KR20240039946A - System and method for predicting response for real time bid requests based on digital advertising - Google Patents

Abstract

The present invention relates to a system and method for predicting response to a real-time bidding request based on digital advertising. As a user accesses a specific digital advertising space, an advertisement for the inventory is provided on a supply-side platform that acts as a medium for selling the inventory of the advertising space. When transmitting a request to an advertising exchange, a bid request data acquisition unit that sequentially acquires bid request data transmitted along with a bid request for the inventory from the advertising exchange to a demand-side platform acting on behalf of advertisers, and the bid request data acquisition unit; A bid response data acquisition unit that sequentially acquires bid response data transmitted from the demand-side platform to the advertising exchange depending on whether the advertiser responds to the request, and a result of comparing the bid request data and the bid response data. Based on this, a training data set creation module generates a training data set by setting some fields of the data to one of an explanatory variable and a response variable, and designates a part of the generated training data set as training data for learning and sets it to a preset value. A regression analysis learning module that performs learning to predict the result value for the response variable by applying it to a regression analysis model, and a product generated according to an advertising request sent by the supply-side platform to the advertising exchange after completing the learning. 1 Characterized by comprising a prediction result output unit that applies the bid request data to the regression analysis model and outputs a predicted result value for the bid request corresponding to the first bid request data.
Accordingly, there is an effect of reducing unnecessary server traffic costs and improving profitability by conducting advertising bidding only for bid requests that are predicted to have a bidding response.

Description

System and method for predicting response to real-time bidding request based on digital advertising {SYSTEM AND METHOD FOR PREDICTING RESPONSE FOR REAL TIME BID REQUESTS BASED ON DIGITAL ADVERTISING}

The present invention provides a response prediction system for digital advertising-based real-time bidding requests that can predict whether or not there will be a bidding response to a specific bidding request by supervised learning of advertiser's bidding response data to real-time bidding requests targeting inventory of digital advertising space. It's about.

Advertising refers to a series of activities conducted for the purpose of informing potential consumers of information about products and services and how to purchase them, and was traditionally provided through media such as newspapers and television.

However, as digital communication technology gradually develops, new advertising media such as advertising using various mobile terminals and digital signage that can output various types of information and advertising data are emerging.

Among these various advertising media, the most common advertising method is to provide advertisements to users by exposing advertisements to the advertisement exposure area on a mobile web page or application through mobile advertising services using mobile devices that are closely related to daily life. It is being used.

In particular, programmatic advertising, which systemizes the automatic purchase of advertising inventory using dedicated software and algorithms that automates digital advertising space such as display advertising, banner advertising, and video advertising, has recently emerged and is widely used as an alternative to traditional marketing. It is becoming.

Among these types of programmatic advertising, Real Time Bidding (RTB), which allows advertisers to purchase advertising inventory on an impression-by-impression basis through bidding in the open market, is receiving the most attention.

In the case of real-time bidding (RTB), real-time bidding or dynamic search bidding is performed per impression or user by centralizing the buying and selling of media through platforms and ad exchanges that act on behalf of media and advertisers, respectively. It has possible characteristics.

However, the existing RTB method sends a bid request for advertising inventory to multiple advertisers every time a user accesses the advertising space, and the advertiser who sends the bidding response participates in the bidding. Regardless, there is a problem of excessive traffic costs because bid request data must be transmitted to all registered advertisers.

KR 10-2015-0130282 A KR 10-2016-0006301 A

The purpose of the present invention is to solve the above problems, and is a digital advertisement that can predict whether or not a bid will respond to a specific bid request by supervised learning of the advertiser's bid response data to real-time bid requests targeting inventory of digital advertising space. The purpose is to provide a response prediction system for real-time bid requests.

In order to achieve the above object, the system for predicting response to a real-time bidding request based on digital advertising according to one aspect of the present invention is a supply side acting as a medium that sells the inventory of the advertising space as a user accesses a specific digital advertising space. When the platform transmits an advertising request for the inventory to an advertising exchange, bid request data is acquired in time series along with a bid request for the inventory from the advertising exchange to a demand-side platform acting on behalf of advertisers. an acquisition unit, a bid response data acquisition unit that sequentially acquires bid response data transmitted from the demand-side platform to the advertising exchange depending on whether the advertiser responds to the bid request, and the bid request data and the bid response. A training data set creation module that generates a training data set by setting some fields of the data as either an explanatory variable or a response variable based on the results of mutual comparison of the data, and a part of the generated training data set for learning training A regression analysis learning module that performs learning to predict the result for the response variable by specifying it as data and applying it to a preset regression analysis model, and after completing the learning, the supply-side platform transmits to the advertising exchange It is characterized by comprising a result output unit that applies the first bid request data generated according to the advertisement request to the regression analysis model and outputs a prediction result for the bid request corresponding to the first bid request data.

In addition, a method for predicting response to a real-time bidding request based on digital advertising according to another aspect of the present invention to achieve the above object includes a supply-side platform acting as a medium for selling inventory of digital advertising space, and advertising in the inventory. In the method of predicting response to a real-time bidding request based on digital advertising using a demand-side platform that acts on behalf of an advertiser who executes an advertisement and an advertising exchange that brokers advertising transactions based on real-time bidding between the medium and the advertiser, the user specifies When accessing an advertising space and sending an advertising request for the inventory of the advertising space from the supply-side platform to the advertising exchange, bid request data transmitted from the advertising exchange to the demand-side platform along with a bid request for the inventory. A step of time-serially acquiring bid response data transmitted from the demand-side platform to the advertising exchange according to whether the advertiser responds to the bid request, and the bid request data and the bid. A step of generating a training data set by setting some fields of the data to one of an explanatory variable and a response variable based on the results of mutual comparison of the response data, and designating a part of the generated training data set as training data for learning. and applying this to a preset regression analysis model to perform learning to predict the result for the response variable, and after completing the learning, the supply-side platform generates an advertisement request sent to the advertising exchange. Characterized by comprising the step of applying first bid request data to the regression analysis model and outputting a prediction result for a bid request corresponding to the first bid request data.

According to the present invention, it is possible to predict in advance whether or not a bid will respond to a specific bid request based on the results of completing supervised learning based on quantitative data collected while performing a real-time bidding advertising transaction that occurred in the past for a predetermined period of time. It works.

In addition, according to the present invention, there is an effect of reducing unnecessary server traffic costs and improving profitability by conducting advertising bidding only for bid requests that are expected to receive a bidding response.

1 is a diagram schematically showing a response prediction system for a digital advertisement-based real-time bidding request according to an embodiment of the present invention;
Figure 2 is a block diagram showing in detail the configuration of a response prediction system for a digital advertisement-based real-time bidding request according to an embodiment of the present invention;
Figures 3 and 4 are diagrams showing examples of bid request data and bid response data, respectively, according to the present invention;
Figure 5 is a diagram showing an example of a training data set generated by the training data set generation module of Figure 2;
Figure 6 is a diagram showing an example of dividing training data for learning and test data for verification in the data division unit of Figure 2;
Figure 7 is a diagram showing the results of comparing the actual response value and predicted response value to the bid request in the performance evaluation unit of Figure 2 in the form of an error matrix;
Figure 8 is a flowchart showing a method for predicting a response to a real-time bidding request based on digital advertising according to an embodiment of the present invention.

Specific details, including the problem to be solved by the present invention, the means for solving the problem, and the effect of the invention, are included in the examples and drawings described below. The advantages and features of the present invention and how to achieve them will become clear by referring to the embodiments described in detail below along with the accompanying drawings. Like reference numerals refer to like elements throughout the specification.

