A method and related apparatus for predicting business conversion rates

Abstract

This application provides a method and related apparatus for predicting business conversion rates. The embodiments of this application can be applied to the field of artificial intelligence technology. The method includes: first, acquiring business data of the target business, object data of the target object, and platform data of the advertisement to be placed; then, classifying and feature mapping based on the data characteristics of each type of data in the acquired business data, object data, and platform data to obtain a first feature vector, a second feature vector, and a third feature vector; then, inputting the first feature vector, the second feature vector, and the third feature vector into an information extraction network, and outputting a target feature vector through the information extraction network; finally, inputting the target feature vector into a fully connected network to obtain the exposure conversion rate of the advertisement on each platform. The business conversion rate prediction method provided by this application predicts the exposure conversion rate by combining platform data from multiple platforms, resulting in a wider range of data sources and improving the accuracy of the exposure conversion rate prediction results.

Description

Technical Field

[0001] This application relates to the field of internet advertising technology, and in particular to a method and apparatus for predicting business conversion rates. Background Technology

[0002] With the development of the internet and the widespread use of computers, the online advertising market has expanded rapidly. When advertisers place ads containing information about their target business on web pages, search engines, browsers, or other online media, they are promoting their target business.

[0003] After a user becomes interested in the content of an advertisement, they will acquire the corresponding product through either click-through conversion or impression conversion. Click-through conversion means that the user first clicks on the advertisement, then enters the download page for the product advertised, downloads the product, and registers as a user after the download is complete. Impression conversion means that the user downloads the product through other means based on the advertisement content, and registers as a user after the download is complete.

[0004] Currently, when calculating the conversion rate of advertisements, only the click-through rate is considered, resulting in inaccurate statistics on advertisement conversion rates. Summary of the Invention

[0005] This application provides a method and related apparatus for predicting business conversion rates. When calculating and predicting the conversion rate of advertisements, it predicts the exposure conversion rate by combining platform data from multiple platforms, thus broadening the data sources and improving the accuracy of the exposure conversion rate prediction results.

[0006] One aspect of this application provides a method for predicting business conversion rates, comprising:

[0007] Obtain business data of the target business, object data of the target object, and data from K platforms. The business data includes the business type, the object data includes the activity sequence of the target object applying to each business, and the data from the K platforms includes the number of times the target object logs into each of the K platforms, where K is an integer greater than or equal to 1.

[0008] Based on business data, object data, and data from K platforms, a first feature vector, a second feature vector, and a third feature vector are generated. The first feature vector corresponds to dense features, the second feature vector corresponds to sparse features, and the third feature vector corresponds to sequence features.

[0009] The first feature vector, the second feature vector, and the third feature vector are used as inputs to the information extraction network, and the target feature vector is output through the information extraction network. The information extraction network includes a first sub-network and a second sub-network. The inputs of the first sub-network are the first feature vector and the second feature vector, and the inputs of the second sub-network are the first feature vector, the second feature vector, and the third feature vector.

[0010] The target feature vector is used as the input of the first fully connected network, and the first fully connected network outputs the exposure conversion rate of the target business for each of the K platforms.

[0011] Another aspect of this application provides a business conversion rate prediction device, comprising:

[0012] The data acquisition module is used to acquire business data of the target business, object data of the target object, and data from K platforms. Among them, business data includes business type, object data includes the activity sequence of the target object applying each business, and K platform data includes the number of times the target object logs into K platforms respectively, where K is an integer greater than or equal to 1.

[0013] The feature extraction and mapping module is used to generate a first feature vector, a second feature vector, and a third feature vector based on business data, object data, and K platform data. The first feature vector corresponds to dense features, the second feature vector corresponds to sparse features, and the third feature vector corresponds to sequence features.

[0014] The information extraction module is used to take the first feature vector, the second feature vector, and the third feature vector as inputs to the information extraction network, and output the target feature vector through the information extraction network. The information extraction network includes a first sub-network and a second sub-network. The inputs of the first sub-network are the first feature vector and the second feature vector, and the inputs of the second sub-network are the first feature vector, the second feature vector, and the third feature vector.

[0015] The exposure conversion rate output module is used to take the target feature vector as input to the first fully connected network and output the exposure conversion rate of the target service for each of the K platforms through the first fully connected network.

[0016] In another implementation of this application embodiment, the business conversion rate prediction device further includes: a first training module. The first training module includes:

[0017] The first sample data acquisition submodule is used to acquire the first training sample data, which includes the business data of the first historical business, the object data of the first historical object, and the platform data of N historical platforms, where N is an integer greater than or equal to 1.

[0018] The first sample data feature extraction and mapping submodule is used to generate a first training feature vector, a second training feature vector and a third training feature vector based on the first training sample data. The first training feature vector corresponds to dense features, the second training feature vector corresponds to sparse features, and the third training feature vector corresponds to sequence features.

[0019] The first sample data information extraction submodule is used to take the first training feature vector, the second training feature vector and the third training feature vector as input to the information extraction network, and output the target training feature vector through the information extraction network. The information extraction network includes a first subnetwork and a second subnetwork. The input of the first subnetwork is the first training feature vector and the second training feature vector, and the input of the second subnetwork is the first training feature vector, the second training feature vector and the third training feature vector.

[0020] The predicted exposure conversion rate output submodule is used to take the target training feature vector as the input of the first fully connected network and output the predicted exposure conversion rate of the historical business for each of the N historical platforms through the first fully connected network.

[0021] The first loss result calculation submodule is used to calculate the first loss result based on N predicted exposure conversion rates;

[0022] The first training submodule is used to train the information extraction network and the first fully connected network using the first loss result.

[0023] In another implementation of the embodiments of this application, the predicted exposure conversion rate is associated with the predicted conversion rate and the predicted exposure count;

[0024] The first loss result calculation submodule is also used for:

[0025] The result of the loss of exposure is calculated based on the predicted conversion rate and the predicted number of impressions;

[0026] The loss in exposures is taken as the first loss result.

[0027] In another implementation of this application embodiment, the business conversion rate prediction device further includes:

[0028] The click conversion rate output module is used to take the target feature vector as input to the second fully connected network and output the click conversion rate of the target business for each of the K platforms through the second fully connected network.

[0029] In another implementation of this application embodiment, the business conversion rate prediction device further includes: a second training module. The second training module includes:

[0030] The second sample data acquisition submodule is used to acquire the second training sample data and the first business conversion data. The second training sample data includes the business data of the second historical business, the object data of the second historical object, and the platform data of P historical platforms. The first business conversion data includes the actual click-through rate and the actual click-through conversion rate obtained from the placement of advertisements corresponding to the historical business on P historical platforms, where P is an integer greater than or equal to 1.

[0031] The second sample data feature extraction and mapping submodule is used to generate a first training feature vector, a second training feature vector and a third training feature vector based on the second training sample data. The first training feature vector corresponds to dense features, the second training feature vector corresponds to sparse features, and the third training feature vector corresponds to sequence features.

[0032] The second sample data information extraction submodule is used to take the first training feature vector, the second training feature vector and the third training feature vector as inputs to the information extraction network, and output the target training feature vector through the information extraction network. The information extraction network includes a first subnetwork and a second subnetwork. The inputs of the first subnetwork are the first training feature vector and the second training feature vector, and the inputs of the second subnetwork are the first training feature vector, the second training feature vector and the third training feature vector.

[0033] The predicted exposure conversion rate output submodule is used to take the target training feature vector as the input of the first fully connected network and output the predicted exposure conversion rate of the historical business for each of the P historical platforms through the first fully connected network.

[0034] The predicted click-through rate output submodule is used to take the target training feature vector as input to the second fully connected network and output the predicted click-through rate of the historical business for each of the P historical platforms through the second fully connected network.

[0035] The second loss result calculation submodule is used to calculate the second loss result based on the first business conversion data, P predicted exposure conversion rates and P predicted click conversion rates;

[0036] The second training submodule is used to train the information extraction network, the first fully connected network, and the second fully connected network using the second loss result.

[0037] In another implementation of the embodiments of this application, the predicted click conversion rate is associated with the predicted click-through rate, the predicted conversion rate and the predicted number of clicks, and the predicted exposure conversion rate is associated with the predicted click-through rate and the predicted number of exposures.

[0038] The second loss result calculation submodule is also used for:

[0039] The click-through rate loss is calculated based on the actual click-through rate and the predicted click-through rate.

[0040] The click-through rate loss is calculated based on the actual click-through rate, the predicted click-through rate, and the predicted conversion rate.

[0041] The click loss result is calculated based on the predicted conversion rate and the predicted number of clicks;

[0042] The result of the loss of exposure is calculated based on the predicted conversion rate and the predicted number of impressions;

[0043] The second loss result is obtained by summing the results of click-through rate loss, click-to-conversion rate loss, click count loss, and impression count loss.

[0044] In another implementation of this application embodiment, the business conversion rate prediction device further includes:

[0045] The global conversion rate output module is used to take the target feature vector as input to the third fully connected network and output the global conversion rate of the target business for K platforms through the third fully connected network.

[0046] In another implementation of this application embodiment, the business conversion rate prediction device further includes: a third training module. The third training module includes:

[0047] The third sample data acquisition submodule is used to acquire the third training sample data and the second business conversion data. The third training sample data includes the business data of the third historical business, the object data of the third historical object, and the platform data of M historical platforms. The second business conversion data includes the actual click-through rate and actual click-through conversion rate obtained from the M historical platforms after placing the corresponding advertisements of the historical business, and the actual global conversion rate of the historical business for the M historical platforms, where M is an integer greater than or equal to 1.

[0048] The third sample data feature extraction and mapping submodule is used to generate a first training feature vector, a second training feature vector, and a third training feature vector based on the third training sample data. The first training feature vector corresponds to dense features, the second training feature vector corresponds to sparse features, and the third training feature vector corresponds to sequence features.

[0049] The third sample data information extraction submodule is used to take the first training feature vector, the second training feature vector and the third training feature vector as input to the information extraction network, and output the target training feature vector through the information extraction network. The information extraction network includes a first subnetwork and a second subnetwork. The input of the first subnetwork is the first training feature vector and the second training feature vector, and the input of the second subnetwork is the first training feature vector, the second training feature vector and the third training feature vector.

[0050] The predicted exposure conversion rate output submodule is used to take the target training feature vector as the input of the first fully connected network and output the predicted exposure conversion rate of the historical business for each of the M historical platforms through the first fully connected network.

[0051] The predicted click-through rate output submodule is used to take the target training feature vector as input to the second fully connected network and output the predicted click-through rate of historical services for each of the M historical platforms through the second fully connected network.

[0052] The global conversion rate prediction output submodule is used to take the target training feature vector as input to the third fully connected network and output the predicted global conversion rate of historical services for M historical platforms through the third fully connected network.

[0053] The third loss result calculation submodule is used to calculate the third loss result based on the second business conversion data, M predicted click conversion rates, M predicted exposure conversion rates and predicted global conversion rate;

[0054] The third training submodule is used to train the information extraction network, the first fully connected network, the second fully connected network, and the third fully connected network using the third loss result.

