Some content of this application is unavailable at the moment.
If this situation persist, please contact us atFeedback&Contact
1. (US09092510) Modifying search result ranking based on a temporal element of user feedback
Note: Text based on automatic Optical Character Recognition processes. Please use the PDF version for legal matters

BACKGROUND

      The present disclosure relates to ranking of search results.
      Internet search engines aim to identify documents or other items that are relevant to a user's needs and to present the documents or items in a manner that is most useful to the user. Such activity often involves a fair amount of mind-reading—inferring from various clues what the user wants. Certain clues may be user specific. For example, knowledge that a user is making a request from a mobile device, and knowledge of the location of the device, can result in much better search results for such a user.
      Clues about a user's needs may also be more general. For example, search results can have an elevated importance, or inferred relevance, if a number of other search results link to them. If the linking results are themselves highly relevant, then the linked-to results may have a particularly high relevance. Such an approach to determining relevance, generally associated with the GOOGLE® PageRank technology, is premised on the assumption that, if authors of web pages felt that another web site was relevant enough to be linked to, then web searchers would also find the site to be particularly relevant. In short, the web authors “vote up” the relevance of the sites.
      Other various inputs may be used instead of, or in addition to, such techniques for determining and ranking search results. For example, user reactions to particular search results or search result lists may be gauged, so that results on which users often click will receive a higher ranking. The general assumption under such an approach is that searching users are often the best judges of relevance, so that if they select a particular search result, it is likely to be relevant, or at least more relevant than the presented alternatives.

SUMMARY

      Systems, methods, and apparatus including computer program products for ranking search results of a search query are described. In general, particular inputs may be generated or analyzed to affect the presentation of search results. For example, such inputs may increase the relevance that a system will assign to a particular result in a particular situation, thus boosting the score or other indicator of relevance for the result (and perhaps the relevance of the result in the context of a particular query). Such an approach may benefit a user by providing them with search results that are more likely to match their needs. As a result, users can learn more using the interne, can find more relevant information more quickly, and will thus achieve more in their work or elsewhere, and will be more likely to use such a system again. A provider of such services may also benefit, by providing more useful services to users, and by thereby inducing more traffic to their search services. Such additional traffic may provide an operator with additional revenue, such as in the form of advertising that accompanies the searching and the delivery of search results.
      The subject matter described in this specification can be embodied in a computer-implemented method that includes obtaining user feedback associated with quality of an electronic document; adjusting a measure of relevance for the electronic document based on a temporal element of the user feedback; and outputting the measure of relevance to a ranking engine for ranking of search results, including the electronic document, for a search for which the electronic document is returned. Obtaining user feedback can include receiving user selections of documents presented by a document search service, and the method can include evaluating the user selections in accordance with an implicit user feedback model to determine the measure of relevance, and adjusting the measure of relevance can include adjusting the measure of relevance in accordance with the implicit user feedback model.
      Adjusting the measure of relevance can include comparing change over time in user selections of the electronic document with change over time in user selections of the documents. The implicit user feedback model can include a background population click trend model and adjusting the measure of relevance can include: determining a likelihood ratio for the electronic document and the documents with respect to the background population click trend model; and modifying the measure of relevance for the electronic document based on a difference between a document click trend model for the electronic document and the background population click trend model.
      Evaluating the user selections in accordance with the implicit user feedback model can include determining the measure of relevance for the electronic document within a context of a search query for which the electronic document is returned by the document search service. Adjusting the measure of relevance can include adjusting contributions of the user selections to the measure of relevance based on temporal distance between a current time and times of the user selections. Moreover, adjusting the measure of relevance can include weighting the user selections based on recency of document viewing grouped into time span categories.
      Weighting of the user selections can be initiated only after a total number of selections within a context of a search query surpasses a threshold. The threshold can be one hundred selections within the context of the search query. In addition, the time span categories can include less than two weeks old, two to four weeks old, four to six weeks old, six to eight weeks old, and more than eight weeks old, and the weighting can include: applying no weighting to selections that are less than two weeks old; applying at least a fifteen percent reduction weighting to selections that are two to four weeks old; applying at least a fifty percent reduction weighting to selections that are four to six weeks old; applying at least a seventy five percent reduction weighting to selections that are six to eight weeks old; and applying at least a ninety percent reduction weighting to selections that are more than eight weeks old.
      The subject matter described in this specification can also be embodied in various systems, apparatus and corresponding computer program products (encoded on a computer-readable medium and operable to cause data processing apparatus to perform method operations). For example, a system can include a tracking component and a rank modifier engine structured to perform the operations described. Moreover, a system can include various means for performing the operations described, as detailed below, and equivalents thereof.
