Patent Grant
8938463
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.
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 internet, 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 receiving multiple features, including a first feature indicative of presentation bias that affects document result selection for search results presented in a user interface of a document search service; obtaining, based on the multiple features, information regarding document result selections for searches performed using the document search service, the information spanning multiple different queries; generating a prior model using the information spanning the multiple different queries, the prior model representing a query-independent probability of document result selection given values of the multiple features; and outputting the prior model to a ranking engine for ranking of search results to reduce influence of the presentation bias.
The receiving can include receiving the first feature including a position of a document result in the user interface of the document search service within a context of a search query for which the document result is returned. The document result can be a first document result, and the receiving can include receiving a second feature including a relevance score of a second document result within the context of the search query for which the first document result and the second document result are returned. Moreover, the receiving can include receiving a third feature comprising the relevance score of the first document result within the context of the search query for which the first document result and the second document result are returned.
The receiving can include receiving a second feature indicative of relevance for a second document result returned along with a first document result within a context of a search query, and the obtaining can include collecting the information regarding document result selections, including selection of the first document result, for multiple combinations of values for the multiple features. The receiving can include receiving a first set of features including the first feature and a second set of features including the second feature, the generating can include creating a first user selection model based on the first set of features and creating a second user selection model based on the second set of features, and the outputting can include outputting the first user selection model and the second user selection model. The receiving can include receiving the second set of features being a proper subset of the first set of features. In addition, the obtaining can include identifying, based on frequency of document result selection, a reduced set of document result selection records from one or more logs of user selections for the document search service; and collecting the information regarding document result selections from the reduced set of document result selection records.
The subject matter described in this specification can be embodied in a computer-implemented method that includes obtaining, in accordance with a first implicit user feedback model, a first signal for a document result of a search query for which the document result is returned, wherein the first signal corresponds to an aspect of document relevance; obtaining, in accordance with a second implicit user feedback model, a second signal for the document result, wherein the second signal corresponds to an aspect of search result presentation; combining the first signal and the second signal to form a relevance signal for the document result; and outputting the relevance signal to a ranking engine for ranking of search results including the document result.
The obtaining the first signal can include obtaining, in accordance with the first implicit user feedback model, a measure of relevance for the document result within a context of the search query for which the document result is returned; and the obtaining the second signal can include obtaining, in accordance with the second implicit user feedback model, a measure of presentation bias for the document result independent of the context of the search query for which the document result is returned. The operations can further include identifying an additional document result of the search query that is not included in the first implicit user feedback model; obtaining, in accordance with a third implicit user feedback model, an additional signal for the additional document result, wherein the additional signal corresponds to an aspect of document relevance independent of the context of the search query; and outputting the additional signal to the ranking engine for ranking of the search results including the additional document result.
The obtaining the first signal can include obtaining, in accordance with the first implicit user feedback model, a measure of relevance for the document result independent of a context of the search query for which the document result is returned; and the obtaining the second signal can include obtaining, in accordance with the second implicit user feedback model, a measure of presentation bias for the document result independent of the context of the search query for which the document result is returned.
The subject matter described in this specification can be embodied in a computer-implemented method that includes receiving multiple features, including a first feature indicative of presentation bias that affects document result selection for search results presented in a user interface of a document search service, and a second feature indicative of relevance for a second document result returned along with a first document result within a context of a search query; obtaining, based on the multiple features, information regarding document result selections for searches performed using the document search service, including selection of the first document result; generating a prior model using the information, the prior model representing a probability of document result selection given values of the multiple features; and outputting the prior model to a ranking engine for ranking of search results to reduce influence of the presentation bias.
The receiving can include receiving a first set of features including the first feature and a second set of features including the second feature, the generating can include creating a first user selection model based on the first set of features and creating a second user selection model based on the second set of features, and the outputting can include outputting the first user selection model and the second user selection model. The receiving can include receiving the second set of features being a proper subset of the first set of features. The obtaining can include identifying, based on frequency of document result selection, a reduced set of document result selection records from one or more logs of user selections for the document search service; and collecting the information regarding document result selections from the reduced set of document result selection records.
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 transformed into a background probability of user selection in light of one or more presentation bias features, and this background probability can be used to adjust future search result rankings to reduce the influence of presentation bias on implicit user feedback. A prior model can be designed to assist in separating the effects of result quality and presentation bias on user selections, and the prior model can be used to normalize the output of a separate implicit user feedback model in view of the effects of presentation bias. Moreover, a prior model can be used to estimate what the implicit feedback should be for results that are too infrequent, or too recent to have sufficient historical records of implicit feedback applicable to them.
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.
Like reference numbers and designations in the various drawings indicate like elements.