[0055] In another implementation of the embodiments of this application, the predicted click conversion rate is associated with the predicted click-through rate, the predicted conversion rate and the predicted number of clicks, and the predicted exposure conversion rate is associated with the predicted conversion rate and the predicted number of exposures.

[0056] The third loss result calculation submodule is also used for:

[0057] The click-through rate loss is calculated based on the actual click-through rate and the predicted click-through rate.

[0058] The click-to-conversion loss is calculated based on the actual click-to-conversion rate, the predicted click-to-conversion rate, and the predicted conversion rate.

[0059] The global conversion loss is calculated based on the actual global conversion rate and the predicted global conversion rate.

[0060] The click loss result is calculated based on the predicted conversion rate and the predicted number of clicks;

[0061] The result of the loss of exposure is calculated based on the predicted conversion rate and the predicted number of impressions;

[0062] The third loss result is obtained by summing the click-through rate loss result, click-to-conversion loss result, global conversion loss result, click loss result, and impression loss result.

[0063] In another implementation of this application embodiment, the information extraction module is further used for:

[0064] The first feature vector and the second feature vector are used as inputs to the Factorization Machine (FM) network, and the first target feature vector is output through the Factorization Machine (FM) network. The first sub-network includes the Factorization Machine (FM) network, the Field Factorization Machine (FFM) network, and the Deep Factorization Machine (DeepFM) network. The Factorization Machine (FM) network is used to realize feature crossing.

[0065] The first feature vector, the second feature vector, and the third feature vector are used as inputs to the Progressive Hierarchical Extraction (PLE) network. The PLE network outputs the second target feature vector, wherein the second sub-network includes the Progressive Hierarchical Extraction (PLE) network and the Custom Gate Control (CGC) network.

[0066] The first target feature vector and the second target feature vector are concatenated to obtain the target feature vector.

[0067] Another aspect of this application provides a computer device, comprising:

[0068] Memory, transceivers, processors, and bus systems;

[0069] The memory is used to store programs;

[0070] The processor is used to execute programs in memory, including methods for performing the aspects mentioned above;

[0071] Bus systems are used to connect memory and processors to enable communication between them.

[0072] Another aspect of this application provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the methods described above.

[0073] Another aspect of this application provides a computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the methods provided in the above aspects.

[0074] As can be seen from the above technical solutions, the embodiments of this application have the following advantages:

[0075] This application provides a method and related apparatus for predicting business conversion rates. The method includes: first, acquiring business data of the target business, object data of the target object, and platform data of the advertisement to be placed; then, classifying and feature mapping each type of data in the acquired business data, object data, and platform data according to their data characteristics to obtain a first feature vector, a second feature vector, and a third feature vector; next, inputting the first feature vector, the second feature vector, and the third feature vector into an information extraction network, and outputting a target feature vector through the information extraction network; finally, inputting the target feature vector into a fully connected network to obtain the exposure conversion rate of the advertisement on each platform. The business conversion rate prediction method provided in one embodiment of this application, when predicting the conversion rate of advertisements, combines platform data from multiple platforms to predict the exposure conversion rate, resulting in a wider range of data sources and improved accuracy of the exposure conversion rate prediction results. Attached Figure Description

[0076] Figure 1 A schematic diagram of the architecture of a business conversion rate prediction system provided in a certain embodiment of this application;

[0077] Figure 2 A flowchart illustrating a business conversion rate prediction method provided in one embodiment of this application;

[0078] Figure 3 A flowchart of a business conversion rate prediction method provided in another embodiment of this application;

[0079] Figure 4 A flowchart of a business conversion rate prediction method provided in another embodiment of this application;

[0080] Figure 5 A flowchart of a business conversion rate prediction method provided in another embodiment of this application;

[0081] Figure 6 A flowchart of a business conversion rate prediction method provided in another embodiment of this application;

[0082] Figure 7 A flowchart of a business conversion rate prediction method provided in another embodiment of this application;

[0083] Figure 8 A flowchart of a business conversion rate prediction method provided in another embodiment of this application;

[0084] Figure 9 A flowchart of a business conversion rate prediction method provided in another embodiment of this application;

[0085] Figure 10 A flowchart of a business conversion rate prediction method provided in another embodiment of this application;

[0086] Figure 11 A flowchart illustrating a method for predicting business conversion rates, as provided in yet another embodiment of this application;

[0087] Figure 12(a) is a schematic diagram of the factorization machine (FM) network processing the first feature vector according to a certain embodiment of this application;

[0088] Figure 12(b) is a schematic diagram of the factorization machine (FM) network processing the second feature vector according to a certain embodiment of this application;

[0089] Figure 13 A schematic diagram of feature vector processing using a progressive hierarchical extraction (PLE) network provided in a certain embodiment of this application;

[0090] Figure 14 A schematic diagram illustrating a method for predicting the advertising conversion rate of game A according to a certain embodiment of this application;

[0091] Figure 15 A schematic diagram of a business conversion rate prediction device provided in one embodiment of this application;

[0092] Figure 16 A schematic diagram of a business conversion rate prediction device provided in another embodiment of this application;

[0093] Figure 17 A schematic diagram of a business conversion rate prediction device provided in another embodiment of this application;

[0094] Figure 18 A schematic diagram of a business conversion rate prediction device provided in another embodiment of this application;

[0095] Figure 19 A schematic diagram of a business conversion rate prediction device provided in another embodiment of this application;

[0096] Figure 20 A schematic diagram of a business conversion rate prediction device provided in yet another embodiment of this application;

[0097] Figure 21 This is a schematic diagram of a server structure provided in one embodiment of this application. Detailed Implementation

[0098] This application provides a method for predicting business conversion rates. When calculating the conversion rate of advertisements, it predicts the exposure conversion rate by combining platform data from multiple platforms, thus broadening the data sources and improving the accuracy of the exposure conversion rate prediction results.

[0099] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “corresponding,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0100] To facilitate understanding of the technical solutions provided in the embodiments of this application, some key terms used in the embodiments of this application will be explained below:

[0101] Advertisers: Advertisers are the initiators of advertising campaigns, and are businesses that sell or promote their products and services online. Any business that promotes or sells its products or services can act as an advertiser. Advertisers publish advertising campaigns and pay the advertising platform a fee based on the total number of marketing results achieved in the campaign and the unit price per result.

[0102] Clicks: refers to the number of times an ad is clicked within a certain period of time.

[0103] Click-through rate (CTR) is the ratio of the number of times an ad is clicked to the number of times it is displayed within a certain period. It is a metric used to measure the effectiveness of advertising. CTR = Clicks / Impressions.

[0104] Exposure: refers to the number of times an advertisement is exposed within a certain period of time.

[0105] Exposure rate: This refers to the ratio of the number of times an advertisement is displayed to the number of page views on a media platform within a certain period. It is a metric used to measure the effectiveness of advertising promotion. For example, for internet advertising, if a certain online media platform has 10,000 page views per day, and an advertisement occupies a single ad slot, then the exposure rate is 1:1, and the number of page views is 10,000. If the ad slot displays three advertisements in rotation, then the exposure rate of that advertisement is 1 / 3, and the number of page views is 10,000 / 3.

[0106] Conversion: A conversion is defined as when a user becomes a valid activated, registered, or even paying user.

[0107] Conversion Rate (CVR): The ratio of the number of valid activated / registered / paying users to the number of ad impressions within a certain period. Conversion Rate = Conversions / Impressions.

[0108] Click-through conversion rate (CTCVR): The conversion rate of users who become active / registered / paying users after clicking on an ad.

[0109] Impression Conversion Rate (ICVR): The conversion rate of users who become active / registered / paying users after being exposed to an ad, in the form of a non-click ad.

[0110] Global Action Conversion Rate (GCVR): The total conversion rate of an ad across all advertising platforms.

[0111] Dense feature: Data whose values ​​have practical significance have a dense feature. For example, if the length, width and height of a table are 80cm, 40cm and 100cm respectively, then the data set [80, 40, 100] has a dense feature.

[0112] Sparse feature: Data formed by one-hot encoding of non-data information has a sparse feature. For example, the data structure after one-hot encoding of the category ["cat", "puppy", "bear"] is [[1,0,0,0],[0,1,0,0], [0,0,1,0]]. Then [[1,0,0,0], [0,1,0,0], [0,0,1,0]] has a sparse feature.

[0113] Sequence feature: Numerical data with a certain order. For example: the conversion sequence of an application [social application 001, game application 002, shopping application 003].

[0114] Advertisers place ads targeting their business on websites, search engines, browsers, and other online media platforms. Whenever a target audience views the ad through a website, search engine, browser, or other online media platform, the ad's exposure increases. Whenever a target audience clicks on the ad, the click-through rate increases. For example, with mobile app ads, whenever a target audience clicks on an ad through a media platform, downloads the app within the ad, and then registers, the platform's click-through conversion rate increases. Currently, when calculating conversion rates for ads on a platform, only click-through conversion rates are considered, ignoring exposure conversion rates—that is, the percentage of target audiences who download the app through an app store after viewing or clicking the ad on a platform. This results in inaccurate conversion rate statistics for that ad.

[0115] The method provided in this application provides for statistically predicting the click-through rate and exposure conversion rate of advertisements on each advertising platform, thereby improving the accuracy of the statistical prediction results of advertisement conversion rates.

[0116] For easier understanding, please refer to Figure 1 , Figure 1 This is an application environment diagram of the business conversion rate prediction method in the embodiments of this application, such as... Figure 1 As shown, in this embodiment, the information recommendation method is applied to a business conversion rate prediction system. The business conversion rate prediction system includes: an advertiser terminal, a server, and target user terminal devices. The server can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms. The terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, etc., but is not limited to these. The advertiser terminal and the target user terminal are collectively referred to as terminals. The terminals and the server can be directly or indirectly connected via wired or wireless communication, which is not limited in this embodiment.

[0117] like Figure 1As shown, advertisers can open the ad creation page provided by the advertising platform on their own devices and fill in the relevant information for their target business to complete the creation of the target business ad. The advertising platform then pushes the target business ad to various media platforms via its server. When a target audience views the target business ad on a media platform and clicks on it, the target audience's device sends an ad click request to the server. The server then sends the target business details and a download link to the target audience's device. The target audience can choose to download the target business via the download link, or they can choose to search for and download the target business through their device's app store or other app download channels.

[0118] To reduce advertising costs, advertisers need to predict the promotion effect of each media platform where the target business advertisement is to be launched before promoting the target business advertisement. This requires the server to predict the business conversion rate on each media platform where the target business advertisement is to be launched, and adjust the advertising volume on different media platforms based on the predicted business conversion rate on the media platform.

