Loading...
Please wait, while we are loading the content...
Similar Documents
Entropy-Driven Online Active Learning for Ranking Functions
| Content Provider | Semantic Scholar |
|---|---|
| Author | Weber, Julie S. |
| Copyright Year | 2006 |
| Abstract | We present a new algorithm for active learning within an interactive calendar management system that learns its users' scheduling preferences. This application imposes certain constraints on the active learning process. Most notably, arbitrary examples cannot be presented for labeling by the user; instead, labeling opportunities arise only when the system receives a request for a new meeting, and the only possibilities for presentation are valid schedules that incorporate the requested meeting. Previous work on active learning of preferences in this context has incorporated a single technique that is applied uniformly after every meeting request to select the solution schedules to present for labeling. The results of prior work show variation amongst the quality of the techniques in different contexts. In this paper, we develop an alternative: an online approach that determines what technique to use in response to each meeting request. To do this, our approach makes use of the notion of entropy, a measure of the diversity of the solutions that might be presented. Our results indicate that choosing an active learning technique dynamically through use of the knowledge of entropy among a set of solutions increases the speed of learning and arrives at a more accurate assessment of user preferences. |
| File Format | PDF HTM / HTML |
| Alternate Webpage(s) | http://www.eecs.umich.edu/~weberjs/QualsPaper.pdf |
| Language | English |
| Access Restriction | Open |
| Content Type | Text |
| Resource Type | Article |
