Exploiting memory cues in personal lifelog retrieval

• Authors: Yi Chen
• Venue: Proceedings of the 32nd international ACM SIGIR conference on Research and development in information retrieval
• Year: 2009

In recent years personal lifelogs (PLs) have become an emerging field of research. PLs are collections of digital data taken from an individual's life experiences, gathered from both digital and physical worlds. PLs collections can potentially be collected over periods of years and thus can be very large. In our research group, four researchers have been engaged in collection of individual one-year-long PLs data sets. These collections include logs of computer and mobile phone activities, digital photos, in particular passively captured Microsoft SenseCam images, geo-location (via GPS), surrounding people or objects (via Bluetooth), and various biometric data. The complex data types and heterogeneous structure of this corpus brings great challenges to traditional content based information retrieval (IR). Yet, the rich connections integral to personal experience offer exciting potential opportunities to leverage features from human memory and associated models to support retrieval. My PhD project aims to develop an interface to assist IR from PLs. In doing this I plan to exploit features in human memory, in particular the mechanisms in associative memory models. Previous studies in personal information re-finding have explored the use of generally well-remembered attributes or metadata of the search targets, such as date, item type/format, authors of documents [1]. There have also been systems which utilize associated computer items or real life events (e.g. [2, 3]) to assist re-finding tasks. However, few of them looked into exactly what types of associated items/events people tend to recall. I plan to explore associations among PL items, as well as their attributes regarding their role in an individual's memory, since I believe that some associations and types of metadata which are available and feasible for use, may have been omitted in existing systems; due to the methods used in previous research where the users' behaviour may have been guided by the searching or management tools available to them. As indicated by some information seeking studies (e.g. [4]), different search context, search motivation, or personal differences such as habits, may lead to varied recall of contents and information seeking behaviours. For this reason, I will also investigate: the influences on personal information re-finding behaviour of context, lifestyle, and differences in prior personal experiences of IR tools. Results from these studies will be used to explore personalisation in search, e.g. to dynamically increase the importance of geo-location in scoring of search results for subjects who travel frequently. As indicated by [4], people tend to make small steps to approach the targets they are looking for, rather than trying to do this in a single search with a "perfect query" comprising all of the relevant details, partially because of their trouble in recalling them. To relieve users from the heavy cognitive burden of recalling the exact target, and on the other hand to reduce the rate of inaccurate queries caused by false recall, my proposed interface will be based on browsing and recognizing, instead of traditional recalling and searching. For example, a user will be able to browse and narrow results by recognizing landmarks and estimating the target activities' time range from the user's digital or physical life [5]. An important issue in my work will be to consider the challenges of evaluating my work with only a very limited number of PL datasets. To partially address this issue, I am currently in engaged in a number of smaller scale diary studies of searching experiences for larger numbers of subjects.