[0119] To predict business conversion rates, the server first acquires business data of the target business, object data of the target object, and data from multiple platforms. Next, it classifies and maps the data based on the characteristics of each type of data in the acquired business data, object data, and platform data, obtaining a first feature vector, a second feature vector, and a third feature vector. Then, the first, second, and third feature vectors are input into an information extraction network, which outputs a target feature vector. Finally, the target feature vector is input into a fully connected network to obtain the ad exposure conversion rate for each platform.

[0120] The following section will introduce the business conversion rate prediction method in this application from the server's perspective. Please refer to [link / reference]. Figure 2 The business conversion rate prediction method provided in this application embodiment includes steps S110 to S140. Specifically:

[0121] S110. Obtain business data of the target business, object data of the target object, and data from K platforms.

[0122] The business data includes business types, the object data includes the activity sequence of the target object applying to each business, and the K platform data includes the number of times the target object logs into each of the K platforms, where K is an integer greater than or equal to 1.

[0123] It should be noted that business data also includes one or more of the following: business installation package data, business update time data, and business art style. Business type refers to the type of the target business. For example, when the target business is an application (APP), its business type includes: social, news, game, online shopping, etc.; when the target business is a game application, its business type includes: role-playing game (RPG), first-person shooter (FPS), strategy game (SLG), multiplayer online battle arena (MOBA), etc. Business installation package data refers to the size of the target business application's installation package. Business update time data refers to the application's most recent update time and the time elapsed since the last update. When the target business is a game, the business art style includes: cartoon style and realistic style, etc.

[0124] Object data also includes the object's age and gender. The target object's activity sequence across various services refers to the sequence of the target object's activity levels across different services within a certain period. This can be represented by the sequence of the target object's usage time across each service. For example, if the services are various games, and within 3 days, the first game's usage time is 50 hours (h), the second game's usage time is 60 hours (h), and the third game's usage time is 40 hours (h), then the target object's activity sequence across various services is [50, 60, 40].

[0125] The K platform data also includes K platform installation package data, K platform types, and sequence data of the target business corresponding to each of the K platforms. A platform refers to a media platform, specifically a mobile application. The number of times the target object logs into each of the K platforms refers to the number of times the target object logs into different media platforms within a certain period. For example, in one day, the target object logs into the first media platform 5 times and the second media platform once, etc. Platform installation package data refers to the size of the application's installation package on the platform. Platform types include: social, news, games, online shopping, etc. The sequence data of the target business corresponding to each platform refers to the sequence of the target business on different platforms; for example, on the first platform, the sequence of the first game is 001, and the sequence of the second game is 002.

[0126] It is understood that in the specific embodiments of this application, user information (such as gender, age) and related data are involved. When the above embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. Based on this, the user information (such as gender, age) involved in this application is obtained with full authorization under the premise of meeting the principle of minimum necessity and complying with laws and regulations on the collection of personal information, and has been anonymized when used.

[0127] S120. Based on business data, object data, and K platform data, generate the first feature vector, the second feature vector, and the third feature vector.

[0128] Among them, the first feature vector corresponds to dense features, the second feature vector corresponds to sparse features, and the third feature vector corresponds to sequence features.

[0129] It should be noted that, firstly, based on the characteristics of each type of data in business data, object data, and platform data, the data are divided into dense data, sparse data, and sequential data. Secondly, based on the data characteristics of each type of data in business data, object data, and platform data, feature mapping is performed on each type of data to obtain the first feature vector, the second feature vector, and the third feature vector. For example, if the target object logs into the first platform 3 times, the second platform 4 times, and the third platform 5 times, the login count data is [3,4,5]. The login count data is a dense feature. After mapping the login count data to four dimensions, we get [[0,0,0,3], [0,0,0,4], [0,0,0,5]]. Multiplying this by the value of each data point yields the first feature vector [[0,0,0,9], [0,0,1,6], [0,0,2,5]]. The business type is MOBA, and after one-hot encoding, the data is

[101] . The business type is sparse feature, and feature mapping is performed on

[101] to map it to 4 dimensions, resulting in the second feature vector [[0,0,0,1], [0,0,0,0], [0,0,0,1]]. The activity sequence of the target object's first, second, and third applications is [123,456,789], which is a sequence feature. Feature mapping is performed on [123,456,789] to map it to 4 dimensions, resulting in the third feature vector [[0,0,0,1], [0,0,0,2], [0,0,0,3], [0,0,0,4], [0,0,0,5], [0,0,0,6], [0,0,0,7], [0,0,0,8], [0,0,0,9]].

[0130] It is understandable that each type of data in business data, object data, and platform data has one of the following characteristics: dense feature, sparse feature, or sequence feature. That is, each type of data in business data, object data, and platform data can be corresponding to a first feature vector, a second feature vector, and a third feature vector.

[0131] S130. The first feature vector, the second feature vector, and the third feature vector are used as inputs to the information extraction network, and the target feature vector is output through the information extraction network.

[0132] The information extraction network includes a first sub-network and a second sub-network. The inputs of the first sub-network are the first feature vector and the second feature vector, and the inputs of the second sub-network are the first feature vector, the second feature vector, and the third feature vector.

[0133] It should be noted that the first and second subnetworks are parallel structures, meaning that the input and output of the first subnetwork do not affect the input and output of the second subnetwork.

[0134] Understandably, the concatenated vector of the first and second feature vectors is used as the input to the first sub-network, which outputs the first target feature vector. The concatenated vector of the first, second, and third feature vectors is used as the input to the second sub-network, which outputs the second target feature vector. The first and second target feature vectors together constitute the target feature vector.

[0135] S140. The target feature vector is used as the input of the first fully connected network, and the first fully connected network outputs the exposure conversion rate of the target business for each of the K platforms.

[0136] It should be noted that the first fully connected network is used to output the exposure conversion rate.

[0137] This application provides a method for predicting business conversion rates. When predicting the conversion rate of an advertisement, the method combines platform data from multiple platforms to predict the exposure conversion rate, thus providing a wider range of data sources and improving the accuracy of the exposure conversion rate prediction results.

[0138] In this application Figure 2 In one optional embodiment of the business conversion rate prediction method provided in the corresponding implementation, please refer to... Figure 3 The information extraction network and the first fully connected network are trained through steps S210 to S260. Specifically:

[0139] S210. Obtain the first training sample data.

[0140] The first training sample data includes business data of the first historical business, object data of the first historical object, and platform data of N historical platforms, where N is an integer greater than or equal to 1.

[0141] It should be noted that the business data for the first historical business includes one or more of the following: business type, business installation package data, business update time data, and business art style. The object data for the first historical object includes one or more of the following: the activity sequence of the historical object applying various businesses, the age of the historical object, and the gender of the historical object. The platform data for the historical platforms includes one or more of the following: the number of times the historical object logged into N historical platforms, the installation package data of N historical platforms, the types of N historical platforms, and the sequence data of the target business corresponding to each of the N historical platforms.

[0142] It is understandable that the first training sample data is obtained from the training sample set, the business data of the first historical business in the first sample data corresponds to the business data of the target business in step S110, the object data of the first historical object corresponds to the object data of the target object in step S110, and the platform data of the historical platform corresponds to the platform data in step S110.

[0143] S220. Generate a first training feature vector, a second training feature vector, and a third training feature vector based on the first training sample data.

[0144] Among them, the first training feature vector corresponds to dense features, the second training feature vector corresponds to sparse features, and the third training feature vector corresponds to sequence features.

[0145] It should be noted that, firstly, based on the characteristics of each type of data in the business data, object data, and platform data, the business data, object data, and platform data are divided into dense data, sparse data, and sequential data; secondly, based on the data characteristics of each type of data in the business data, object data, and platform data, feature mapping is performed on each type of data to obtain the first training feature vector, the second training feature vector, and the third training feature vector.

[0146] It is understandable that each type of data in business data, object data, and platform data has one of the following features: dense features, sparse features, or sequence features. That is, each type of data in business data, object data, and platform data can be corresponding to a first training feature vector, a second training feature vector, and a third training feature vector.

[0147] S230. The first training feature vector, the second training feature vector, and the third training feature vector are used as inputs to the information extraction network, and the target training feature vector is output through the information extraction network.

[0148] The information extraction network includes a first sub-network and a second sub-network. The inputs of the first sub-network are the first training feature vector and the second training feature vector, and the inputs of the second sub-network are the first training feature vector, the second training feature vector, and the third training feature vector.

[0149] It should be noted that the first and second subnetworks are parallel structures, meaning that the input and output of the first subnetwork do not affect the input and output of the second subnetwork.

[0150] Understandably, the concatenated vector of the first and second training feature vectors is used as the input to the first sub-network, which outputs the first target training feature vector. Similarly, the concatenated vector of the first, second, and third training feature vectors is used as the input to the second sub-network, which outputs the second target training feature vector. The first and second target training feature vectors together constitute the target training feature vector.

[0151] S240. The target training feature vector is used as the input of the first fully connected network, and the first fully connected network outputs the predicted exposure conversion rate of the historical business for each of the N historical platforms.

[0152] S250. Calculate the first loss result based on N predicted exposure conversion rates.

[0153] Understandably, a loss function is constructed by predicting exposure conversion rates, and the first loss result is calculated.

[0154] S260. The information extraction network and the first fully connected network are trained using the first loss result.

[0155] It should be noted that, in order to make the predicted click-through rate and predicted exposure conversion rate closer to the actual click-through rate and actual exposure conversion rate, the parameters in the information extraction network, the second fully connected network, and the first fully connected network need to be adjusted.

[0156] It is understandable that steps S210 to S260 constitute a training process for the information extraction network and the first fully connected network. In actual training, multiple repeated training sessions are required, and each training session is based on the network model after the previous training session. The training of the network model is completed when the loss result of the network model meets the expected preset result.

[0157] The business conversion rate prediction method provided in this application uses historical data as training samples to train the network model, thereby training and adjusting the network model parameters so that the output predicted exposure conversion rate is closer to the actual exposure conversion rate, thus improving the accuracy of the exposure conversion rate prediction results.

[0158] In this application Figure 3 In one optional embodiment of the business conversion rate prediction method provided in the corresponding embodiment, the predicted exposure conversion rate is associated with the predicted conversion rate and the predicted number of exposures.

[0159] Please see Figure 4 Step S250 further includes steps S2501 to S2505. Specifically:

[0160] S2501. Based on the predicted conversion rate and the predicted number of exposures, calculate the result of the exposure loss.

[0161] S2502, take the exposure loss result as the first loss result.

[0162] It should be noted that the first loss result can be obtained by calculating the following formula:

[0163] ;

[0164] in, As the first loss outcome, As a result of the loss of exposure, To predict conversion rates, To predict the number of exposures, To adjust parameters for conversion rate, This is the first training sample data.

[0165] The business conversion rate prediction method provided in this application obtains a first loss function by predicting the conversion rate and the number of exposures, and trains the network model using the first loss function, thereby improving the accuracy of the exposure conversion rate prediction results.