A user 1002 (1002a, 1002b, 1002c) can interact with the system 1000 through a client device 1004 (1004a, 1004b, 1004c) 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 1002a can connect to a search engine 1030 within a server system 1014 to submit a query 1015. When the user 1002a submits the query 1015 through an input device attached to a client device 1004a, a client-side query signal 1010a is sent into a network 1012 and is forwarded to the server system 1014 as a server-side query signal 1010b. 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 1010b 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 GOGGLE® 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 prior models, or one or more measures of relevance for the documents based on one or more prior models, which can be used by the ranking engine 1052 to improve the search results' ranking provided to the user 1002. In general, a prior model represents a background probability of document result selection given the values of multiple selected features, as described further below. The rank modifier engine 1056 can perform one or more of the operations described below to generate the one or more prior models, or the one or more measures of relevance based on one or more prior models.
The search engine 1030 can forward the final, ranked result list within a server-side search results signal 1028a through the network 1012. Exiting the network 1012, a client-side search results signal 1028b can be received by the client device 1004a 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 1002a.
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
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.
Nonetheless, despite all the above techniques for developing a robust implicit user feedback model, presentation (or display) bias can unduly influence the effectiveness of implicit user feedback. Presentation bias includes various aspects of presentation, such as an attractive title or snippet provided with the document result, and where the document result appears in the presented ranking (position). Note that users tend to click results with good snippets, or that are higher in the ranking, regardless of the real relevance of the document to the query as compared with the other results.
As shown in
In general, the presentation of the results in the page can have a big impact on the observed click rates, and various aspects of presentation are often independent of the quality of a given result. The bolding 4030 of the results (both in the title and the snippet), length of titles 4040, indenting 4050, presence of ads 4060 (or similar objects that appear above the first result for a given query) can affect the observed click rates. Thus, in order to accurately predict the quality of a result based on click signals, the effects of presentation bias should be factored out. Note that the presented rank (the position) of a document result is really a mixed feature since clicks based on position are influenced by both the presentation bias aspect and the result quality aspect, and thus in some implementations, position can be treated as a quality feature (limiting the bias features to those features that are unconnected with result quality, e.g., bolding, length of title, length of snippet, presence of ads, etc.).
Based on the multiple features, information regarding document result selections for searches performed using the document search service can be obtained 4120, where the information spans multiple different queries. This information can be used to generate 4130 a prior model, where the prior model represents a probability of document result selection (e.g., a query-independent probability of selection) given values of the multiple features. For example,
One or more user selection log(s) 4300 include click through data, such as described above. These user selection log(s) 4300 can be understood conceptually as a set of queries, including a first query 4310 through query N, where many user searches, and document result selections, have been observed within the context of each query. A query 4310 includes a document result 4320 that was selected by a user, and the document result 4320 can have an associated relevance feature 4322 (e.g., the IR score of the document) and an associated display bias feature 4324 (e.g., the position of the document).
In addition, another document result 4330, which was returned along with first document result 4320 within the context of the query 4310, has its own associated relevance feature and display bias feature. These features of the document result 4330, presented along with the document result 4320, can be considered features of the document result 4320. Thus, for example, even if the document result 4330 was never selected, the IR score of the document result 4330 can be saved as a feature of the document result 4320, which was selected, in a single log entry including the IR score of the result 4320, the position of result 4320, and the IR score of the result 4330 returned along with result 4320. Note that being “returned along with” does not indicate the two documents were necessarily displayed in the same page of the user interface, but only that the two documents were included in the search results for the given query.
Thus, given a set of selected features, user selections can be collected 4350 from the user selection log(s) 4300 across queries (e.g., independent of query) for multiple combinations of values of the selected features, where those features can include a feature related to another document returned along with a selected document within the context of a query. This collecting can involve accumulating counts observed in the log(s) 4300, including separately accumulating short, medium and long clicks, as well as the event of a result being shown, but another result being clicked.
A query-independent selection probability can be determined 4360 given the selected features. For example, in the case of the features being position and language, the number of long, short and medium clicks at position 1 on an English language search site can be extracted from the log(s) 4300, as well as the number of times other results have been clicked while a result at position 1 on the English language search site has been shown. These numbers can be found for every combination of position and language to determine the query-independent selection probability. In general, a mapping can be computed (based on the log(s) 4300) between the selected feature values and a predicted click through rate to produce a prior model for the selected features.
If the number of feature value combinations is too large (e.g., due to limitations on data storage or processing resources), the features can be split into two groups. The first group can include a reduced set of more important features, and the second group can contain the rest (i.e., two disjoint groups forming the full set of features). The first group can be used for accumulating click counts based on all possible combinations of feature values in the first group. With the click counts collected, the prior model can be trained on the values of the second group of features. Note that the prior model can be a standard linear or logistic regression, or other model types. In general though, the prior model can essentially represent the statistics of historical click data, indicating what percentage of people clicked on a result given presentation of the result and the set of features. Moreover, splitting the features into two groups can allow the use of a larger number of parameters in the model (for every combination of feature values in group one), but not so many parameters that the model cannot be trained.
