Patent Application
20140089106
The embodiments described herein relate to Internet advertising exchanges. Specifically, the embodiments provide a technique for optimizing advertiser bids that incorporates forecasts of the future state of the market within the exchange.
The Internet is now a ubiquitous medium of communication in most parts of the world. The emergence of the Internet has opened a new forum for the creation and placement of advertisements promoting products, services, and brands. Internet content providers rely on advertising revenue to drive the production of free or low-cost content. Advertisers, in turn, increasingly view Internet content portals and online publications as a critically important medium for the placement of advertising. Thus, the sale and purchase of Internet advertising impressions has evolved into a multibillion-dollar a year industry.
Early in the history of Internet advertising, advertisers seeking to purchase inventory sought out individual publishers. Advertisers considered a publisher's inventory (impressions per unit time) and typical population demographics when deciding whether to make a bulk purchase. As advertisers' appetite for volume and reach (fraction of some population) increased, publishers were no longer able to singly fulfill these needs. This gave rise to advertising networks, where multiple advertisers and publishers came together to buy and sell inventory. Advertising networks facilitate various types of transactions, including guaranteed and non-guaranteed contracts.
Typically, advertisers (e.g., buyers of ad space) and online publishers (sellers of ad space) have agreements with one or more advertising networks, which provide for serving an advertiser's banner or advertisement across multiple publishers, and concomitantly provide for each publisher having access to a large number of advertisers. Advertising networks, which may also manage payment and reporting, may also attempt to target certain Internet users with particular advertisements to increase the likelihood that the user will take an action with respect to the advertisement. From an advertiser's perspective, effective targeting is important for achieving a high return on investment (ROI).
As advertisers' need for volume and reach increased further still, internetwork purchases started to take shape. Further, rather than buying a fixed amount of inventory in bulk, many advertisers only wanted to buy inventory that met certain performance goals. Advertising exchanges were developed to address this need. Advertising exchanges aim to provide a high degree of trading efficiency by bringing together a large number of buyers and sellers. Such exchanges are typically focused on directly matching the bids and offers of buyers and sellers. Conventional transactions on these exchanges are typically between (i) buyers and sellers, (ii) intermediaries (e.g., brokers, which may be a buyer or seller), or (iii) buyers or sellers and intermediaries. One such exchange, the Right Media Exchange (RMX), processes billions of impressions per day.
The emergence of such exchanges has presented advertisers, brokers, and other entities seeking to purchase advertising inventory with the challenge of formulating an effective bidding strategy. Advertisers typically have limited advertising budgets, and decisions on allocating advertising resources must be made carefully and intelligently. If bids are excessively high, the advertiser's budget will be exhausted prematurely. If bids are excessively low, the advertiser will not achieve an effective amount of exposure for the product or brand. In addition, the pricing patterns of advertising exchanges are similar to those of any open market for goods and services: prices often fluctuate based on available quantity and demand. Thus, there is a need for a technique that allows advertisers to formulate an optimal bidding strategy based on forecasts of the future state of the market within the advertising exchange. As will be shown, the embodiments disclosed herein provide such a technique in an elegant manner.
[BASED ON FINAL CLAIM SET]
The embodiments of the advertising system are described using a number of terms. In order to aid in clarity, some definitions of the terms used to describe these embodiments follow. However, these terms define general concepts, and thus are not to be construed narrowly. A publisher is generally defined as a Web site that has inventory for the delivery of advertisements. As such, advertisements are displayed on the Web pages of the publisher's Web site. Users are generally defined as those individuals that access Web pages through use of a browser. However, the term user may also be used to describe entities that use the advertising exchange system, such as users that access an application on the advertising exchange system to set parameters. Various participants of the advertising exchange system are referred to as “entities.” Thus, the term entity is generally used to describe any number of participants of the advertising exchange system. Those participants include advertisers, publishers, advertising networks and integrator networks.
