next up previous
Next: The user model manager Up: Pic1 Module Structure Previous: The user interface

The user model

As the search progresses, Pic1 is trying to guess which area of the search space deserves more attention. In a sense, it's hypothesizing about which image the user is really searching for. Therefore, each image Pic1 displays can be thought of as a representative of a hypothesis about the target image. The actual selection the user then makes provides both positive information about that particular hypothesis and negative information about the others presented. A variety of attempts to second-guess the user are possible; here are a few:

Naive Guessing:
A simple user model would merely use a weighted average of the points corresponding to the selected images to generate a `guess.' Each dimension in the guess would have a range of certainty determined by the deviations from that guess in the history of selections.
Dimensional Analysis:
A dimension analyzer should examine the history of images selected and not selected on each cycle and learn which particular dimensions seemed most important in the user's selections. Emphasizing these dimensions in the future (for example, by increasing their weight in the distance metric used by other user modelers and the search engine) should improve performance.

Vector Trajectory Analysis:
The user's target image may evolve over time, since the selection process itself could give the user more detailed ideas about ever smaller details in the target image. So we can imagine the user as `moving toward' it in vector space. This user modeler could take the history of past selections not simply as a sequence to be averaged, like the naive guesser, but as a `vector' to attempt to extrapolate toward likely future areas to explore. Such a scheme should allow dynamic, as well as static, information to be exploited.

User History:
Items chosen by the user in previous uses of Pic1, particularly recent ones, are likely to be the targets of subsequent searches. Retaining state information about past searches and presenting those images simplifies queries which are similar or identical to previous queries.

To increase flexibility and robustness, a variety of user models may be employed. In general, many different systems for generating hypotheses about the desired image based on the sequence of user selections and, in more sophisticated modelers, state information from previous sessions with this user and other users. A simple approach would permit each of the user modeling modules a certain number of images in each presentation. Those presenting images which are chosen by the user have their allotment increased by the user model manager, creating competition and increasing the quality of subsequent presentations.

The prototype user model consists only of the simple guesser; as discussed above. It maintains a current guess $\vec{g}$and a current range of certainty $\vec{r}$.For each dimension i, gi is initially set to the mean of the set of points, and ri is set large enough to include all the points. Let $\vec{s}$ be the vector corresponding to the image selected and passed in by the user interface. Over time, $\vec{g}$ is just a weighted average of the series of selected vectors with exponential decay. Range is a bit more subtle; for each dimension i, ri is initialized to a value just large enough to contain all instances in the database, and is updated by the equation:

next up previous
Next: The user model manager Up: Pic1 Module Structure Previous: The user interface
Gregory J. E. Rawlins