Results (
Thai) 1:
[Copy]Copied!
With explicit profiling, the users themselves do the profiling work by either specifying search preferences up front, or by providing personal relevance feedback such as rating returned search results. Chirita et al [22] use individual user profiles which are defined by the searcher through ODP1 web directory categories to re-rank results according to the distance between the profile and ODP categories for each result. They investigate a number of different distance metrics, and report the findings of a live user evaluation that shows that their personalized approach is capable of more relevant result rankings than standard Google search. One of the drawbacks of relying on ODP categories in this way however is that only a small proportion of the web is categorised in the ODP and so many of the returned search results have no category information to base the re-ranking on. Ma et al [48] propose a similar approach whereby user profiles are explicitly expressed through ODP categories, except they re-rank search results based on the cosine similarity between result page content and the ODP directory category profiles. In this way the search results themselves are not required to be categorised in the ODP. In contrast, ifWeb [2] builds user profiles using a less structured approach through keywords, free-text descriptions, and web page examples provided by the user to express their specific information needs, which are stored as a weighted semantic network of concepts. ifWeb also takes advantage of explicit relevance feedback where the searcher provides result ratings that are used to refine and update their profile. A similar approach is used by the Wifs system [55] in which profiles initially built using terms selected from a list can be subsequently improved with feedback on viewed documents provided by the users. The major drawback with these types of explicit approaches to profiling is that the majority of users are reluctant to make the extra effort in providing feedback [16]. Furthermore, searchers may find it difficult to categorise their information needs and preferences accurately in the first place. A potentially more successful approach to profiling is to infer user preferences implicitly (implicit profiling). As in the work of [22], Liu et al [47] also use hierarchical categories from the ODP to represent a searcher’s profile, except in this work the categories are chosen automatically based on past search behaviour such as previously submitted queries and the content of selected result documents. A number of different learning algorithms are analysed for mapping this search behaviour onto the ODP categories, including those based on Linear Least Squares Fit (LLSF) [107], the Rocchio relevance feedback algorithm [78], and k-Nearest Neighbor (kNN) [28]. In a related approach, [103] use statistical language methods to mine contextual information from this type of long-term search history to build a language model based profile, and [69] also infer user preferences based on past behaviour, this time using the browser cache of visited pages to infer subject areas that the user is interested in. These subject areas, or categories, are combined into a hierarchical user profile where each category is also weighted according to the length of time the user spent viewing the pages corresponding to the category.
The above are all examples of long-term user profiles that seek to capture information about the user’s preferences over an extended period of time, certainly
Being translated, please wait..