[0166] In this application Figure 2 In one optional embodiment of the business conversion rate prediction method provided in the corresponding implementation, please refer to... Figure 5 Following step S130, the prediction method also includes step S150. Specifically:

[0167] S150. The target feature vector is used as the input of the second fully connected network, and the second fully connected network outputs the click conversion rate of the target business for each of the K platforms.

[0168] It should be noted that the first fully connected network and the second fully connected network are parallel structures, meaning that the input and output of the first fully connected network do not affect the input and output of the second fully connected network. The first fully connected network is used to output the click-through rate, and the second fully connected network is used to output the impression conversion rate.

[0169] It is understandable that steps S140 and S150 are parallel steps.

[0170] This application provides a method for predicting business conversion rates. When statistically predicting the advertising conversion rate of each platform, it simultaneously statistically predicts the click conversion rate and the exposure conversion rate of each platform, thereby improving the accuracy of the advertising conversion rate statistical prediction results.

[0171] In this application Figure 5 In one optional embodiment of the business conversion rate prediction method provided in the corresponding implementation, please refer to... Figure 6 The information extraction network, the first fully connected network, and the second fully connected network are trained through steps S310 to S370. Specifically:

[0172] S310. Obtain the second training sample data and the first business conversion data.

[0173] The second training sample data includes business data of the second historical business, object data of the second historical object, and platform data of P historical platforms. The second business conversion data includes the actual click-through rate and actual click-through conversion rate obtained after placing ads corresponding to the historical business on P historical platforms, where P is an integer greater than or equal to 1.

[0174] It should be noted that the business data for the second historical business includes one or more of the following: business type, business installation package data, business update time data, and business art style. The object data for the second historical object includes one or more of the following: the activity sequence of the historical object applying various businesses, the age of the historical object, and the gender of the historical object. The platform data for the historical platform includes one or more of the following: the number of times the historical object logged into P historical platforms, the installation package data of P historical platforms, the type of P historical platforms, and the sequence data of the target business corresponding to each of the P historical platforms.

[0175] It is understandable that the second training sample data is obtained from the training sample set. The business data of the second historical business in the second sample data corresponds to the business data of the target business in step S110, the object data of the second historical object corresponds to the object data of the target object in step S110, and the platform data of the historical platform corresponds to the platform data in step S110.

[0176] S320. Generate a first training feature vector, a second training feature vector, and a third training feature vector based on the second training sample data.

[0177] Among them, the first training feature vector corresponds to dense features, the second training feature vector corresponds to sparse features, and the third training feature vector corresponds to sequence features.

[0178] It should be noted that, firstly, based on the characteristics of each type of data in the business data, object data, and platform data, the business data, object data, and platform data are divided into dense data, sparse data, and sequential data; secondly, based on the data characteristics of each type of data in the business data, object data, and platform data, feature mapping is performed on each type of data to obtain the first training feature vector, the second training feature vector, and the third training feature vector.

[0179] It is understandable that each type of data in business data, object data, and platform data has one of the following features: dense features, sparse features, or sequence features. That is, each type of data in business data, object data, and platform data can be corresponding to a first training feature vector, a second training feature vector, and a third training feature vector.

[0180] S330. The first training feature vector, the second training feature vector, and the third training feature vector are used as inputs to the information extraction network, and the target training feature vector is output through the information extraction network.

[0181] The information extraction network includes a first sub-network and a second sub-network. The inputs of the first sub-network are the first training feature vector and the second training feature vector, and the inputs of the second sub-network are the first training feature vector, the second training feature vector, and the third training feature vector.

[0182] It should be noted that the first and second subnetworks are parallel structures, meaning that the input and output of the first subnetwork do not affect the input and output of the second subnetwork.

[0183] Understandably, the concatenated vector of the first and second training feature vectors is used as the input to the first sub-network, which outputs the first target training feature vector. Similarly, the concatenated vector of the first, second, and third training feature vectors is used as the input to the second sub-network, which outputs the second target training feature vector. The first and second target training feature vectors together constitute the target training feature vector.

[0184] S340. The target training feature vector is used as the input of the first fully connected network, and the first fully connected network outputs the predicted exposure conversion rate of the historical business for each of the P historical platforms.

[0185] S350. The target training feature vector is used as the input of the second fully connected network, and the predicted click-through rate of the historical business for each of the P historical platforms is output through the second fully connected network.

[0186] It should be noted that the first fully connected network and the second fully connected network are parallel structures, meaning that the input and output of the first fully connected network do not affect the input and output of the second fully connected network. The first fully connected network is used to output the predicted exposure conversion rate, and the second fully connected network is used to output the predicted click-through rate.

[0187] S360. Based on the first business conversion data, P predicted exposure conversion rates, and P predicted click conversion rates, calculate the second loss result.

[0188] Understandably, a loss function is constructed by predicting click-through rate, predicting exposure conversion rate, actual click-through rate, and actual click-through rate, and a second loss result is calculated.

[0189] S370. The information extraction network, the first fully connected network, and the second fully connected network are trained using the second loss result.

[0190] It should be noted that, in order to make the predicted click-through rate and predicted exposure conversion rate closer to the actual click-through rate and actual exposure conversion rate, the parameters in the information extraction network, the first fully connected network, and the second fully connected network need to be adjusted.

[0191] It is understandable that steps S310 to S370 constitute a training process for the information extraction network, the first fully connected network, and the second fully connected network. In actual training, multiple repeated training sessions are required, and each training session is based on the network model after the previous training session. The training of the network model is completed when the loss result of the network model meets the expected preset result.

[0192] The business conversion rate prediction method provided in this application uses historical data as training samples to train the network model, thereby training and adjusting the network model parameters so that the output predicted click conversion rate and predicted exposure conversion rate are closer to the actual click conversion rate and actual exposure conversion rate, thus improving the accuracy of the advertising conversion rate prediction results.

[0193] In this application Figure 6 In one optional embodiment of the business conversion rate prediction method provided in the corresponding embodiment, the predicted click conversion rate is associated with the predicted click-through rate, the predicted conversion rate and the predicted number of clicks, and the predicted exposure conversion rate is associated with the predicted conversion rate and the predicted number of exposures.

[0194] It's important to note that predicted conversion rate includes predicted click-through rate (CTR) and predicted impression rate (PIR). Predicted CTR can be calculated by subtracting predicted click-through rate (CTR) from predicted conversion rate. Similarly, predicted PIR can be calculated by subtracting predicted conversion rate from predicted CTR. For example, Predicted CTR = Predicted CTR × Predicted Conversion Rate.

[0195] Please see Figure 7 Step S360 further includes steps S3601 to S3605. Specifically:

[0196] S3601. Calculate the click-through rate loss result based on the actual click-through rate and the predicted click-through rate.

[0197] S3602. Based on the actual click-through rate, the predicted click-through rate, and the predicted conversion rate, calculate the click-through rate loss result.

[0198] S3603. Based on the predicted conversion rate and predicted click count, calculate the click loss result.

[0199] S3604. Based on the predicted conversion rate and the predicted number of exposures, calculate the result of the exposure loss.

[0200] S3605. Sum the click-through rate loss result, click-to-conversion rate loss result, click count loss result, and impression count loss result to obtain the second loss result.

[0201] It should be noted that the second loss result can be obtained by calculating the following formula:

[0202] ;

[0203] in, This is the second loss outcome. For the result of click-through rate loss, For click-through rate loss results, For the result of click loss, As a result of the loss of exposure, This represents the actual click-through rate. This represents the actual click-through rate. To predict click-through rate, To predict conversion rates, To predict the number of clicks, To predict the number of exposures, Adjust parameters for click-through rate. To adjust parameters for conversion rate, This is the second training sample data.

[0204] The business conversion rate prediction method provided in this application embodiment obtains a second loss function by fitting the predicted click-through rate, predicted conversion rate, predicted number of clicks, and predicted number of impressions with the actual click-through rate and actual conversion rate. The network model is then trained using the second loss function, thereby improving the accuracy of the advertising conversion rate prediction results.

[0205] In this application Figure 5 In one optional embodiment of the business conversion rate prediction method provided in the corresponding implementation, please refer to... Figure 8 Following step S130, the prediction method further includes step S160. Specifically:

[0206] S160. The target feature vector is used as the input of the third fully connected network, and the global conversion rate of the target service across K platforms is output through the third fully connected network.

[0207] It's important to note that the global conversion rate refers to the overall conversion rate achieved across K platforms when ads are deployed to those platforms. The first, second, and third fully connected networks are parallel structures; that is, the inputs and outputs of the first, second, and third fully connected networks do not affect each other. The first fully connected network outputs the impression conversion rate, the second fully connected network outputs the click conversion rate, and the third fully connected network outputs the global conversion rate.

[0208] It is understandable that steps S140, S150 and S160 are parallel steps.

[0209] This application provides a method for predicting business conversion rates, which statistically analyzes the click-through conversion rate of advertisements on each platform, the exposure conversion rate on each platform, and the global conversion rate across all platforms, thereby improving the accuracy of the statistical prediction results of advertisement conversion rates.

[0210] In this application Figure 8 In one optional embodiment of the business conversion rate prediction method provided in the corresponding implementation, please refer to... Figure 9 The information extraction network, the first fully connected network, the second fully connected network, and the third fully connected network are trained through steps S410 to S480. Specifically:

[0211] S410, Obtain the third training sample data and the second business conversion data.

[0212] The third training sample data includes business data of the third historical business, object data of the third historical object, and platform data of M historical platforms. The second business conversion data includes the actual click-through rate and actual click-through conversion rate obtained after placing ads corresponding to the historical business on M historical platforms, and the actual global conversion rate of the historical business on M historical platforms, where M is an integer greater than or equal to 1.

[0213] It should be noted that the business data for the third historical business includes one or more of the following: business type, business installation package data, business update time data, and business art style. The object data for the third historical object includes one or more of the following: the activity sequence of the historical object applying various businesses, the age of the historical object, and the gender of the historical object. The platform data for the historical platform includes one or more of the following: the number of times the historical object logged into M historical platforms, the installation package data of the M historical platforms, the type of the M historical platforms, and the sequence data of the target business corresponding to each of the M historical platforms.

[0214] It is understandable that the third training sample data is obtained from the training sample set. The business data of the third historical business in the third sample data corresponds to the business data of the target business in step S110. The object data of the third historical object corresponds to the object data of the target object in step S110. The platform data of the historical platform corresponds to the platform data in step S110.

[0215] S420. Generate the first training feature vector, the second training feature vector, and the third training feature vector based on the third training sample data.

[0216] Among them, the first training feature vector corresponds to dense features, the second training feature vector corresponds to sparse features, and the third training feature vector corresponds to sequence features.

[0217] It should be noted that, firstly, based on the characteristics of each type of data in the business data, object data, and platform data, the business data, object data, and platform data are divided into dense data, sparse data, and sequential data; secondly, based on the data characteristics of each type of data in the business data, object data, and platform data, feature mapping is performed on each type of data to obtain the first training feature vector, the second training feature vector, and the third training feature vector.