      Particular embodiments of the described subject matter can be implemented to realize one or more of the following advantages. A ranking sub-system can include a rank modifier engine that uses implicit user feedback to cause re-ranking of search results in order to improve the final ranking presented to a user of an information retrieval system. User selections of search results (click data) can be tracked and evaluated in accordance with an implicit user feedback model, where age of the user selections is taken into account. For example, once a given query accumulates enough clicks to be considered frequent (e.g., 100 clicks) weights can then be assigned to each click according to its age. This can reduce the self-reinforcing effects in frequent queries, where the longer a given document is gathering clicks, the more likely it is to place well in future searches and to thereby gather even more clicks. Infrequent queries, which typically have less click data, can be left unchanged. Thus, the implicit user feedback model can be unchanged initially, but once a significant amount of user selections have been observed, historic user selections can be down-weighted based on their age, which can result in improved re-ranking of search results. In addition, statistical techniques can be used to identify trends in historic user selections and improve re-ranking of search results.
      The details of one or more implementations are set forth in the accompanying drawings and the description below. Other features, aspects, and advantages will become apparent from the description, the drawings, and the claims.

DESCRIPTION OF THE DRAWINGS

       FIG. 1 shows an example information retrieval system in which the relevance of results obtained for submitted search queries can be improved.
       FIG. 2 shows example components of an information retrieval system.
       FIG. 3 shows another example information retrieval system.
       FIG. 4A shows an example process of using a temporal element of user feedback to improve search result rankings.
       FIG. 4B shows an example process of comparing changes in user selections over time to improve search result rankings.
       FIG. 4C shows an example process of using temporal distance between a current time and times of past user selections to improve search result rankings.
       FIG. 5 is a schematic diagram of an example computer system.
      Like reference numbers and designations in the various drawings indicate like elements.

DETAILED DESCRIPTION

       FIG. 1 shows an example system 1000 for improving the relevance of results obtained from submitting search queries as can be implemented in an internet, intranet, or other client/server environment. The system 1000 is an example of an information retrieval system in which the systems, components and techniques described below can be implemented.
      Although several components are illustrated, there may be fewer or more components in the system 1000. Moreover, the components can be distributed on one or more computing devices connected by one or more networks or other suitable communication mediums.
      A user 1002 ( 1002 a, 1002 b, 1002 c) can interact with the system 1000 through a client device 1004 ( 1004 a, 1004 b, 1004 c) or other device. For example, the client device 1004 can be a computer terminal within a local area network (LAN) or wide area network (WAN). The client device 1004 can include a random access memory (RAM) 1006 (or other memory and/or a storage device) and a processor 1008. The processor 1008 is structured to process instructions within the system 1000. In some implementations, the processor 1008 is a single-threaded processor. In other implementations, the processor 1008 is a multi-threaded processor. The processor 1008 can include multiple processing cores and is structured to process instructions stored in the RAM 1006 (or other memory and/or a storage device included with the client device 1004) to display graphical information for a user interface.
      A user 1002 a can connect to a search engine 1030 within a server system 1014 to submit a query 1015. When the user 1002 a submits the query 1015 through an input device attached to a client device 1004 a, a client-side query signal 1010 a is sent into a network 1012 and is forwarded to the server system 1014 as a server-side query signal 1010 b. Server system 1014 can be one or more server devices in one or more locations. A server device 1014 includes a memory device 1016, which can include the search engine 1030 loaded therein. A processor 1018 is structured to process instructions within the system 1014. These instructions can implement one or more components of the search engine 1030. The processor 1018 can be a single-threaded processor or a multi-threaded processor, and can include multiple processing cores. The processor 1018 can process instructions stored in the memory 1016 related to the search engine 1030 and can send information to the client device 1004, through the network 1012, to create a graphical presentation in a user interface of the client device 1004 (e.g., a search results web page displayed in a web browser).
      The server-side query signal 1010 b is received by the search engine 1030. The search engine 1030 uses the information within the user query 1015 (e.g. query terms) to find relevant documents. The search engine 1030 can include an indexing engine 1020 that actively searches a corpus (e.g., web pages on the Internet) to index the documents found in that corpus, and the index information for the documents in the corpus can be stored in an index database 1022. This index database 1022 can be accessed to identify documents related to the user query 1015. Note that, an electronic document (which for brevity will simply be referred to as a document) does not necessarily correspond to a file. A document can be stored in a portion of a file that holds other documents, in a single file dedicated to the document in question, or in multiple coordinated files.
      The search engine 1030 can include a ranking engine 1052 to rank the documents related to the user query 1015. The ranking of the documents can be performed using traditional techniques for determining an information retrieval (IR) score for indexed documents in view of a given query. The relevance of a particular document with respect to a particular search term or to other provided information may be determined by any appropriate technique. For example, the general level of back-links to a document that contains matches for a search term may be used to infer a document's relevance. In particular, if a document is linked to (e.g., is the target of a hyperlink) by many other relevant documents (e.g., documents that also contain matches for the search terms), it can be inferred that the target document is particularly relevant. This inference can be made because the authors of the pointing documents presumably point, for the most part, to other documents that are relevant to their audience.
