Research On Context-Based Personal Information Management

With the rapid development of the Internet and information technology, the data storedin the personal computer is growing at an unprecedented rate. According to statistics, from2000to2004, the number of files in personal computer increased by2times with an increaseof75%the average file size and an increase of seven times the capacity of the file system.How to effectively organize and manage these data are important issues for users. For thisreason, Personal Information Management receives growing attention. Personal informationis stored by individuals and used for their own, which is completely controlled by theindividual. Personal Information Management is that individuals manage persnonalinformation for reuse through the process of storage, organization and retrieval etc.A good understanding of a user’s working contexts provides the basis for improveddesktop information management, as well as for personalized service for informationretrieval. The task is defined as the actions that people aim to achieve a goal. In the field ofpersonal information management, user task is embodied as at least two documents tocomplete target. Many studies use contextual information associated with the file to improvethe capacity of task identification. However, these studies lack the comprehensive utilizationof the information about reading time, window switching, file content and metadata. Toaddress the issue, we propose a novel approach to identify user’s task with information aboutboth user behavior-based context and text-based context. The reading time and switchfrequency of files are taken into consideration, which construct the user behavior-basedcontext. We also utilize the information about file content and metadata, including file name,size, access date, creation date and modification date, to build the text-based context. Afterthe task identification problem converted to a multi-class classification problem solved bysupport vector machine (SVM) method with contextual information, we can infer which filesbelong to which tasks. The experiments show that the user behavior, especially reading timeand window switching, is an important factor to improve the ability to identify the tasks.Web search makes it possible for users to quickly find the information they need frommassive Web pages on the Internet. Meanwhile, traditional file systems with the ability of management and organization fail to provide retrieval service for a large amount of localfiles, which bring opportunities to use information retrieval technology for quick file search.Content-based desktop search tools are such systems that they index local files and returnranked result list after users issue queries via a search interface. The currentcontext-enhanced desktop search extends the function of content-based ones usingcontextual information for ranking. However, they do not take full advantage of thecontextual information on desktop, or lack efficient ranking mechanism. To address the issue,we propose a novel context-based desktop search framework Hsearch, which provideretrieval service through the collection of users’ information about reading activity and querybehavior on the user interface (UI) layer. After content-based search, the system re-ranks theresults with contextual information and learning-to-rank method. Compared with theinformation about user behavior collected in the file system layer, the UI layer informationcan accurately reflect the true intentions of the user, and reduce the background noisegenerated by the current application, operating system and background processes.While analyzing user behavior on desktop computers, the role of Web interaction israrely taken into consideration by personal information retrieval tools. To capture the benefitof contextual information on desktop computers while considering the impact of Webinteraction, we propose a novel ranking method combining data content, metadata of desktopitems, as well as user activity, which is implemented in Hsearch system. The contextualinformation is defined as the data about user activity, file content and metadata. According tothe information about window switching and reading time, we propose a new conceptiondocument-window switch tree for describing the dynamic behavior of users to access files.After content-only search, our method based on Ranking SVM method re-ranks the resultswith the contextual information. Using user’s information about UI layer and Ranking SVMmethod, our tool outperforms a state-of-the-art context-enhanced search system.