Click rate estimation method and device, storage medium and electronic equipment
By extracting features from user search text and processing it with a click-through rate (CTR) prediction model, semantic, discrete, and statistical feature data are generated, which solves the problem of insufficient accuracy in CTR prediction and improves the accuracy of CTR prediction and ranking of recommended words.
Patent Information
- Application Number
- CN202410985328.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-22
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2044-07-22
AI Technical Summary
In existing technologies, the accuracy of click-through rate (CTR) prediction is insufficient, which affects the effectiveness of search engine keyword recommendations and ad display.
By extracting features from user search text, semantic feature data, discrete feature data, and statistical feature data are generated. Then, a click-through rate (CTR) prediction model is used to predict the CTR of promotional information. Combined with conversion rate prediction, the final CTR is calibrated.
It improved the accuracy of click-through rate prediction for candidate recommended terms and enhanced the accuracy of the recommended order of candidate recommended terms under the semantic meaning of user search text.
Smart Images

Figure CN118982389A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of computer technology, and in particular to a click rate prediction method, device, storage medium and electronic device. Background Art
[0002] In the related technology, search advertising is a common form of online advertising. Search engines can mix search advertising and natural search results for display exposure. As a link between advertising companies and users, search engines can bring personalized experience to users on the one hand, and help advertising companies to expand their brands, promote products and increase sales on the other hand. Click-through rate (CTR) is an important indicator in online advertising systems. By estimating CTR, search engines can adjust the user search experience based on the estimated CTR, while also ensuring that advertising companies have more advertising exposure opportunities. Summary of the invention
[0003] The present application provides a method, device, computer storage medium and electronic device for estimating click rate. The technical solution is as follows:
[0004] In a first aspect, an embodiment of the present application provides a click rate prediction method, the method comprising:
[0005] Acquire a user search text input by a user, and perform feature extraction processing based on the user search text to obtain semantic feature data, discrete feature data, and statistical feature data;
[0006] Based on the semantic feature data, the discrete feature data, and the statistical feature data, a click-through rate prediction model is used to perform click-through rate prediction processing on the promotion information for the candidate recommendation words to obtain the click-through rate of the promotion information corresponding to the candidate recommendation words.
[0007] In some possible implementations, the feature extraction process based on the user search text to obtain semantic feature data, discrete feature data, and statistical feature data includes:
[0008] Performing semantic feature extraction processing based on the user search text to obtain semantic feature data;
[0009] Performing discrete feature query processing based on the semantic feature data to obtain discrete feature data;
[0010] Feature calculation processing is performed based on the semantic feature data to obtain statistical feature data.
[0011] In some possible implementations, the performing semantic feature extraction processing based on the user search text to obtain semantic feature data includes:
[0012] Performing a search word extraction process based on the user search text to obtain a search word, and performing a recommendation word matching process based on the search word to obtain a candidate recommendation word;
[0013] Obtain a set of historical click search terms corresponding to the candidate recommendation terms;
[0014] Determine the target promotion information title corresponding to the candidate recommendation word;
[0015] Generate semantic feature data including the search term, the candidate recommendation term, the historical click search term set, and the target promotion information title.
[0016] In some possible implementations, the performing discrete feature query processing based on the semantic feature data to obtain discrete feature data includes:
[0017] Acquire an information recommendation source mapping relationship, and query the recommendation source corresponding to the candidate recommendation word and the recommendation type corresponding to the candidate recommendation word based on the information recommendation source mapping relationship;
[0018] Obtaining a search application identifier corresponding to the user search text;
[0019] Generate discrete feature data including the recommendation source, the recommendation type, and the search application identifier.
[0020] In some possible implementations, the performing feature calculation processing based on the semantic feature data to obtain statistical feature data includes:
[0021] Performing similarity calculation based on the search term and the candidate recommendation term to obtain similarity data;
[0022] Determine the historical click rate and normalized click count corresponding to the candidate recommendation word within a preset historical period, calculate the product of the normalized click count and the historical click rate, and obtain the normalized click rate data corresponding to the candidate recommendation word;
[0023] Statistical feature data including the similarity data and the normalized click data is generated.