An advertising network typically integrates entities, such as advertisers and publishers. An advertising network typically operates in conjunction with advertisers and publishers in order to deliver ads, from one or more advertisers, to Web pages of one or more publishers. For example, Yahoo! Inc, the assignee of the present invention, operates such an advertising network.
An integrator network entity generally defines a participant of the advertising exchange system that represents or integrates one or more entities on the advertising exchange system (e.g., advertisers, publishers, advertising networks, etc.). For example, an integrator network may represent advertisers on the advertising exchange system in order to deliver advertisements to publishers, advertising networks and other integrator networks. In some embodiments, the integrator networks are referred to as the “users” of the advertising exchange system. The integrated networks may comprise third party agents that operate on behalf of or are part of the integrator network. The term “third party agent” is used to generally describe an agent or customer that participates in transactions on the advertising exchange system. Similarly, the term “third party recipient” is used to describe a user or participant of the advertising exchange system that receives information from the system, such as bid requests. However, the terms integrator networks, third party agents and third party recipients are intended to represent a broad class of entities, including publishers, advertisers and networks, as well as the agents that represent them, that operate on the advertising exchange system.
The user 103 accesses information and/or content provided by the publisher 104. One form of access may include a browser 105 that has inventory locations 107 for the presentation of advertising. In one embodiment, an advertisement call is generated that requests an advertisement, from advertisements 112, 120 and 121, for placement with the inventory location 107. The corresponding advertisement may be delivered to publisher 104 by one or more networks. For instance, in one example, the network 106 is coupled to the publisher 104, and the network 108 is coupled to the advertiser 110. For this example, the networks 106 and 108 are coupled to each other. The advertiser 110 generally has one or more ad campaigns each comprising one or more advertisements 112 that the advertiser 110 wishes to place with the inventory of publishers such as, for example, the inventory location 107 of the publisher 104 that is presented to the user 103 via the browser application 105.
For the example illustrated in
As noted above, Internet advertising is typically sold by impression, click, or action. An impression is a single instance of an advertisement being displayed before an Internet user. In the context of an advertising exchange, an impression is an opportunity to transmit an advertisement to a user. The impression, including its attributes (e.g., page content, demographic data of the user, etc), is presented to the entities that participate in the exchange. An advertiser or other entity can place a bid on the impression. The exchange evaluates all of the bids and awards the impression to the winning bid. Due to the sheer volume of impressions and publishers within an exchange, advertisers often employ bidding agents that automatically place bids and purchase impressions based on a pre-determined strategy. In many prior art techniques for formulating a bidding strategy, bids are determined solely on the basis of the advertiser's goals and the current market state within the advertising exchange.
However, the market rates of the various types of impressions that are offered for sale within an advertising exchange may fluctuate wildly based on availability, bidding activity, and various other factors. A particular type of impression may currently command a higher or lower bid than it would in the future. If the impression is likely to command a lower bid in the future, the advertiser may wish to reduce its current bid. If the impression is likely to command a higher bid in the future, the advertiser may wish to increase its current bid. Thus, an awareness of this future state of the market would allow an advertiser to further optimize its bidding strategy and thereby allocate its advertising budget more efficiently. The embodiments disclosed herein provide a technique for forecasting the future state of the market and generating bids that take this forecast into account.
An example embodiment is illustrated by the flow diagram 300 depicted in
The costs of Internet advertising campaign are usually analyzed in terms of cost per impression (CPM), cost per click (CPC), or cost per action (CPA). Internet advertising campaigns may be classified by payment type and goal type. Payment type refers to an event for which the advertiser is willing to pay: advertisers may pay per impression in CPM, per click in CPC, and per action in CPA. Goal type refers to the set of events that advertisers value: clicks in CPC and actions in CPA. For instance, in a CPM with CPC goals campaign, the advertiser pays per impression but profits on a click. Advertisers specify a maximum payment associated with the payment event and a value associated with each goal event. Multiple goal events and associated values can be specified.
