论文部分内容阅读
People use Web search engines to look for information on the Web However，current Web search engines do not fully satisfy the needs of different individuals having unique search goals for the same query We examine the variations in Web search results individual’s find to be relevant for the same query by using explicit relevance judgments and clicks as a source of evidence for relevance Besides，We identify truly ambiguous queries and examine how-both explicit relevance judgments and implicit measures of relevance vary for these queries.
In this thesis，we propose an enhanced framework for personalized search，which contains a new component that can assist users in a decision making process as to whether the query is a potential candidate for personalization or not It considers the potential category of a query and then applies personalization techniques depending on the category to which the query belongs Besides，the new framework considers the differences in interest among different individuals for the results of the same query,which is useful for identifying the type of individuals benefiting from personalization.
We then investigate the variations in explicit relevance judgments for Web search results of same query from the view point of different query categories The result we get shows that there are great variations in judgment among different individuals who evaluate the search results of the same query For those queries for which have relevance judgments from multiple people，it is also possible to find the best possible result ranking for an individual and for different sized groups of individuals.With an increase in the number of people in group，the gap between user satisfaction with the individual search result rankings and group search result ranking increases. This observed gap is the potential for personalizing search and quantified using the normalized discounted gain. This verifies the potential benefit of personalizing search. In this section，we also examine the potential for personalizing search from the perspectives of different query categories. Firstly,we classify queries into clear and unclear,and investigate their potential for personalization. We found that unclear queries have greater potential for personalization than clear queries. This might be a result of the various search intents people have for the same unclear queries when they look for information on the Web Secondly,we further classify unclear queries into broad and ambiguous and examine their potential for personalization separately. We found that generally ambiguous queries have greater potential for personalization than broad queries，while clear queries have the least potential for personalization The result of our experiment reveals that there is the need to examine the potential for personalization from the perspectives of different query categories and this opens up a new research dimension for further investigation in which personalization algorithms should be applied to different query categories in different manners.
We then try to explore how different people perceive search results for the same query by mining clicks，which act as a proxy for relevance，from a large query log data. In an attempt to investigate the variability in what people are searching for when they issue the same query，we only select queries from query log for which we have clicks from at least eight individuals. Extensive analysis of clicks on search result for the same query reveals that there are variations in implicit judgments across individuals indicated in the potential for personalization curve constructed by considering different group sizes. It shows that there is an observable gap between individual preferences and best group preferences for the results of the same query and this gap increases as the number of people in the group increases. This tells US how much room is there to improve the search results through personalization.
We finally try to examine the potential for personalization of truly ambiguous queries by extracting them from query log. Firstly，we use user entropy and its derivates as input features for classifiers to distinguish informational and ambiguous queries characterized by similar click distribution. We then use potential for personalization curve to investigate the potential of ambiguous queries for personalization at different group sizes as measured by the average normalized discounted gain. The result we get shows that ambiguous queries have greater potential for personalization for all low，medium and high frequency classes.Therefore，we suggest that identifying query ambiguity is the way forward for personalization. If we are able to identify truly ambiguous queries beforehand，we can devote all our resources to apply personalization only to queries identified to be truly ambiguous than uniformly applying to all queries.
In this thesis，we propose an enhanced framework for personalized search，which contains a new component that can assist users in a decision making process as to whether the query is a potential candidate for personalization or not It considers the potential category of a query and then applies personalization techniques depending on the category to which the query belongs Besides，the new framework considers the differences in interest among different individuals for the results of the same query,which is useful for identifying the type of individuals benefiting from personalization.
We then investigate the variations in explicit relevance judgments for Web search results of same query from the view point of different query categories The result we get shows that there are great variations in judgment among different individuals who evaluate the search results of the same query For those queries for which have relevance judgments from multiple people，it is also possible to find the best possible result ranking for an individual and for different sized groups of individuals.With an increase in the number of people in group，the gap between user satisfaction with the individual search result rankings and group search result ranking increases. This observed gap is the potential for personalizing search and quantified using the normalized discounted gain. This verifies the potential benefit of personalizing search. In this section，we also examine the potential for personalizing search from the perspectives of different query categories. Firstly,we classify queries into clear and unclear,and investigate their potential for personalization. We found that unclear queries have greater potential for personalization than clear queries. This might be a result of the various search intents people have for the same unclear queries when they look for information on the Web Secondly,we further classify unclear queries into broad and ambiguous and examine their potential for personalization separately. We found that generally ambiguous queries have greater potential for personalization than broad queries，while clear queries have the least potential for personalization The result of our experiment reveals that there is the need to examine the potential for personalization from the perspectives of different query categories and this opens up a new research dimension for further investigation in which personalization algorithms should be applied to different query categories in different manners.
We then try to explore how different people perceive search results for the same query by mining clicks，which act as a proxy for relevance，from a large query log data. In an attempt to investigate the variability in what people are searching for when they issue the same query，we only select queries from query log for which we have clicks from at least eight individuals. Extensive analysis of clicks on search result for the same query reveals that there are variations in implicit judgments across individuals indicated in the potential for personalization curve constructed by considering different group sizes. It shows that there is an observable gap between individual preferences and best group preferences for the results of the same query and this gap increases as the number of people in the group increases. This tells US how much room is there to improve the search results through personalization.
We finally try to examine the potential for personalization of truly ambiguous queries by extracting them from query log. Firstly，we use user entropy and its derivates as input features for classifiers to distinguish informational and ambiguous queries characterized by similar click distribution. We then use potential for personalization curve to investigate the potential of ambiguous queries for personalization at different group sizes as measured by the average normalized discounted gain. The result we get shows that ambiguous queries have greater potential for personalization for all low，medium and high frequency classes.Therefore，we suggest that identifying query ambiguity is the way forward for personalization. If we are able to identify truly ambiguous queries beforehand，we can devote all our resources to apply personalization only to queries identified to be truly ambiguous than uniformly applying to all queries.