My Blog List

Tuesday, January 24, 2017

Week 4: Location-based and Preference-aware recommendation using sparse geo-spatial networking data

Title: Location-based and Preference-aware recommendation using sparse geo-spatial networking data, ACM SIGPATIAL GIS 2012.
Author: Jie Bao, Yu Zheng, Mohamed F. Mokbel

Background: Location-based social networking services (LBSNs), such as Foursquare, Yelp and GeoLife, allow users to leave their mobility and activity records. For example, an user can leave a comment for a restaurant, or check-in a museum. Actually, such location history encapsulates much knowledge about an user's interests and preferences, which can be further exploited to improve location recommendation systems.

System Overview: In this paper, they propose a location recommendation system which is location-based and preference aware. The location-based property take users' current location as an important aspect and only recommend points of interest that are near to him/her. Besides, the preference aware property takes care of target user's preferences as well as social knowledge mainly from local experts.

As shown in Figure 1, the overall system contains two major components - offline modeling and online recommendation. Offline modeling is a static module which precompute each user's preferences and local experts' knowledge. Online recommendation is a dynamic module requiring the realtime inputs of users' queries and current locations. Offline module provide precomputed results into online module to facilitate the efficiency of realtime recommendation.

Figure 1. System Overview.
In offline module, the system implements two tasks: (1) social knowledge modeling, and (2) personal preference modeling. Social knowledge modeling is to identify local experts by calculating people's expertise in each category of venue in each city. To achieve it, they build a user-location matrix and apply HITS (or Hypertext Induced Topic Search) to such matrix. HITS assumes that each user holds a score indicating its knowledge and each location is associated with an authority score to show its level of interest. Personal preference modeling aims to extract a user's preference from his/her visited location history. Here they use TF-IDF (Term Frequency - Inverted Document Frequency) method by taking location history as a document and categories as terms in the document. A weighted category hierarchy (WCH) is constructed to show an user's preference at various levels of granularity.

In online module, it first selects a set of candidate local expert and venues matching target user's preferences within a limited geographical range, and then infers a score of candidate locations based on the opinions of selected local experts. The realtime online recommendation is similar to collaborative filtering (CF), yet, much more efficient than CF, since the procedure of candidate selection stage avoids computing similarities with the full set of users in the city.

Results: They collect users' tips in New York City and Los Angeles from Foursquare. They take New Jersey users as target users to study the recommendation effectiveness and efficiency. The baseline approaches consist of (1) Most Preferred Category (MPC) recommendation [1], Location-based CF (LCF) [2] and Preference-based CF (PCF). As show in Figure 2, their proposed system outperforms competing baselines in precision and recall. Besides, their method maintains a relatively short processing time.
Figure 2. Performance in precision, recall and processing time.

[1] Y. Zheng, L. Zhang, X. Xie, and W.Y. Ma. Mining interesting locations and travel sequences from gps trajectories. In WWW, pages 791–800. ACM, 2009.
[2]V.W. Zheng, Y. Zheng, X. Xie, and Q. Yang. Collaborative location and activity recommendations with gps history data. In WWW, pages 1029–1038. ACM, 2010.

No comments:

Post a Comment