[0024] In some possible implementations, the process of estimating the click rate of promotion information for candidate recommendation words using a click rate estimation model based on the semantic feature data, the discrete feature data, and the statistical feature data to obtain the click rate of promotion information corresponding to the candidate recommendation word includes:
[0025] The semantic feature data, the discrete feature data, and the statistical feature data are processed by a click-through rate prediction model to predict the click-through rate of the promotion information for the candidate recommendation word, so as to obtain the click-through rate of the reference promotion information corresponding to the candidate recommendation word;
[0026] The click rate prediction model is used to perform promotion information conversion rate prediction processing on the semantic feature data and the discrete feature data for the candidate recommendation words to obtain the promotion information conversion rate corresponding to the candidate recommendation words;
[0027] The click rate prediction model is used to perform click rate calibration processing on the reference promotion information click rate and the estimated promotion information conversion rate to obtain the promotion information click rate corresponding to the candidate recommendation word.
[0028] In some possible implementations, the process of performing click-through rate estimation processing on the semantic feature data, the discrete feature data, and the statistical feature data for the candidate recommendation word to obtain the click-through rate of the reference promotion information corresponding to the candidate recommendation word includes:
[0029] Performing semantic cross processing on the semantic feature data to obtain a semantic cross feature vector, and performing embedding processing on the discrete feature data to obtain a discrete feature vector;
[0030] Performing vector concatenation processing based on the semantic cross feature vector and the discrete feature vector to obtain a comprehensive feature vector;
[0031] The click rate prediction model is used to predict the click rate of the promotion information for the candidate recommendation word by processing the comprehensive feature vector and the statistical feature data to obtain the click rate of the reference promotion information corresponding to the candidate recommendation word.
[0032] In some possible implementations, the estimating process of the promotion information conversion rate of the candidate recommendation word on the semantic feature data and the discrete feature data by the click rate estimation model to obtain the promotion information conversion rate corresponding to the candidate recommendation word includes:
[0033] The promotion information conversion rate of the candidate recommendation word is estimated by the comprehensive splicing vector through a click-through rate estimation model to obtain the promotion information conversion rate corresponding to the candidate recommendation word.
[0034] In some possible implementations, the click rate prediction model is used to perform click rate calibration processing on the reference promotion information click rate and the estimated promotion information conversion rate to obtain the promotion information click rate corresponding to the candidate recommendation word, including:
[0035] The product of the reference promotion information click-through rate and the estimated promotion information conversion rate is calculated by the click-through rate prediction model to obtain the promotion information click-through rate corresponding to the candidate recommendation word.
[0036] In some possible implementations, the method further comprises:
[0037] Create an initial click rate prediction model for promotion information click rate prediction scenarios;
[0038] Perform feature extraction processing based on sample user search text to obtain sample semantic feature data, sample discrete feature data, and sample statistical feature data, and annotate the sample semantic feature data, the sample discrete feature data, and the sample statistical feature data with a sample promotion information click-through rate label corresponding to the sample candidate recommendation word;
[0039] The sample semantic feature data, the sample discrete feature data, and the sample statistical feature data are used to perform at least one round of model training on the initial click-through rate prediction model. During the model training process, the sample semantic feature data, the sample discrete feature data, and the sample statistical feature data are used to perform promotion information click-through rate estimation processing for sample candidate recommendation words to obtain the predicted sample promotion information click-through rate corresponding to the sample candidate recommendation words, and a model loss value is calculated based on the sample promotion information click-through rate label and the predicted sample promotion information click-through rate. The model loss value is used to adjust the model parameters of the initial click-through rate prediction model to obtain a click-through rate prediction model after training.
[0040] In some possible implementations, the calculating the model loss value based on the sample promotion information click-through rate label and the predicted sample promotion information click-through rate includes:
[0041] The sample promotion information click rate label and the predicted sample promotion information click rate are calculated using a loss calculation formula to obtain a model loss value;
[0042] The loss calculation formula satisfies the following formula:
[0043]
[0044] L(θ cvr ,θ ctr ) represents the model loss value, l() represents the cross loss function, y i Indicates the click-through rate label of the sample promotion information, f(x i θ ctr ) represents the click-through rate of the reference sample promotion information, f(x i θ ctr ×f(x iθ cvr )) represents the predicted click rate of the sample promotion information, f(x i θ cvr ) represents the predicted sample promotion information conversion rate, represents the logical AND operation result of the sample promotion information click rate label and the sample promotion information conversion rate label, θ ctr and θ cvr Represents the network parameters of the initial click rate prediction model.