1 is a diagram schematically showing a system for predicting responses to real-time bidding requests based on digital advertising according to an embodiment of the present invention, and FIG. 2 is predicting responses to real-time bidding requests based on digital advertising according to an embodiment of the present invention. It is a block diagram showing the configuration of the system in detail, Figures 3 and 4 are diagrams showing examples of bid request data and bid response data, respectively, according to the present invention, and Figure 5 is a diagram generated by the training data set generation module of Figure 2. It is a diagram showing an example of a training data set, and Figure 6 is a diagram showing an example of dividing training data for learning and test data for verification by the data division unit of Figure 2, and Figure 7 is a bid request from the performance evaluation unit of Figure 2. It is a diagram showing the actual response value and the predicted response value in the form of an error matrix by comparing them, and Figure 8 is a flowchart showing a method of predicting a response to a real-time bidding request based on digital advertising according to an embodiment of the present invention.

Hereinafter, a system and method for predicting a response to a real-time bidding request based on digital advertising according to a preferred embodiment of the present invention will be described with reference to the above-described drawings.

The method for predicting response to a real-time bidding request based on digital advertising according to the present invention includes, as shown in FIG. 1, a supply-side platform 10 acting on behalf of a medium 1 that sells inventory of digital advertising space, and advertising in the inventory. In the demand-side platform 20, which acts on behalf of the advertiser 2, and the advertising exchange 30, which brokers advertising transactions based on real-time bidding between the media 1 and the advertiser 2, It may be performed by the system 50 that predicts the response to a real-time bid request based on the data.

Here, the supply side platform 10 corresponds to a platform (Supply Side Platform, SSP) used by the media 1 to sell advertising space, and is a programmatic software for the media that coordinates the sale of advertising impressions through advertising exchange. The demand side platform 20 corresponds to a platform (Demand Side Platform, DSP) that advertisers use when purchasing advertising space, and can be built based on programmatic software for advertisers.

The system for predicting response to a real-time bidding request based on digital advertising according to the present invention is, when a user accesses a specific digital advertising space, the inventory is generated from the supply-side platform (10) acting on behalf of the medium (1) that sells the inventory of the advertising space. When an advertisement request for is transmitted to the advertising exchange 30, the advertising exchange 30 transmits the bid request data along with the bid request for the inventory to the demand-side platform 20 acting on behalf of the advertiser 2, and It can be applied when a real-time bidding process is performed by transmitting bid response data from the demand-side platform 20 to the advertising exchange 30 depending on whether the advertiser 2 responds to the bid request, and the advertising exchange 30 It may be connected to enable data transmission and reception, thereby obtaining data on bid requests and bid responses from the advertising exchange 30, respectively.

In this regard, referring to FIG. 2, the response prediction system 50 for a digital advertisement-based real-time bid request according to an embodiment of the present invention largely includes a bid request data acquisition unit 100 and a bid response data acquisition unit 200. ), a training data set creation module (300), a regression analysis learning module (400), a learning performance evaluation module (500), a result output unit (600), and a control unit (800).

When an advertisement request for inventory of digital advertising space occurs, the bid request data acquisition unit 100 sends a bid request from one advertising exchange 30 that received the advertisement request to one demand-side platform 20. Obtain the transmitted bid request data (B_req) (S100).

The bid request data acquisition unit 100 transmits an advertising request for the inventory of a specific advertising space accessed by a user from the supply-side platform 10 to one advertising exchange 30, and in response thereto, the advertising exchange 30 In the case of transmitting bid request data (B_req) along with a bid request for the inventory to one demand-side platform 20, bid request data (B_req) transmitted from the advertising exchange 30 during a predetermined first time It may be obtained by accumulating in time series.

Here, the bid request data (B_req) is based on the object model (object mode) for the preset bid request as shown in Table 1 below, one top-level object that is mandatory exposed when requesting a bid, and a specific context is optional. At least one or more sub-objects exposed may be configured in a hierarchical structure.

At this time, it is preferable that the bid request data acquisition unit 100 stores the acquired bid request data (B_req) in the form of a csv file.

object explanation BidRequest Top-level object Imp Container for the description of a specific impression;
at least 1 per request. App Details of the application calling for the impression. Device Details of the device on which the content and impressions are displayed.

Referring to Table 1, the BidRequest object corresponds to the top-level object, the Imp object is a sub-object representing the description container of a specific exposure, the App object is a sub-object representing the details of the app requesting exposure, and the Device The object may be a sub-object representing device information on which content and exposure are displayed.

In addition, Tables 2 to 5 below table a description of the attribute fields of each of the plurality of objects constituting the object model of Table 1 and the data format for each field. Table 2 above provides information on the BidRequest object. The contents are summarized, and Table 3 above summarizes the contents of the Imp object, Table 4 above summarizes the contents of the App object, and Table 5 above summarizes the contents of the Device object.

attribute field data format explanation id string;
required Unique ID of the bid request, provided by the exchange. imp object array;
required Array of Imp objects representing the impressions offered. site object;
recommended Details via a site object about the publisher's website. app object;
recommended Details via an app object about the publisher's app. device object;
recommended Details via a device object about the user’s device to which the impression will be delivered. user object;
recommended Details via a User object about the human user of the device; the advertising audience. test integer;
default 0 Indicator of test mode in which auctions are not billable, where 0 = live mode, 1 = test mode. at integer;
default 2 Auction type, where 1 = First Price,
2 = Second Price Plus. tmax integer Maximum time in milliseconds the exchange allows for bids to be received including Internet latency to avoid timeout. wseat string array white list of buyer seats allowed to bid on this impression. bseat string array Block list of buyer seats restricted from bidding on this impression. allimps integer;
default 0 Flag to indicate if Exchange can verify that the impressions offered represent all of the impressions available in context to support road-blocking. cur string array Array of allowed currencies for bids on this bid request using ISO-4217 alpha codes. wlang string array White list of languages for creatives using ISO-639-1-alpha-2. bcat string array Blocked advertiser categories using the IAB content categories. badv string array Block list of advertisers by their domains. bapp string array Block list of applications by their platform-specific exchange-independent application identifiers. source object A Sorce object that provides data about the inventory source and which entity makes the final decision. regs object A Regs object that specifies any industry, legal, or governmental regulations in force for this request. ext object Placeholder for exchange-specific extensions to OpenRTB.

Referring to Table 2 above, in the case of a BidRequest object, an attribute field (id) corresponding to the unique ID of the bid request provided by the advertising exchange, and an attribute field (imp) corresponding to an array of Imp objects indicating the number of impressions provided ), an attribute field (site) corresponding to the details of the site object for the website of the medium (publisher), and an attribute field (app) corresponding to the details of the app object for the app of the medium (publisher), An attribute field (device) corresponding to the details of the device object for the user device to which the impression is to be delivered, an attribute field (user) corresponding to the details of the user object for the user of the device, and an indication of a test mode that does not charge an auction. an attribute field (test) corresponding to the auction type, an attribute field (at) corresponding to the auction type, an attribute field (tmax) corresponding to the maximum reception time for which bidding is allowed in the advertising exchange, and an attribute field (tmax) corresponding to the maximum reception time allowed for bidding on the advertisement exposure. The attribute field (wseat) corresponding to the whitelist of the buyer sheet, the attribute field (bseat) corresponding to the blacklist of the buyer sheet where bidding for advertising exposure is restricted, and the exposure provided from the ad exchange are used to support the load blocking function. an attribute field (allimps) corresponding to a flag indicating whether to check whether the context represents all available impressions, and an attribute field (cur) corresponding to an array of currencies allowed for bidding in bid requests using ISO-4217 alpha codes. ), an attribute field corresponding to an acceptable whitelist for advertising languages using ISO-639-1-alpha-2 (wlang), and an attribute field corresponding to blocked advertiser categories using IAB content categories (bcat). , an attribute field corresponding to a domain-specific advertiser blocklist (badv), an attribute field corresponding to an application blocklist based on a platform-specific exchange-independent app identifier (bapp), and data about the inventory source and makes the final decision. An attribute field (source) corresponding to the source object that serves as the issuing entity, and an attribute field (regs) corresponding to the Regs object that specifies any industry, law, or legal regulations in effect for the request, and for exchange only for OpenRTB. It can contain an attribute field (ext) that is a placeholder for expansion.

