Patent Application
20090254397
The invention relates generally to computer systems, and more particularly to an improved system and method for optimizing online keyword auctions subject to budget and estimated query volume constraints.
A major problem faced by an online advertising publisher is to decide how to allocate advertisement space to advertisers in real-time. This problem is complicated by the fact that many advertisers are constrained by their budget, and also by the fact that the total volume of available advertisement space in a given day is unknown ahead of the time. Such is the case for query volumes in sponsored search advertising for instance.
One method for solving this problem is to use the historical data such as bids and query volumes from previous days to obtain an estimate for the volume of available advertising space, and then based on this data, solve a linear program (LP) to allocate advertisement space to advertisers optimally. If the predicted advertising space volume is accurate, this algorithm yields an approximately optimal allocation. However, often the estimates are slightly inaccurate, and this can lead to a few percentage points decrease in the revenue compared to the optimal allocation.
Competitive analysis for evaluating the performance of an online algorithm has been applied to online keyword auctions by Mehta for optimally allocating online advertisement space to budget-constrained advertisers. See Aranyak Mehta, Amin Saberi, Umesh Vazirani, and Vijay Vazirani, Adwords and Generalized On-line Matching, In Proceedings of the 46th Annual IEEE Symposium on Foundations of Computer Science, pages 264-273, 2005. Mehta developed an algorithm with a competitive ratio of 1-1/e where at most one advertisement may be shown for each query. Mehta showed that this is the best possible ratio an online algorithm can achieve in a worst-case analysis where the query volume may be unknown beforehand. Unfortunately, the approach of worst-case analysis for online algorithms favors algorithms that make extremely conservative choices.
What is needed is a way for allocating online advertisement space that reduces the loss when estimates of query volumes are inaccurate. Such a system and method should perform well in the case that the estimates are accurate as well as when the estimates are inaccurate. Moreover, the system and method should be applicable for displaying multiple advertisements for each query and for payments based on a generalized second price auction.
Briefly, the present invention may provide a system and method for optimizing online keyword auctions subject to budget and estimated query volume constraints. In various embodiments, a client having a web browser may be operably coupled to a query processing server for sending a query request. The query processing server may include a query forecasting engine for predicting the query volume during multiple time periods of a time span and a model generator for creating a linear programming model using the predicted query volumes to provide a candidate set of advertisements for keywords of query requests for multiple time periods. The query processing server may also include an operably coupled linear programming analysis engine for optimizing the linear programming model offline to generate slates of advertisements for keywords of a query request for multiple time periods and to generate a frequency for each slate to indicate how often the slate of advertisements should be displayed. The query processing server may additionally include a dynamic programming engine for determining a slate of advertisements for a keyword of a query request that may maximize a weighted sum of the prices in order to compute a near optimal solution when the estimates of query volume are inaccurate. The query processing server may also include a competitive analysis engine for choosing either the slate generated by the linear program or the slate generated by dynamic programming based on whether the weighted sum of prices for the slate of advertisements computed by dynamic programming may be within a factor of the weighted sum of the prices for the slate of advertisements computed by the linear program. The query processing server may then serve the chosen slate of advertisements to accompany the search results of a query request to the web browser.
In an embodiment to compute a near optimal slate of advertisements in the event the estimates of query volume are inaccurate, dynamic programming may be applied to recursively determine a slate of advertisements that may maximize the sum of revenue for allocating advertisements to multiple web page placements calculated for a generalized second price auction using a parameterized discount factor for each bid based on a budget spent by an advertiser. First, advertisers may be sorted in decreasing order by expected value. The maximum value of revenue may then be computed for each advertiser for a generalized second price auction using a parameterized discount factor for each bid based on the budget spent by the advertiser. A slate of advertisements may be constructed in decreasing order by maximum value of revenue computed for each advertiser, and the slate of advertisements may be output.
Advantageously, the present invention may effectively use a forecast of the frequency and sequence of keywords to compute a slate of advertisements and may also compute a near optimal slate of advertisements to use in the event estimates of query volume are inaccurate. Other advantages will become apparent from the following detailed description when taken in conjunction with the drawings, in which:
The invention may be described in the general context of computer-executable instructions, such as program modules, being executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, and so forth, which perform particular tasks or implement particular abstract data types. The invention may also be practiced in distributed computing environments where tasks are performed by remote processing devices that are linked through a communications network. In a distributed computing environment, program modules may be located in local and/or remote computer storage media including memory storage devices.
With reference to
The computer system 100 may include a variety of computer-readable media. Computer-readable media can be any available media that can be accessed by the computer system 100 and includes both volatile and nonvolatile media. For example, computer-readable media may include volatile and nonvolatile computer storage media implemented in any method or technology for storage of information such as computer-readable instructions, data structures, program modules or other data. Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium which can be used to store the desired information and which can accessed by the computer system 100. Communication media may include computer-readable instructions, data structures, program modules or other data in a modulated data signal such as a carrier wave or other transport mechanism and includes any information delivery media. The term “modulated data signal” means a signal that has one or more of its characteristics set or changed in such a manner as to encode information in the signal. For instance, communication media includes wired media such as a wired network or direct-wired connection, and wireless media such as acoustic, RF, infrared and other wireless media.