      If the pointing documents are in turn the targets of links from other relevant documents, they can be considered more relevant, and the first document can be considered particularly relevant because it is the target of relevant (or even highly relevant) documents. Such a technique may be the determinant of a document's relevance or one of multiple determinants. The technique is exemplified in the GOOGLE® PageRank system, which treats a link from one web page to another as an indication of quality for the latter page, so that the page with the most such quality indicators is rated higher than others. Appropriate techniques can also be used to identify and eliminate attempts to cast false votes so as to artificially drive up the relevance of a page.
      To further improve such traditional document ranking techniques, the ranking engine 1052 can receive an additional signal from a rank modifier engine 1056 to assist in determining an appropriate ranking for the documents. The rank modifier engine 1056 provides one or more measures of relevance for the documents, which can be used by the ranking engine 1052 to improve the search results' ranking provided to the user 1002. The rank modifier engine 1056 can perform one or more of the operations described further below to generate and adjust the one or more measures of relevance.
      The search engine 1030 can forward the final, ranked result list within a server-side search results signal 1028 a through the network 1012. Exiting the network 1012, a client-side search results signal 1028 b can be received by the client device 1004 a where the results can be stored within the RAM 1006 and/or used by the processor 1008 to display the results on an output device for the user 1002 a.
       FIG. 2 shows example components of an information retrieval system. These components can include an indexing engine 2010, a scoring engine 2020, a ranking engine 2030, and a rank modifier engine 2070. The indexing engine 2010 can function as described above for the indexing engine 1020. In addition, the scoring engine 2020 can generate scores for document results based on many different features, including content-based features that link a query to document results, and query-independent features that generally indicate the quality of document results. The content-based features can include aspects of document format, such as query matches to title or anchor text in an HTML (Hyper Text Markup Language) page. The query-independent features can include aspects of document cross-referencing, such as the PageRank of the document or the domain. Moreover, the particular functions used by the scoring engine 2020 can be tuned, to adjust the various feature contributions to the final IR score, using automatic or semi-automatic processes.
      The ranking engine 2030 can produce a ranking of document results 2040 for display to a user based on IR scores received from the scoring engine 2020 and one or more signals from the rank modifier engine 2070. A tracking component 2050 can be used to record information regarding individual user selections of the results presented in the ranking 2040. For example, the tracking component 2050 can be embedded JavaScript code included in a web page ranking 2040 that identifies user selections (clicks) of individual document results and also identifies when the user returns to the results page, thus indicating the amount of time the user spent viewing the selected document result. In other implementations, the tracking component 2050 can be a proxy system through which user selections of the document results are routed, or the tracking component can include pre-installed software at the client (e.g., a toolbar plug-in to the client's operating system). Other implementations are also possible, such as by using a feature of a web browser that allows a tag/directive to be included in a page, which requests the browser to connect back to the server with message(s) regarding link(s) clicked by the user.
      The recorded information can be stored in result selection log(s) 2060. The recorded information can include log entries that indicate, for each user selection, the query (Q), the document (D), the time (T) on the document, the language (L) employed by the user, and the country (C) where the user is likely located (e.g., based on the server used to access the IR system). Other information can also be recorded regarding user interactions with a presented ranking, including negative information, such as the fact that a document result was presented to a user, but was not clicked, position(s) of click(s) in the user interface, IR scores of clicked results, IR scores of all results shown before the clicked result, the titles and snippets shown to the user before the clicked result, the user's cookie, cookie age, IP (Internet Protocol) address, user agent of the browser, etc. Sill further information can be recorded, such as described below during discussion of the various features that can be used to build a prior model. Moreover, similar information (e.g., IR scores, position, etc.) can be recorded for an entire session, or multiple sessions of a user, including potentially recording such information for every click that occurs both before and after a current click.
      The information stored in the result selection log(s) 2060 can be used by the rank modifier engine 2070 in generating the one or more signals to the ranking engine 2030. In general, a wide range of information can be collected and used to modify or tune the click signal from the user to make the signal, and the future search results provided, a better fit for the user's needs. Thus, user interactions with the rankings presented to the users of the information retrieval system can be used to improve future rankings.
      The components shown in FIG. 2 can be combined in various manners and implemented in various system configurations. For example, the scoring engine 2020 and the ranking engine 2030 can be merged into a single ranking engine, such as the ranking engine 1052 of FIG. 1. The rank modifier engine 2070 and the ranking engine 2030 can also be merged, and in general, a ranking engine includes any software component that generates a ranking of document results after a query. Moreover, a ranking engine can be included in a client system in addition to (or rather than) in a server system.
       FIG. 3 shows another example information retrieval system. In this system, a server system 3050 includes an indexing engine 3060 and a scoring/ranking engine 3070. A client system 3000 includes a user interface for presenting a ranking 3010, a tracking component 3020, result selection log(s) 3030 and a ranking/rank modifier engine 3040. For example, the client system 3000 can include a company's enterprise network and personal computers, in which a browser plug-in incorporates the ranking/rank modifier engine 3040. When an employee in the company initiates a search on the server system 3050, the scoring/ranking engine 3070 can return the search results along with either an initial ranking or the actual IR scores for the results. The browser plug-in can then re-rank the results locally based on tracked page selections for the company-specific user base.