attribute field data format explanation id string;
reqired A unique identifier for this impression within the context of the bid request. metric object array An array of metric objects. banner object A banner object. video object A video object. audio object An audio object. native object A native object. pmp object A pmp object containing any private marketplace deals in effect for this impression. display
manager string Name of mediation partner, SDK technology,
or player responsible for rendering ad. display
managerver string Version of ad mediation partner, SDK technology, or player responsible for rendering ad. inst1 integer;
default 0 1 = the ad is interstitial of full screen,
0 = not interstitial. tagid string Identifier for specific ad placement or ad tag that was used to initiate the auction. bid floor float;
default 0 Minimum bid for this impression expressed in CPM. bidfloorcurr string;
default "USD" Currency specified using ISO-4217 alpha codes. clickbrowser integer Indicates the type of browser opened upon clicking the creative in an app, where 0 = embedded, 1 = native. secure integer Flag to indicate if the impression requires secure HTTPS URL creative assets and markup,
where 0 = non-secure, 1 = secure. iframebuster string array Flag to indicate if the impression requires secure HTTPS URL creative assets and markup,
where 0 = non-secure, 1 = secure. exp integer Advisory as to the number of seconds that may elapse between the auction and the actual impression. ext object Placeholder for exchange-specific extensions to OpenRTB.

Referring to Table 3 above, in the case of an Imp object, an attribute field (id) corresponding to a unique identifier for the imp object within the context of a bid request, an attribute field (metric) corresponding to a metric object, and an advertising exposure banner an attribute field (banner) corresponding to the object required when advertising is exposed as video (video), and an attribute corresponding to the object required when advertising exposure is audio. A field (audio), an attribute field (native) corresponding to an object required when displaying an advertisement natively, an attribute field (pmp) corresponding to a pmp object containing a valid private market transaction for the exposure, and an advertising brokerage. An attribute field corresponding to the name of the partner, SDK technology, or player rendering the ad (displaymanager); an attribute field corresponding to the version of the ad intermediary partner, SDK technology, or player rendering the ad (displaymanagerver); and the ad being an interstitial. An attribute field (inst1) indicating whether it is an advertisement or interstitial, an attribute field (tagid) corresponding to the identifier for the specific ad placement or ad tag used to initiate the auction, and an attribute field corresponding to the minimum bid for the impression. (bidfloor), a corresponding attribute field (bidfloorcur) corresponding to a currency unit specified using an ISO-4217 alpha code, an attribute field indicating the type of browser opened when an ad is clicked in the app (clickbrowser), and a secure HTTPS connection for that impression. An attribute field corresponding to a URL creative asset and a flag indicating whether markup is required (secure), an attribute field corresponding to a list of supported extended iFrame file names (iframebuster), and the amount of time that can elapse between the auction and the actual impression. It may contain an attribute field (exp) corresponding to a recommendation for , and an attribute field (ext) corresponding to a placeholder for an exchange-only extension to OpenRTB.

attribute field data format explanation id string;
recommended Exchange-specific app ID. name string App name. bundle string A platform-specific application identifier intended to be unique to the app and independent of the exchange. domain string Domain of the app. storeurl string App store URL for an installed app; for IQG 2.1 compliance. cat string array Array of IAB content categories of the app. sectioncat string array Array of IAB content categories that describe the current section of the app. pagecat string array Array of IAB content categories that describe the current page or view of the app. ver string Application version. privacypolicy integer Indicates if the app has a privacy policy, where 0 = no, 1 = yes. paid integer 0 = the app is free, 1 = the app is a paid version publisher object Details about the Publisher of the app. content object Details about the Content within the app. keywords string Comma separated list of keywords about the app. ext object Placeholder for exchange-specific extensions to OpenRTB.

Referring to Table 4 above, in the case of an App object, an attribute field (id) corresponding to the app ID specified in the exchange, an attribute field (name) corresponding to the app name, and an application bundle or package name (Android) or number ID In the attribute field (bundle) corresponding to (iOS), the attribute field (domain) corresponding to the app's domain, the attribute field (storeurl) corresponding to the store URL where the app is uploaded, and the IAB content category array for the app. a corresponding attribute field (cat), an attribute field (sectioncat) that corresponds to an IAB content category array for the current section of the app, and an attribute field (pagecat) that corresponds to an IAB content category array for a depiction of the current page or view of the app. ), an attribute field (ver) corresponding to the app version, an attribute field (privacypolicy) indicating whether the app displays a privacy policy, and an attribute field (paid) indicating whether the app is a free or paid version, An attribute field representing details about the publisher (publisher), an attribute field corresponding to details about the content within the app (content), and an attribute field corresponding to the result of a comma-separated list of keywords for the app (keywords). Wow, you can include an attribute field (ext) that is a placeholder for an exchange-only extension to OpenRTB.

attribute field data format explanation ua string;
recommended Browser user agent string. geo object;
recommended Location of the device assumed to be the user's current location defined by a geo object. dnt integer;
recommended Standard "Do Not Track" flag as set in the header by the browser, where 0 = tracking is unrestricted, 1 = do not track. lmt integer;
recommended “Limit Ad Tracking” signal commercially endorsed,
where 0 = tracking is unrestricted, 1 = tracking must be limited per commercial guidelines. ip string;
recommended IPv4 address closest to device. ipv6 string IP address closet to device as IPv6. devicetype integer The general type of device. make string Device make. model string Devide model. os string Device operating system. osv string Device operating system version. hwv string Hardware version of the device. h ingeter Physical height of the screen in pixels. w integer Physical width of the screen in pixels. ppi integer Screen size as pixels per linear inch. pxratio float The ratio of physical to device independent pixels. js integer Support for JavaScript, where 0 = no, 1 = yes. geofetch ingeter Indicates if the geolocation API will be available to JavaScript code running in the banner,
where 0 = no, 1 = yes. flashver string Version of Flash supported by the browser. language string Browser language using ISO-639-1-alpha-2. carrier string Carrier or ISP using exchange curated string names which should be published to bidders a priori. mccmnc string Mobile carrier as the concatenated MCC-MNC code. connectiontype integer Network connection type. ifa string ID sanctioned for advertiser use in the clear. didshall string Hardware device ID; hashed via SHA1. didmd5 string Hardware device ID;hashed via MD5. dpidsha1 string Platform device ID; hashed via SHA1 dpidmd5 string Platform device ID; hashed via MD5. macsha1 string MAC address of the device; hashed via SHA1 macmd5 string MAC address of the device; hashed via MD5 ext object Placeholder for exchange-specific extensions to OpenRTB.

Referring to Table 5 above, in the case of a Device object, an attribute field (ua) corresponding to browser user agent information, an attribute field (geo) corresponding to the current location of the user device defined as a geo object, and a header in the browser An attribute field (dnt) indicating the standard Do Not Track flag you have set, an attribute field (lmt) corresponding to a commercially approved advertising tracking limit signal, an attribute field (ip) corresponding to the device's IPv4 address, and an attribute field (ip) corresponding to the device's IPv6 address. An attribute field (ipv6) corresponding to the address, an attribute field (devicetype) corresponding to the standard type of the device, an attribute field (make) corresponding to the device manufacturer, an attribute field (model) corresponding to the device model, and the device An attribute field corresponding to the OS (os), an attribute field corresponding to the device OS version (osv), an attribute field corresponding to the device hardware version (hwv), and an attribute field corresponding to the physical screen height in pixels ( h), an attribute field (w) corresponding to the physical screen width in pixels, an attribute field (ppi) corresponding to the number of pixels per inch of the screen, and an attribute field (js) indicating whether JavaScript is supported. , an attribute field (geofetch) indicating whether the banner will use JavaScript's geolocation API, an attribute field (flashver) corresponding to the Flash version supported by the browser, and the browser language using ISO-639-1-alpha-2. an attribute field (language) corresponding to the carrier or ISP using the exchange-curated string name that must first be revealed to the bidder, and an attribute field corresponding to the carrier as the associated MCC-MNC code ( mccmnc), an attribute field corresponding to the network connection type (connectiontype), an attribute field corresponding to an ID that the advertiser is authorized to use (ifa), and an attribute field representing the SHA1 hashed value as the hardware device ID (didsha1). ), an attribute field (didmd5) indicating a value hashed with MD5 as a hardware device ID, an attribute field (dpidsha1) indicating a value hashed with SHA1 as a platform device ID, and an attribute field (dpidsha1) indicating a value hashed with MD5 as a platform device ID. An attribute field (dpidmd5), representing a value hashed with SHA1 as the device's address, and an attribute field (macmd5) representing a value hashed with SHA1 as the device's address, and an attribute field (macmd5) representing a value hashed with MD5 as the device's address, for exchange only for OpenRTB. It can contain an attribute field (ext) that is a placeholder for expansion.