[0045] In a second aspect, an embodiment of the present application provides a click rate estimation device, the device comprising:
[0046] A feature extraction module is used to obtain a user search text input by a user, and perform feature extraction processing based on the user search text to obtain semantic feature data, discrete feature data, and statistical feature data;
[0047] The model estimation module is used to use a click-through rate estimation model to perform click-through rate estimation processing on the promotion information of the candidate recommendation words based on the semantic feature data, the discrete feature data, and the statistical feature data to obtain the click-through rate of the promotion information corresponding to the candidate recommendation words.
[0048] Optionally, the feature extraction module includes:
[0049] A first extraction unit, configured to perform semantic feature extraction processing based on the user search text to obtain semantic feature data;
[0050] A second extraction unit, configured to perform discrete feature query processing based on the semantic feature data to obtain discrete feature data;
[0051] The third extraction unit is used to perform feature calculation processing based on the semantic feature data to obtain statistical feature data.
[0052] Optionally, the first extraction unit is specifically used to:
[0053] Performing a search word extraction process based on the user search text to obtain a search word, and performing a recommendation word matching process based on the search word to obtain a candidate recommendation word;
[0054] Obtain a set of historical click search terms corresponding to the candidate recommendation terms;
[0055] Determine the target promotion information title corresponding to the candidate recommendation word;
[0056] Generate semantic feature data including the search term, the candidate recommendation term, the historical click search term set, and the target promotion information title.
[0057] Optionally, the second extraction unit is specifically used to:
[0058] Acquire an information recommendation source mapping relationship, and query the recommendation source corresponding to the candidate recommendation word and the recommendation type corresponding to the candidate recommendation word based on the information recommendation source mapping relationship;
[0059] Obtaining a search application identifier corresponding to the user search text;
[0060] Generate discrete feature data including the recommendation source, the recommendation type, and the search application identifier.
[0061] Optionally, the third extraction unit is specifically used to:
[0062] Performing similarity calculation based on the search term and the candidate recommendation term to obtain similarity data;
[0063] Determine the historical click rate and normalized click count corresponding to the candidate recommendation word within a preset historical period, calculate the product of the normalized click count and the historical click rate, and obtain the normalized click rate data corresponding to the candidate recommendation word;
[0064] Statistical feature data including the similarity data and the normalized click data is generated.
[0065] Optionally, the model estimation module includes a first estimation unit, a second estimation unit and a third estimation unit, which are specifically used to:
[0066] A first estimation unit is used to estimate the click rate of the promotion information for the candidate recommendation word by using a click rate estimation model on the semantic feature data, the discrete feature data, and the statistical feature data to obtain a click rate of the reference promotion information corresponding to the candidate recommendation word;
[0067] A second estimation unit is used to estimate the conversion rate of promotion information for candidate recommendation words on the semantic feature data and the discrete feature data through the click rate estimation model to obtain the conversion rate of promotion information corresponding to the candidate recommendation word;
[0068] The third estimation unit is used to perform click rate calibration processing on the reference promotion information click rate and the estimated promotion information conversion rate through the click rate estimation model to obtain the promotion information click rate corresponding to the candidate recommendation word.
[0069] Optionally, the first estimation unit is specifically configured to:
[0070] Performing semantic cross processing on the semantic feature data to obtain a semantic cross feature vector, and performing embedding processing on the discrete feature data to obtain a discrete feature vector;
[0071] Performing vector concatenation processing based on the semantic cross feature vector and the discrete feature vector to obtain a comprehensive feature vector;
[0072] The click rate prediction model is used to predict the click rate of the promotion information for the candidate recommendation word by processing the comprehensive feature vector and the statistical feature data to obtain the click rate of the reference promotion information corresponding to the candidate recommendation word.
[0073] Optionally, the second estimation unit is specifically configured to:
[0074] The promotion information conversion rate of the candidate recommendation word is estimated by the comprehensive splicing vector through a click-through rate estimation model to obtain the promotion information conversion rate corresponding to the candidate recommendation word.
[0075] Optionally, the third estimation unit is specifically configured to:
[0076] The product of the reference promotion information click-through rate and the estimated promotion information conversion rate is calculated by the click-through rate prediction model to obtain the promotion information click-through rate corresponding to the candidate recommendation word.