       FIG. 4A shows an example process of using a temporal element of user feedback to improve search result rankings. User feedback associated with document quality can be obtained 4010. This user feedback can be explicit user feedback (e.g., a user's express rating of a web page through a feedback interface provided with a web browser), implicit user feedback (e.g., a measure of page quality derived from how long the user views the web page), or a combination of these, which can be used to derive a measure of relevance for documents. For example, obtaining user feedback can involve receiving user selections of documents presented by a document search service, such as described above.
      User selections can be evaluated 4020 in accordance with an implicit user feedback model to determine a measure of relevance in accordance with the implicit user feedback model. Various different implicit user feedback models can be employed. In general, individual selections of document results can be tracked. For example, in a web based information retrieval system, user's click data on web page search results can be gathered and stored in log(s), which can be kept for all user queries. When a user clicks on a search result, the click can be tracked via JavaScript code embedded in the search results page, an embedded browser tag, etc. This code can track when and on what a user clicks in the main search results page, and can track when the user returns to that main page.
      Post-click behavior can also be tracked via pre-installed software at the client (e.g., a toolbar plug-in to the client's operating system). Provided the user opts into fully sharing their browsing behavior, the toolbar software can track all the pages that the user visits, both before and after the search results page is delivered.
      The information gathered for each click can include: (1) the query (Q) the user entered, (2) the document result (D) the user clicked on, (3) the time (T) on the document, (4) the interface language (L) (which can be given by the user), (5) the country (C) of the user (which can be identified by the host that they use, such as www.google.co.uk to indicate the United Kingdom), and (6) additional aspects of the user and session. The time (T) can be measured as the time between the initial click through to the document result until the time the user comes back to the main page and clicks on another document result. Moreover, an assessment can be made about the time (T) regarding whether this time indicates a longer view of the document result or a shorter view of the document result, since longer views are generally indicative of quality for the clicked through result. This assessment about the time (T) can further be made in conjunction with various weighting techniques.
      Document views resulting from the selections can be weighted based on viewing length information to produce weighted views of the document result. Thus, rather than simply distinguishing long clicks from short clicks, a wider range of click through viewing times can be included in the assessment of result quality, where longer viewing times in the range are given more weight than shorter viewing times. This weighting can be either continuous or discontinuous.
      A continuous function can be applied to the document views resulting from the selections. Thus, the weight given to a particular click through time can fall within a continuous range of values, as defined by the specified function. Alternatively, a discontinuous function can be applied to the document views resulting from the selections. For example, there can be three viewing time categories, each having a corresponding weight. Note that such functions can be explicitly defined, or merely implicit in the software implementation.
      In the case of discontinuous weighting, the individual selections of the document result can be classified into viewing time categories, and weights can be assigned to the individual selections based on results of the classifying. For example, a short click can be considered indicative of a poor page and thus given a low weight (e.g., −0.1 per click), a medium click can be considered indicative of a potentially good page and thus given a slightly higher weight (e.g., 0.5 per click), a long click can be considered indicative of a good page and thus given a much higher weight (e.g., 1.0 per click), and a last click (where the user doesn't return to the main page) can be considered as likely indicative of a good page and thus given a fairly high weight (e.g., 0.9). Note that the click weighting can also be adjusted based on previous click information. For example, if another click preceded the last click, the last click can be considered as less indicative of a good page and given only a moderate weight (e.g., 0.3 per click).
      The various time frames used to classify short, medium and long clicks, and the weights to apply, can be determined for a given search engine by comparing historical data from user selection logs with human generated explicit feedback on the quality of search results for various given queries, and the weighting process can be tuned accordingly. Furthermore, these time frames and weights can be adjusted based on one or more viewing length differentiators, as is described further below.
      The weighted views of the document result can be combined to determine a number to be used in determining a measure of relevance. For example, the weighted clicks described above can be summed together for a given query-document pair. Note that safeguards against spammers (users who generate fraudulent clicks in an attempt to boost certain search results) can be taken to help ensure that the user selection data is meaningful, even when very little data is available for a given (rare) query. These safeguards can include employing a user model that describes how a user should behave over time, and if a user doesn't conform to this model, their click data can be disregarded. The safeguards can be designed to accomplish two main objectives: (1) ensure democracy in the votes (e.g., one single vote per cookie and/or IP for a given query-URL (Universal Resource Locator) pair), and (2) entirely remove the information coming from cookies or IP addresses that do not look natural in their browsing behavior (e.g., abnormal distribution of click positions, click durations, clicks per minute/hour/day, etc.). Suspicious clicks can be removed, and the click signals for queries that appear to be spammed need not be used (e.g., queries for which the clicks feature a distribution of user agents, cookie ages, etc. that do not look normal).
      A measure of relevance for the document result can be determined within the context of the search query for which the document result is returned. This measure of relevance can be calculated as a fraction, which can be directly applied to IR scores of the search results, thereby boosting the documents in the resulting ranking that have implicit user feedback indicating document quality. For example, a traditional click fraction, which takes into consideration the other results for the given query, has been defined as follows:
          BASE=[#WC( Q,D)]/[#WC( Q)+ S0]
where #WC(Q,D) is the sum of weighted clicks for a query-URL pair, #WC(Q) is the sum of weighted clicks for the query (summed over all results for the query), and S0 is a smoothing factor.
