Representing queries and determining similarity based on an ARIMA model

Many search engine services, such as Google and Yahoo, provide for searching for information that is accessible via the Internet. These search engine services allow users to search for display pages, such as web pages, that may be of interest to users. After a user submits a search request (i.e., a query) that includes search terms, the search engine service identifies web pages that may be related to those search terms. To quickly identify related web pages, the search engine services may maintain a mapping of keywords to web pages. This mapping may be generated by “crawling” the web (i.e., the World Wide Web) to identify the keywords of each web page. To crawl the web, a search engine service may use a list of root web pages to identify all web pages that are accessible through those root web pages. The keywords of any particular web page can be identified using various well-known information retrieval techniques, such as identifying the words of a headline, the words supplied in the metadata of the web page, the words that are highlighted, and so on. The search engine service identifies web pages that may be related to the search request based on how well the keywords of a web page match the words of the query. The search engine service then displays to the user links to the identified web pages in an order that is based on a ranking that may be determined by their relevance to the query, popularity, importance, and/or some other measure.

Search engine services obtain revenue by placing advertisements along with search results. These paid-for advertisements are commonly referred to as “sponsored links,” “sponsored matches,” or “paid-for search results.” An advertiser who wants to place an advertisement (e.g., a link to their web page) along with certain search results provides a search engine service with an advertisement and one or more bid terms. When a search request is received, the search engine service identifies the advertisements whose bid terms match the terms of the search request. The search engine service then selects advertisements to display based on the closeness of their match, the amount of money that the advertisers are willing to pay for placing the advertisement, and other factors. The search engine service then adds a sponsored link to the search result that points to a web page of the advertiser. The search engine services typically either charge for placement of each advertisement along with search results (i.e., cost per impression) or charge only when a user actually selects a link associated with an advertisement (i.e., cost per click).

Advertisers would like to maximize the effectiveness of their advertising dollars used to pay for advertisements. Thus, advertisers try to identify bid term, advertisement, and bid amount combinations that result in the highest benefits (e.g., most profit) to the advertiser. Advertisers may analyze query trends to identify bid terms, timing for placing advertisements based on those bid terms, bid amounts for those bid terms, and so on. Query trend analysis studies how the frequency of queries changes over time so that future frequency of queries can be predicted. If query trends can be accurately predicted, then advertisers can adjust their placement of advertisements in an attempt to maximize the advertising effectiveness. For example, if the frequency of a query is likely to increase in the near future, an advertiser may want to increase the bid amount for terms of that query. It has been, however, difficult to accurately model the frequency of queries and thus difficult to accurately predict the frequency of queries.

Because of the popularity of search engine services, the query logs generated by search engine services tend to be very large. A query log may include millions of entries, each of which identifies a query that was submitted by a searcher and a time of submission. Because of their size, the query logs consume vast amounts of storage. To reduce their storage requirements, the query logs are often compressed into query frequency information by storing, for each query and for each interval (e.g., a day), the frequency (i.e., count of submissions) of that query for that interval, rather an entry for each individual query submission. Nevertheless, because millions of different queries can be submitted by searchers, even the query frequency information consumes large amounts of storage.

Representing queries and determining similarity of queries based on an autoregressive integrated moving average (“ARIMA”) model is provided by a query analysis system. The query analysis system represents each query by its ARIMA coefficients. The query analysis system may estimate the frequency information for a desired past or future interval based on frequency information for some initial intervals. To estimate the frequency, the query analysis system applies the ARIMA model to the initial intervals to generate the frequency of the next interval. The query analysis system can then repeat the process for each subsequent interval using the newly generated frequency information until the frequency of the desired interval is estimated. The query analysis system may also determine the similarity of a pair of queries based on the similarity of their ARIMA coefficients. The query analysis system may use various metrics, such as a correlation metric, to determine the similarity of the ARIMA coefficients.

This Summary is provided to introduce a selection of concepts in a simplified form that are further described below in the Detailed Description. This Summary is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used as an aid in determining the scope of the claimed subject matter.

FIG. 1 is a block diagram that illustrates components of a query analysis system in some embodiments.

FIG. 2 is a flow diagram that illustrates high-level processing of a determine similarity of queries component of the query analysis system in some embodiments.

FIG. 3 is a flow diagram that illustrates low-level processing of a determine similarity of queries component of the query analysis system in some embodiments.

FIG. 4 is a flow diagram that illustrates the processing of a generate ARIMA coefficients component of the query analysis system in some embodiments.

FIG. 5 is a flow diagram that illustrates the processing of a calculate similarity score component of the query analysis system in some embodiments.

