Citation Context

...e leave further investigation of the right labeling strategy as future work. Given that than 2/3 of proactive impressions in our logs have only one card with a positive label (Hybrid-SAT = true), we decided to pick mean reciprocal rank (MRR) as our evaluation metric E. RR is defined as the reciprocal rank of the highest-ranked relevant card in an impression, and is considered as 0 in the absence of positive labels. MRR is computed by averaging RR over the test impressions. Our choice is in line with several previous works on search personalization that also used MRR as their evaluation metric [12, 38, 39, 42]. Training. To train Carré, we sampled proactive impressions from the logs of a commercial search engine as described in Section 3. Each impression consists of a unique and persistent user identifier that can be matched against reactive logs, a set of cards that were presented to the user, and the related click statistics and time-stamps. The relevance labels are assigned to the cards based on clicks and viewport duration as explained above. We then train a model that learns to rank the (pseudo) relevant cards on top based on available features and penalizes their demotion. Table 1 illustrates...

Citation Context

...t Topics, Search Personalisation, Re-ranking. 1 Introduction As one of the key components in advanced search engines (e.g., Google and Bing), Search Personalisation has attracted increasing attention =-=[1,9,12,15,16,19]-=-. The personalisation is expected to improve the usefulness of search algorithms. Unlike the search methods which don’t use personalisation, personalised search engines utilise the personal data of ea...

Citation Context

...what follows, we assume that M,N S, since in the personalization framework it is usually very easy to collect a training set of arbitrarily large size, and, typically, in studies on personalization =-=[3, 4, 15]-=-, the number of samples in the training set is several orders of magnitude larger than the number of features. Under these assumptions we estimate the complexity of our algorithm with a sigmoid weight...

Citation Context

...n to their attention. On the other hand, users who are satisfied with the first results they find are less likely to continue clicking when they return. A search engine could instead provide these users with query suggestions or information related to the clicked result to support further exploration on the same topic. Our results also reveal an opportunity to provide new, relevant content to searchers when they return to a result list. Thus far efforts to contextualize results have focused on using information from the initial queries in a session to improve the ranking of subsequent queries [20]. Our findings suggest it may also be possible to identify new content for users without their ever having to issue a new query. This could be done using implicit feedback from the user’s initial interactions (e.g., dwell time and click position), or by taking more than a few hundred milliseconds to process the initial query [26]. However, there appears to be a risk to capitalizing on this opportunity, in that changing the result ranking during a search may cause confusion. When ranking results mid-query, a search engine must account for the user’s initial experiences. Clicked results appear t...

Citation Context

...f search systems that help searchers struggle less and succeed more. Categories and Subject Descriptors H.3.3 [Information Search and Retrieval]: Information Storage and Retrieval: search process, selection process. Keywords Information retrieval; Struggling 1. INTRODUCTION When searchers experience difficulty in finding information, their struggle may be apparent in search behaviors such as issuing numerous search queries or visiting many results within a search session [5]. Such long sessions are prevalent and time consuming (e.g., around half of Web search sessions contain multiple queries [37]). Long sessions occur when searchers are exploring or learning a Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a...

Citation Context

...eevan et al. [27] analyzed the queries issued by 114 anonymous users in one year’s data from Yahoo search logs, and observed that users re-found previously found URLs for 40% of all queries. In a larger study involving tens of millions of users, Tyler and Teevan [31] further examined re-finding using both search and browser logs from Microsoft’s Bing search engine and Windows Live Toolbar. They observed distinct query patterns and behaviors relating to re-finding that differ from other search patterns, and explored differences between same-session and cross-session re-finding. Shokouhi et al. [21] analyzed repeated result use within a session, and showed that users engage differently with repeats than with results shown for the first time. Our study builds on previous work to characterize how people re-find not just previously viewed URLs, but also previously viewed domains. We study the re-finding of individual URLs as a subset of domain re-finding behavior. Although domain preference has not been well studied in the context of Web search, researchers have explored the consistency of people’s domain use in the context of Web browsing. Tauscher and Greenberg [26] found that the majorit...