recommendation with Deep Semantic Analysis) to exploit recommendation-oriented deep neural networks to extract dense, abstract, low dimensional, and effective feature representations for users, news, and loca-tions. Experimental results show that LP-ESA and LP-DSA both significantly outperform the state-of-the-art baselines. In addition, LP-DSA offers more ef-fective news recommendation with much lower time cost than LP-ESA.3) We present an algorithm of multi-context-based mobile users' news preferences acquisition and mobile news recommendation. Mobile users' news preferences are usually affected by various contexts, e.g., geographical context,social context, news recency and so on. However, the social context utilized in a lot of research work is virtual, not the real social relationship between users,which degrades the recommendation performance to a great extent. In addition,the news recency is usually measured by time filtering or time modeling with a manual setting threshold in most related methods, but they don't consider the effects of news recency on mobile users' news preferences, and manual set-ting threshold will bring certain errors into recency measurement. Therefore,to solve the aforementioned problems, we propose to infer the real friendship between mobile users according to their context-aware telecommunication be-haviors, and obtain the news recency based on clickthrough data objectively.Finally, the social contexts of mobile users, users' interest similarities and news recency are integrated for mobile news recommendation. Experimental results demonstrate that our algorithm significantly enhances the recommendation per-formance and news recency in multi-context-based mobile news recommenda-tion.4) We propose an algorithm for novelty-based mobile users' news prefer-ences acquisition and mobile news recommendation. Personalized recommen-dation systems usually recommend to users the most related news according to their preferences learned from their history data. However, in the news do-main, there is a wide range of news sources, it is quite common to see more than one news article about an event, resulting in the redundancy problem in news candidate set. Given this status quo, if offer news recommendation mere-ly based on users' history preferences, the news articles have been read by users will rank top in the recommendation list. Thus, novelty detection is important to achieve a good personalized news recommendation. However, most news novelty detection approaches use either geometric distances or distributional similarities, which need to compare an incoming news article with all the pre-viously read ones. This is very time-consuming, and can't meet the need of real-time response of mobile news recommendation. Therefore, to overcome the aforementioned problems, we exploit a topic model, LDA, to extract the latent topics from news articles, where news articles are seen as samples and their latent topics are sample attributes. Then, the rough set and information entropy theories are applied to news samples to obtain weight for each attribute,based on which the total information of each news article is measured; so the novelty of a given news article regarding the previously seen ones of a given user can be measured by the differences between their total information rapid-ly. We further propose a matrix factorization model regularizing with obtained news novelty and users' interest similarities to recommend news. Experimental results show that our algorithm greatly improves the efficiency of novelty de-tection, the recommendation performance and news novelty in novelty-based mobile news recommendation.