In this case, as shown in FIG. 3, the bid request data (B_req) includes at least one attribute field belonging to the objects (BidRequest, Imp, App, Device) listed in Tables 2 to 5 and information about the field. The value may be stored according to the data format for each preset attribute field.

The bid response data acquisition unit 200 transmits bid response data from the demand-side platform 20 to the advertising exchange 30 depending on whether the advertiser 2 responds to the bid request transmitted from the advertising exchange 30 ( B_res) is obtained (S200).

The bid response data acquisition unit 200 receives bid response data from the advertising exchange 30 during a second time corresponding to the time from the start of the first time to a preset threshold time after the end of the first time ( B_res) may be acquired by accumulating time series.

Here, the bid response data (B_res) is based on the object model for the preset bid response as shown in Table 6 below, one top-level object that is necessarily exposed during the bid response, and a specific context is optional. At least one or more sub-objects exposed may be configured in a hierarchical structure.

At this time, it is preferable that the bid response data acquisition unit 200 stores the acquired bid response data (B_res) in the form of a csv file.

object explanation BidResponse Top-level object. SeatBid Collection of bids made by the bidder on behalf of a specific seat. Bid An offer to buy a specific impression under certain business terms.

Referring to Table 6 above, the BidResponse object corresponds to the top-level object, the SeatBid object is a child object representing the bidding meeting of a specific seat on behalf of the bidder, and the Bid object is the act of purchasing advertising exposure according to specific business conditions. It may be a sub-object representing .

In addition, Tables 7 to 9 below provide a description of the attribute fields of each of the plurality of objects constituting the object model of Table 6 and the data format for each field. Table 7 provides the information for the BidResponse object. The contents are summarized, Table 8 above summarizes the contents of the SeatBid object, and Table 9 above summarizes the contents of the Bid object.

attribute field data format explanation id string; required ID of the bid request to which this is a response. seatbid object array Array of seatbid objects. bidid string Bidder generated response ID to assist with logging/tracking. cur string;
default "USD" Bid Currency using ISO-4217 alpha codes. customdata string Optional feature to allow a bidder to set data in the exchange's cookie. nbr integer Reason for not bidding. ext object Placeholder for bidder-specific extentions to OpenRTB.

Referring to Table 7 above, in the case of a BidResponse object, an attribute field (id) corresponding to the ID value of the bid request for the response, an attribute field (seatbid) corresponding to an array of seatbid objects, and logging/tracking of bidders An attribute field (bidid) corresponding to the response ID created to support it, an attribute field (cur) corresponding to the currency of the bid expressed using ISO-4217 alpha code, and an attribute field (cur) that allows the bidder to set the exchange's cookie data. It may contain an attribute field (customdata) corresponding to an optional feature, an attribute field indicating a reason not to bid (nbr), and an attribute field corresponding to a placeholder for an exchange-only extension to OpenRTB (ext).

attribute field data format explanation bid object array;
required Array of 1+ Bid objects each related to an impression. seat string ID of the buyer seat on whose behalf this bid is made. group ingeter;
default 0 0 = impressions can be won individually;
1 = impressions must be won or lost as a group ext object Placeholder for bidder-specific extensions to OpenRTB.

Referring to Table 8 above, in the case of a SeatBid object, an attribute field (bid) corresponding to an array of one or more bidding objects added in association with each exposure, and an attribute field (seat) corresponding to the seat ID of the buyer making the bid ) and an attribute field (group) corresponding to a value that sets the bidding unit individually or as a group when there are multiple impressions included in one bid, and a placeholder for an exchange-only extension to OpenRTB. Can include attribute fields (ext).

attribute field data format explanation id string;required Bidder generated bid ID to assist with logging/tracking. impid string;required ID of the Imp object in the related bid request. price float;required Bid price expressed as CPM although the actual transaction is for a unit impression only. nur1 string Win notice URL called by the exchange if the bid wins. bur1 string Billing notice URL called by the exchange when a winning bid becomes billable based on exchange-specific business policy. lur1 string Loss notice URL called by the exchange when a bid is known to have been lost. adm string Optional means of conveying ad markup in case the bid wins. adi string ID of a preloaded ad to be served if the bid wins. domain string array Advertiser domain for block list checking. bundle string A platform-specific application identifier intended to be unique to the app and independent of the exchange. iurl string URL without cache-busting to an image that is representative of the content of the campaign for ad quality/safety checking. cid string Campaign ID to assist with ad quality checking; the collection of creative for which iurl should be representative. crid string creative ID to assist with ad quality checking. tactic string Tactic ID to enable buyers to label bids for reporting to the exchange the tactic through which their bid was submitted. cat string array IAB content categories of the creative. attr integer array Set of attributes describing the creative. api integer API required by the markup if applicable. protocol integer Video response protocol of the markup if applicable. qagmedia
rating integer Creative media rating per IQG guidelines. language string Language of the creative using ISO-639-1-alpha-2. dealid string Reference to the deal.id from the bid request if this bid pertains to a private marketplace direct deal. w integer Width of the creative in device independent pixels. h integer Height of the creative in device independent pixels. wratio integer Relative width of the creative when expressing size as a ratio. hratio integer Relative height of the creative when expressing size as a ratio. exp integer Advisory as to the number of seconds the bidder is willing to wait between the auction and the actual impression. ext object Placeholder for bidder-specific extentions to OpenRTB.

Referring to Table 9 above, for a Bid object, an attribute field (id) corresponding to the bid ID generated to support logging/tracking of bidders, and an attribute field (id) corresponding to the ID of the imp object in the related bid request ( impid), an attribute field (price) corresponding to the bid price expressed in CPM, an attribute field (nur1) corresponding to the URL that will receive a Win Notice by the exchange if the bid is successful, and according to the business policy of each exchange. An attribute field (bur1) corresponding to the billing notice URL called by the exchange when claiming the winning bid price, an attribute field (lur1) corresponding to the URL that will receive a Loss Notice call by the exchange when the bid fails, and a bid An attribute field (adm) indicating the option to deliver advertising markup if the bid is successful, an attribute field (adid) corresponding to the advertising ID to be sent if the bid is successful, and the purpose of checking whether the advertiser's domain is on the block list. An attribute field (adomain) used as an attribute field (bundle) corresponding to the application bundle or package name (Android) or number ID (iOS), and a cache for images representing the content of the campaign to check advertising quality/safety. An attribute field (iurl) corresponding to a non-busting URL, a corresponding attribute field (cid) corresponding to a campaign ID supporting ad quality checking, and an attribute field corresponding to an advertisement ID supporting ad quality checking ( crid) and an attribute field (tactic) corresponding to a tactic ID that allows buyers to label bids to report to the exchange the tactic for which the bid was submitted, and an attribute field corresponding to the creative expressed through the IAB content category. (cat), an attribute field (attr) corresponding to the value that sets the attribute describing the advertisement, an attribute field (api) corresponding to the API technology required by the markup, and an attribute field corresponding to the video response protocol ( protocol), an attribute field (qagmediarating) corresponding to advertising media evaluation according to the IQG guidelines, an attribute field (language) corresponding to the language of the advertisement using ISO-639-1-alpha-2, and the bidding is private. If relevant to a market transaction, an attribute field (dealid) corresponding to a reference to dead.id in the bid request; an attribute field (w) corresponding to the width of the advertisement in device-independent pixels (DIPS); An attribute field (h) corresponding to the height of the advertisement based on pixels (DIPS), an attribute field (wratio) corresponding to the relative width of the advertisement when the size is expressed as a ratio, and an attribute field (wratio) corresponding to the relative width of the advertisement when the size is expressed as a ratio. An attribute field (hratio) corresponding to the relative height, an attribute field (exp) corresponding to a recommendation on how long a bidder can wait between the auction and the actual exposure, and a placeholder for an exchange-only extension to OpenRTB. It can include a property field (ext) corresponding to .