The system memory 104 includes computer storage media in the form of volatile and/or nonvolatile memory such as read only memory (ROM) 106 and random access memory (RAM) 110. A basic input/output system 108 (BIOS), containing the basic routines that help to transfer information between elements within computer system 100, such as during start-up, is typically stored in ROM 106. Additionally, RAM 110 may contain operating system 112, application programs 114, other executable code 116 and program data 118. RAM 110 typically contains data and/or program modules that are immediately accessible to and/or presently being operated on by CPU 102.
The computer system 100 may also include other removable/non-removable, volatile/nonvolatile computer storage media. By way of example only,
The drives and their associated computer storage media, discussed above and illustrated in
The computer system 100 may operate in a networked environment using a network 136 to one or more remote computers, such as a remote computer 146. The remote computer 146 may be a personal computer, a server, a router, a network PC, a peer device or other common network node, and typically includes many or all of the elements described above relative to the computer system 100. The network 136 depicted in
The present invention is generally directed towards a system and method for optimizing online keyword auctions subject to budget and estimated query volume constraints. A linear programming model of slates of advertisements may be created offline for predicting the frequency and sequence of keywords occurring for multiple time periods throughout a time span for use in online scheduling of bidders to auctions that may optimize revenue of an auctioneer. Each slate of advertisements may represent a candidate set of advertisements in order of optimal revenue to an auctioneer. Upon receiving a query request, a slate of advertisements may be chosen online according to the generated frequencies. To improve the solution provided by the linear program, a slate of advertisements may be selected online that may maximize a weighted sum of the prices in order to compute a near optimal solution when the estimates of query volume are inaccurate. Then one of these two slates may be chosen based on whether the weighted sum of prices for the second slate may be within a factor of the weighted sum of the prices for the first slate provided by the linear program. And the chosen slate of advertisements may then be output for sending to a web browser for display.
As will be seen, dynamic programming may be applied to compute a near optimal slate of advertisements to hedge against inaccurate estimates of query volume used by the linear program. As will be understood, the various block diagrams, flow charts and scenarios described herein are only examples, and there are many other scenarios to which the present invention will apply.
Turning to
In various embodiments, a client computer 202 may be operably coupled to one or more servers 210 by a network 208. The client computer 202 may be a computer such as computer system 100 of
The server 210 may be any type of computer system or computing device such as computer system 100 of
The server 210 may be operably coupled to a database of advertisements such as advertisement store 224 that may include any type of advertisements 230 that may be associated with an advertisement ID 228. In an embodiment, several bidders 226 may be associated with an advertisement ID 228 for one or more advertisements 230. The advertisement store 224 may also include a collection of advertisement slates 232 that may be generated as part of the linear programming model, each advertisement slate representing an ordered candidate set of advertisements for keywords of a query request.
For each query qi, consider Ri,jt=Ai,jt·Qi,jt to be the ranking function used to rank the j-th offer in an auction instance, where Qi,jt may be generally a time-dependent weighting factor, or “quality score” for the i-th query and j-th bidder for time period t. The ranking function Ri,jt may be equal to zero for any bidder bj that may not participate in an auction instance. For a keyword i in time period t, consider Ti,j
A linear programming model may be created for this defined marketplace by estimating query volumes.
Using the method described in related copending U.S. patent application Ser. No. 11/497,085, entitled “SYSTEM AND METHOD FOR SCHEDULING ONLINE KEYWORD AUCTIONS SUBJECT TO BUDGET CONSTRAINTS,” assigned to the assignee of the present invention, a linear programming model of slates of advertisements within bidders' budgets may be created at step 302 for a time span and the linear program may be solved at step 304 using column generation. At step 306, a query having a keyword may be received in a time period, and a slate of advertisements for the keyword in the time period may be selected at step 308 from a set of slates generated by the linear program in an embodiment using the method described in related copending U.S. patent application Ser. No. 11/497,085, entitled “SYSTEM AND METHOD FOR SCHEDULING ONLINE KEYWORD AUCTIONS SUBJECT TO BUDGET CONSTRAINTS,” assigned to the assignee of the present invention.
