International Symposium on Ubiquitous VR 2007 1 Recommendation of preferable photo contents Jeong-min Yu, Sang-wook Lee, and Moon-gu Jeon Abstract — This paper presents a recommendation module II. RECOMMENDATION SYSTEM which provides preferable photo contents to the user from among A. Context descriptions the huge amount of photo contents in UMPC database. To A user preference is stored in the user profile which is extract the preferable photo to user, we use a hybrid approach that is combined with context and content-base approach using composed of two parts, human relations and keywords. The data mining concept. reason why we divide two parts in the user profile is that the Using two vectors concerned with the user profile and the photo human relations have relative important factors. A user metadata, we can calculate the cosine similarity between them. preference in the user profile is represented as a vector of The higher cosine similarity indicates that relevant photo content . is more preferable. The experimental results showed that our The photo metadata contains context and content information proposed algorithm has high potential to give high satisfaction to of the photo contents. Context information includes a time and the user. location information. As we get the context information, we assign the dependent information such as light status and Index Terms —user profile, data mining, similarity measure, season information. Content information describes the smart phone, UMPC (Ultra-Mobile Person Computers) background of focus of photo image. As the same manner with the user profile, we assign relative important factors to the photo metadata for the photo identification. I. INTRODUCTION R ECENT years, it have seen a lot of recommendation photo systems in a lot of fields. Most of them could show the list of photo contents based on the context information B. Data mining for recommendation of photo contents When one tries to seek specific photos from the UMPC database, one usually wants a system which automatically of photos [1][2]. recommends photos with considering the user preference. To However, when it is growing the technology of personalizing implementing the automatic recommendation system, we use mobile machine, it should be needed to personalize the data mining methodology which can extract useful recommendation system by considering a user preference [3]. information (preferable photo) from UMPC database. In This research focuses on the method how the photo contents details, we adopt a hybrid approach which is combined with can be recommended to user in terms of user preference in context and content base approach in the data mining personalize mobile interface (UMPC). Moreover, we consider methodology. Fig.1 shows the process of photo not only user preference, but also photo metadata which recommendation system using data mining concept. contain a context and content information based on ontology- based photo annotation [4]. In this paper, we adopt a hybrid approach which is combined with context and content User profile Photo meta data information with the user preference. ... Clean Transform Context Integrate Integration Load Context data Content data Contents Algorithms Generation (Similarity measure) F. Author is with the Gwangju Institute of Science and Technology(GIST), MS student, Korea (corresponding author to provide phone: 82-62-970-2410; Fax: 82-62-970-2384; e-mail: estevan119@gist.ac.kr). S. Author is with the Gwangju Institute of Science and Technology(GIST), ... Visualization Ph. D. candidate, Korea (corresponding author to provide phone: 82-62-970- User 2410; Fax: 82-62-970-2384; e-mail: yashine96@gist.ac.kr). T. Author is with the Gwangju Institute of Science and Technology(GIST), Fig.1 Process of the photo recommendation system Professor in Mechatronics, Korea (corresponding author to provide phone: 82- C. Hybrid recommendation approach 62-970-2406; Fax: 82-62-970-2384; e-mail: mgjeon@gist.ac.kr). International Symposium on Ubiquitous VR 2007 2 We define a user profile as a vector U = ( u1 , … un ), where ui get the high score of precision value, it means that the means a user preference represented between -1 and 1. The recommending photo system has high accuracy and high user value -1 indicates least preferred and 1 indicates most satisfaction. preferred. An example of user preference is as follows: User preference = {(smith, 0.5), (honeymoon, 0.9)}. We also define a photo metadata as a vector P = ( w1,…,wn ), ACKNOWLEDGMENT where wi is weight represented between -1 and 1. We assign this weight to photo metadata for representing the relative This research is supported by the UCN Project, the MCI 21 important factors. The value of 1 means more important Century Frontier R&D Program in Korea factors and -1 means less important factors. In order to find a preferable photo, the relationship between REFERENCES two vectors, photo metadata and user preference profile, must be investigated. As a method for investigating the relationship, [1] Naaman, M., Harada, S., Wang, Q.Y., Garcia-Molina, and H., Paepcke, A. Context Data in Geo-Referenced Digital Photo Collections. In Proc. ACM we adopt cosine similarity measure as below. Multimedia (MM 2004) (New York, NY, October 10-16, 2004). ACM Press, U ⋅P New York, NY, 2004, 196-203. Similarity (U , P ) = [2] O’Hare, N., Jones, G., Gurrin, C., and Smeaton, A., Combination of || U || × || P || content analysis and context features for digital photograph retrieval. in IEE European Workshop on the Integration of Knowledge, Semantic and Digital n ∑u w Media Technologies, (2005). i i [3] Yu, Z., Zhou, X., Zhang, D., Chin, C.-Y., Wang, X., Men, J., 2006. = i =1 Supporting Context-Aware Media Recommendations for Smart Phones. IEEE Pervasive Magazine 5 (3), 68-75. n n ∑u ∑w i =1 i 2 i =1 i 2 [4] A. T. Schreiber, B. Dubbeldam, J. Wielemaker, and B. J. Wielinga. Ontology-based photo annotation. IEEE Intelligent Systems, 16:66-74, May/June 2001. The following figure describes the process of cosine similarity calculation. (smith,0.5) (chanmi,0.5) (joshua,-0.5) (temple,0.1) (tower,0.3) (night,0.7) (honeymoon,0.9) (music concert,0.8) … Photo_ID 0.878 U ⋅P Similarity (U , P ) = Photo_ID 0.787 || U || × || P || Photo_ID 0.334 Photo_ID 0.122 … (20,30) smith chanmi music concert … Fig.2 The process of cosine similarity calculation. The format of output is Result = ( Photo_ID , 0.878 ). It means the user’s interest of the Photo_ID photo is 0.878. III. CONCLUSION The proposed recommending photo system incorporating context and content information with the user preference is preliminary crucial system in personalized Smart phone. As future work, for indicating the quality of recommending photo system, we will adopt the precision measure. As much as we