In this case, as shown in FIG. 4, the bid response data (B_res) includes at least one attribute field belonging to the objects (BidResponse, Seatbid, Bid) shown in Tables 7 to 9 and a value for the field. It may be stored according to the data format for each preset attribute field.

The training data set creation module 300 is used to generate a training data set (D _SET ) for machine learning learning by acquiring accumulated data on whether or not there is a bid response to a bid request that occurred over a predetermined period of time and the bid-related information accordingly. .

The training data set creation module 300 converts some fields of the data into one of the explanatory variable (X) and the response variable (Y) based on the results of comparing the bid request data (B_req) and the bid response data (B_res). Create a training data set (D _SET ) according to the settings (S300).

As shown in FIG. 2, the training data set generation module 300 specifically includes a data comparison unit 310, a response variable setting unit 320, an explanatory variable setting unit 330, and a training data set generating unit 340. It can be included.

The data comparison unit 310 compares the fields of the bid request data (B_req) and the bid response data (B_res) to check whether the unique ID information corresponding to the bid request matches.

The data comparison unit 310 stores the first attribute field (id) corresponding to the unique ID of the bid request provided by the advertising exchange in the BidRequest object of the bid request data (B_req) and the App object of the bid request data (B_req). After selecting one of the second attribute fields (id) corresponding to the specified app ID of the exchange, the third attribute field ( id), it is possible to check whether the unique ID information matches.

The response variable setting unit 320 sets one of a specific field value and a preset variable value of the bid response data (B_res) as the response variable (Y) according to the confirmation result by the data comparison unit 310.

If the unique ID information matches based on the confirmation result, the response variable setting unit 320 sets one of the preset first variable value and the field value corresponding to the bid price included in the bid response data (B_res). It can be set as a response variable (Y), and if the unique ID information does not match, a second variable value that is preset differently from the first variable value can be set as the response variable (Y).

Here, the field value corresponding to the bid price may represent the value of an attribute field (price) corresponding to the bid price expressed as CPM in the Bid object of bid response data (B_res).

If the unique ID information matches, the response variable setting unit 320 sets one of the first variable value and the value for the specific attribute field (price) as the response variable (Y) according to the type of object to be classified or predicted. Specifically, if you want to predict a response to a bid request, you can set the preset first variable value as the response variable (Y1), and if you want to predict the bid price according to a bid request, you can set a specific attribute. The value of the field (price) can be set as the response variable (Y2).

At this time, when predicting whether to respond to a bid request, the corresponding response variable (Y1) is a discrete variable with a probability distribution between 0 and 1, and the first variable value is set to the value of '1'. The second variable value is preferably set to '0'.

Meanwhile, when the object to be classified or predicted is a buyer ID according to a bid request, the response variable setting unit 320 performs a bidding in the SeatBid object of the bid response data (B_res) if the unique ID information matches. The value of the attribute field (seat) corresponding to the seat ID of the buyer's demand-side platform 20 is set as the response variable (Y3), and if the unique ID information does not match, 'NA (Non Available)' indicating a missing value is set. ' can also be set as the response variable (Y).

The explanation variable setting unit 330 sets at least one field value of the bid request data (B_req) as an explanation variable (X).

The explanatory variable setting unit 330 includes at least one attribute field belonging to each of a plurality of objects (BidRequest, Imp, App, Device) included in the bid request data (B_req), as shown in Tables 2 to 6. The value can be set as an explanatory variable (X).

The training data set generator 340 collects the bid request data (B_req) and bid response data (B_res) obtained by the bid request data acquisition unit 100 and the bid response data acquisition unit 200, respectively, according to the acquisition time. The training data set ( D _SET ) is created.

Here, the training data generator 340 may generate a training data set (D _SET ) using Scikit-learn, which is a machine learning library based on the Python language.

In this case, the training data generator 340 uses the explanatory variable ( It is expressed in the data form 51 of a feature matrix (Feature Maxtrix) composed of features (n_features), and the response variable (Y) set by the response variable setting unit 320 is a target corresponding to the label that should be predicted from the input data. After expressing it in the data form 52 of a vector (Target Vector), the characteristic matrix 51 and the target vector 52 are combined into one data frame as shown in FIG. 5 to create a training data set (D _SET ) can be created.

At this time, it is preferable that the feature matrix and the target vector are formed to have the same number of rows (n_samples).

The regression analysis learning module 400 designates a part of the training data set (D _SET ) generated by the training data set generation module 300 as training data (D _L ) for learning and applies it to a preset regression analysis model (RA). Apply learning to predict the result for the response variable (Y) (S400).

The regression analysis learning module 400 performs regression prediction learning to predict whether to respond to a bid request based on the type of object to be classified or predicted through the response variable (Y) set by the response variable setting unit 320. One of (L1), regression prediction learning (L2) to predict the bid price according to the bid request, and classification prediction learning (L3) to classify and predict the buyer ID according to the bid request can be performed.

The regression analysis learning module 400 may specifically include a data division unit 410 and a learning unit 420 as shown in FIG. 2 .

The data division unit 410 randomly extracts and divides the training data set (D _SET ) into one of training data for learning (D _L ) and test data for verification (D _T ) in a reproducible range according to a preset ratio.

As shown in FIG. 6, the data division unit 410 sorts the training data set (D _SET ) generated based on the bid request data and bid response data accumulated over a predetermined period of time in chronological order and then sorts them at a preset ratio. Accordingly, it can be divided in the order of training data (Training; 61) and test data (Test; 62), and the ratio is set to 6 to 4 for training data (61) and test data (62), respectively. desirable.

Here, the data splitter 410 can split the training data (D _L ) and the test data (D _T ) using the train_test_split function included in the Scikit-learn library.

In this case, the division ratio can be set using the test_size parameter of the function, and the initial value of the random number can be set so that the same result can be reproduced every time it is performed using the random_state parameter of the function.

For example, if the ratio of dividing the training data set (D _SET ) is set to a ratio of 6 to 4 for training data (D _L ) and test data (D _T ), the test_size parameter can be set to '0.4'. , you can set any number value to the random_state parameter.

The learning unit 420 performs the learning by applying the training data (D _L ) divided for learning by the data dividing unit 410 to a regression analysis model (RA) based on a decision tree ensemble algorithm using random characteristics. Perform.

Here, the decision tree ensemble algorithm may include a Random Forest algorithm, XGBoost algorithm, and LightBGM algorithm.

In this case, the learning unit 420 uses one of random forest, You can select and perform the learning through a regression analysis model (RA) based on this.

Specifically, if the response variable (Y1) for predicting whether to respond to a bid request is set by the response variable setting unit 320, the learning unit 420 considers that the response variable (Y1) is a discrete variable. The learning may be performed by selecting an algorithm that shows optimal performance when it is a discrete variable from Random Forest, XGBoost, and LightBGM, and then applying training data (D _L ) to a regression analysis model (RA) based on the selected algorithm. And, when the learning is completed, a first regression analysis model (RA_L1) that predicts whether to respond to a bid request may be constructed.

