Mining Time-Aware Urban Living Styles via Latent Semantic Concept Analysis with Service Recommendation Application LAM, Wai Department of Systems Engineering and Engineering Management The Chinese University of Hong Kong A proposal for Microsoft Research Asia April 9, 2014 1 / 12
Introduction Technology has changed the way we live in urban centers. Two kinds of social environment - real world and online social networking world. Social networking sites such as Foursquare, Facebook, etc. tell others what one is doing, where one is, what meals one is about to have, and a plethora of other activities. In a way, people leave their “digital footprints” there. What does this mammoth information mean to data miners? Exploiting such vast source of human information via computational approaches would be very beneficial to sociologists, business corporations, data mining scientists, etc. April 9, 2014 2 / 12
Basic Idea about our Project Our framework can learn living style patterns of urban users based on latent semantic concept analysis using textual information in publicly shared heterogeneous sources of social network information such as user profiles and check-ins. We take into account temporal patterns such as the time of the day when a person visits a particular shop or check-ins a particular place of interest. Dynamic nature of the human lifestyle over larger periods of time. Recommends services to the user based on the user interests and living styles. April 9, 2014 3 / 12
Related Work LifeSpec (Yuan et al. 2013) mainly study human lifestyle patterns for a group of individuals in a large area, but it does not capture temporal lifestyle patterns and also does not recommend services to those groups of people. Han et al. 2013 presented a context-dependent search engine that considers the granularity of information. The authors developed a search engine to represent user context in a knowledge base context model. Liu et al, 2013 proposed a matrix factorization technique for Point-of-Interest recommendation based on topic and location, but do not exploit the textual information such as profiles and check-ins descriptions. Li et al., 2011 proposed an interesting economic model for product search which attempts to rank products with best value for money. April 9, 2014 4 / 12
Our Proposed Framework We propose a new framework for mining time-aware urban living styles and providing service recommendations based on the living patterns of a person or a group of people. We make use of publicly available data from location-based social networks. We exploit the textual information present in these resources and conduct latent semantic topic analysis to detect more useful latent concepts. Temporal nature in the data which could be the everyday lifestyle of an individual and what the individual does at a particular time, and also the changes in the lifestyle patterns over longer periods of time. We adopt nonparametric Bayesian topic modeling which detects latent concepts from textual information with the capability of automatically determining the number of latent concepts based on the characteristics of data. April 9, 2014 5 / 12
Some useful Concepts Footprint - A footprint of an individual consists of domain-specific tags which describe the behaviour of a person on a certain domain at a particular time. Living pattern - Some frequently co-occurring footprints. Lifestyle spectrum - A tree structured hierarchy summarizing the living patterns of an individual or groups of individuals. Lifestyle - A sequence of living patterns that is obtained by traversing the lifestyle spectrum tree from root to the leaf. Point-of-Interest (POI) - A specific point location that someone may find useful or interesting. April 9, 2014 6 / 12
Components in our Proposed Framework Organize the Points-of-Interest (POIs) and generate latent 1 semantic POI categories by analyzing the content from location-based social network sites. Detect the living patterns of an individual or a group of individuals 2 based on the time-aware latent concept detection from footprint data found in location-based social networks. Handle the long range temporal dynamics in the lifestyle spectrum 3 such as changes in patterns over seasons or months. The service recommendation based on the user’s or a group of 4 users’ past footprint patterns. April 9, 2014 7 / 12
P .I.s Previous Works Jameel, Shoaib, and Wai Lam. “An unsupervised topic segmentation model incorporating word order.” Proceedings of the 36th International ACM SIGIR Conference on Research and Development in Information Retrieval. ACM, 2013. Bing, Lidong, Wai Lam, and Tak-Lam Wong. “Robust detection of semi-structured web records using a DOM structure-knowledge-driven model.” ACM Transactions on the Web (TWEB) 7.4 (2013): 21. Jameel, Shoaib, and Wai Lam. “An N-gram topic model for time-stamped documents.” Advances in Information Retrieval (ECIR). Springer Berlin Heidelberg, 2013. 292-304. Jameel, Shoaib, Wai Lam, Ching-man Au Yeung, and Sheaujiun Chyan. “An unsupervised ranking method based on a technical difficulty terrain.” In Proceedings of the 20th ACM CIKM International Conference on Information and Knowledge Management, pp. 1989-1992. ACM, 2011. Lu, Chunliang, Lidong Bing, and Wai Lam. “Structured positional entity language model for enterprise entity retrieval.” Proceedings of the 22nd ACM CIKM International Conference on Information and Knowledge Management. ACM, 2013. April 9, 2014 8 / 12
Latent Semantic Point-of-Interest Construction The input data is obtained from online location-based social networks such as POI profiles of individuals or groups of individuals. We design a nonparametric hierarchical topic model to organize the POIs and generate a semantic POI category. Words in each document are footprints of the users. This component first generates the distribution over the general concepts, and each of these general concept is characterized by more specific concepts as we move down the tree. We will use a Nested Chinese Restaurant Franchise (nCRF) (Ahmed et al, 2013) prior in the Hierarchical Dirichlet Processes (Teh et al. 2006). April 9, 2014 9 / 12
Capturing Temporal Dynamics over Long Time Span High Level Framework We assume a continuous distribution over time. The model uses a first order Markovian assumption on the distribution over the change in the lifestyle spectrum over time. Two samples of the lifestyle spectrum at close times have a higher probability of sharing the same underlying model parameters than parameters drawn at times which are distant apart. η ψ g t g t +1 µ t µ t +1 µ t − 1 v t +1 v t d d θ t − 1 θ t θ t +1 α t − 1 α t α t +1 April 9, 2014 10 / 12
Service Recommendation Learn the daily social lifestyle pattern of an individual or groups of individuals. Recommend a service based on the pattern learnt. Best Deal Concept - Motivating Example When we learn that a person has lunch at KFC everyday for USD 50, but that day a 5-star hotel runs a promotion offer for USD 60 which is located close to the person’s location, then the system can recommend through targeted advertising some time before the person goes to the KFC about this promotion where the usual day lunch cost would have been USD 300. We will extend the “economic theory model” in order to accomplish such personalized recommendation system. April 9, 2014 11 / 12
References B. Li, A. Ghose, and P . G. Ipeirotis. Towards a theory model for product search. In Proc. of the WWW, pages 327-336, 2011. B. Liu, H. Xiong, B. Liu, and H. Xiong. Point-of-interest recommendation in location based social networks with topic and location awareness. Proc. of SDM13, pages 396-404, 2013. N. J. Yuan, F . Zhang, D. Lian, K. Zheng, S. Yu, and X. Xie. We know how you live: exploring the spectrum of urban lifestyles. In Proc. of the OSN, pages 3-14, 2013. J. Han, H. R. Schmidtke, X. Xie, and W. Woo. Adaptive content recommendation for mobile users: Ordering recommendations using a hierarchical context model with granularity. Pervasive and Mobile Computing, 2013. Ahmed, A., Hong, L., and Smola, A. Nested Chinese Restaurant Franchise Process: Applications to User Tracking and Document Modeling. In Proc. of the ICML, pp. 1426-1434, 2013. Y. W. Teh, M. I. Jordan, M. J. Beal, and D. M. Blei. Hierarchical dirichlet processes. JASA, 101(476):1566-1581, 2006. April 9, 2014 12 / 12
Recommend
More recommend
