#1 topic/interest tracking and discovery

[Wu Xuan]

make search result more concise by only expanding those that "related" with the query (show the first one that seem to be not related, and stop expanding from there), and learn from user's expand action to adjust interest prediction (predict failures, see Fig 2, WebAdapt's introspective process and knowledge)
<---- this is to determine the relevance or similarity or topic/interest tracking, and can be reused for recommendation

[Wu Xuan]

each time a tree is EXPANDED, learn from top to bottom: predict if next one is related/recommended, if not don't expand since this node, and get feedback of opening a link that was "thought" to be not recommended; the learning result can be used to recommend based on the current page title (and maybe plus where-from info, if it's accessible)

[Wu Xuan]

for a folded node, when select "open in new tab", or db click, same effect as expanding (user show interest)