Referring again to
For example, for the position prior model described above, the ranking score of a search result can be multiplied by a number smaller than one if the traditional click fraction is smaller than the click fraction predicted by the prior model, or by a number greater than one otherwise. Thus, the prior model can be used to clean the signal of an implicit user feedback model that gives a measure of relevance for a document result within the context of the search query for which the document result is returned, thereby reducing the affects of presentation bias. In general, various uses can be made of, or transforms applied to the output signal of the prior model before application to search result ranking for a new search. For example, the following formulas can be used:
Boost=C(a/p),
where p is the predicted click fraction, a is the actual click fraction, and C is a tuned parameter;
Boost=max(min(1+Z,m0),m1), where
Z=(a−p)^k1 if a≧p,
Z=−1*abs((a−p)^k2) if a<p,
p is the predicted click fraction, a is the actual click fraction, and m0, m1, k1 and k2 are tuned parameters;
Boost=max(min(1+Z,m0),m1), where
Z=C[1/(1+e^(−k(a−p)))]−½,
p is the predicted click fraction, a is the actual click fraction, and m0, m1, k and C are tuned parameters.
In any event, the manner in which the output signal of the prior model is used can be adjusted based on the specific implementation 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 prior model 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).
For example, the second set of features can include position, IR score, IR score of the top result, IR score of the previous result(s), IR score of the next result(s), country, language, and number of words in the query; and the first set of features can include the above features, plus the number of bold terms in the title of the current result, the number of bold terms in the title of the previous result(s), the number of bold terms in the title of next result(s), the number of bold terms in the snippet of the current result, the number of bold terms in the snippet of the previous result(s), the number of bold terms in the snippet of the next result(s), the length of the title of the current result, the length of the title of the previous result(s), the length of the title of next result(s), the length of the snippet of the current result, the length of the snippet of the previous result(s), the length of the snippet of the next result(s), a Boolean indicator of whether any pornographic terms were shown (e.g., in the current result, the previous result(s) and the next result(s)), a Boolean indicator of whether an ad was shown, and a Boolean indicator of whether a special result was shown.
Two prior models can be created 4420 using the two sets of features: a first user selection model based on the first set of features and a second user selection model based on the second set of features, using the techniques described above. Thus, the second user selection model can be built on features that are highly correlated with the quality of the result, and the first user selection model can be built on those features and presentation bias features. Any difference between these two models then is likely to be due only to presentation bias.
Both the first and second user selection models can be output 4430 to a ranking engine for ranking of search results. For example, for every search event, the ratio of the click fraction predicted by the second prior model over the click fraction predicted by the first prior model can be computed (based on the observed feature values) for every document result, and the ranking score of each document result can be multiplied by its corresponding ratio, or by a monotonic function of this ratio. The basic rationale embodied by this approach is that, if a result is expected to have a higher click rate due to presentation bias, this result's click evidence should be discounted; and if the result is expected to have a lower click rate due to presentation bias, this result's click evidence should be over-counted.
It should be appreciated that many different prior models can be generated using the techniques described above. The various features that can be used to generate prior models can include all those described above, plus any of the additional features detailed in the following table:
In addition, multiple different types of prior models can be built and used in conjunction with each other.
Information regarding document result selections can be collected 4520 from this reduced set, and a prior model can be built on this information using the techniques described above. For example, a prior model can be built on the extracted results and features such as position, country, language and IR score. If a result for a new search does not appear at all in the regular implicit user feedback model (e.g., the traditional click fraction is undefined for the result because the result has not been selected before) this result can be assigned a click fraction as predicted by this additional prior model for ranking purposes.
A second signal can be obtained 4620, in accordance with a second implicit user feedback model, for the document result, where the second signal corresponds to an aspect of search result presentation. For example, the second signal can include a measure of presentation bias for the document result independent of the context of the search query for which the document result is returned (e.g., the second signal can be the output of a prior model built using the techniques described above in connection with
Furthermore, an additional document result of the search query can be identified 4650 that is not included in the first implicit user feedback model (e.g., a document result for which the traditional click fraction is not defined). An additional signal can be obtained 4670, in accordance with a third implicit user feedback model, for the additional document result, where the additional signal corresponds to an aspect of document relevance independent of the context of the search query (e.g., the additional signal can be the output of the prior model described above in connection with
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
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 traditional click fraction model described. 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 prior model(s) can be applied in varying manners. For example, a prior model can be applied at run time as an adjustment to the ranking boost given to a document in accordance with the implicit user feedback model since the set of features used for the prior model can be available for direct input at run time. Alternatively, the prior model can be applied at model building time, where features are fetched from the log(s), which can result in improved response time during searches. In addition, when the model is applied at building time, the implicit feedback can be adjusted per each click record before aggregating the feedback from multiple clicks into a signal. This adjustment can be for instance a weighting of the clicks according to how much they were affected by display bias before the clicks are aggregated. At run time, the signal is typically only adjusted after the clicks were already aggregated, which can result in some loss of precision.