In addition, if the response variable (Y2) for predicting the bid price according to a bid request is set by the response variable setting unit 320, the learning unit 420 considers that the response variable (Y2) is a continuous variable. When the variable is an algorithm, the learning may be performed by selecting an algorithm that _shows optimal performance among Random Forest, , When the learning is completed, a second regression analysis model (RA_L2) that predicts the bid price according to the bid request may be constructed.

In addition, if the response variable (Y3) for predicting the buyer ID according to a bid request is set by the response variable setting unit 320, the learning unit 420 includes a missing value (NA). Considering that it is a nominal variable, the learning may be performed by applying training data (D _L ) to a regression analysis model (RA) based on either XGBoost or LightBGM algorithm, and when the learning is completed, a bid request is made. A third regression analysis model (RA_L3) that predicts the buyer ID according to may be built.

The learning performance evaluation module 500 serves to verify the performance of a regression analysis model (RA) trained through learning using training data ( _DL ) (S500).

The learning performance evaluation module 500 performs regression analysis on the test data (D _T ) after the learning by the learning unit 420 is completed on the entire training data (D _L ) divided for learning by the data dividing unit 410. When applied to the model (RA), the performance of the learning result is evaluated based on the results of calculating a plurality of preset evaluation indicators.

As shown in FIG. 2, the learning performance evaluation module 500 may specifically include a performance evaluation unit 510, a cutoff determination unit 520, and a parameter adjustment unit 530.

The performance evaluation unit 510 compares the actual response value for the response variable (Y) included in the test data (D _T ) with the predicted response value output when applying the test data (D _T ) to the regression analysis model (RA). After calculating the error by comparison, the performance of the results learned by the learning unit 420 is evaluated according to the results of calculating a plurality of preset evaluation indices using this.

The performance evaluation unit 510 is a case where the first regression analysis model (RA_L1) based on the response variable (Y1) for predicting whether to respond to a bid request has been learned by the learning unit 420, or the purchaser according to the bid request. In the case where the third regression analysis model (RA_L3) for predicting buyer ID based on the response variable (Y3) for classifying and predicting ID has been trained by the learning unit 420, accuracy and precision are calculated through the error matrix. The performance of the results learned by the learning unit 420 can be evaluated according to the results of calculating at least one evaluation index among Precision, Recall, and F1 Score.

Here, the performance evaluation unit 510 cross-tabs the classification categories of the actual response values and predicted response values for the bid request using the confusion_matrix function included in the Scikit-learn library. After organizing in the form of a cross table, it can be visualized as a confusion matrix as shown in FIG. 7.

At this time, referring to FIG. 7, the confusion matrix includes true negative data 71 corresponding to a case that is actually negative when predicted to be negative (True Negative, TN), and true negative data 71 corresponding to a case that is actually negative when predicted to be positive. False negative data (72) corresponding to (False Negative, FN), false positive data (73) corresponding to a case that is actually positive when predicted to be negative (False Positive, FP), and a case that is actually positive when predicted to be positive (True It may include true positive data 74 corresponding to Positive, TP).

Here, the plurality of evaluation indicators may include at least one of Accuracy, Precision, Recall, and F1 Score.

Specifically, accuracy refers to the proportion of correctly classified data out of the entire data, that is, the portion corresponding to the diagonal (71,74) in the confusion matrix, and can be calculated according to Equation 1 below.

In addition, precision refers to the proportion of positive actual values among those determined to be positive, that is, the portion (72,74) corresponding to the right column direction in the confusion matrix, according to Equation 2 below: It can be calculated.

In addition, recall refers to the proportion of those with positive actual values that are correctly classified as positive, that is, the portion (73,74) corresponding to the lower row direction in the confusion matrix. Equation 3 below: It can be calculated according to .

Additionally, F1 Score represents the harmonic average of precision and recall, and can be calculated according to Equation 4 below.

In this case, the result values calculated according to Equation 1 to Equation 4 all have values between 0 and 1, and the performance evaluation unit 510 determines that the closer the result value is to 1, the better the performance. It can be evaluated as good.

Meanwhile, when the second regression analysis model (RA_L2) based on the response variable (Y2) for predicting the bid price according to the bid request has been learned by the learning unit 420, the performance evaluation unit 510 calculates the MSE (Mean) The performance of the results learned by the learning unit 420 can be evaluated according to the results of calculating at least one evaluation index of Squared Error and coefficient of determination (R ² , R-Square).

)의 차이를 제곱해 평균한 값을 나타내는 것으로서, 아래의 수학식 5에 따라 계산할 수 있다.In this regard, the MSE is divided into the actual value (y) and the predicted value (

), which represents the average value of squaring the difference, and can be calculated according to Equation 5 below.

In addition, the coefficient of determination (R ² ) is an indicator of how much the independent explanatory variable (X) explains the dependent response variable (Y) in the regression model (explanatory power), and is the explanatory variable among the variance of the response variable (Y). It represents the ratio explained by (X), and can be calculated according to Equation 6 below.

) 간의 차이를 제곱하여 합산한 총 제곱합(SST, Total Sum of squares), 즉, 반응변수(Y)의 분산을 나타내는 것으로서, 이를 수식으로 정리하면 아래의 수학식 7과 같다.Here, SST is the actual value (y _i ) of the response variable (Y) and the average value (

), which represents the total sum of squares (SST), which is the sum of the squares of the difference between the squares, that is, the variance of the response variable (Y). When summarized in a formula, it is as shown in Equation 7 below.

) 간의 차이(잔차)를 제곱하여 합산한 잔차 제곱합(SSR, Residual Sum of Squares), 즉, 에러의 분산을 나타내는 것으로서, 이를 수식으로 정리하면 아래의 수학식 8과 같다.In addition, SSR is the actual value (y _i ) for the response variable (Y) and the predicted value according to the regression model (

), which represents the residual sum of squares (SSR), which is the sum of the squares of the differences (residuals) between the squares, that is, the variance of the error, which can be summarized as Equation 8 below.

At this time, it is preferable that the result calculated according to Equation 6 has a value between 0 and 1.

In this case, the performance evaluation unit 510 can evaluate the performance as being better as the result value (R ² ) calculated according to Equation 6 is closer to 1, and the result value calculated according to Equation 5 The lower the (MSE), the better the performance.

The cutoff determination unit 520 determines a cutoff value that provides optimal performance for the test data D _T based on the results of evaluating the performance by the performance evaluation unit 510.

Here, the cutoff value corresponds to a threshold value that serves as a standard for binary classification of the actual response value for the response variable (Y) included in the test data (D _T ) according to whether or not there is a bid response, and the cutoff determination unit 520 ) may provide the determined cutoff value to the training data set generator 340.

At this time, the training data set generator 340 may reflect the result of binary classification based on the cutoff value in the generated training data set (D _SET ) and then provide it as the data to be divided by the data splitter 410. there is.

In this case, since the results of binary classification according to the cutoff value are used as learning data, accuracy, one of the evaluation indicators calculated by the performance evaluation unit 510, is improved to optimize the prediction performance of the learning model. You can.

The parameter adjustment unit 530 adjusts specific parameter values of the regression analysis model (RA) to provide optimal performance for the test data (D _T ) based on the results of evaluating the performance by the performance evaluation unit 510. .

The parameter adjustment unit 530 includes a regression analysis model (RA) based on the random forest algorithm, an n_estimators parameter that sets the number of decision trees according to the random forest, and a max_features parameter that sets the number of features to be used when generating the tree. parameter, a max_depth parameter that sets the maximum depth of the tree, a max_leaf_nodes parameter that specifies the maximum number of leaf nodes in the tree, and a min_samples_leaf parameter that specifies the sample number of the minimum leaf node in the tree. It could be.

At this time, the max_depth parameter, max_leaf_nodes parameter, and min_samples_leaf parameter relate to pre-pruning options and are used to improve overfitting. The smaller the value of the max_features parameter or the smaller the value of the n_estimators parameter, the more it is possible to prevent overfitting.

