(Liveblogged from SIGIR’09 Boston)
In the first session of the first day of SIGIR’09 Boston, three papers presented approaches at using implicit user actions (activities) for improving information retrieval precision. Here is a brief summary:
Queries from Activities in Daily Living – Takuya Maekawa (NTT Japan) presented work describing using RFID sensors embedded in devices in a home to automatically formulate queries to give just-in-time information pertaining to things people were doing in their homes. For example, he said, while making green tea you might want feedback on how long to optimally steep the tea; while cleaning your coffee pot you might benefit from knowing that vinegar removes stains effectively. He showed that using wearable accelerometer plus an RFID sensor, he could effectively segment times of interaction with certain devices around the home, and the identity of those devices; then he used 5 different strategies mixing query expansion and past interaction history to issue queries and re-rank results returned from a query engine. Interesting and fun paper.
Global Ranking by Exploiting User Clicks – Shihao Ji (Yahoo! Labs) discussed the use of multiple result hit visitations (what Jared Spool would call pogo-sticking between engine results) in the prediction of page relevance. He used user click stream history features with three different classifiers and compared their results. This was interesting to me as it provoked the idea that a sequence of article-clicks could contain information about how relevant the first hit(s) were towards serving the information needs of the user. An audience member asked an interesting question pertaining to whether it would be possible to disambiguate query intent with click-through patterns; the author was not able to answer this question.
- Segment-Level Display Time as Implicit Feedback: A Comparison to Eye Tracking – Georg Buscher (DFKI) conducted a study that compared textual passage display time (e.g., how long a particular paragraph was visible on the screen) to eyegaze fixation and baseline query enrichment for improving query retrieval performance (through term expansion). He showed that eyegaze and passage display time improved poorly formed queries equally well in his experiments, and did no harm to queries that were already effective. These findings are very exciting, as it suggests that we should conisder greater use of textual display passage time of personalizing search.