[0218] It is understandable that each type of data in business data, object data, and platform data has one of the following features: dense features, sparse features, or sequence features. That is, each type of data in business data, object data, and platform data can be corresponding to a first training feature vector, a second training feature vector, and a third training feature vector.

[0219] S430. The first training feature vector, the second training feature vector, and the third training feature vector are used as inputs to the information extraction network, and the target training feature vector is output through the information extraction network.

[0220] The information extraction network includes a first sub-network and a second sub-network. The inputs of the first sub-network are the first training feature vector and the second training feature vector, and the inputs of the second sub-network are the first training feature vector, the second training feature vector, and the third training feature vector.

[0221] It should be noted that the first and second subnetworks are parallel structures, meaning that the input and output of the first subnetwork do not affect the input and output of the second subnetwork.

[0222] Understandably, the concatenated vector of the first and second training feature vectors is used as the input to the first sub-network, which outputs the first target training feature vector. Similarly, the concatenated vector of the first, second, and third training feature vectors is used as the input to the second sub-network, which outputs the second target training feature vector. The first and second target training feature vectors together constitute the target training feature vector.

[0223] S440. The target training feature vector is used as the input of the first fully connected network, and the first fully connected network outputs the predicted exposure conversion rate of the historical business for each of the M historical platforms.

[0224] S450. The target training feature vector is used as the input of the second fully connected network, and the predicted click-through rate of the historical business for each of the M historical platforms is output through the second fully connected network.

[0225] S460. The target training feature vector is used as the input of the third fully connected network, and the predicted global conversion rate of historical services for M historical platforms is output through the third fully connected network.

[0226] It should be noted that the first, second, and third fully connected networks are parallel structures; that is, the inputs and outputs of the first, second, and third fully connected networks do not affect each other. The first fully connected network is used to output the impression conversion rate, the second fully connected network is used to output the click conversion rate, and the third fully connected network is used to output the global conversion rate.

[0227] S470. Based on the second business conversion data, M predicted click-through rates, M predicted exposure conversion rates, and the predicted global conversion rate, calculate the third loss result.

[0228] Understandably, a loss function is constructed by predicting click-through rate, predicting exposure conversion rate, predicting global conversion rate, actual click-through rate, actual click-through rate, and actual global conversion rate, and a third loss result is calculated.

[0229] S480. The information extraction network, the first fully connected network, the second fully connected network, and the third fully connected network are trained using the third loss result.

[0230] Understandably, in order to make the predicted click-through rate, predicted exposure conversion rate, and predicted global conversion rate closer to the actual click-through rate, actual exposure conversion rate, and actual global conversion rate, it is necessary to adjust the parameters in the information extraction network, the first fully connected network, the second fully connected network, and the third fully connected network.

[0231] It is understandable that steps S410 to S480 constitute a training process for the information extraction network, the first fully connected network, the second fully connected network, and the third fully connected network. In actual training, multiple repeated training sessions are required, and each training session is based on the network model after the previous training session. The training of the network model is completed when the loss result of the network model meets the expected preset result.

[0232] The business conversion rate prediction method provided in this application uses historical data as training samples to train the network model. By adjusting the network model parameters through training, the output predicted click conversion rate, predicted exposure conversion rate, and predicted global conversion rate are made closer to the actual click conversion rate, actual exposure conversion rate, and actual global conversion rate, thereby improving the accuracy of the advertising conversion rate prediction results.

[0233] In this application Figure 9 In one optional embodiment of the business conversion rate prediction method provided in the corresponding embodiment, the predicted click conversion rate is associated with the predicted click-through rate, the predicted conversion rate and the predicted number of clicks, and the predicted exposure conversion rate is associated with the predicted conversion rate and the predicted number of exposures.

[0234] It's important to note that predicted conversion rate includes predicted click-through rate (CTR) and predicted impression rate (PIR). Predicted CTR can be calculated by subtracting predicted click-through rate (CTR) from predicted conversion rate. Similarly, predicted PIR can be calculated by subtracting predicted conversion rate from predicted CTR. For example, Predicted CTR = Predicted CTR × Predicted Conversion Rate.

[0235] Please see Figure 10 Step S470 further includes steps S4701 to S4706. Specifically:

[0236] S4701. Calculate the click-through rate loss result based on the actual click-through rate and the predicted click-through rate.

[0237] S4702. Based on the actual click-through rate, the predicted click-through rate, and the predicted conversion rate, calculate the click-through loss result.

[0238] S4703. Calculate the global conversion loss result based on the actual global conversion rate and the predicted global conversion rate.

[0239] S4704. Based on the predicted conversion rate and predicted click count, calculate the click loss result.

[0240] S4705. Based on the predicted conversion rate and the predicted number of exposures, calculate the result of the exposure loss.

[0241] S4706. Sum the click-through rate loss result, click-to-conversion loss result, global conversion loss result, click count loss result, and impression count loss result to obtain the third loss result.

[0242] It should be noted that the third loss result can be obtained by calculating the result using the following formula:

[0243] ;

[0244] in, As a third loss outcome, For the result of click-through rate loss, For click-through rate loss results, For the global transformation loss result, For the result of click loss, As a result of the loss of exposure, This represents the actual click-through rate. This represents the actual click-through rate. This represents the actual global conversion rate. To predict click-through rate, To predict conversion rates, To predict the overall conversion rate, To predict the number of clicks, To predict the number of exposures, Adjust parameters for click-through rate. To adjust parameters for conversion rate, Adjust parameters for global conversion rate. This is the third training sample data.

[0245] The business conversion rate prediction method provided in this application embodiment obtains a third loss function by fitting the predicted click-through rate, predicted conversion rate, predicted number of clicks, predicted number of impressions, and predicted global conversion rate with the actual click-through rate, actual conversion rate, and actual global conversion rate. The network model is then trained using the third loss function, thereby improving the accuracy of the advertising conversion rate prediction results.

[0246] In this application Figure 2 In one optional embodiment of the business conversion rate prediction method provided in the corresponding implementation, please refer to... Figure 11 Step S130 further includes steps S1301 to S1303. Specifically:

[0247] S1301. The first feature vector and the second feature vector are used as inputs to the factorization machine (FM) network, and the first target feature vector is output through the factorization machine (FM) network.

[0248] The first sub-network includes a Factor Machine (FM) network, a Field-aware Factorization Machines (FFM) network, and a Deep Factor Machine (DeepFM) network. The Factor Machine (FM) network is used to achieve feature crossing.

[0249] It should be noted that the Factorization Machine (FM) network is a general prediction method that can still estimate reliable parameters for prediction even when the data is very sparse. The Factorization Machine (FM) network applied in the embodiments of this application includes: a sparse feature layer, a dense feature layer, a computation layer, and an output layer.

[0250] It is understood that the first feature vector includes X first feature sub-vectors, and the first feature vector has dense features; the second feature vector includes Y second feature sub-vectors, and the second feature vector has sparse features. Please refer to Figures 12(a) and 12(b). Figure 12(a) is a schematic diagram of the factorization machine (FM) network processing the first feature vector according to an embodiment of this application, and Figure 12(b) is a schematic diagram of the factorization machine (FM) network processing the second feature vector according to an embodiment of this application. The factorization machine (FM) network processing of the first feature vector includes: firstly, performing an inner product operation on the X first feature sub-vectors to obtain X first inner product values; then, operating the X first inner product values ​​according to the Sigmoid function to obtain the first target feature sub-vector.

[0251] The Factorization Machine (FM) network processes the second feature vector as follows: First, it performs feature mapping on the Y second feature vectors to obtain Y mapped feature vectors, which have dense features; next, it sums the Y second feature vectors to obtain the second feature vector sum; then, it performs inner product operation on the Y mapped feature vectors to obtain Y second inner product values; finally, it operates on the second feature vector sum and the Y second inner product values ​​according to the Sigmoid function to obtain the second target feature vector.

[0252] The first target feature vector and the second target feature vector are the outputs of the first sub-network, that is, the first target feature vector includes the first target feature vector and the second target feature vector.

[0253] S1302. The first feature vector, the second feature vector, and the third feature vector are used as inputs to the Progressive Hierarchical Extraction (PLE) network, and the second target feature vector is output through the Progressive Hierarchical Extraction (PLE) network.

[0254] The second sub-network includes a progressive hierarchical extraction (PLE) network and a custom gate control (CGC) network.

[0255] It should be noted that the progressive hierarchical extraction (PLE) network applied in the embodiments of this application includes: an input layer, a first dedicated network layer, a second dedicated network layer, a third dedicated network layer, a shared network layer, a first gating layer, a second gating layer, a third gating layer, a first task tower layer, a second task tower layer, a third task tower layer, a first output layer, a second output layer, and a third output layer.

[0256] Understandably, please refer to Figure 13 , Figure 13 This is a schematic diagram illustrating how the Progressive Hierarchical Extraction (PLE) network, as provided in this embodiment, processes the first feature vector, the second feature vector, and the third feature vector. The process by which the Progressive Hierarchical Extraction (PLE) network processes the first feature vector, the second feature vector, and the third feature vector includes:

[0257] Step 1: Input the concatenated vector composed of the first, second, and third feature vectors into the first dedicated network layer, which outputs the first feature vector. Then, input the concatenated vector composed of the first, second, and third feature vectors into the second dedicated network layer, which outputs the second feature vector. Finally, input the concatenated vector composed of the first, second, and third feature vectors into the third dedicated network layer, which outputs the third feature vector.

[0258] Step 2: Input the concatenated vector consisting of the first feature vector, the second feature vector, and the third feature vector into the shared network layer. The shared network layer outputs a shared feature vector set, which includes the first shared feature vector, the second shared feature vector, and the third shared feature vector. The first shared feature vector is the concatenated vector of the first and second feature vectors, the second shared feature vector is the concatenated vector of the first and third feature vectors, and the third shared feature vector is the concatenated vector of the second and third feature vectors.

[0259] Step 3: Use the first feature vector, the first shared feature vector, and the second shared feature vector as inputs to the first gating layer, and output the first gating processing feature vector through the first gating layer; input the first gating processing feature vector to the first task tower layer, and output the first task feature sub-vector through the first task tower layer, wherein the first task feature sub-vector includes the first feature vector, the first feature vector in the first shared feature vector, and the first feature vector in the second shared feature vector.

[0260] Step 4: Use the second feature vector, the first shared feature vector, and the third shared feature vector as inputs to the second gating layer, and output the second gating processing feature vector through the second gating layer; input the second gating processing feature vector to the second task tower layer, and output the second task feature sub-vector through the second task tower layer, wherein the second task feature sub-vector includes the second feature vector, the second feature vector in the first shared feature vector, and the second feature vector in the third shared feature vector.

[0261] Step 5: Use the third feature vector, the second shared feature vector, and the third shared feature vector as inputs to the third gating layer, and output the third gating processing feature vector through the third gating layer; input the third gating processing feature vector to the third task tower layer, and output the third task feature sub-vector through the third task tower layer, wherein the third task feature sub-vector includes the third feature vector, the third feature vector in the second shared feature vector, and the third feature vector in the third shared feature vector.