In a regression analysis model (RA) based on the This may be adjusting the value of at least one of the learning_rate parameter that controls whether to make corrections, the max_depth parameter that sets the maximum depth of the tree, and the max_leaf_nodes parameter that specifies the maximum number of leaf nodes in the tree.

At this time, the max_depth parameter, max_leaf_nodes parameter, and min_samples_leaf parameter relate to pre-pruning options and are used to improve overfitting. The smaller the n_estimators parameter value or the smaller the learning_rate parameter value, the more overfitting can be prevented.

After completing the learning by the learning unit 420, the result output unit 600 sends the first bid request data (B_req1) generated according to the advertisement request sent by the supply-side platform 10 to the advertising exchange 30. The trained regression analysis model (RA) is applied to output prediction results for the bid request corresponding to the first bid request data (B_req1) (S600).

When the response variable (Y1) for predicting whether or not a bid will respond to a bid request is set by the response variable setting unit 320, the result output unit 600 uses the first bid request data (B_req1) as a regression analysis model (RA ) and compare the output result value (r ₁ ) with the first variable value (e.g., the value of '1') and the second variable value (e.g., the value of '0') and compare the result value (r ₁ ) and The closest variable value can be output as the prediction result.

When the response variable (Y2) for predicting the bid price according to the bid request is set by the response variable setting unit 320, the result output unit 600 uses the first bid request data (B_req1) as a regression analysis model (RA). The profitability judgment result is obtained by comparing the predicted bid price corresponding to the result value (r ₁ ) output by applying to the attribute field (bidfloor) corresponding to the minimum bid price of exposure in the Imp object of the bid request data (B_req). The prediction result can be output.

When the first response variable (Y1) and the second response variable (Y2) for predicting the bid response and bid price according to the bid request are set by the response variable setting unit 320, the result output unit 600 performs the learning process. It is predicted that there will be a bidding response in the predicted results (r ₁₁ , r ₁₂ ) output by applying the first bid request data (B_req1) generated for a predetermined period of time to the regression analysis model (RA) after completing. If the bid price according to the response is predicted to be profitable, you can select the corresponding predicted result value and then apply the corresponding expected bid price to the preset formula to output the results of calculating the expected sales and expected profit. there is.

Here, the case where a bidding response is predicted and the bidding price according to the response is predicted to be profitable means that the predicted result value (r ₁₁ ) output in response to the first response variable (Y1) is the preset value. 1 It has a value close to the variable value (e.g., the value of '1'), and the predicted result value (r ₁₂ ) output in response to the second response variable (Y2) is exposed in the Imp object of the bid request data (B_req). This may be the case when it is greater than the value of the attribute field (bidfloor) corresponding to the minimum bid price.

At this time, the preset formula calculates the expected sales amount by adding up all the expected bid prices corresponding to the predicted result values selected by the result output unit 600, and calculates the first bid request data (B_req1) from the calculated expected sales amount. ) may be expressed as a formula to calculate the expected profit by deducting the traffic cost during the time it was created.

The control unit 800 controls according to the prediction result value (r ₁ ) output by the result output unit 600 based on the type of object to be classified or predicted through the response variable set by the response variable setting unit 320. Perform the action.

When the response variable (Y1) for predicting whether to respond to a bid request is set by the response variable setting unit 320, the control unit 800 is set if the predicted result value (r ₁ ) is the first variable value (e.g. , a value of '1'), as it is predicted that there will be a bidding response, the advertising exchange 30 controls to transmit the first bid request data (B_req1) to the demand-side platform 20, and if the predicted result value (r ₁₎ ) is a second variable value (e.g., a value of '0'), it is predicted that there will be no bidding response, so it can be controlled not to transmit the first bid request data (B_req1).

When a response variable (Y2) for predicting a bid price according to a bid request is set by the response variable setting unit 320, the control unit 800 sets the prediction result value (r ₁ ) to the value of the attribute field (bidfloor). As it is predicted that it will be profitable, the advertising exchange 30 controls to transmit the first bid request data (B_req1) to the demand-side platform 20, and if the predicted result value (r ₁ ) is transmitted to the attribute field (bidfloor) If it is smaller than the value, it is predicted that it will not be profitable, so it can be controlled not to transmit the first bid request data (B_req1).

Meanwhile, the response prediction system according to the present invention may further include a priority determination unit 700 as shown in FIG. 2.

The priority determination unit 700 serves to determine the priority of a plurality of bid requests generated during a predetermined time.

After completing the learning by the learning unit 420, the priority determining unit 700 generates a plurality of bid request data according to the advertising request transmitted from the supply-side platform 10 to the advertising exchange 30 for a predetermined period of time. (B_req) is applied to each regression analysis model (RA) to determine the priority for each bid request based on the plurality of output results (r ₁ , r ₂ , r ₃ ,…,r _n ).

In this case, the control unit 800 determines a transmission target to transmit the bid request data based on the priority for each bid request determined by the priority determination unit 700, and performs the above-described control operation on the determined transmission target. You can.

Specifically, if learning of the first regression analysis model (RA_L1) for predicting whether to respond to a bid request has been completed by the learning unit 420, the control unit 800 provides a plurality of bid request data (B_req). 1 Result values output by applying the regression analysis model (RA_L1) are selected according to preset priorities, and the first bid request data (B_req) corresponding to each of the selected result values is sent to the demand side in the advertising exchange 30. It can be controlled to transmit to the platform 20.

In addition, if learning of the second regression analysis model (RA_L2) for predicting the bid price according to the bid request is completed by the learning unit 420, the control unit 800 sends a plurality of bid request data (B_req) to the second The results output by applying the regression analysis model (RA_L2) are selected according to preset priorities, and the first bid request data (B_req) corresponding to each of the selected results is sent to the demand-side platform at the advertising exchange 30. It can be controlled to transmit to (20).

At this time, the priority can be set to the top 20%, and when only results selected according to priority are used, precision, which is one of the evaluation indicators calculated by the performance evaluation unit 510, is improved. You can get the effect.

Although the present invention has been described in detail through preferred embodiments, the present invention is not limited thereto and may be implemented in various ways within the scope of the patent claims.

In particular, the foregoing has described the features and technical strengths of the present invention rather broadly to enable a better understanding of the claims of the invention to be described later. Therefore, the concept and specific embodiments of the present invention described above are intended to serve a similar purpose as the present invention. It should be recognized by those skilled in the art that it can be immediately used as a basis for the design or modification of other shapes for use.

In addition, the embodiment described above is only one embodiment according to the present invention, and can be implemented in various modified and changed forms by those skilled in the art within the scope of the technical idea of the present invention. You will understand. Accordingly, the disclosed embodiments should be considered from an explanatory rather than a limiting perspective, and various modifications and changes thereof are also included in the scope of the technical spirit of the present invention and are indicated in the claims of the present invention, and the scope equivalent thereto is thereto. All differences therein should be construed as being included in the present invention.