To improve upon the solution provided by the linear program in the event that the estimates of query volumes are inaccurate, a slate of advertisements for the keyword in the time period may be selected at step 310 by dynamic programming using a parameterized discount factor for each bid based on a budget spent by an advertiser. In an embodiment, the search engine may rank the advertisements in a slate according to Ri,jt=Ai,jt·Qi,jt. For a keyword i in time period t, consider Ti,j
In applying dynamic programming to recursively determine a slate of advertisements that may maximize the sum of revenue for allocating advertisements to multiple web page placements in a generalized second price auction, a parameterized discount factor may be used for each bid based on a budget spent by an advertiser. The function φα(x)=1−eα·(x−1) may be defined to calculate a parameterized discount factor for each bid based on a budget spent by an advertiser. Let ƒj be the fraction of the budget spent for advertiser j. Consider slate k for keyword i where the ordering of the advertisers in this slate is 1, . . . , m. Then Φα(k) may be defined as:
where Φα(k) may represent the sum of revenue for allocating the slate of advertisements to multiple web page placements which is calculated for a generalized second price auction using a parameterized discount factor for each bid based on a budget spent by an advertiser. A description of the steps applying dynamic programming to recursively determine a slate of advertisements that may maximize the sum of revenue for allocating advertisements to multiple web page placements calculated for a generalized second price auction using a parameterized discount factor for each bid based on a budget spent by an advertiser may be described in more detail in conjunction with
The algorithm may be fine tuned by adjusting a parameter α≧1 that may control the extent of relying on the solution provided by the linear program using estimated query volumes. If α=1, then the algorithm may ignore the solution provided by the linear program using estimated query volumes and will use the solution provided by dynamic programming which will have the same performance as the competitive ratio of (1-1/e). If α=∞, the algorithm will ignore the solution provided by dynamic programming and use the solution provided by the linear program. In an embodiment, α may be set based on the relative accuracy of the estimates of query volume over time.
At step 312, one of the two slates may be chosen by comparing revenue for each one of the slates of advertisements calculated using a parameterized discount factor for a bid based on a budget spent by an advertiser. Accordingly, consider S1 to denote the slate of advertisements selected at step 308 by the linear program and consider S2 to denote the slate of advertisements selected at step 310 by dynamic programming using a parameterized discount factor for each bid based on a budget spent by an advertiser for the dynamic programming solution. If αΦα(S1)≧Φα(S2), then the slate of advertisements S1 may be chosen; otherwise, the slate of advertisements S2 may be chosen. And the chosen slate of advertisements may then be served at step 314, for instance, to a web browser for display with query results.
At step 404, the maximum value of revenue may be computed for each advertiser for a generalized second price auction using a parameterized discount factor for each bid based on the budget spent by the advertiser. Assume the advertisers may be numbered in such a way that advertiser number 1 has the highest bid times quality score, then advertiser number 2, and so on, until advertiser number j, which has the lower value of the product of a bid and a quality score. An j by p matrix M may be defined where j is the number of advertisers, and p is the number of slots. The value of M(j,p) may be defined as the maximum value of the function φα(ƒj) restricted to the first p slots, assuming that advertiser j is placed in the p+1 slot. The following recurrence relation may be used in an embodiment to compute the entries of matrix M:
Note that by using the above recurrence, all of the entries of the matrix M may be computed, starting from the first column and proceeding column by column to the right. Intuitively, the above recurrence tries all possible advertisers for slot p, and for each choice, computes the value of φα(ƒj) using a previously-computed entry in the matrix, and it then chooses the advertiser that gives the maximum value for the function φα(ƒj).
Accordingly, a slate of advertisements may be constructed at step 406 in decreasing order by maximum value of revenue computed for each advertiser for a generalized second price auction using a parameterized discount factor for each bid based on the budget spent by the advertiser, and the slate of advertisements may be output at step 408.
Although estimates of query volume may often be accurate, occasionally the estimates may be inaccurate due to unexpected events. In response to the occurrence of an event in the world, there may be a surge in the frequency of search terms. Any such increase in query volumes of a search term may be valuable for advertisers but are extremely difficult to predict. Thus, the present invention may provide a considerably better competitive ratio for an optimal allocation of advertisements in keyword auctions if the future events are actually close to the estimates, while still maintaining a reasonable competitive ratio if an unpredicted event changes the sequence of events drastically.
As can be seen from the foregoing detailed description, the present invention provides an improved system and method for optimizing online keyword auctions subject to budget and estimated query volume constraints. Such a system and method may effectively use a forecast of the frequency and sequence of keywords to compute a slate of advertisements and may also compute a near optimal slate of advertisements to use in the event estimates of query volume are inaccurate. Upon receiving a query request, the slate generated by the linear program or the slate generated by dynamic programming may be chosen based on whether the weighted sum of prices for the slate of advertisements computed by dynamic programming may be within a factor of the weighted sum of the prices for the slate of advertisements computed by the linear program. This may increase revenue when unexpected changes in query volume occur. As a result, the system and method provide significant advantages and benefits needed in contemporary computing and in online applications.
While the invention is susceptible to various modifications and alternative constructions, certain illustrated embodiments thereof are shown in the drawings and have been described above in detail. It should be understood, however, that there is no intention to limit the invention to the specific forms disclosed, but on the contrary, the intention is to cover all modifications, alternative constructions, and equivalents falling within the spirit and scope of the invention.
