CS598CXZ Course Project Predicting User’s Site Preferences in Web Search

Nowadays search engines are so powerful that they have largely taken over the functionality of bookmarks; instead of managing a lot of bookmarks for fast access to frequently visited web sites, people would send simple queries to search engines and look in the returned results for the web pages they want. Yet the typical query is very short (usually only one or two keywords) and does not carry enough context information, which makes it difficult for search engines to return the desired results. For example, since I do not keep a bookmark of the DAIS group’s home page, I would search in Google when I need to access it. If I am willing to use the query “DAIS UIUC”, then I would get what I want at the top of the ranked results. But usually I would just type “DAIS”, and have to skip one or two pages of irrelevant search results before I can find it. I will not blame Google for this search inefficiency, as the term “DAIS” without any context can mean many things. However, Google could have done better. There are plenty of clues in my past search activities; it should have associated ”DAIS” with the web page http://dais.cs.uiuc.edu/ since I always click on that and nothing else. Generally, people trust some web sites (in the above example, dais.cs.uiuc.edu over others, and prefer to see search results of those sites. In my case, I prefer cnn.com for news, wunderground.com for weather, portal.acm.org and citeseer.ist.psu.edu for papers. Existing search engines do not provide ranked results tailored to different user’s need; their search services are not personalized. For example, Google always ranks my favorite weather web site wunderground.com below weather.com, regardless of how many times I have clicked the former site, but not the latter. This project is motivated by the above scenarios. It aims to learn a user’s site preferences from the user’s search history (queries and clicked results) and promote the search result coming from a web site that is most likely to be preferred by the user. Site preferences should be topic-dependent. It would be meaningless to argue whether the user likes cnn.com better or wunderground.com better overall. It makes more sense to say the user likes cnn.com for news and wunderground.com for weather. When a user submits a query, we can determine the most appropriate topic and select the sites preferred for that topic.