1: Media 2: Advertiser
10: Supply-side platform 20: Demand-side platform
30: Advertising exchange 50: Response prediction system
100: Bid request data acquisition unit 200: Bid response data acquisition unit
300: Training data set generation module 310: Data comparison unit
320: Response variable setting unit 330: Explanation variable setting unit
340: Training data set generation unit 400: Regression analysis learning module
410: Data division unit 420: Learning unit
500: Learning performance evaluation module 510: Performance evaluation unit
520: Cutoff determination unit 530: Parameter adjustment unit
600: Result output unit 700: Priority decision unit
800: Control unit

Claims (14)

When a user accesses a specific digital advertising space, the supply-side platform that acts on behalf of the media that sells the inventory of the advertising space transmits an advertising request for the inventory to the advertising exchange, and the demand-side platform that acts on behalf of the advertiser in the advertising exchange A bid request data acquisition unit that sequentially acquires bid request data transmitted together with a bid request for the inventory;
a bid response data acquisition unit that acquires bid response data transmitted from the demand-side platform to the advertising exchange in time series according to whether the advertiser responds to the bid request;
A training data set generation module that generates a training data set by setting some fields of the data to one of an explanatory variable and a response variable based on a result of comparing the bid request data and the bid response data;
a regression analysis learning module that designates a portion of the generated training data set as training data for learning and applies it to a preset regression analysis model to perform learning to predict result values for the response variable; and
After completing the learning, the first bid request data generated according to the advertisement request sent by the supply-side platform to the advertising exchange is applied to the regression analysis model to predict the bid request corresponding to the first bid request data. A response prediction system for a real-time bidding request based on digital advertising, comprising a result output unit that outputs. According to paragraph 1,
The training data set creation module is,
a data comparison unit that checks whether unique ID information corresponding to the bid request matches each other by comparing fields of the bid request data and the bid response data;
A response variable setting unit that sets one of a specific field value of the bid response data and a preset variable value as the response variable according to the confirmation result;
an explanation variable setting unit that sets at least one field value of the bid request data as the explanation variable; and
The bid request data and the bid response data are divided into data pairs corresponding to the bid response for each bid request according to the acquisition time, and the response variable and the explanatory variable are set using the divided data pair as a unit. A response prediction system for digital advertising-based real-time bidding requests, comprising a training data set generator that generates an accumulated training data set as it repeats for each unit. According to paragraph 1,
The regression analysis learning module is,
a data division unit that randomly extracts and divides the generated training data set into one of training data for learning and test data for verification according to a preset ratio within a reproducible range; and
A learning unit that performs the learning by applying the segmented training data to the regression analysis model based on a decision tree ensemble algorithm using random characteristics; a response to a real-time bidding request based on digital advertising, comprising a. Prediction system. According to paragraph 3,
After the learning is completed for the entire divided training data, the test data is applied to the regression analysis model, and the learning performance is evaluated based on the results of calculating a plurality of preset evaluation indices. A response prediction system for digital advertising-based real-time bidding requests, further comprising an evaluation module. According to paragraph 4,
The learning performance evaluation module is,
According to the results of calculating a plurality of preset evaluation indices based on a comparison result between the actual response value for the response variable included in the test data and the predicted response value output when applying the test data to the regression analysis model, A performance evaluation department that evaluates performance on learning outcomes;
a cutoff determination unit that determines a cutoff value that provides optimal performance for the test data based on the results of evaluating the performance; and
A parameter adjustment unit that adjusts specific parameter values of the regression analysis model to provide optimal performance for the test data based on the results of evaluating the performance; a response to a real-time bidding request based on digital advertising, comprising a. Prediction system. According to paragraph 2,
The response variable setting unit,
If the unique ID information matches based on the confirmation result, one of the first variable value preset according to the type of object to be classified or predicted and the field value corresponding to the bid price included in the bid response data Set as the above response variable,
If the unique ID information is inconsistent, a digital advertising-based real-time method is set as the response variable, a second variable value that is preset differently from the first variable value or a missing value depending on the type of object to be classified or predicted. Response prediction system for bid requests. According to clause 6,
The regression analysis learning module is,
When the first and second response variables for predicting the bid response and bid price according to the bid request are set by the response variable setting unit, the training data set generated in response to the first and second response variables is used for the regression. By applying it to the analysis model, learning is performed to predict bid response and bid price.
The result output unit,
In the predicted results output by applying the first bid request data generated for a predetermined time after completing the learning to the regression analysis model, it is predicted that there will be a bidding response, and the bidding price according to the response is profitable. A real-time bidding request based on digital advertising, characterized in that the predicted result values corresponding to the cases expected to be selected are selected and the corresponding expected bid prices are applied to a preset formula to output the results of calculating the expected sales and expected profits. Response prediction system for. A supply-side platform that acts on behalf of media that sells inventory of digital advertising space, a demand-side platform that acts on behalf of advertisers that place advertisements on the inventory, and advertising that brokers advertising transactions based on real-time bidding between the media and the advertisers. In a method of predicting response to a digital advertising-based real-time bidding request using an exchange,
When a user accesses a specific advertising space and the supply-side platform transmits an advertising request for the inventory of the advertising space to the advertising exchange, the advertising exchange transmits to the demand-side platform along with a bid request for the inventory. Obtaining bid request data in time series;
Time-serially acquiring bid response data transmitted from the demand-side platform to the advertising exchange according to whether the advertiser responds to the bid request;
Creating a training data set by setting some fields of the data as one of an explanatory variable and a response variable based on a result of comparing the bid request data and the bid response data;
Designating a portion of the generated training data set as training data for learning and applying it to a preset regression analysis model to perform learning to predict result values for the response variable; and
After completing the learning, the first bid request data generated according to the advertisement request sent by the supply-side platform to the advertising exchange is applied to the regression analysis model to predict the bid request corresponding to the first bid request data. A method for predicting a response to a real-time bidding request based on digital advertising, comprising: outputting. According to clause 8,
The step of generating the training data set is,
Comparing fields of the bid request data and the bid response data to check whether unique ID information corresponding to the bid request matches each other;
Setting one of a specific field value and a preset variable value of the bid response data as a response variable according to the confirmation result;
Setting at least one field value of the bid request data as an explanatory variable; and
The bid request data and the bid response data are divided into data pairs corresponding to the bid response for each bid request according to the acquisition time, and the response variable and the explanatory variable are set using the divided data pair as a unit. A method for predicting response to a real-time bidding request based on digital advertising, comprising the step of generating an accumulated training dataset through repetition. According to clause 8,
The steps for performing the learning are:
Randomly extracting and dividing the generated training data set into one of training data for learning and test data for verification in a reproducible range according to a preset ratio; and
Predicting responses to real-time bidding requests based on digital advertising, comprising: performing the learning by applying the segmented training data to the regression analysis model based on a decision tree ensemble algorithm using random characteristics. method. According to clause 10,
After performing the learning,
After the learning is completed for all of the divided training data, evaluating performance of the learning result based on results of calculating a plurality of preset evaluation indices by applying the test data to the regression analysis model; A method for predicting response to a digital advertising-based real-time bidding request, further comprising: According to clause 11,
The step of evaluating the performance is,
According to the results of calculating a plurality of preset evaluation indices based on a comparison result between the actual response value for the response variable included in the test data and the predicted response value output when applying the test data to the regression analysis model, evaluating performance on learning outcomes;
determining a cutoff value that provides optimal performance for the test data based on the results of evaluating the performance; and
Predicting responses to real-time bidding requests based on digital advertising, comprising: adjusting specific parameter values of the regression analysis model to provide optimal performance for the test data based on the results of evaluating the performance. method. According to clause 9,
The step of setting the response variable is,
If the unique ID information matches based on the confirmation result, one of the first variable value preset according to the type of object to be classified or predicted and the field value corresponding to the bid price included in the bid response data Set as the above response variable,
If the unique ID information is inconsistent, a digital advertising-based real-time method is set as the response variable, a second variable value that is preset differently from the first variable value or a missing value depending on the type of object to be classified or predicted. How to predict response to bid requests. According to clause 13,
The steps for performing the learning are:
In the step of setting the response variable, if the first and second response variables for predicting the bid response and bid price according to the bid request are set, the training dataset generated in response to the first and second response variables is By applying the regression analysis model, learning is performed to predict bid response and bid price.
The step of outputting the prediction result is,
In the predicted results output by applying the first bid request data generated for a predetermined time after completing the learning to the regression analysis model, it is predicted that there will be a bidding response, and the bidding price according to the response is profitable. A real-time bidding request based on digital advertising, characterized in that the predicted result values corresponding to the cases expected to be selected are selected and the corresponding expected bid prices are applied to a preset formula to output the results of calculating the expected sales and expected profits. How to predict response to.

Images