| Number | Name | Date | Kind |
|---|---|---|---|
| 5265065 | Turtle | Nov 1993 | A |
| 5488725 | Turtle | Jan 1996 | A |
| 5696962 | Kupiec | Dec 1997 | A |
| 6006222 | Culliss | Dec 1999 | A |
| 6014665 | Culliss | Jan 2000 | A |
| 6026388 | Liddy et al. | Feb 2000 | A |
| 6067565 | Horvitz | May 2000 | A |
| 6078916 | Culliss | Jun 2000 | A |
| 6078917 | Paulsen et al. | Jun 2000 | A |
| 6088692 | Driscoll | Jul 2000 | A |
| 6182068 | Culliss | Jan 2001 | B1 |
| 6185559 | Brin et al. | Feb 2001 | B1 |
| 6249252 | Dupray | Jun 2001 | B1 |
| 6285999 | Page | Sep 2001 | B1 |
| 6321228 | Crandall et al. | Nov 2001 | B1 |
| 6327590 | Chidlovskii et al. | Dec 2001 | B1 |
| 6341283 | Yamakawa et al. | Jan 2002 | B1 |
| 6353849 | Linsk | Mar 2002 | B1 |
| 6363378 | Conklin et al. | Mar 2002 | B1 |
| 6370526 | Agrawal et al. | Apr 2002 | B1 |
| 6421675 | Ryan et al. | Jul 2002 | B1 |
| 6480843 | Li | Nov 2002 | B2 |
| 6490575 | Berstis | Dec 2002 | B1 |
| 6526440 | Bharat | Feb 2003 | B1 |
| 6529903 | Smith et al. | Mar 2003 | B2 |
| 6539377 | Culliss | Mar 2003 | B1 |
| 6560590 | Shwe et al. | May 2003 | B1 |
| 6567103 | Chaudhry | May 2003 | B1 |
| 6587848 | Aggarwal et al. | Jul 2003 | B1 |
| 6615209 | Gomes | Sep 2003 | B1 |
| 6623529 | Lakritz | Sep 2003 | B1 |
| 6658423 | Pugh et al. | Dec 2003 | B1 |
| 6671681 | Emens et al. | Dec 2003 | B1 |
| 6678681 | Brin et al. | Jan 2004 | B1 |
| 6701309 | Beeferman et al. | Mar 2004 | B1 |
| 6725259 | Bharat | Apr 2004 | B1 |
| 6738764 | Mao et al. | May 2004 | B2 |
| 6754873 | Law et al. | Jun 2004 | B1 |
| 6792416 | Soetarman et al. | Sep 2004 | B2 |
| 6795820 | Barnett | Sep 2004 | B2 |
| 6816850 | Culliss | Nov 2004 | B2 |
| 6853993 | Ortega et al. | Feb 2005 | B2 |
| 6877002 | Prince | Apr 2005 | B2 |
| 6882999 | Cohen et al. | Apr 2005 | B2 |
| 6901402 | Corston-Oliver et al. | May 2005 | B1 |
| 6944611 | Flank et al. | Sep 2005 | B2 |
| 6944612 | Roustant et al. | Sep 2005 | B2 |
| 6954750 | Bradford | Oct 2005 | B2 |
| 6990453 | Wang et al. | Jan 2006 | B2 |
| 7016939 | Rothwell et al. | Mar 2006 | B1 |
| 7028027 | Zha et al. | Apr 2006 | B1 |
| 7072886 | Salmenkaita et al. | Jul 2006 | B2 |
| 7085761 | Shibata | Aug 2006 | B2 |
| 7113939 | Chou et al. | Sep 2006 | B2 |
| 7117206 | Bharat et al. | Oct 2006 | B1 |
| 7136849 | Patrick | Nov 2006 | B2 |
| 7146361 | Broder et al. | Dec 2006 | B2 |
| 7222127 | Bem et al. | May 2007 | B1 |
| 7231399 | Bem et al. | Jun 2007 | B1 |
| 7243102 | Naam et al. | Jul 2007 | B1 |
| 7266765 | Golovchinsky et al. | Sep 2007 | B2 |
| 7293016 | Shakib et al. | Nov 2007 | B1 |
| 7379951 | Chkodrov et al. | May 2008 | B2 |
| 7382358 | Kushler et al. | Jun 2008 | B2 |
| 7395222 | Sotos | Jul 2008 | B1 |
| 7426507 | Patterson | Sep 2008 | B1 |
| 7451487 | Oliver et al. | Nov 2008 | B2 |
| 7505964 | Tong et al. | Mar 2009 | B2 |
| 7516146 | Robertson et al. | Apr 2009 | B2 |
| 7526470 | Karnawat et al. | Apr 2009 | B1 |
| 7533092 | Berkhin et al. | May 2009 | B2 |
| 7533130 | Narayana et al. | May 2009 | B2 |
| 7552112 | Jhala et al. | Jun 2009 | B2 |
| 7565363 | Anwar | Jul 2009 | B2 |
| 7565367 | Barrett et al. | Jul 2009 | B2 |
| 7574530 | Wang et al. | Aug 2009 | B2 |
| 7584181 | Zeng et al. | Sep 2009 | B2 |
| 7610282 | Datar et al. | Oct 2009 | B1 |
| 7636714 | Lamping et al. | Dec 2009 | B1 |
| 7657626 | Zwicky | Feb 2010 | B1 |
| 7680775 | Levin et al. | Mar 2010 | B2 |
| 7693818 | Majumder | Apr 2010 | B2 |
| 7716225 | Dean et al. | May 2010 | B1 |
| 7747612 | Thun et al. | Jun 2010 | B2 |
| 7756887 | Haveliwala | Jul 2010 | B1 |
| 7783632 | Richardson et al. | Aug 2010 | B2 |
| 7801885 | Verma | Sep 2010 | B1 |
| 7809716 | Wang et al. | Oct 2010 | B2 |