[0077] Optionally, the click rate estimation device further includes:
[0078] The first training module is used to create an initial click rate prediction model for the promotion information click rate prediction scenario;
[0079] The second training module is used to perform feature extraction processing based on sample user search text to obtain sample semantic feature data, sample discrete feature data, and sample statistical feature data, and to label the sample semantic feature data, the sample discrete feature data, and the sample statistical feature data with a sample promotion information click-through rate label corresponding to the sample candidate recommendation word;
[0080] The third training module is used to use the sample semantic feature data, the sample discrete feature data, and the sample statistical feature data to perform at least one round of model training on the initial click-through rate prediction model. During the model training process, the initial click-through rate prediction model is used to perform promotion information click-through rate estimation processing on the sample semantic feature data, the sample discrete feature data, and the sample statistical feature data for the sample candidate recommendation words to obtain the predicted sample promotion information click-through rate corresponding to the sample candidate recommendation words, and the model loss value is calculated based on the sample promotion information click-through rate label and the predicted sample promotion information click-through rate. The model loss value is used to adjust the model parameters of the initial click-through rate prediction model to obtain the click-through rate prediction model after training.
[0081] Optionally, the third training module is specifically used to:
[0082] The sample promotion information click rate label and the predicted sample promotion information click rate are calculated using a loss calculation formula to obtain a model loss value;
[0083] The loss calculation formula satisfies the following formula:
[0084]
[0085] L(θ cvr ,θ ctr ) represents the model loss value, l() represents the cross loss function, y i Indicates the click-through rate label of the sample promotion information, f(x i θ ctr ) represents the click-through rate of the reference sample promotion information, f(x i θ ctr ×f(x i θ cvr )) represents the predicted click rate of the sample promotion information, f(x i θ cvr ) represents the predicted sample promotion information conversion rate, represents the logical AND operation result of the sample promotion information click rate label and the sample promotion information conversion rate label, θ ctr and θ cvr Represents the network parameters of the initial click rate prediction model.
[0086] In a third aspect, an embodiment of the present application provides a computer storage medium, wherein the computer storage medium has a plurality of instructions, and the instructions are suitable for being loaded by a processor and executing the above-mentioned method.
[0087] In a fourth aspect, an embodiment of the present application provides an electronic device, which may include: a memory and a processor; wherein the memory stores a computer program, and the computer program is suitable for being loaded by the memory and executing the above method.
[0088] The beneficial effects brought by the technical solution provided by the embodiment of the present application include at least:
[0089] The click-through rate prediction method provided in the embodiment of the present application performs feature extraction processing on the user search text to obtain semantic feature data, discrete feature data, and statistical feature data, and then uses the click-through rate prediction model to perform click-through rate prediction processing on the above features for the promotion information of the candidate recommendation words, thereby obtaining the click-through rate of the promotion information corresponding to the candidate recommendation words. In this way, the use of the trained click-through rate prediction model can improve the prediction accuracy of the click-through rate of the promotion information corresponding to the candidate recommendation words (such as the click-through rate of search ads), and the candidate recommendation words are obtained based on the matching of the user search text, and the click-through rate of the promotion information can determine the recommended ranking order of the corresponding candidate recommendation words among all candidate recommendation words. Therefore, the embodiment of the present application can improve the accuracy of the recommended ranking order of the candidate recommendation words under the semantics of the current user search text. BRIEF DESCRIPTION OF THE DRAWINGS
[0090] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without paying any creative work.
[0091] Figure 1 It is a scenario diagram of a click rate estimation method provided in an embodiment of the present application;
[0092] Figure 2 is a schematic diagram of a model structure of a click rate prediction model provided in an embodiment of the present application;
[0093] Figure 3 is a flow chart of another click rate estimation method provided in an embodiment of the present application;
[0094] Figure 4 is a structural schematic diagram of a click rate estimation device provided in an embodiment of the present application;
[0095] Figure 5 It is a structural diagram of a model estimation module provided in an embodiment of the present application;
[0096] Figure 6 It is a structural schematic diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION
[0097] In order to make the purpose, features, and advantages of the embodiments of the present application more obvious and easy to understand, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of this application.
[0098] In the description of the present application, it should be understood that the terms "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance. In the description of the present application, it should be noted that, unless otherwise clearly specified and limited, "including" and "having" and any of their variations are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but optionally also includes steps or units that are not listed, or optionally also includes other steps or units inherent to these processes, methods, products or devices. For those of ordinary skill in the art, the specific meanings of the above terms in the present application can be understood in specific circumstances. In addition, in the description of the present application, unless otherwise specified, "multiple" refers to two or more. "And / or" describes the association relationship of associated objects, indicating that there can be three relationships, for example, A and / or B, which can represent: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the associated objects before and after are an "or" relationship.