      The click fraction can also employ per-language and per-country fractions (with smoothing there between):
          LANG=[#WC( Q,D,L)+ S1·BASE]/[#WC( Q,L)+ S1]
          COUNTRY=[#WC( Q,D,L,C)+ S2·LANG]/[#WC( Q,L,C)+ S2]
where LANG incorporates language specific click data, plus BASE, and COUNTRY incorporates country (and language) specific click data, plus LANG. In this manner, if there is less data for the more specific click fractions, the overall fraction falls back to the next higher level for which more data is available.
      Furthermore, it should be noted that different smoothing factors S0, S1 and S2 can be used, or one or more of these can be the same smoothing factor. The smoothing factors used can be determined based on how much traffic is received within the context of the click fraction. For example, for a given country-language tuple, the smoothing factor can be raised concordant with the amount of traffic received (e.g., a larger smoothing factor can be used for US-English queries if a good deal more of such queries are received). In addition, the smoothing factor can be increased for query sources that have historically generated more spamming activity (e.g., queries from Russia).
      In addition, as mentioned above, one or more viewing length differentiators (e.g., query category and user type) can be identified for use in the weighting. A viewing length differentiator can include a factor governed by a determined category of the search query, a factor governed by a determined type of a user generating the individual selections, or a combination of them. The document views can be weighted based on the viewing length information in conjunction with the viewing length differentiator(s), such as the determined category of the search query and the determined type of the user. Thus, in the discontinuous weighting case (and the continuous weighting case), the threshold(s) (or formula) for what constitutes a good click can be evaluated on query and user specific bases. For example, the query categories can include “navigational” and “informational”, where a navigational query is one for which a specific target page or site is likely desired (e.g., a query such as “BMW”), and an informational query is one for which many possible pages are equally useful (e.g., a query such as “George Washington's Birthday”). Note that such categories may also be broken down into sub-categories as well, such as informational-quick and informational-slow: a person may only need a small amount of time on a page to gather the information they seek when the query is “George Washington's Birthday”, but that same user may need a good deal more time to assess a result when the query is “Hilbert transform tutorial”.
      The query categories can be identified by analyzing the IR scores or the historical implicit feedback provided by the click fractions. For example, significant skew in either of these (meaning only one or a few documents are highly favored over others) can indicate a query is navigational. In contrast, more dispersed click patterns for a query can indicate the query is informational. In general, a certain category of query can be identified (e.g., navigational), a set of such queries can be located and pulled from the historical click data, and a regression analysis can be performed to identify one or more features that are indicative of that query type (e.g., mean staytime for navigational queries versus other query categories; the term “staytime” refers to time spent viewing a document result, also known as document dwell time).
      Traditional clustering techniques can also be used to identify the query categories. This can involve using generalized clustering algorithms to analyze historic queries based on features such as the broad nature of the query (e.g., informational or navigational), length of the query, and mean document staytime for the query. These types of features can be measured for historical queries, and the threshold(s) can be adjusted accordingly. For example, K means clustering can be performed on the average duration times for the observed queries, and the threshold(s) can be adjusted based on the resulting clusters.
      User types can also be determined by analyzing click patterns. For example, computer savvy users often click faster than less experienced users, and thus users can be assigned different weighting functions based on their click behavior. These different weighting functions can even be fully user specific (a user group with one member). For example, the average click duration and/or click frequency for each individual user can be determined, and the threshold(s) for each individual user can be adjusted accordingly. Users can also be clustered into groups (e.g., using a K means clustering algorithm) based on various click behavior patterns.
      Moreover, the weighting can be adjusted based on the determined type of the user both in terms of how click duration is translated into good clicks versus not-so-good clicks, and in terms of how much weight to give to the good clicks from a particular user group versus another user group. Some user's implicit feedback may be more valuable than other users due to the details of a user's review process. For example, a user that almost always clicks on the highest ranked result can have his good clicks assigned lower weights than a user who more often clicks results lower in the ranking first (since the second user is likely more discriminating in his assessment of what constitutes a good result). In addition, a user can be classified based on his or her query stream. Users that issue many queries on (or related to) a given topic (e.g., queries related to law) can be presumed to have a high degree of expertise with respect to the given topic, and their click data can be weighted accordingly for other queries by them on (or related to) the given topic.
      Various other factors and approaches can be employed in an implicit user feedback model, but in general, user selections of documents are evaluated in accordance with an overall model of how user activity with respect to search results indicates the quality of those results. This indication of quality can be condensed into a measure of relevance within some defined context, which can include the context of the search query, e.g., the implicit user feedback model can generate a measure of relevance that is applicable only within the context of a given query. For example, evaluating the user selections can involve determining the measure of relevance for the electronic document within the context of the search query for which the electronic document is returned by the document search service, and that same document will thus have a different measure of relevance score when returned for a different query.