| 7818320 | Makeev | Oct 2010 | B2 |
| 7836058 | Chellapilla | Nov 2010 | B2 |
| 7844589 | Wang et al. | Nov 2010 | B2 |
| 7849089 | Zhang et al. | Dec 2010 | B2 |
| 7853557 | Schneider et al. | Dec 2010 | B2 |
| 7877404 | Achan et al. | Jan 2011 | B2 |
| 7895177 | Wu | Feb 2011 | B2 |
| 7925498 | Baker et al. | Apr 2011 | B1 |
| 7953740 | Vadon et al. | May 2011 | B1 |
| 7974974 | Tankovich et al. | Jul 2011 | B2 |
| 7987185 | Mysen et al. | Jul 2011 | B1 |
| 8001136 | Papachristou et al. | Aug 2011 | B1 |
| 8019650 | Donsbach et al. | Sep 2011 | B2 |
| 8024326 | Tong et al. | Sep 2011 | B2 |
| 8024330 | Franco et al. | Sep 2011 | B1 |
| 8027439 | Zoldi et al. | Sep 2011 | B2 |
| 8037042 | Anderson et al. | Oct 2011 | B2 |
| 8037043 | Zoeter et al. | Oct 2011 | B2 |
| 8037086 | Upstill et al. | Oct 2011 | B1 |
| 8051061 | Niu et al. | Nov 2011 | B2 |
| 8060497 | Zatsman et al. | Nov 2011 | B1 |
| 8065296 | Franz et al. | Nov 2011 | B1 |
| 8069182 | Pieper | Nov 2011 | B2 |
| 8073263 | Hull et al. | Dec 2011 | B2 |
| 8073772 | Bishop et al. | Dec 2011 | B2 |
| 8086599 | Heymans | Dec 2011 | B1 |
| 8090717 | Bharat et al. | Jan 2012 | B1 |
| 8156111 | Jones et al. | Apr 2012 | B2 |
| 8224827 | Dean et al. | Jul 2012 | B2 |
| 20010000356 | Woods | Apr 2001 | A1 |
| 20020034292 | Tuoriniemi et al. | Mar 2002 | A1 |
| 20020042791 | Smith et al. | Apr 2002 | A1 |
| 20020049752 | Bowman et al. | Apr 2002 | A1 |
| 20020103790 | Wang et al. | Aug 2002 | A1 |
| 20020123988 | Dean et al. | Sep 2002 | A1 |
| 20020133481 | Smith et al. | Sep 2002 | A1 |
| 20020165849 | Singh et al. | Nov 2002 | A1 |
| 20030009399 | Boerner | Jan 2003 | A1 |
| 20030018707 | Flocken | Jan 2003 | A1 |
| 20030028529 | Cheung et al. | Feb 2003 | A1 |
| 20030037074 | Dwork et al. | Feb 2003 | A1 |
| 20030078914 | Witbrock | Apr 2003 | A1 |
| 20030120654 | Edlund et al. | Jun 2003 | A1 |
| 20030135490 | Barrett et al. | Jul 2003 | A1 |
| 20030149704 | Yayoi et al. | Aug 2003 | A1 |
| 20030167252 | Odom et al. | Sep 2003 | A1 |
| 20030195877 | Ford et al. | Oct 2003 | A1 |
| 20030204495 | Lehnert | Oct 2003 | A1 |
| 20030220913 | Doganata et al. | Nov 2003 | A1 |
| 20030229640 | Carlson et al. | Dec 2003 | A1 |
| 20040006456 | Loofbourrow et al. | Jan 2004 | A1 |
| 20040006740 | Krohn et al. | Jan 2004 | A1 |
| 20040034632 | Carmel et al. | Feb 2004 | A1 |
| 20040059708 | Dean et al. | Mar 2004 | A1 |
| 20040083205 | Yeager | Apr 2004 | A1 |
| 20040093325 | Banerjee et al. | May 2004 | A1 |
| 20040119740 | Chang et al. | Jun 2004 | A1 |
| 20040122811 | Page | Jun 2004 | A1 |
| 20040153472 | Rieffanaugh, Jr. | Aug 2004 | A1 |
| 20040158560 | Wen et al. | Aug 2004 | A1 |
| 20040186828 | Yadav | Sep 2004 | A1 |
| 20040186996 | Gibbs et al. | Sep 2004 | A1 |
| 20040199419 | Kim et al. | Oct 2004 | A1 |
| 20040215607 | Travis, Jr. | Oct 2004 | A1 |
| 20050015366 | Carrasco et al. | Jan 2005 | A1 |
| 20050027691 | Brin et al. | Feb 2005 | A1 |
| 20050033803 | Vleet et al. | Feb 2005 | A1 |
| 20050050014 | Gosse et al. | Mar 2005 | A1 |
| 20050055342 | Bharat et al. | Mar 2005 | A1 |
| 20050055345 | Ripley | Mar 2005 | A1 |
| 20050060290 | Herscovici et al. | Mar 2005 | A1 |
| 20050060310 | Tong et al. | Mar 2005 | A1 |
| 20050060311 | Tong et al. | Mar 2005 | A1 |
| 20050071741 | Acharya et al. | Mar 2005 | A1 |
| 20050102282 | Linden | May 2005 | A1 |
| 20050125376 | Curtis et al. | Jun 2005 | A1 |
| 20050160083 | Robinson | Jul 2005 | A1 |
| 20050192946 | Lu et al. | Sep 2005 | A1 |
| 20050198026 | Dehlinger et al. | Sep 2005 | A1 |
| 20050222987 | Vadon | Oct 2005 | A1 |
| 20050222998 | Driessen et al. | Oct 2005 | A1 |
| 20050240576 | Piscitello et al. | Oct 2005 | A1 |
| 20050240580 | Zamir et al. | Oct 2005 | A1 |