[0099] In related technologies, CTR is an important indicator in online advertising systems. Search engines can adjust recommended keywords, advertising display methods, advertising display positions, advertising display frequency, recommended keywords and other advertising parameters based on the estimated CTR, and ensure that after the adjustment, not only the user's search experience can be improved, but also the advertising company can have more advertising exposure opportunities. Therefore, how to improve the accuracy of the estimated CTR is a technical problem that needs to be solved urgently.
[0100] In order to solve the above technical problems, the present application is described in detail below in conjunction with specific embodiments.
[0101] In one embodiment, Figure 1 As shown, a click rate prediction method is proposed, which can be implemented by a computer program and can be run on a click rate prediction device based on the von Neumann system. The computer program can be integrated into an application or run as an independent tool application.
[0102] Specifically, the click rate prediction method is executed by an electronic device, and the click rate prediction method includes:
[0103] S101, obtaining a user search text input by a user, and performing feature extraction processing based on the user search text to obtain semantic feature data, discrete feature data, and statistical feature data.
[0104] User search text refers to the search text entered by the user in the search scenario. Taking the search scenario in the search engine as an example, the text entered by the user in the search input box is the user search text.
[0105] Semantic feature data may include search terms, candidate recommendation terms, historical click search terms, and target promotion information titles. Semantic feature data can be simply understood as feature data with semantic information.
[0106] Discrete feature data may include recommendation sources, recommendation types, and search application identifiers. Discrete feature data can be understood as feature data with limited values, which can be used to characterize categories or classifications. Specifically, the recommendation source can be a limited number of recommendation sources, the recommendation type can be a limited number of recommendation types, and the search application identifier can be a limited number of application identifiers.
[0107] The statistical feature data may include similarity data and normalized click-through rate data. The statistical feature data may be understood as features that can be represented by numerical values. Specifically, the similarity data may include similarity values, and the normalized click-through rate data may include click-through rate values.
[0108] In some embodiments, a search request sent by a user terminal may be received, and the search request may be parsed to obtain a user search text input by the user. The search request may carry the user input text. Further, semantic feature processing may be performed on the user search text to obtain semantic feature data, and then discrete feature query processing may be performed based on the semantic feature data to obtain discrete feature data, and feature calculation processing may be performed based on the semantic feature data to obtain statistical feature data.
[0109] S102, based on the semantic feature data, the discrete feature data, and the statistical feature data, a click rate prediction model is used to perform click rate prediction processing on the promotion information for the candidate recommendation words to obtain the click rate of the promotion information corresponding to the candidate recommendation words.
[0110] The click-through rate prediction model can be a model obtained by a multi-objective joint modeling method and used to predict the click-through rate of promotional information corresponding to candidate recommendation words. Taking the promotional information as a search advertisement in a search engine as an example, in the search engine, the candidate recommendation words are recommendation words used to trigger the display of search advertisements. The candidate recommendation words can be keywords obtained by performing recommendation processing based on the search words extracted from the user's search text. Therefore, the candidate recommendation words are associated with the search advertisements they trigger. Each candidate recommendation word can correspond to a search advertisement, and the click-through rate of promotional information is the click-through rate of the search advertisement.
[0111] For details, see Figure 2 , Figure 2 A schematic diagram of the model structure of a click rate prediction model provided in an embodiment of the present application. Figure 2 As shown, the click-through rate prediction model is a dual-tower model, which includes a click-through rate prediction tower and a conversion rate prediction tower. By inputting semantic feature data, discrete feature data, and statistical feature data into the click-through rate prediction model, the click-through rate of the reference promotion information corresponding to the candidate recommendation word can be predicted using the click-through rate prediction tower in the click-through rate prediction model. At the same time, the conversion rate prediction tower in the click-through rate prediction model can be used to predict the conversion rate of the promotion information corresponding to the candidate recommendation word. Finally, the prediction results of the two towers are multiplied to obtain the output result of the click-through rate prediction model, that is, the predicted click-through rate of the promotion information corresponding to the candidate recommendation word. Among them, the promotion information conversion rate refers to the proportion of users who click on the promotion information to complete the conversion behavior, and the conversion behavior refers to the user's purchase, registration, download and other behaviors based on the promotion information.