      In addition, the implicit user feedback model can be changed over time. For example, one or more user selection log(s) can be maintained to store click through data, such as described above. These user selection log(s) can store values of many different features for each document result selection observed. The implicit user feedback model can use selected subsets of these features in determining the measure of relevance, and the selected subsets can be changed, as well as the manner in which they are used, even while the user selections log(s) remain the same (albeit growing larger as more user selections are observed, subject to a cutoff time limit, such as twelve months, after which log data need not be retained).
      The measure of relevance for an electronic document can be adjusted 4030 based on a temporal element (e.g., a temporal aspect of implicit user feedback). More recent user feedback can be given higher priority over older user feedback. This can be of particular value in the context of web search, where documents change frequently. In addition, the temporal element can include historic trend(s) in user selections, which can be measured statistically.
      The adjustment can be applied to explicit user feedback and to implicit user feedback. The adjustment can be considered part of an implicit user feedback model, in that the temporal element can form part of the model that connects user actions with document quality. Thus, adjusting the measure of relevance can include adjusting in accordance with the implicit user feedback model.
      The measure of relevance can be output 4040 to a ranking engine for ranking of search results for a new search. The adjusted measure of relevance can be saved to a computer-readable medium for use by the ranking engine, or otherwise made available to the ranking engine. The adjusted measure of relevance can be directly applied to IR scores for documents, or the measure can be passed through a transform to create a boosting factor that can be applied to the IR scores. For example, the adjusted measure of relevance can be further adjusted based on the given context and historical data combined with human generated relevance ratings (e.g., employed in a tuning process to select an appropriate boosting transform for a given implementation). Moreover, the adjusted measure of relevance can be used to modify and improve the ranking of search results generated for a given query, and the modified ranking can be presented to a user (e.g., on a display device in a web browser user interface).
       FIG. 4B shows an example process of comparing changes in user selections over time to improve search result rankings. In this example, a statistical method is used to identify trends in implicit user feedback by comparing individual click trends for a candidate result against the overall trends across one or more queries (the background population). The candidate result is the document that is being considered for ranking adjustment. The background population is a set of documents and features of their prior selection, e.g., (query, document) pairs that establish the baseline trends. The set of documents can contain one or more queries.
      A click trend corresponds to temporal change in clicks for a given set of data pulled from the log(s). For example, a click trend can be a set of (t, c) pairs, where each t is a date interval and each c is the volume of clicks accumulated by a result or set of results within that date interval (e.g., the click trend for the candidate result and the click trend for the background population). In addition, a model class is a parametric model family that can be used to model the click trend of a population. The example below uses a linear model, but it should be appreciated that any family of functions (e.g., polynomials of degree k) can be used.
      A background population click model (M 1) can be generated 4110. This can involve fitting the model class to the background population click trend to minimize squared residual error. For example, the least squares procedure can be used to fit a line to the background population's click trend, and the average squared residual error (E) of the model M 1 can be computed. A likelihood for documents in the background population can be computed 4120 with respect to the background population click model. For example, the average log likelihood of all members of the background population can be computed under a Gaussian whose mean is determined by the model M 1 and whose variance is E, the average squared residual error. A likelihood for a given document (the candidate result) can be computed 4130 with respect to the background population click model. For example, the average log likelihood of all members of the click trend for the candidate result can be computed under a Gaussian whose mean is determined by the model M 1 and whose variance is E.
      The likelihood ratio can be computed 4140 for the two likelihood values calculated above, and if the ratio is significantly less than one, then the candidate result behaves differently than the background population. Note that the test to determine if the ratio is significant can be determined empirically and will vary with specific applications. For example, one method to determine a ratio significance test can be to collect and histogram the likelihood ratios for a large set of randomly chosen (query, document) pairs. One can then use human generated relevance ratings from selections within each histogram bin to determine where the appropriate cutoff threshold should be located.
      When the ratio passes a selected threshold, the candidate result is considered to behave differently than the background population, and a document click trend model can be generated 4150. The model class can be fit to the candidate result's click trend to obtain the document click trend model (M 2), and the rate of change of models M 1 and M 2 can be compared. For example, if the models M 1 and M 2 are linear, then the slopes of the two lines can be compared.
      Finally, the measure of relevance for the given document (candidate result) can be modified 4160 based on one or more differences in change over time of the two click trend models. For example, if the rate of change of M 2 is larger than that of M 1, the candidate result can be promoted, and if the candidate's rate of change is less, then it can be demoted. To help ensure that only the most significant trends are captured, the applied boost can be relatively conservative. An example boost formula can be:
           B*sqrt( M2 /M1)
where B is the normal boost value, M 2 is the rate of change of the document click trend model, and M 1 is the rate of change of the background model. This type of approach can be used in order to dampen the effects and produce a relatively conservative boost.
