Next: Adapting to Dynamic Up: Specializing to User Previous: Experiment 2

Results

The results are shown in figure . The graph is a precision vs. recall graph at different time points. The graph does show an improvement over time. However, this time around, the improvement is much slower. This is primarily because the previously learnt ``Chinese'' terms are still in the profile which affects the system performance from the ``Russian'' point of view.

Another potential problem is that in the transition period, the profile is neither specialized to Chinese articles nor Russian articles. Besides, there is an implicit ``AND''ing of terms - an article which refers to both China and Russia will be rated higher than one referring to either one alone. This might not be what the user desired.

When an empty initial profile is used, the system converges relatively quickly to the desired set of interests. However, if the initial profile has some terms which are not related to the new user interests, then the profile takes longer to converge. This is because the profile needs to unlearn the terms which are already in there, but are of no use.

As can be seen from the two preceding experiments, a profile is capable of narrowing down the search and focusing on the current user interest. It specializes quite well. However, if user interests change, the profile has strong inertia to adapting to the change. Also, in the transition phase the profile could be specialized to neither the old nor the new interest. This approach using relevance feedback appears to work well for specializing but not as much for adapting when user interests change.


MIT Media Lab - Autonomous Agents Group - agentmaster@media.mit.edu