Representing queries and determining similarity of queries based on an autoregressive integrated moving average (“ARIMA”) model is provided by a query analysis system. An ARIMA model represents a time series using autoregressive terms, integration terms, and moving average terms. The autoregressive terms represent the importance of past values of a time series to the current value. The integration terms represent differences between values with their trend and drift removed. The moving average terms represent a moving average of the values with random noise removed. The ARIMA model may be represented as follows:

ARIMA( p,d,q )  (1)

where p represents the number of autoregressive terms, d represents the number of integration terms, and q represents the number of moving average terms. The number of these terms represents the order of the ARIMA model. An ARIMA (1, 1, 1) model with an order of p, d, and q each equal to 1 may be represented as follows:

X t =μ+X t-1 +φ( X t-1 −X t-2 )−βε t-1   (2)

where X t represents the estimated value at time or interval t, μ, φ, and β represent the ARIMA coefficients, and ε t-1 represents background noise with a mean of zero and a certain variance. ARIMA model with an order of (X,0,0), where X is greater than zero, represents an auto regressive (“AR”) model, and an ARIMA model with an order of (0,0,X) represents a moving average (“MA”) model.

In some embodiments, the query analysis system represents each query by its ARIMA coefficients. To generate the ARIMA coefficients, the query analysis system may use an Akaike Information Criterion (“AIC”) to determine the order of the ARIMA model for the query. (See Box, G. and Jenkins, G. M., “Time Series Analysis: Forecasting and Control,” Prentice Hall, 1994.) Alternatively, the query analysis system may use a fixed order of the ARIMA model (e.g., (1, 1, 1)) for each query or may use a fixed order for some of the terms and a variable order for other terms. The query analysis system then uses a least squares method to determine the ARIMA coefficients for the query. (See Box, G. and Jenkins, G. M., “Time Series Analysis: Forecasting and Control,” Prentice Hall, 1994.) Thus, the ARIMA coefficients of a query represent a compressed form of the query with a reduced dimensionality that is based on the order of the ARIMA model for the query. By representing each query by its ARIMA coefficients, the query analysis system reduces the amount of storage needed to store information describing the frequencies of the queries.

In some embodiments, the query analysis system regenerates the frequency information for a desired interval based on frequency information for some initial intervals. The query analysis system applies the ARIMA model to the initial intervals to generate the frequency of the next interval. The query analysis system can then repeat the process for each subsequent interval using the newly generated frequency information until the frequency of the desired interval is regenerated. For example, for an ARIMA (1, 1, 1) model, the query analysis system would start out with the frequencies of intervals 1 and 2 and then calculate the frequency for interval 3 using Equation 2. The query analysis system would then calculate the frequency for interval 4 using the frequencies of intervals 2 and 3. The query analysis system may also store with the ARIMA coefficients the frequencies for periodic intervals such as the frequency for the first two days of each month. The query analysis system could then quickly regenerate the frequency for a given interval by starting with the closest stored frequency information. The query analysis system can also continue the process of regenerating, or actually generating, for intervals in the future for predicting or forecasting the frequencies. For example, the query analysis system may start out with frequency information for some recent intervals and then, using Equation 2, predict frequencies for intervals in the future.

In some embodiments, the query analysis system determines the similarity of a pair of queries based on the similarity of their ARIMA coefficients. The query analysis system may use various metrics to determine the similarity of the ARIMA coefficients. For example, the query analysis system may use a correlation metric, a Euclidean distance metric, a cosine similarity metric, and so on. The similarity between pairs of queries can be represented by similarity scores. The similarity scores can then be used to analyze query trends of similar queries, which may be helpful in determining the timing of advertisement placements, selecting of keywords for advertisements, determining bid amounts, and so on.

In some embodiments, the query analysis system may represent the ARIMA coefficients of queries p and q as a vector as follows:

X p ARIMA =( X p,1 A , . . . , X p,n p A )

X q ARIMA =( X z,1 A , . . . , X q,n q A )  (3)

where n p and n q represent the number of ARIMA coefficients or dimensions of the vector for queries p and q, respectively. When determining the similarity of a pair of queries, the query analysis system may normalize the dimensions of the vectors by adding coefficients with a value of zero to the shorter vector. Thus, the normalized vectors may be represented as follows:

X p ARIMA =( X p,1 A , . . . , X p,n A )

X q ARIMA =( X q,1 A , . . . , X q,n A )  (4)

where n represents the maximum of n p and n q . The query analysis system may normalize the numbers of p, d, and q coefficients of the queries independently. Thus, if one query has two p coefficients and the other query has only one p coefficient, then the query analysis system adds another p coefficient with a zero value to the vector for the other query. The query analysis system may represent the similarity between a pair of vectors and thus a pair of queries based on a correlation metric as represented by the following:

1 n ⁢ ∑ i ⁢ ( X p , i A - μ ⁡ ( X p A ) σ ⁡ ( X p A ) ) ⁢ ( X q , i A - μ ⁡ ( X q A ) σ ⁡ ( X q A ) ) ( 5 )

where μ(X p A ) represents the mean of the coefficients for query p and σ(X P A ) represents the standard deviation of the coefficients for query p. The value calculated using Equation 5 represents a similarity score indicating the similarity between queries p and q. Such a similarity score ranges between −1 and 1 with a similarity score for a query and itself being 1. One skilled in the art will appreciate that many different variations of determining the correlation of two series of variables may be used to generate a similarity score of a pair of queries.