According to one series of embodiments, for each matched opportunity to bid on an impression, a bidding agent generates a bid for a campaign. The bid is generated through an optimization process that has both online and offline components. The offline component generates a set of bid parameters, akin to dual variables, and the online component generates a bid from the bid parameters, campaign constraints, and action-rate predictions. In this series of embodiments, advertisers select bids (or bid policies) in order to maximize value subject to a budget constraint and (possibly) a maximum bid constraint. Because advertisers specify the value they receive for each goal, the expected value per impression can be calculated as:
where λg is the prediction of the rate at which action g will occur. In this embodiment, an additive function is used for the total expected value. This is because either of the following assumptions may be made: a) the goal actions are independent, or b) if the goal actions are not independent, then marginal values are specified for each action.
One objective in this series of embodiments is to generate a family of bidding policies that produce bids given information about goal actions, their values, predicted rates, and the predicted rate of the payment action for each impression. Thus, a bid policy, b(v,λ) is a function of the expected-value-per-impression, v, and the payment rate, λ>0, for each impression. In one such embodiment, the auction is assumed to be a second-price auction where the winner pays the second-highest bid and a prediction of the highest competitive bid p is available. An advertiser's effective bid (eCPM) must exceed p in order to win the auction. The advertiser pays p (the expected value of p if it is a distribution) if it wins the auction. The expected profit of a bidding policy may be given by the functional:
Π[b]q∫0∞∫0∞0λb(v,λ)(v−p)f(v,λ,p)dpdvdλ (Equation 2)
where f is the joint density function of v, λ, the highest competitive bid p, and the quantity q. The deterministic approximation problem (DAP) is formulated as:
where bmax is the maximum-bid constraint and B is the budget. Solving the Euler-Lagrange equations characterizes the optimal bid policy as a single parameter, b*:
where b*ε[0,1]. The optimal b* solves
S(b*)=B (Equation 5)
where S(b*)=S[b(b*)], if the constraints are active at the maximum. In order to avoid overshooting the budget B, according to one embodiment, an online budget management system is employed to monitor spending and ensure that it does not exceed the budget. According to another embodiment, the algorithm iteratively solves for the optimal policy over time; as the remaining budget shrinks, the overshoot becomes smaller and smaller. Any technique for ensuring that the budget is not exceeded may be used without deviating from the spirit or scope of the invention.
A procedure for formulating a bid policy according to an embodiment is illustrated by the flow diagram 600 depicted in
An advertised optimal bidding policy may thus be represented by the single parameter b*. According to the embodiments disclosed herein, this parameter may be optimally set by accounting for the supply quantity, value, and forecasts of the future state of the market. Using these forecasts, an estimate of S(•) is derived, which is known as the spend curve. Given an estimate of the spend curve, b* is determined such that either S(b*)=B or b*=1.
According to one series of embodiments, the spend curve for a target time period is estimated using a sampling approach. As noted above, the future state of the market within an advertising exchange may be forecast by analyzing the previous impressions of an advertising campaign. This historic data is sampled and then annotated with quantity, value, and highest competitor bid (or price in the second-price auction). The quantity and highest competitor bid are forecast using independent models. Because the target time period and the time period over which historic impressions are sampled do not overlap, the models need to be able to predict how quantity and price will change over time. In this embodiment, supply quantity and prices are forecast while action rates are assumed to remain stationary conditioned on other features.
According to one such embodiment, this is performed using a Monte Carlo model, a computational approach that relies on repeated random sampling of a data set to simulate phenomena with a high degree of uncertainty. Monte Carlo models are often used in contexts with computational constraints that make exhaustive or exact calculations infeasible. In the context of an advertising exchange, each impression is unique and comprises potentially hundreds of attributes. Furthermore, an advertiser may purchase millions of impressions throughout the history of an advertising campaign. Thus, an exhaustive analysis of previous campaign impressions would require complex calculations involving massive amounts of data each time a current impression arrives for bidding. Storage and resource limitations make such calculations infeasible to perform repeatedly during the milliseconds that pass between the arrival of an impression and the placement of a bid. A Monte Carlo model allows for a compact representation of an advertising campaign's optimization parameters in the form of a single value.