[0262] Step 6: The first task feature vector, the second task feature vector, and the third task feature vector are the outputs of the second sub-network; that is, the second target feature vector includes the first task feature vector, the second task feature vector, and the third task feature vector.

[0263] Steps one and two can be executed simultaneously, as can steps three, four, and five.

[0264] S1303. Concatenate the first target feature vector and the second target feature vector to obtain the target feature vector.

[0265] The business conversion rate prediction method provided in this application improves multi-task processing by using a factor decomposition machine (FM) network and a progressive hierarchical extraction (PLE) network, that is, it can simultaneously predict the ad click conversion rate and the exposure conversion rate, thereby improving the accuracy of the ad conversion rate prediction results.

[0266] To facilitate understanding, the following will combine... Figure 14This paper introduces a method for predicting the conversion rate of ads for game G. Advertisers plan to place ads for game G on media platforms A, B, and C. Before placing ads for game G, the paper predicts the click-through rate and impression conversion rate for each media platform, as well as the overall conversion rate for the three media platforms.

[0267] Figure 14 A schematic diagram illustrating the method for predicting ad conversion rates for Game A, including:

[0268] Step 1: Information gathering.

[0269] Specifically: acquire business data for game G, object data for target user objects, and platform data for three media platforms.

[0270] Understandably, the business data for Game G includes one or more of the following: Game G's game type, Game G's installation package data, Game G's release date, the difference between Game G's most recent release date and the current date, and Game G's art style. The target user's data includes one or more of the following: the target user's activity sequence using Game G and other games within a certain period, the target user's age, and the target user's gender. The media platform's platform data includes one or more of the following: the number of times the target user logged into each media platform within a certain period, the installation package data for each media platform, the type of each media platform, and one or more of the following sequences of Game G and other games on each media platform.

[0271] Step 2: Feature vectorization.

[0272] Specifically: First, all the information collected in step 1 is divided into dense feature data, sparse feature data, and sequence feature data according to the characteristics of each type of information. Then, according to the characteristics of each type of information, the feature data corresponding to each type of information is mapped to obtain a first feature vector, a second feature vector, and a third feature vector; wherein, the first feature vector corresponds to dense features, the second feature vector corresponds to sparse features, and the third feature vector corresponds to sequence features.

[0273] It is understandable that all the information collected in step 1 is classified according to features as shown in Table 1.

[0274] Table 1

[0275] Business data object data Platform data Dense features GameG's installation package data, and the difference between GameG's most recent release date and the current date. Age of target users The number of times the target user logged into each media platform, and the installation package data for each media platform. Sparse features Game G's game genres and art style Gender of target user Types of each media platform Sequence features Game G Release Date The target user group uses the activity sequence of game G and other games. Sequences of Game G and other games on each media platform

[0276] Step 3: Vector concatenation.

[0277] Specifically: the first feature vector and the second feature vector are concatenated to obtain the first concatenated vector; the first feature vector, the second feature vector, and the third feature vector are concatenated to obtain the second concatenated vector.

[0278] It is understandable that the first feature vector includes X first feature sub-vectors; the second feature vector includes Y second feature sub-vectors; and the third feature vector includes Z third feature sub-vectors.

[0279] Step 4: Information extraction network.

[0280] Specifically: The first concatenated vector is used as input to the FM network, which outputs the first target feature vector. The second concatenated vector is used as input to the PLE network, which outputs the second target feature vector. The first and second target feature vectors are then concatenated to obtain the target feature vector.

[0281] The FM network processing of the first concatenated vector consists of two parts: processing the first feature vector and processing the second feature vector.

[0282] Step 4.1.1: Processing the first feature vector: First, perform an inner product operation on the X first feature sub-vectors to obtain X first inner product values; then, perform an operation on the X first inner product values ​​according to the Sigmoid function to obtain the first target feature sub-vector.

[0283] Step 4.1.2, Processing the Second Feature Vector: First, perform feature mapping on the Y second feature vectors to obtain Y mapped feature vectors, where the mapped feature vectors have dense features; next, sum the Y second feature vectors to obtain the second feature vector sum; then, perform inner product operation on the Y mapped feature vectors to obtain Y second inner product values; finally, perform the Sigmoid function operation on the second feature vector sum and the Y second inner product values ​​to obtain the second target feature vector.

[0284] Step 4.1.3: The concatenated vector of the first target feature vector and the second target feature vector is the first target feature vector.

[0285] The PLE network's processing of the second concatenated vector includes:

[0286] Step 4.2.1: Input the concatenated vector composed of the first, second, and third feature vectors into the first dedicated network layer, and output the first feature vector through the first dedicated network layer. Input the concatenated vector composed of the first, second, and third feature vectors into the second dedicated network layer, and output the second feature vector through the second dedicated network layer. Input the concatenated vector composed of the first, second, and third feature vectors into the third dedicated network layer, and output the third feature vector through the third dedicated network layer.

[0287] Step 4.2.2: Input the concatenated vector composed of the first feature vector, the second feature vector, and the third feature vector into the shared network layer. The shared network layer outputs a shared feature vector set, which includes the first shared feature vector, the second shared feature vector, and the third shared feature vector. The first shared feature vector is the concatenated vector of the first feature vector and the second feature vector, the second shared feature vector is the concatenated vector of the first feature vector and the third feature vector, and the third shared feature vector is the concatenated vector of the second feature vector and the third feature vector.

[0288] Step 4.2.3: Use the first feature vector, the first shared feature vector, and the second shared feature vector as inputs to the first gating layer, and output the first gating processing feature vector through the first gating layer; input the first gating processing feature vector to the first task tower layer, and output the first task feature sub-vector through the first task tower layer, wherein the first task feature sub-vector includes the first feature vector, the first feature vector in the first shared feature vector, and the first feature vector in the second shared feature vector.

[0289] Step 4.2.4: Use the second feature vector, the first shared feature vector, and the third shared feature vector as inputs to the second gating layer, and output the second gating processing feature vector through the second gating layer; input the second gating processing feature vector to the second task tower layer, and output the second task feature sub-vector through the second task tower layer, wherein the second task feature sub-vector includes the second feature vector, the second feature vector in the first shared feature vector, and the second feature vector in the third shared feature vector.

[0290] Step 4.2.5: Use the third feature vector, the second shared feature vector, and the third shared feature vector as inputs to the third gating layer, and output the third gating processing feature vector through the third gating layer; input the third gating processing feature vector to the third task tower layer, and output the third task feature sub-vector through the third task tower layer, wherein the third task feature sub-vector includes the third feature vector, the third feature vector in the second shared feature vector, and the third feature vector in the third shared feature vector.

[0291] Step 4.2.6: The first task feature vector, the second task feature vector, and the third task feature vector are the outputs of the second sub-network; that is, the second target feature vector includes the first task feature vector, the second task feature vector, and the third task feature vector.

[0292] Step 5: Independent Task Network.

[0293] Specifically: The target feature vector is used as the input to the click-through rate task network, the exposure conversion rate task network, and the global conversion rate task network, respectively. The click-through rate task network outputs the click-through rate of each media platform, the exposure conversion rate task network outputs the exposure conversion rate of each media platform, and the global conversion rate task network outputs the global conversion rate of the three platforms.

[0294] This application provides a method for predicting the advertising conversion rate of game G. When calculating the advertising conversion rate of each media platform, it simultaneously calculates the click conversion rate of each platform, the exposure conversion rate of each platform, and the global conversion rate of all media platforms, thereby improving the accuracy of the statistical results of the advertising conversion rate of game G.

[0295] The map data processing apparatus in this application is described in detail below. Please refer to [link / reference]. Figure 15 . Figure 15 This is a schematic diagram of an embodiment of the business conversion rate prediction device 10 in this application. The business conversion rate prediction device 10 includes:

[0296] The data acquisition module 110 is used to acquire business data of the target business, object data of the target object, and data of K platforms. The business data includes business type, the object data includes the activity sequence of the target object applying each business, and the data of K platforms includes the number of times the target object logs into K platforms respectively, where K is an integer greater than or equal to 1.

[0297] The feature extraction and mapping module 120 is used to generate a first feature vector, a second feature vector, and a third feature vector based on business data, object data, and K platform data. The first feature vector corresponds to dense features, the second feature vector corresponds to sparse features, and the third feature vector corresponds to sequence features.

[0298] The information extraction module 130 is used to take the first feature vector, the second feature vector and the third feature vector as input to the information extraction network, and output the target feature vector through the information extraction network. The information extraction network includes a first sub-network and a second sub-network. The input of the first sub-network is the first feature vector and the second feature vector, and the input of the second sub-network is the first feature vector, the second feature vector and the third feature vector.

[0299] The exposure conversion rate output module 140 is used to take the target feature vector as the input of the first fully connected network and output the exposure conversion rate of the target service for each of the K platforms through the first fully connected network.

[0300] This application provides a business conversion rate prediction device. When predicting the conversion rate of an advertisement, it combines platform data from multiple platforms to predict the exposure conversion rate, thus providing a wider range of data sources and improving the accuracy of the exposure conversion rate prediction results.

[0301] In this application Figure 15 In one optional embodiment of the business conversion rate prediction device provided in the corresponding embodiment, please refer to... Figure 16 The business conversion rate prediction device 10 further includes: a first training module 200. The first training module 200 further includes:

[0302] The first sample data acquisition submodule is used to acquire the first training sample data and the first business conversion data. The first training sample data includes the business data of the first historical business, the object data of the first historical object, and the platform data of N historical platforms. The first business conversion data includes the actual click-through rate and the actual click-through conversion rate obtained from the placement of advertisements corresponding to the historical business on N historical platforms, where N is an integer greater than or equal to 1.

[0303] The first sample data feature extraction and mapping submodule is used to generate a first training feature vector, a second training feature vector and a third training feature vector based on the first training sample data. The first training feature vector corresponds to dense features, the second training feature vector corresponds to sparse features, and the third training feature vector corresponds to sequence features.

[0304] The first sample data information extraction submodule is used to take the first training feature vector, the second training feature vector and the third training feature vector as input to the information extraction network, and output the target training feature vector through the information extraction network. The information extraction network includes a first subnetwork and a second subnetwork. The input of the first subnetwork is the first training feature vector and the second training feature vector, and the input of the second subnetwork is the first training feature vector, the second training feature vector and the third training feature vector.

[0305] The predicted exposure conversion rate output submodule is used to take the target training feature vector as the input of the first fully connected network and output the predicted exposure conversion rate of the historical business for each of the N historical platforms through the first fully connected network.

[0306] The first loss result calculation submodule is used to calculate the first loss result based on N predicted exposure conversion rates;

[0307] The first training submodule is used to train the information extraction network and the first fully connected network using the first loss result.