FIG. 1 is a block diagram that illustrates components of a query analysis system in some embodiments. The query analysis system 110 may be connected to search engine servers 150 , web servers 160 , and user computing devices 170 via a communications link 140 . The search engine servers generate query logs based on queries submitted by users using the user computing devices. The search engine servers provide to the user computing device search results that include links to web pages of the web servers. The query analysis system may include a query log 111 , a query frequency store 112 , an ARIMA coefficients store 113 , a similarity score store 114 , and a calculate query frequencies component 115 . The query log contains query logs received from the various search engine servers. The calculate query frequencies component counts the number of query submissions for each query for each interval and stores the information in the query frequency store. The query frequency store has an entry for each query identified in the query log. Each entry contains a sub-entry for each interval containing the frequency of the submissions of that query during the interval. The ARIMA coefficients store contains an entry for each query along with its ARIMA coefficients and some initial frequencies for regenerating the frequencies. The similarity score store contains an entry for each pair of queries with a similarity score for that pair of queries.

The query analysis system also includes a determine similarity of queries component 121 and an estimate frequency component 131 . The determine similarity of queries component generates the ARIMA coefficients for the queries and calculates a similarity score for each pair of queries. The determine similarity of queries component invokes a generate ARIMA coefficients component 122 for each query and a calculate similarity score component 123 for each pair of queries. The generate ARIMA coefficients component invokes an identify number of ARIMA coefficients component 124 and a calculate ARIMA coefficients component 125 to generate the ARIMA coefficients for a query. The identify number of ARIMA coefficients component may use an AIC algorithm to determine the order of the ARIMA model for a query. The calculate ARIMA coefficients component calculates the ARIMA coefficients for an ARIMA model of that order using a least squares method. The calculate similarity score component calculates the similarity score for a pair of queries based on a correlation between the ARIMA coefficients of the queries. The estimate frequency component uses the ARIMA coefficients to estimate (e.g., regenerate or predict) frequencies for a query.

The computing device on which the query analysis system is implemented may include a central processing unit, memory, input devices (e.g., keyboard and pointing devices), output devices (e.g., display devices), and storage devices (e.g., disk drives). The memory and storage devices are computer-readable media that may be encoded with computer-executable instructions that implement the query analysis system, which means a computer-readable medium that contains the instructions. In addition, the instructions, data structures, and message structures may be stored or transmitted via a data transmission medium, such as a signal on a communication link. Various communication links may be used, such as the Internet, a local area network, a wide area network, a point-to-point dial-up connection, a cell phone network, and so on.

Embodiments of the system may be implemented in and used with various operating environments that include personal computers, server computers, hand-held or laptop devices, multiprocessor systems, microprocessor-based systems, programmable consumer electronics, digital cameras, network PCs, minicomputers, mainframe computers, computing environments that include any of the above systems or devices, and so on.

The query analysis system may be described in the general context of computer-executable instructions, such as program modules, executed by one or more computers or other devices. Generally, program modules include routines, programs, objects, components, data structures, and so on that perform particular tasks or implement particular abstract data types. Typically, the functionality of the program modules may be combined or distributed as desired in various embodiments. In addition, the functions of the query analysis may be performed offline. For example, the query analysis system may generate new ARIMA coefficients for the queries on a periodic basis. Also, the generating of the ARIMA coefficients and the determining of similarity may be performed by computing devices separate from computing devices that use the ARIMA coefficients to estimate frequencies.

FIG. 2 is a flow diagram that illustrates high-level processing of the determine similarity of queries component of the query analysis system in some embodiments. The component generates ARIMA coefficients for each query and then calculates a similarity score for each pair of queries. In block 201 , the component selects the next query. In decision block 202 , if all the queries have already been selected, then the component continues at block 204 , else the component continues at block 203 . In block 203 , the component invokes the generate ARIMA coefficients component to determine the number of ARIMA coefficients and calculate the ARIMA coefficients for the selected query. The component then loops to block 201 to select the next query. In block 204 , the component selects the next pair of queries. In decision block 205 , if all the pairs of queries have already been selected, then the component completes, else the component continues at block 206 . In block 206 , the component calculates a similarity score based on the ARIMA coefficients for the selected pair of queries. The component then loops to block 204 to select the next pair of queries.

