Location-Aware News Recommendation Using Deep Localized Semantic Analysis

Abstract

With the popularity of mobile devices and the quick growth of the mobile Web, users can now browse news wherever they want, so their news preferences are usually strongly correlated with their geographical contexts. Consequently, many research efforts have been put on location-aware news recommendation; the explored approaches can mainly be divided into physical distance-based and geographical topic-based ones. As for geographical topic-based location-aware news recommendation, ELSA is the state-of-the-art geographical topic model: it has been reported to outperform many other topic models, e.g., BOW, LDA, and ESA. However, the Wikipedia-based topic space in ELSA suffers from the problems of high dimensionality, sparsity, and redundancy, which greatly degrade the recommendation performance of ELSA. Therefore, to overcome these problems, in this work, we propose three novel geographical topic feature models, CLSA, ALSA, and DLSA, which integrate clustering, autoencoders, and recommendation-oriented deep neural networks, respectively, with ELSA to extract dense, abstract, low dimensional, and effective topic features from the Wikipedia-based topic space for the representation of news and locations. Experimental results show that (i) CLSA, ALSA, and DLSA all greatly outperform the state-of-the-art geographical topic model, ELSA, in location-aware news recommendation in terms of both the recommendation effectiveness and efficiency; (ii) Deep Localized Semantic Analysis (DLSA) achieves the most significant improvements: its precision, recall, MRR, and MAP are all about 3 times better than those of ELSA; while its recommendation time-cost is only about 1/29 of that of ELSA; and (iii) DLSA, ALSA, and CLSA can also remedy the “cold-start” problem by uncovering users’ latent news preferences at new locations.