More particularly, the main advantages of Monte Carlo sampling for spend-curve generation, compared to a parametric approach that explicitly characterizes each type of supply impression for a campaign, are that sampling (i) provides opportunities that are fully-qualified with the necessary attributes to apply the derived prediction algorithms (which are trained using the same set of attributes), and (ii) requires no assumptions on how the supply landscape is partitioned by the different attribute combinations.
A parameter optimization process incorporating Monte Carlo sampling is illustrated according to an embodiment by the flow diagram 700 depicted in
Steps 701 and 702 in the flow diagram 700 correspond to step 304 in the flow diagram 300 of
In step 702 a sample's future weight is predicted given 1) its current weight from the stratified sampling in step 701 and 2) historic data capturing changes in supply quantity for each campaign. Several algorithms, parametric and machine-learned, may be used in forecasting weights. A few such algorithms are described in U.S. patent application Ser. No. 12/242,709, entitled “Displaying Advertising Inventory Estimation”, filed Sep. 30, 2008, inventors Sergei Vassilvikkii et al., which is incorporated by reference herein.
An example algorithm for forecasting these weights is illustrated by the flow diagram 800 depicted in
In steps 703 and 704, annotation is performed use machine learning feature-based models that, respectively, predict action-rate probabilities and prices. Any such model may be used without deviating from the spirit or scope of the invention. Step 705 uses a parametric model based on the input price distributions.
Based on the campaign definition (goal values) and the action rates, a value per impression v can be generated for each sample. Based on the first bid and reserve price estimates, estimates for p can be constructed. If the first bid is a point estimate, then p is simply the maximum of the first bid and reserve price estimates. Otherwise, if the first bid is represented as a distribution F, then the win distribution W can be constructed such that:
where r is the reserve-price estimate. Therefore for point prices, the technique uses the set of samples
ζ={(qi,vi,λi,pi)}i=1N, (Equation 7)
and for distributional prices, the technique uses the set of samples
ζ={(qi,vi,λi,Wi(p))}i=1N, (Equation 8)
where q is the quantity of supply of this type.
Given ζ, Monte Carlo estimates of the spend curve may be constructed. For point prices,
where 1 is the indicator function. For distribution prices,
Thus, S(b*) is simply the summation over each sample, whether point or distributional prices. When the samples are constructed, each campaign may match up to hundreds of thousands of samples. In addition, there may be hundreds of thousands of campaigns over which to optimize.
Because of the additive nature of the spend curve, the unit interval can simply be partitioned into M pieces and, for each sample, its contribution to the curve at each of the M points can be computed. Because the curve is monotonic, a simple binary search algorithm can be used to find the optimal b*.
According to one series of embodiments, the steps described above are executed as offline processes. The output is a single bid parameter (b*) per campaign. In the online system, a bid generation server (BGS) constructs a bid for each matching campaign from the action-rate predictions (λs), campaign definition (goals and maximum bids), and the bid parameter (b*) according to Equation 4. In one such embodiment, the steps occur asynchronously. For instance, a campaign table, which contains the goals and maximum bid information, may be updated frequently (perhaps every 15 minutes), whereas the sampling process much less so (perhaps daily). Any technique for executing asynchronous operations may be used without deviating from the spirit or scope of the embodiments described herein.
This may present a synchronization issue; a b* value is desired that ensures that the online campaign definition is the same version as was used in optimization. However, changes in campaign definitions that are incorporated on a relatively short timescale must be supported. To facilitate this check, a checksum is introduced that hashes relevant campaign definition values. This value is stored in a campaign index accessible to the bid generation server. In addition, the checksum used to optimize the campaign is stored in a parameter index. When both the campaign index checksum and the parameter index checksum coincide, a bid is placed according to Equation 4. When the checksums do not match, a bid is placed according to a default bidding strategy (b*=1).