[0308] The business conversion rate prediction device provided in this application embodiment trains the network model through the first training module, thereby training and adjusting the network model parameters so that the output predicted exposure conversion rate is closer to the actual exposure conversion rate, thus improving the accuracy of the exposure conversion rate prediction result.

[0309] In this application Figure 16 In one optional embodiment of the business conversion rate prediction device provided in the corresponding embodiment, the predicted exposure conversion rate is associated with the predicted click-through rate and the predicted number of impressions. The first loss result calculation submodule is further used for:

[0310] The result of the loss of exposure is calculated based on the predicted conversion rate and the predicted number of impressions;

[0311] The loss in exposures is taken as the first loss result.

[0312] The business conversion rate prediction device provided in this application embodiment obtains a first loss function by predicting the conversion rate and the number of exposures, and trains the network model using the first loss function, thereby improving the accuracy of the exposure conversion rate prediction results.

[0313] In this application Figure 15 In one optional embodiment of the business conversion rate prediction device provided in the corresponding embodiment, please refer to... Figure 17 The business conversion rate prediction device 10 also includes:

[0314] Click conversion rate output module 150 is used to take the target feature vector as input to the second fully connected network and output the click conversion rate of the target business for each of the K platforms through the second fully connected network.

[0315] This application provides a business conversion rate prediction device that, when predicting the advertising conversion rate for each platform, simultaneously predicts the click-through conversion rate and the exposure conversion rate for each platform, thereby improving the accuracy of the advertising conversion rate prediction results.

[0316] In this application Figure 17 In one optional embodiment of the business conversion rate prediction device provided in the corresponding embodiment, please refer to... Figure 18 The business conversion rate prediction device 10 also includes a second training module 300. The second training module 300 specifically includes:

[0317] The second sample data acquisition submodule is used to acquire the second training sample data and the first business conversion data. The second training sample data includes the business data of the second historical business, the object data of the second historical object, and the platform data of P historical platforms. The first business conversion data includes the actual click-through rate and the actual click-through conversion rate obtained from the placement of advertisements corresponding to the historical business on P historical platforms, where P is an integer greater than or equal to 1.

[0318] The second sample data feature extraction and mapping submodule is used to generate a first training feature vector, a second training feature vector and a third training feature vector based on the second training sample data. The first training feature vector corresponds to dense features, the second training feature vector corresponds to sparse features, and the third training feature vector corresponds to sequence features.

[0319] The second sample data information extraction submodule is used to take the first training feature vector, the second training feature vector and the third training feature vector as inputs to the information extraction network, and output the target training feature vector through the information extraction network. The information extraction network includes a first subnetwork and a second subnetwork. The inputs of the first subnetwork are the first training feature vector and the second training feature vector, and the inputs of the second subnetwork are the first training feature vector, the second training feature vector and the third training feature vector.

[0320] The predicted exposure conversion rate output submodule is used to take the target training feature vector as the input of the first fully connected network and output the predicted exposure conversion rate of the historical service for each of the P historical platforms through the first fully connected network.

[0321] The predicted click-through rate output submodule is used to take the target training feature vector as input to the second fully connected network and output the predicted click-through rate of the historical business for each of the P historical platforms through the second fully connected network.

[0322] The second loss result calculation submodule is used to calculate the second loss result based on the first business conversion data, P predicted exposure conversion rates and P predicted click conversion rates;

[0323] The second training submodule is used to train the information extraction network, the first fully connected network, and the second fully connected network using the second loss result.

[0324] The business conversion rate prediction device provided in this application uses historical data as training samples to train the network model, thereby training and adjusting the network model parameters so that the output predicted click conversion rate and predicted exposure conversion rate are closer to the actual click conversion rate and actual exposure conversion rate, thus improving the accuracy of the advertising conversion rate prediction results.

[0325] In this application Figure 18In one optional embodiment of the business conversion rate prediction device provided in the corresponding embodiment, the predicted click conversion rate is associated with the predicted click-through rate, the predicted conversion rate and the predicted number of clicks, and the predicted exposure conversion rate is associated with the predicted conversion rate and the predicted number of exposures.

[0326] The second loss result calculation submodule is also used for:

[0327] The click-through rate loss is calculated based on the actual click-through rate and the predicted click-through rate.

[0328] The click-through rate loss is calculated based on the actual click-through rate, the predicted click-through rate, and the predicted conversion rate.

[0329] The click loss result is calculated based on the predicted conversion rate and the predicted number of clicks;

[0330] The result of the loss of exposure is calculated based on the predicted conversion rate and the predicted number of impressions;

[0331] The second loss result is obtained by summing the results of click-through rate loss, click-to-conversion rate loss, click count loss, and impression count loss.

[0332] The business conversion rate prediction method provided in this application embodiment obtains a second loss function by fitting the predicted click-through rate, predicted conversion rate, predicted number of clicks, and predicted number of impressions with the actual click-through rate and actual conversion rate. The network model is then trained using the second loss function, thereby improving the accuracy of the advertising conversion rate prediction results.

[0333] In this application Figure 17 In one optional embodiment of the business conversion rate prediction device provided in the corresponding embodiment, please refer to... Figure 19 The business conversion rate prediction device 10 also includes:

[0334] The global conversion rate output module 160 is used to take the target feature vector as the input of the third fully connected network and output the global conversion rate of the target business for K platforms through the third fully connected network.

[0335] This application provides a business conversion rate prediction device that calculates the click-through conversion rate of ads on each platform, the exposure conversion rate on each platform, and the global conversion rate across all platforms, thereby improving the accuracy of the ad conversion rate prediction results.

[0336] In this application Figure 19 In one optional embodiment of the business conversion rate prediction device provided in the corresponding embodiment, please refer to... Figure 20 The business conversion rate prediction device 10 also includes a third training module 400. The third training module 400 specifically includes:

[0337] The third sample data acquisition submodule is used to acquire the third training sample data and the second business conversion data. The third training sample data includes the business data of the third historical business, the object data of the third historical object, and the platform data of M historical platforms. The third business conversion data includes the actual click-through rate and actual click-through conversion rate obtained from the M historical platforms after placing the corresponding advertisements of the historical business, and the actual global conversion rate of the historical business for the M historical platforms, where M is an integer greater than or equal to 1.

[0338] The third sample data feature extraction and mapping submodule is used to generate a first training feature vector, a second training feature vector, and a third training feature vector based on the third training sample data. The first training feature vector corresponds to dense features, the second training feature vector corresponds to sparse features, and the third training feature vector corresponds to sequence features.

[0339] The third sample data information extraction submodule is used to take the first training feature vector, the second training feature vector and the third training feature vector as input to the information extraction network, and output the target training feature vector through the information extraction network. The information extraction network includes a first subnetwork and a second subnetwork. The input of the first subnetwork is the first training feature vector and the second training feature vector, and the input of the second subnetwork is the first training feature vector, the second training feature vector and the third training feature vector.

[0340] The predicted exposure conversion rate output submodule is used to take the target training feature vector as the input of the first fully connected network and output the predicted exposure conversion rate of the historical business for each of the M historical platforms through the first fully connected network.

[0341] The predicted click-through rate output submodule is used to take the target training feature vector as input to the second fully connected network and output the predicted click-through rate of the historical business for each of the M historical platforms through the second fully connected network.

[0342] The global conversion rate prediction output submodule is used to take the target training feature vector as input to the third fully connected network and output the predicted global conversion rate of historical services for M historical platforms through the third fully connected network.

[0343] The third loss result calculation submodule is used to calculate the third loss result based on the second business conversion data, M predicted click conversion rates, M predicted exposure conversion rates and predicted global conversion rate;

[0344] The third training submodule is used to train the information extraction network, the first fully connected network, the second fully connected network, and the third fully connected network using the third loss result.

[0345] The business conversion rate prediction device provided in this application embodiment trains the network model through a third training module. By adjusting the network model parameters through training, the output predicted click conversion rate, predicted exposure conversion rate, and predicted global conversion rate are made closer to the actual click conversion rate, actual exposure conversion rate, and actual global conversion rate, thereby improving the accuracy of the advertising conversion rate prediction results.

[0346] In this application Figure 20 In one optional embodiment of the business conversion rate prediction device provided in the corresponding embodiment, the predicted click conversion rate is associated with the predicted click-through rate, the predicted conversion rate and the predicted number of clicks, and the predicted exposure conversion rate is associated with the predicted conversion rate and the predicted number of exposures.

[0347] The third loss result calculation submodule is also used for:

[0348] The click-through rate loss is calculated based on the actual click-through rate and the predicted click-through rate.

[0349] The click-to-conversion loss is calculated based on the actual click-to-conversion rate, the predicted click-to-conversion rate, and the predicted conversion rate.

[0350] The global conversion loss is calculated based on the actual global conversion rate and the predicted global conversion rate.

[0351] The click loss result is calculated based on the predicted conversion rate and the predicted number of clicks;

[0352] The result of the loss of exposure is calculated based on the predicted conversion rate and the predicted number of impressions;

[0353] The third loss result is obtained by summing the click-through rate loss result, click-to-conversion loss result, global conversion loss result, click loss result, and impression loss result.

[0354] The business conversion rate prediction device provided in this application embodiment obtains a third loss function by fitting the predicted click-through rate, predicted conversion rate, predicted number of clicks, predicted number of impressions, and predicted global conversion rate with the actual click-through rate, actual conversion rate, and actual global conversion rate. The network model is then trained using the third loss function, thereby improving the accuracy of the advertising conversion rate prediction results.

[0355] Figure 21This is a schematic diagram of a server structure provided in an embodiment of this application. The server 300 can vary significantly due to different configurations or performance. It may include one or more central processing units (CPUs) 322 (e.g., one or more processors) and memory 332, and one or more storage media 330 (e.g., one or more mass storage devices) for storing application programs 342 or data 344. The memory 332 and storage media 330 can be temporary or persistent storage. The program stored in the storage media 330 may include one or more modules (not shown in the diagram), each module may include a series of instruction operations on the server. Furthermore, the CPU 322 may be configured to communicate with the storage media 330 and execute the series of instruction operations stored in the storage media 330 on the server 300.

[0356] Server 300 may also include one or more power supplies 326, one or more wired or wireless network interfaces 350, one or more input / output interfaces 358, and / or one or more operating systems 341, such as Windows Server. TM Mac OS X TM Unix TM Linux TM FreeBSD TM etc.

[0357] The steps performed by the server in the above embodiments can be based on this Figure 21 The server structure shown.