FIG. 3 is a flow diagram that illustrates low-level processing of the determine similarity of queries component of the query analysis system in some embodiments. The component generates ARIMA coefficients for each query and then calculates a similarity score for each pair of queries. In blocks 301 - 304 , the component loops generating the ARIMA coefficients for each query and calculating the mean and the standard deviation of the ARIMA coefficients for each query. In block 301 , the component selects the next query. In decision block 302 , if all the queries have already been selected, then the component continues at block 305 , else the component continues at block 303 . In block 303 , the component invokes the generate ARIMA coefficients component to determine the number of ARIMA coefficients and calculate that number of ARIMA coefficients. In block 304 , the component calculates the mean and the standard deviation of the ARIMA coefficients for the selected query for use in calculating the similarity scores. The component then loops to block 301 to select the next query. In blocks 305 - 309 , the component loops generating a similarity score for each pair of queries. In block 305 , the component selects the next query. In decision block 306 , if all the queries have already been selected, then the component completes, else the component continues at block 307 . In blocks 307 - 309 , the component loops selecting each other query for which the similarity score of the selected query and the other query has not yet been calculated. In block 307 , the component selects the next other query. In decision block 308 , if all such other queries have already been selected, then the component loops to block 305 to select the next query, else the component continues at block 309 . In block 309 , the component invokes the calculate similarity score component passing an indication of the selected query and the selected other query to calculate the similarity score for the queries. The component then loops to block 307 to select the next other query.

FIG. 4 is a flow diagram that illustrates the processing of the generate ARIMA coefficients component of the query analysis system in some embodiments. The component is passed a query and determines the numbers of ARIMA coefficients and calculates the value of the ARIMA coefficients. The component may use an AIC algorithm to determine the numbers of coefficients p, d, and q for ARIMA (p, d, q) and their values by minimizing the following objective function:

AIC=2 k+n ln(RSS/ n )  (6)

where AIC represents the value to be minimized, k represents the number of coefficients, n represents the number of intervals of the query, RSS represents the residual sum of squares, and k may be equal to the sum of p, d, and q. The component may successively select combinations of numbers for p, d, and q and then use a least squares algorithm to select values for the coefficients with that number of coefficients. The least squares algorithm selects values that minimize a sum of the squares of the differences between each actual frequency and the estimated frequency using the selected numbers of coefficients. In blocks 401 - 403 , the component loops selecting different combinations of numbers of coefficients and calculating their values until a solution is found that tends to minimize the AIC value. In block 401 , the component selects the next combination of numbers for p, d, and q. The component may use a comprehensive approach by selecting each possible combination of p, d, and q up to a maximum number for each of p, d, and q. Alternatively, the component may attempt to select the next combination based on expected convergence on a solution. In block 402 , the component calculates values for the selected numbers of ARIMA coefficients using a least squares method. One skilled in the art will appreciate that many well-known optimization techniques such as linear regression, gradient descent, and so on can be used to determine the number and value of the coefficients. In decision block 402 , if the numbers of coefficients have converged on a solution (e.g., the AIC value is not changing much), then the component returns, else the component loops to block 401 to select the next combination of the numbers of coefficients.

FIG. 5 is a flow diagram that illustrates the processing of the calculate similarity score component of the query analysis system in some embodiments. The component is passed the ARIMA coefficients of a pair of queries and returns a similarity score indicating similarity of the queries based on a correlation of the ARIMA coefficients. In blocks 501 - 502 , the component calculates the mean and the standard deviation of the ARIMA coefficients if they have not yet been calculated. In block 503 , the component normalizes the number of ARIMA coefficients for the queries. In blocks 504 - 509 , the component loops aggregating the similarity score. In block 504 , the component initializes the similarity score. In block 505 , the component selects the next ARIMA coefficient of the passed query. In decision block 506 , if all the ARIMA coefficients have already been selected, then the component returns the aggregated similarity score, else the component continues at block 507 . In block 507 , the component calculates a correlation factor for one of the passed queries for the selected coefficient. In block 508 , the component calculates a correlation factor for the other of the passed queries for the selected coefficient. In block 509 , the component multiplies the factors and aggregates them into the similarity score. The component then loops to block 505 to select the next ARIMA coefficient.

Although the subject matter has been described in language specific to structural features and/or methodological acts, it is to be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or acts described above. Rather, the specific features and acts described above are disclosed as example forms of implementing the claims. Accordingly, the invention is not limited except as by the appended claims.