       FIG. 4C shows an example process of using temporal distance between a current time and times of past user selections to improve search result rankings. Contributions of user selections to a relevance measure can be adjusted 4210. This adjustment can involve application of continuous weightings or discontinuous weightings. In addition, a threshold can be used (e.g., 100 clicks) before which no weightings are applied. Multiple variations on the threshold approach are also possible. For example, a recency weighting can be applied when the number of clicks on a query exceeds a first threshold (e.g., 100 clicks) or when the number of clicks on a (query, document) pair exceeds a second threshold (e.g., 60 clicks). Taking multifaceted threshold approach such as this can create a more flexible recency trend model, e.g., such as to allow the system to more quickly catch a very popular, new document.
      Gathering user selection data over a large time range (e.g., the trailing twelve months) can assist in capturing queries that are repeated, but only infrequently. By adjusting the measure of relevance based on recency, more recent developing trends can be found. Very popular queries that gather a large number of clicks during the early stage of the data collection interval need not have their developing trends in the latter portion of the interval obscured.
      Frequent queries can fall prey to self-reinforcement, where the longer a document is gathering clicks, the more likely it is to place well in future searches and thus to gather more clicks. Such self-reinforcement can be avoided using recency based adjustment. New documents on frequent queries need not be penalized due to their having less time to accumulate clicks, since larger numbers of older clicks can be discounted in the adjustment of the relevance measure. Moreover, because self-reinforcement can be more of a concern for frequent queries, the weighting can be applied when a query accumulates enough clicks to be deemed frequent (e.g., 100 clicks), after which, weights can be assigned to clicks according to age (calculated as distance in the past).
      For example, no weighting (a weight multiple of 1.0) can be applied 4220 to user selections that are less than two weeks old. At least a fifteen percent reduction weighting (a weight multiple of 0.75) can be applied 4230 to selections that are two to four weeks old. At least a fifty percent reduction weighting (a weight multiple of 0.5) can be applied 4240 to selections that are four to six weeks old. At least a seventy five percent reduction weighting (a weight multiple of 0.25) can be applied 4250 to selections that are six to eight weeks old. At least a ninety percent reduction weighting (a weight multiple of 0.1) can be applied 4260 to selections that are more than eight weeks old.
      This approach can reduce self-reinforcing effects in frequent queries. Infrequent queries can be left unchanged, since in these cases, the click counts will generally be small to begin with and any down-weighting can obscure the information content of the user feedback. Other sets of weights and time span categories are also possible. For example, an exponential decay can be used, such as:
          weight=exp(−0.05 *t)
where exp is the exponential function and t is the age of the click in weeks. In general, the use of declining weights give more significance to recent information, but still allow historical data to have some contribution to the final measure of relevance. Moreover, this example describes weighting implicit user feedback based on recency, but it should be recognized that similar techniques can be applied to explicit user feedback as well.
       FIG. 5 is a schematic diagram of an example computer system 6050. The system 6050 can be used for practicing operations described above. The system 6050 can include a processor 6018, a memory 6016, a storage device 6052, and input/output devices 6054. Each of the components 6018, 6016, 6052, and 6054 are interconnected using a system bus 6056. The processor 6018 is capable of processing instructions within the system 6050. These instructions can implement one or more aspects of the systems, components and techniques described above. In some implementations, the processor 6018 is a single-threaded processor. In other implementations, the processor 6018 is a multi-threaded processor. The processor 6018 can include multiple processing cores and is capable of processing instructions stored in the memory 6016 or on the storage device 6052 to display graphical information for a user interface on the input/output device 6054.
      The memory 6016 is a computer readable medium such as volatile or non volatile that stores information within the system 6050. The memory 6016 can store processes related to the functionality of the search engine 1030, for example. The storage device 6052 is capable of providing persistent storage for the system 6050. The storage device 6052 can include a floppy disk device, a hard disk device, an optical disk device, or a tape device, or other suitable persistent storage mediums. The storage device 6052 can store the various databases described above. The input/output device 6054 provides input/output operations for the system 6050. The input/output device 6054 can include a keyboard, a pointing device, and a display unit for displaying graphical user interfaces.
      The computer system shown in FIG. 5 is but one example. In general, embodiments of the subject matter and the functional operations described in this specification can be implemented in digital electronic circuitry, or in computer software, firmware, or hardware, including the structures disclosed in this specification and their structural equivalents, or in combinations of one or more of them. Embodiments of the subject matter described in this specification can be implemented as one or more computer program products, i.e., one or more modules of computer program instructions encoded on a computer readable medium for execution by, or to control the operation of, data processing apparatus. The computer readable medium can be a machine-readable storage device, a machine-readable storage substrate, a memory device, a composition of matter effecting a machine-readable propagated signal, or a combination of one or more of them. The term “data processing apparatus” encompasses all apparatus, devices, and machines for processing data, including by way of example a programmable processor, a computer, or multiple processors or computers. The apparatus can include, in addition to hardware, code that creates an execution environment for the computer program in question, e.g., code that constitutes processor firmware, a protocol stack, a database management system, an operating system, or a combination of one or more of them. A propagated signal is an artificially generated signal, e.g., a machine-generated electrical, optical, or electromagnetic signal, that is generated to encode information for transmission to suitable receiver apparatus.