[0358] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0359] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0360] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0361] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0362] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0363] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for predicting business conversion rates, characterized in that, include: Acquire business data of the target business, object data of the target object, and data from K platforms. The business data includes business types, the object data includes the activity sequence of the target object applying each business, and the K platform data includes the number of times the target object logs into each of the K platforms, where K is an integer greater than or equal to 1. Based on the business data, the object data, and the K platform data, a first feature vector, a second feature vector, and a third feature vector are generated, wherein the first feature vector corresponds to dense features, the second feature vector corresponds to sparse features, and the third feature vector corresponds to sequence features; The first feature vector, the second feature vector, and the third feature vector are used as inputs to an information extraction network, and the target feature vector is output through the information extraction network. The information extraction network includes a first sub-network and a second sub-network. The inputs of the first sub-network are the first feature vector and the second feature vector, and the inputs of the second sub-network are the first feature vector, the second feature vector, and the third feature vector. The target feature vector is used as input to a first fully connected network, which then outputs the exposure conversion rate of the target service for each of the K platforms. The step of using the first feature vector, the second feature vector, and the third feature vector as inputs to an information extraction network, and outputting a target feature vector through the information extraction network, includes: The first feature vector and the second feature vector are used as inputs to the Factorization Machine (FM) network, and the first target feature vector is output through the Factorization Machine (FM) network. The first sub-network includes the Factorization Machine (FM) network, the Field Factorization Machine (FFM) network, and the Deep Factorization Machine (DeepFM) network. The Factorization Machine (FM) network is used to realize feature crossing. The first feature vector, the second feature vector, and the third feature vector are used as inputs to the Progressive Hierarchical Extraction (PLE) network, and the second target feature vector is output through the PLE network. The second sub-network includes the PLE network and the Custom Gate Control (CGC) network. The first target feature vector and the second target feature vector are concatenated to obtain the target feature vector.

2. The business conversion rate prediction method as described in claim 1, characterized in that, The information extraction network and the first fully connected network are trained through the following steps: Obtain first training sample data, wherein the first training sample data includes business data of the first historical business, object data of the first historical object, and platform data of N historical platforms, where N is an integer greater than or equal to 1; A first training feature vector, a second training feature vector, and a third training feature vector are generated based on the first training sample data, wherein the first training feature vector corresponds to dense features, the second training feature vector corresponds to sparse features, and the third training feature vector corresponds to sequence features. The first training feature vector, the second training feature vector, and the third training feature vector are used as inputs to the information extraction network, and the target training feature vector is output through the information extraction network. The information extraction network includes a first sub-network and a second sub-network. The inputs of the first sub-network are the first training feature vector and the second training feature vector, and the inputs of the second sub-network are the first training feature vector, the second training feature vector, and the third training feature vector. The target training feature vector is used as the input of the first fully connected network, and the first fully connected network outputs the predicted exposure conversion rate of the historical service for each of the N historical platforms. Calculate the first loss result based on N predicted exposure conversion rates; The information extraction network and the first fully connected network are trained using the first loss result.

3. The business conversion rate prediction method as described in claim 2, characterized in that, The predicted exposure conversion rate is correlated with the predicted conversion rate and the predicted number of exposures; The calculation of the first loss result based on N predicted exposure conversion rates includes: Based on the predicted conversion rate and the predicted number of exposures, the result of the exposure loss is calculated; The exposure loss result is taken as the first loss result.

4. The business conversion rate prediction method as described in claim 1, characterized in that, After outputting the target feature vector through the information extraction network, the process further includes: The target feature vector is used as input to the second fully connected network, which outputs the click-through rate of the target service for each of the K platforms.

5. The business conversion rate prediction method as described in claim 4, characterized in that, The information extraction network, the first fully connected network, and the second fully connected network are trained through the following steps: Obtain second training sample data and first business conversion data, wherein the second training sample data includes business data of the second historical business, object data of the second historical object, and platform data of P historical platforms, and the first business conversion data includes the actual click-through rate and actual click-through conversion rate obtained from the placement of the corresponding advertisements of the historical business on the P historical platforms, where P is an integer greater than or equal to 1; A first training feature vector, a second training feature vector, and a third training feature vector are generated based on the second training sample data, wherein the first training feature vector corresponds to dense features, the second training feature vector corresponds to sparse features, and the third training feature vector corresponds to sequence features. The first training feature vector, the second training feature vector, and the third training feature vector are used as inputs to the information extraction network, and the target training feature vector is output through the information extraction network. The information extraction network includes a first sub-network and a second sub-network. The inputs of the first sub-network are the first training feature vector and the second training feature vector, and the inputs of the second sub-network are the first training feature vector, the second training feature vector, and the third training feature vector. The target training feature vector is used as the input of the first fully connected network, and the first fully connected network outputs the predicted exposure conversion rate of the historical service for each of the P historical platforms. The target training feature vector is used as the input of the second fully connected network, and the predicted click-through rate of the historical service for each of the P historical platforms is output through the second fully connected network. Calculate the second loss result based on the first business conversion data, P predicted exposure conversion rates, and P predicted click conversion rates; The information extraction network, the first fully connected network, and the second fully connected network are trained using the second loss result.

6. The business conversion rate prediction method as described in claim 5, characterized in that, The predicted click-through rate is correlated with the predicted click-through rate, the predicted conversion rate and the predicted number of clicks; the predicted exposure conversion rate is correlated with the predicted conversion rate and the predicted number of exposures. The step of calculating the second loss result based on the first business conversion data, P predicted exposure conversion rates, and P predicted click conversion rates includes: The click-through rate loss result is calculated based on the actual click-through rate and the predicted click-through rate. Based on the actual click-through rate, the predicted click-through rate, and the predicted conversion rate, the click-through rate loss result is calculated. Based on the predicted conversion rate and the predicted number of clicks, the click loss result is calculated; Based on the predicted conversion rate and the predicted number of exposures, the result of the exposure loss is calculated; The second loss result is obtained by summing the click-through rate loss result, the click-to-conversion rate loss result, the click count loss result, and the impression count loss result.

7. The business conversion rate prediction method as described in claim 4, characterized in that, After outputting the target feature vector through the information extraction network, the process further includes: The target feature vector is used as input to a third fully connected network, which then outputs the global conversion rate of the target service across the K platforms.

8. The business conversion rate prediction method as described in claim 7, characterized in that, The information extraction network, the first fully connected network, the second fully connected network, and the third fully connected network are trained through the following steps: Obtain third training sample data and second business conversion data, wherein the third training sample data includes business data of the third historical business, object data of the third historical object and platform data of M historical platforms, and the second business conversion data includes the actual click-through rate and actual click-through conversion rate obtained after placing the corresponding advertisement of the historical business on the M historical platforms, and the actual global conversion rate of the historical business for the M historical platforms, where M is an integer greater than or equal to 1; A first training feature vector, a second training feature vector, and a third training feature vector are generated based on the third training sample data, wherein the first training feature vector corresponds to dense features, the second training feature vector corresponds to sparse features, and the third training feature vector corresponds to sequence features. The first training feature vector, the second training feature vector, and the third training feature vector are used as inputs to the information extraction network, and the target training feature vector is output through the information extraction network. The information extraction network includes a first sub-network and a second sub-network. The inputs of the first sub-network are the first training feature vector and the second training feature vector, and the inputs of the second sub-network are the first training feature vector, the second training feature vector, and the third training feature vector. The target training feature vector is used as the input of the first fully connected network, and the first fully connected network outputs the predicted exposure conversion rate of the historical service for each of the M historical platforms. The target training feature vector is used as the input of the second fully connected network, and the predicted click-through rate of the historical service for each of the M historical platforms is output through the second fully connected network. The target training feature vector is used as the input of the third fully connected network, and the predicted global conversion rate of the historical service for the M historical platforms is output through the third fully connected network. Based on the second business conversion data, M predicted click-through rates, M predicted exposure conversion rates, and the predicted global conversion rate, calculate the third loss result; The information extraction network, the first fully connected network, the second fully connected network, and the third fully connected network are trained using the third loss result.

9. The business conversion rate prediction method as described in claim 8, characterized in that, The predicted click-through rate is correlated with the predicted click-through rate, the predicted conversion rate and the predicted number of clicks; the predicted exposure conversion rate is correlated with the predicted conversion rate and the predicted number of exposures. The calculation of the third loss result based on the second business conversion data, M predicted click-through rates, M predicted exposure conversion rates, and the predicted global conversion rate includes: The click-through rate loss result is calculated based on the actual click-through rate and the predicted click-through rate. The click-through rate loss is calculated based on the actual click-through rate, the predicted click-through rate, and the predicted conversion rate. The global conversion loss result is calculated based on the actual global conversion rate and the predicted global conversion rate. Based on the predicted conversion rate and the predicted number of clicks, the click loss result is calculated; Based on the predicted conversion rate and the predicted number of exposures, the result of the exposure loss is calculated; The third loss result is obtained by summing the click-through rate loss result, the click-to-conversion loss result, the global conversion loss result, the click count loss result, and the impression count loss result.

10. A business conversion rate prediction device, characterized in that, include: The data acquisition module is used to acquire business data of the target business, object data of the target object, and data from K platforms. The business data includes business types, the object data includes the activity sequence of the target object applying various businesses, and the K platform data includes the number of times the target object logs into K platforms respectively, where K is an integer greater than or equal to 1. The feature extraction and mapping module is used to generate a first feature vector, a second feature vector, and a third feature vector based on the business data, the object data, and the K platform data, wherein the first feature vector corresponds to dense features, the second feature vector corresponds to sparse features, and the third feature vector corresponds to sequence features; An information extraction module is used to take the first feature vector, the second feature vector, and the third feature vector as inputs to an information extraction network, and output a target feature vector through the information extraction network. The information extraction network includes a first sub-network and a second sub-network. The inputs of the first sub-network are the first feature vector and the second feature vector, and the inputs of the second sub-network are the first feature vector, the second feature vector, and the third feature vector. The exposure conversion rate output module is used to take the target feature vector as the input of the first fully connected network and output the exposure conversion rate of the target service for each of the K platforms through the first fully connected network. The feature extraction and mapping module is further configured to take the first feature vector and the second feature vector as inputs to the factorization machine (FM) network, and output the first target feature vector through the factorization machine (FM) network. The first sub-network includes the factorization machine (FM) network, the field factorization machine (FFM) network, and the deep factorization machine (DeepFM) network. The factorization machine (FM) network is used to realize feature crossing. The feature extraction and mapping module is further configured to take the first feature vector, the second feature vector, and the third feature vector as inputs to the Progressive Hierarchical Extraction (PLE) network, and output a second target feature vector through the PLE network. The second sub-network includes the PLE network and the Custom Gate Control (CGC) network. The feature extraction and mapping module is further configured to concatenate the first target feature vector and the second target feature vector to obtain the target feature vector.

11. A computer device, characterized in that, include: Memory, transceivers, processors, and bus systems; The memory is used to store programs; The processor is used to execute programs in the memory, including executing the business conversion rate prediction method as described in any one of claims 1 to 9; The bus system is used to connect the memory and the processor to enable communication between the memory and the processor.

12. A computer-readable storage medium comprising instructions, when executed on a computer, causing the computer to perform a method for predicting business conversion rates as claimed in any one of claims 1 to 9.

13. A computer program product, comprising a computer program, characterized in that, The computer program is executed by a processor using the business conversion rate prediction method as described in any one of claims 1 to 9.

Images