LSG: A Unified Multi-dimensional Latent Semantic Graph for Personal Information Retrieval

Huangfu Yang, Kuien Liu, Wen Zhang, Peng Zhou, Yanjun Wu, Qing Wang, Jia Zhu

Abstract: Traditional desktop search engines can merely support keywordbased search as they don’t utilize any other information, such as contextual/ semantic information, which has been commonly used in internet search. We observe that a user usually operates some files to complete a task related to a certain topic and organizes these files in some directories. Inspired by the observation, we propose an approach that considers three relations among personal files to improve desktop search, namely Topic, Task and Location. Each relation is derived from topics of files, user activities log and hierarchy of file system respectively. The heart of our approach is Latent Semantic Graph (LSG), which can measure the three relations with associated score. Based on LSG, we develop a personalized ranking schema to improve traditional keyword- based desktop search and design a novel recommendation algorithm to expand search results semantically. Experiments reveal that the performance of proposed approach is superior to that of traditional keyword-based desktop search.