An exemplary environment within which some embodiments of the invention may operate is illustrated in
According to one series of embodiments, one of the Client Devices 904 requests a webpage from Publisher 901. The web page may include an advertising impression sold by Advertising Exchange 902. The impression includes attributes indicating various characteristics such as web page content, user demographic, etc. This impression arrives at Advertising Exchange 902, from which it is transmitted to Entities 903. Entities 903 formulate a bid for the impression using the procedures of various embodiments disclosed herein and transmit the bid to Advertising Exchange 902. The steps for formulating the bid are executed in Bid Generation Server 910 using data stored in Impression Database 911. When all of the bids are received by Advertising Exchange 902, a winning bid is selected and an advertisement corresponding to the entity from among Entities 903 that submitted the winning bid is transmitted to Publisher 901. Publisher 901 incorporates the advertisement into the web page and transmits the web page to the client device from among Client Devices 904 that requested the web page.
Any node of the network 1000 may comprise a general-purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof capable to perform the functions described herein. A general-purpose processor may be a microprocessor, but in the alternative, the processor may be any conventional processor, controller, microcontroller, or state machine. A processor may also be implemented as a combination of computing devices (e.g. a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors in conjunction with a DSP core, or any other such configuration, etc).
In alternative embodiments, a node may comprise a machine in the form of a virtual machine (VM), a virtual server, a virtual client, a virtual desktop, a virtual volume, a network router, a network switch, a network bridge, a personal digital assistant (PDA), a cellular telephone, a web appliance, or any machine capable of executing a sequence of instructions that specify actions to be taken by that machine. Any node of the network may communicate cooperatively with another node on the network. In some embodiments, any node of the network may communicate cooperatively with every other node of the network. Further, any node or group of nodes on the network may comprise one or more computer systems (e.g. a client computer system, a server computer system) and/or may comprise one or more embedded computer systems, a massively parallel computer system, and/or a cloud computer system.
The computer system 1050 includes a processor 1008 (e.g. a processor core, a microprocessor, a computing device, etc), a main memory 1010 and a static memory 1012, which communicate with each other via a bus 1014. The machine 1050 may further include a display unit 1016 that may comprise a touch-screen, or a liquid crystal display (LCD), or a light emitting diode (LED) display, or a cathode ray tube (CRT). As shown, the computer system 1050 also includes a human input/output (I/O) device 1018 (e.g. a keyboard, an alphanumeric keypad, etc), a pointing device 1020 (e.g. a mouse, a touch screen, etc), a drive unit 1022 (e.g. a disk drive unit, a CD/DVD drive, a tangible computer readable removable media drive, an SSD storage device, etc), a signal generation device 1028 (e.g. a speaker, an audio output, etc), and a network interface device 1030 (e.g. an Ethernet interface, a wired network interface, a wireless network interface, a propagated signal interface, etc).
The drive unit 1022 includes a machine-readable medium 1024 on which is stored a set of instructions (i.e. software, firmware, middleware, etc) 1026 embodying any one, or all, of the methodologies described above. The set of instructions 1026 is also shown to reside, completely or at least partially, within the main memory 1010 and/or within the processor 1008. The set of instructions 1026 may further be transmitted or received via the network interface device 1030 over the network bus 1014.
It is to be understood that embodiments of this invention may be used as, or to support, a set of instructions executed upon some form of processing core (such as the CPU of a computer) or otherwise implemented or realized upon or within a machine- or computer-readable medium. A machine-readable medium includes any mechanism for storing information in a form readable by a machine (e.g. a computer). For example, a machine-readable medium includes read-only memory (ROM); random access memory (RAM); magnetic disk storage media; optical storage media; flash memory devices; electrical, optical or acoustical or any other type of media suitable for storing information.