| 20050256848 | Alpert et al. | Nov 2005 | A1 |
| 20060047643 | Chaman | Mar 2006 | A1 |
| 20060069667 | Manasse et al. | Mar 2006 | A1 |
| 20060089926 | Knepper et al. | Apr 2006 | A1 |
| 20060095421 | Nagai et al. | May 2006 | A1 |
| 20060106793 | Liang | May 2006 | A1 |
| 20060173830 | Smyth et al. | Aug 2006 | A1 |
| 20060195443 | Franklin et al. | Aug 2006 | A1 |
| 20060200476 | Gottumukkala et al. | Sep 2006 | A1 |
| 20060200556 | Brave et al. | Sep 2006 | A1 |
| 20060230040 | Curtis et al. | Oct 2006 | A1 |
| 20060293950 | Meek et al. | Dec 2006 | A1 |
| 20070005575 | Dai et al. | Jan 2007 | A1 |
| 20070005588 | Zhang et al. | Jan 2007 | A1 |
| 20070038659 | Datar et al. | Feb 2007 | A1 |
| 20070061211 | Ramer et al. | Mar 2007 | A1 |
| 20070081197 | Omoigui | Apr 2007 | A1 |
| 20070106659 | Lu et al. | May 2007 | A1 |
| 20070112730 | Gulli et al. | May 2007 | A1 |
| 20070130370 | Akaezuwa | Jun 2007 | A1 |
| 20070156677 | Szabo | Jul 2007 | A1 |
| 20070192190 | Granville | Aug 2007 | A1 |
| 20070208730 | Agichtein et al. | Sep 2007 | A1 |
| 20070214131 | Cucerzan et al. | Sep 2007 | A1 |
| 20070233653 | Biggs et al. | Oct 2007 | A1 |
| 20070255689 | Sun et al. | Nov 2007 | A1 |
| 20070260597 | Cramer | Nov 2007 | A1 |
| 20070266021 | Aravamudan et al. | Nov 2007 | A1 |
| 20070266439 | Kraft | Nov 2007 | A1 |
| 20070288450 | Datta et al. | Dec 2007 | A1 |
| 20080010143 | Kniaz et al. | Jan 2008 | A1 |
| 20080027913 | Chang et al. | Jan 2008 | A1 |
| 20080052219 | Sandholm et al. | Feb 2008 | A1 |
| 20080052273 | Pickens | Feb 2008 | A1 |
| 20080059453 | Laderman | Mar 2008 | A1 |
| 20080077570 | Tang et al. | Mar 2008 | A1 |
| 20080082518 | Loftesness | Apr 2008 | A1 |
| 20080091650 | Fontoura et al. | Apr 2008 | A1 |
| 20080114624 | Kitts | May 2008 | A1 |
| 20080114729 | Raman et al. | May 2008 | A1 |
| 20080114750 | Saxena et al. | May 2008 | A1 |
| 20080140699 | Jones et al. | Jun 2008 | A1 |
| 20080162475 | Meggs et al. | Jul 2008 | A1 |
| 20080183660 | Szulczewski | Jul 2008 | A1 |
| 20080189269 | Olsen | Aug 2008 | A1 |
| 20080208825 | Curtis et al. | Aug 2008 | A1 |
| 20080228442 | Lippincott et al. | Sep 2008 | A1 |
| 20080256050 | Zhang et al. | Oct 2008 | A1 |
| 20080313168 | Liu et al. | Dec 2008 | A1 |
| 20080313247 | Galvin | Dec 2008 | A1 |
| 20090012969 | Rail et al. | Jan 2009 | A1 |
| 20090055392 | Gupta et al. | Feb 2009 | A1 |
| 20090157643 | Gollapudi et al. | Jun 2009 | A1 |
| 20090182723 | Shnitko et al. | Jul 2009 | A1 |
| 20090287656 | Bennett | Nov 2009 | A1 |
| 20100106706 | Rorex et al. | Apr 2010 | A1 |
| 20100131563 | Yin | May 2010 | A1 |
| 20100205541 | Rapaport et al. | Aug 2010 | A1 |
| 20100228738 | Mehta et al. | Sep 2010 | A1 |
| 20100241472 | Hernandez | Sep 2010 | A1 |
| 20110295844 | Sun et al. | Dec 2011 | A1 |
| 20120191705 | Tong et al. | Jul 2012 | A1 |
| Number | Date | Country |
|---|---|---|
| WO 0077689 | Dec 2000 | WO |
| WO 0116807 | Mar 2001 | WO |
| WO 0167297 | Sep 2001 | WO |
| WO 2004059514 | Jul 2004 | WO |
| Entry |
|---|
| Boyan, Freitag, Joachims, “A Machine Learning Architecture for Optimizing Web Search Engines”; Oct. 9, 1996; AAAI Techincal Report Ws-96-06. |
| Joachims T. Evaluating retrieval performance using clickthrough data. Proceedings of the SIGIR Workshop on Mathematical/Formal Methods in Information Retrieval; Aug. 12-15, 2002; Tampere, Finland; 2002. |
| Radlinski, Joachims; “Query Chains : Learnign to Rank for Implict Feedback”; International Conference on Knowledge Discovery and Data Mining archive, Proceedings of the eleventh ACM SIGKDD international conference on Knowledge discovery in data mining; 2005. |