      A computer program (also known as a program, software, software application, script, or code) can be written in any form of programming language, including compiled or interpreted languages, and it can be deployed in any form, including as a stand alone program or as a module, component, subroutine, or other unit suitable for use in a computing environment. A computer program does not necessarily correspond to a file in a file system. A program can be stored in a portion of a file that holds other programs or data (e.g., one or more scripts stored in a markup language document), in a single file dedicated to the program in question, or in multiple coordinated files (e.g., files that store one or more modules, sub programs, or portions of code). A computer program can be deployed to be executed on one computer or on multiple computers that are located at one site or distributed across multiple sites and interconnected by a communication network.
      The processes and logic flows described in this specification can be performed by one or more programmable processors executing one or more computer programs to perform functions by operating on input data and generating output. The processes and logic flows can also be performed by, and apparatus can also be implemented as, special purpose logic circuitry, e.g., an FPGA (field programmable gate array) or an ASIC (application specific integrated circuit).
      Processors suitable for the execution of a computer program include, by way of example, both general and special purpose microprocessors, and any one or more processors of any kind of digital computer. Generally, a processor will receive instructions and data from a read only memory or a random access memory or both. The essential elements of a computer are a processor for performing instructions and one or more memory devices for storing instructions and data. Generally, a computer will also include, or be operatively coupled to receive data from or transfer data to, or both, one or more mass storage devices for storing data, e.g., magnetic, magneto optical disks, or optical disks. However, a computer need not have such devices. Moreover, a computer can be embedded in another device, e.g., a mobile telephone, a personal digital assistant (PDA), a mobile audio player, a Global Positioning System (GPS) receiver, to name just a few. Computer readable media suitable for storing computer program instructions and data include all forms of non volatile memory, media and memory devices, including by way of example semiconductor memory devices, e.g., EPROM, EEPROM, and flash memory devices; magnetic disks, e.g., internal hard disks or removable disks; magneto optical disks; and CD ROM and DVD-ROM disks. The processor and the memory can be supplemented by, or incorporated in, special purpose logic circuitry.
      To provide for interaction with a user, embodiments of the subject matter described in this specification can be implemented on a computer having a display device, e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor, for displaying information to the user and a keyboard and a pointing device, e.g., a mouse or a trackball, by which the user can provide input to the computer. Other kinds of devices can be used to provide for interaction with a user as well; for example, feedback provided to the user can be any form of sensory feedback, e.g., visual feedback, auditory feedback, or tactile feedback; and input from the user can be received in any form, including acoustic, speech, or tactile input.
      Embodiments of the invention can be implemented in a computing system that includes a back-end component, e.g., as a data server, or that includes a middleware component, e.g., an application server, or that includes a front-end component, e.g., a client computer having a graphical user interface or a Web browser through which a user can interact with an implementation of the invention, or any combination of one or more such back-end, middleware, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication, e.g., a communication network. Examples of communication networks include a local area network (“LAN”) and a wide area network (“WAN”), e.g., the Internet.
      The computing system can include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other.
      While this specification contains many specifics, these should not be construed as limitations on the scope of the invention or of what may be claimed, but rather as descriptions of features specific to particular embodiments of the invention. Certain features that are described in this specification in the context of separate embodiments can also be implemented in combination in a single embodiment. Conversely, various features that are described in the context of a single embodiment can also be implemented in multiple embodiments separately or in any suitable subcombination. Moreover, although features may be described above as acting in certain combinations and even initially claimed as such, one or more features from a claimed combination can in some cases be excised from the combination, and the claimed combination may be directed to a subcombination or variation of a subcombination.
      Similarly, while operations are depicted in the drawings in a particular order, this should not be understood as requiring that such operations be performed in the particular order shown or in sequential order, or that all illustrated operations be performed, to achieve desirable results. In certain circumstances, multitasking and parallel processing may be advantageous. Moreover, the separation of various system components in the embodiments described above should not be understood as requiring such separation in all embodiments, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.
      Thus, particular embodiments of the invention have been described. Other embodiments are within the scope of the following claims. For example, the actions recited in the claims can be performed in a different order and still achieve desirable results. Moreover, the server environment, which is configured to provide electronic search service and employ the ranking systems and techniques described, need not be implemented using traditional back-end or middleware components. The server environment can be implemented using a program installed on a personal computing apparatus and used for electronic search of local files, or the server environment can be implemented using a search appliance (such as GOOGLE® in a Box, provided by Google Inc. of Mountain View, Calif.) installed in an enterprise network.
      Other implicit user feedback models can be used in place of the click fraction model described above. For example, an implicit user feedback model employing a large-scale logistic regression model that uses the actual query and url as features can be used. The new prior models can be used to denormalize any query-specific click model. In addition, the systems and techniques described above can be applied in other contexts and to other services, such as a bookmarking service (e.g., if a service tracks the number of times a person has bookmarked a particular URL, the URL bookmarking rate can be weighted by recency).