| Bar-Llan, Keenoy, Levene, Yaari; “Presentation Bias is Significant in Determing User Preseference for Search Results—A User Study”; Journal of the American Society for Information Science and Technology, vol. 60, Issue 1 (p. 135-149), Sep. 2008. |
| Jpachins. Radlinski; “Search Engines that Learn from Implict Feedback”; Aug. 2007, IEEE Computer Society. |
| Bar-Ilan, Mat-Hassan, Levene; “Methods for comparing rankings of search engine results”; Computer Networks: The International Journal of Computer and Telecommunications Networking, vol. 50 , Issue 10 (Jul. 2006). |
| U.S. Appl. No. 10/726,345, filed Dec. 3, 2003, Pearson et al. |
| U.S. Appl. No. 10/878,926, filed Jun. 28, 2004, Battle et al. |
| Agichtein, et al; Improving Web Search Ranking by Incorporating User Behavior Information; Aug. 2006; Proceedings of the Twenty-Ninth Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, p. 19-26. |
| Agichtein, et al; Learning User Interaction Models for Predicting Web Search Result Performances; Aug. 2006; Proceedings of the Twenty-Ninth Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, p. 3-10. |
| Boyan et al.; A Machine Learning Architecture for Optimizing Web Search Engines; Aug. 1996; Internet-based information systems—Workshop Technical Report—American Association for Artificial Intelligence, p. 1-8. |
| Cutrell, et al.; Eye tracking in MSN Search: Investigating snippet length, target position and task types; 2007; Conference on Human Factors in Computing Systems—Proceedings of the SIGCHI Conference on Human Factors in Computing Systems. |
| Joachims; Optimizing search engines using clickthrough data; 2002; Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, p. 133-142. |
| Kelly, et al.; Implicit Feedback for Inferring User Preference: A Bibliography; SIGIR Forum, vol. 37, No. 2 (2003), pp. 18-28. |
| U.S. Appl. No. 12/331,872, filed Dec. 10, 2008, Le et al. |
| W3C, URIs, URLs and URNs: Classification and Recommendations 1.0, Report from the joint W3C/IETF URI Planning Interest Group, Sep. 21, 2001, 8 pages. |
| Joachims, “Evaluating Search Engines Using Clickthrough Data”, Cornell University, Department of Computer Science, Draft, Feb. 19, 2002, 13 pages. |
| Jansen et al., “An Analysis of Web Documents Retrieved and Viewed”, School of Information Sciences and Technology, The Pennsylvania State University, the 4th International Conference on Internet Computing, Las Vegas, Nevada, pp. 65-69, Jun. 23-26, 2003, 5 pages. |
| Boldi, et al.; The Query-flow Graph: Model and Applications; CKIM '08, Oct. 26-30, Napaya Valley, California, USA, pp. 609-617. |
| Burke, Robin, Integrating Knowledge-based and Collaborative-filtering Recommender Systems, AAAI Technical Report WS-99-01. Compilation copyright © 1999, AAAI (www.aaai.org), pp. 69-72. |
| Craswell, et al.; Random Walks on the Click Graph; Jul. 2007; SIGIR '07, Amsterdam, The Netherlands, 8 pages. |
| Diligenti, et al., Users, Queries and Documents: A Unified Representation for Web Mining, wi-iat, vol. 1, 2009 IEEE/WIC/ACM International Joint Conference on Web Intelligence and Intelligent Agent Technology, 2009, pp. 238-244. |
| Hofmann, Thomas, Latent Semantic Models for Collaborative Filtering, ACM Transactions on Information Systems, vol. 22, No. 1, Jan. 2004, pp. 89-115. |
| Google News archive, Jul. 8, 2003, Webmasterworld.com, [online] Retrieved from the Internet http://www.webmasterwolrd.com/forum3/15085.htm [retrieved on Nov. 20, 2009] 3 pages. |
| Gr{hacek over (c)}ar, Miha, User Profiling: Collaborative Filtering, SIKDD 2004, Oct. 12-15, 2004, Ljubljana, Slovenia, 4 pages. |
| U.S. Appl. No. 11/556,143, filed Nov. 2, 2006, in Office Action mailed Jul. 6, 2010, 20 pages. |
| U.S. Appl. No. 11/556,143, filed Nov. 2, 2006, in Office Action mailed Apr. 20, 2011, 18 pages. |
| McDonnell, Philip, A., “Time Based Ranking,” U.S. Appl. No. 11/870,893, filed Oct. 11, 2007, 42 pages. |
| Nicole, Kristen, Heeii is StumbleUpon Plus Google Suggestions, [online], Retrieved from the Internet http://mashable.com/2007/05/15/heeii/, 11 pages. |
| Lemire, Daniel, Scale and Translation Invariant Collaborative Filtering Systems, Published in Information Retrieval, 8(1), pp. 129-150, 2005. |
| U.S. Appl. No. 11/556,086, filed Nov. 2, 2006, in Office Action mailed Jun. 23, 2010, 21 pages. |
| Schwab, et al., Adaptivity through Unobstrusive Learning, 2002, 16(3), pp. 5-9. |
| Stoilova, Lubomira et al., GiveALink: Mining a Semantic Network of Bookmarks for Web Search and Recommendation, LinkKDD '05, Aug. 21, 2005, Chicago, IL, USA, 8 pages. |
| Xiao, et al., Measuring Similarity of Interests for Clustering Web-Users, ADC, 2001, pp. 107-114. |
| Xie et al., Web User Clustering from Access Log Using Belief Function, K-CAP '01, Oct. 22-23, 2001, Victoria, British Columbia, Canada, pp. 202-208. |
| Yu et al., Selecting Relevant Instances for Efficient and Accurate Collaborative Filtering, CIKM '01, Nov. 5-10, 2001, Atlanta, Georgia, pp. 239-246. |
| Zeng et al., Similarity Measure and Instance Selection for Collaborative Filtering, WWW '03, May 20-24, 2003, Budapest, Hungary, pp. 652-658. |
| Zeng, et al., “Learning to Cluster Web Search Results”, SIGIR '04, Proceedings of the 27th Annual International ACM SIGIR conference on research and development in information retrieval, 2004. |
| Soumen Chakrabarti, et al. “Enhanced Topic Distillation using Text, Markup tags, and Hyperlinks” ACM, Sep. 9-12, 2001, pp. 208-216. |
| Gabriel Somlo et al., “Using Web Hepler Agent Profiles in Query Generation”, ACM, Jul. 2003, pp. 812-818. |
| Australian Patent Office Non-Final Office Action in AU App. Ser. No. 2004275274, mailed Feb. 3, 2010, 2 pages. |
| Dan Olsen et al., “Query-by-critique: Spoken Language Access to Large Lists”, ACM, Oct. 2002, pp. 131-140. |
| Susan Gauch et al., “A Corpus Analysis Approach for Automatic Query Expansion and its Extension to Multiple Databases”, ACM, Jul. 1999, pp. 250-269. |
| Nicolas Bruno et al., “Top-K Selection Queries over Relational Databases: Mapping Strategies and Performance Evaluation”, ACM, Jun. 2002, pp. 153-187. |
| Ji-Rong Wen et al., “Query Clustering using User Logs”, ACM, Jan. 2002, pp. 59-81. |
| Brin, S. and L. Page, The Anatomy of a Large-Scale Hypertextual Web Search Engine, Computer Science Department, Apr. 1998. |
| International Search Report and Written Opinion for Application No. PCT/US2004/029615, dated Jan. 19, 2005, 8 pages. |
| Hungarian Patent Office, International Search Report and Written Opinion for Application No. 200806756-3, dated Nov. 19, 2010 12 pages. |
| Authorized Officer Athina Nickitas-Etienne, International Preliminary Report and Written Opinion for Application No. PCT/US2004/029615, mailed Mar. 23, 2006. |
| Indian Office Action in Indian Application No. 686/KOLNP/2006, mailed Jun. 3, 2008, 2 pages. |
| Danish Search Report and Written Opinion for Application No. 200601630-7, dated Jun. 21, 2007, 15 pages. |
| Jones et al., “Pictures of Relevance: A Geometric Analysis of Similarity Measures”, Journal of the American Society for Information Science, Nov. 1987, 23 pages. |
| Kaplan et al., “Adaptive Hypertext Navigation Based on User Goals and Context”, User Modeling and User-Adapted Interaction 2, Sep. 1, 1993; pp. 193-220, 28 pages. |
| Liddy et al., “A Natural Language Text Retrieval System With Relevance Feedback”, 16th National Online, May 2-6, 1995, 3 pages. |
| U.S. Appl. No. 11/556,143, filed Nov. 2, 2006, in Office Action mailed Jan. 25, 2010, 14 pages. |