[0112] The click-through rate prediction method provided in the embodiment of the present application performs feature extraction processing on the user search text to obtain semantic feature data, discrete feature data, and statistical feature data, and then uses the click-through rate prediction model to perform click-through rate prediction processing on the above features for the promotion information of the candidate recommendation words, thereby obtaining the click-through rate of the promotion information corresponding to the candidate recommendation words. In this way, the use of the trained click-through rate prediction model can improve the prediction accuracy of the click-through rate of the promotion information corresponding to the candidate recommendation words (such as the predicted click-through rate of the search advertisement), and the candidate recommendation words are obtained based on the matching of the user search text, and the click-through rate of the promotion information can determine the recommended ranking order of the corresponding candidate recommendation words among all candidate recommendation words. Therefore, the embodiment of the present application can improve the accuracy of the recommended ranking order of the candidate recommendation words under the semantics of the current user search text.
[0113] See next Figure 3 , Figure 3 It is a flowchart of another embodiment of a click rate prediction method proposed in the present application.
[0114] Specifically, the click rate prediction method is executed by an electronic device, and the click rate prediction method includes:
[0115] S301, obtaining a user search text input by a user, and performing semantic feature extraction processing based on the user search text to obtain semantic feature data.
[0116] Semantic feature data may include search terms, candidate recommended terms, historical click search terms set, and target promotion information title. Search terms refer to natural search terms extracted from user search texts. Candidate recommended terms refer to keywords that are recommended to users based on the search terms and are used to trigger the display of search ads. Candidate recommended terms may include two types of recommended terms, one type of recommended terms is the keywords that are ultimately recommended to users, and the other type of recommended terms is the keywords that will not be recommended to users. The historical click search terms set refers to the set of specific search terms used by users when clicking on search ads corresponding to candidate recommended terms. The target promotion information title refers to the latest promotion information title corresponding to the candidate recommended terms. Taking the promotion information as a search ad in a search engine as an example, the target promotion information title may refer to the latest ad title corresponding to the candidate recommended terms; in a search engine, a candidate recommended term may be associated with a corresponding search ad, and the search ad may be displayed in an ad display position on a search page. The search ad displayed in the ad display position may change, and the ad title of the search ad may also change, and the target promotion information title is the latest ad title of the search ad.
[0117] In some embodiments, a search request sent by a user terminal may be received and the search request may be parsed to obtain the user search text.
[0118] The step of performing semantic feature extraction processing based on the user search text to obtain semantic feature data may specifically include: performing search word extraction processing based on the user search text to obtain search words, performing recommendation word matching processing based on the search words to obtain candidate recommendation words; obtaining a set of historical click search words corresponding to the candidate recommendation words; determining a target promotion information title corresponding to the candidate recommendation words; and generating semantic feature data including search words, candidate recommendation words, a set of historical click search words, and a target promotion information title.
[0119] Specifically, the search term extraction process may include text preprocessing, word segmentation, keyword extraction, semantic analysis, keyword filtering and sorting, etc., and the user search text may be subjected to the above-mentioned processing to obtain the search term. Text preprocessing may include text cleaning, capitalization unification, and stop word filtering. Word segmentation may refer to dividing the text into words or phrases. Keyword extraction may refer to extracting keywords from the segmented text according to algorithms and rules. Semantic analysis may refer to further verifying and correcting the extracted keywords based on context and semantic understanding. Keyword filtering and sorting may refer to further filtering and sorting the extracted keywords according to the needs of a specific search engine or a specific search application.
[0120] The recommendation word matching process is performed based on the search word to obtain the candidate recommendation word. Specifically, the search word may be input into the recommendation word matching model to perform the recommendation word matching process to obtain the candidate recommendation word.
[0121] Obtaining the historical click search word set corresponding to the candidate recommendation word can specifically be: from the historical interaction data between the user and the search engine, querying the specific search words used by the user when clicking the search advertisement corresponding to the candidate recommendation word, the set of these search words is the historical click search word set.
[0122] Determine the target promotion information title corresponding to the candidate recommendation word, specifically: in the search advertisement database, query the target search advertisement corresponding to the candidate recommendation word, the target search advertisement is set to be triggered and displayed after the user clicks on the candidate recommendation word, query the latest advertisement title of the target search advertisement, and use the latest advertisement title as the target promotion information title.
[0123] Finally, semantic feature data including search terms, candidate recommendation terms, historical click search term sets, and target promotion information titles are generated.
[0124] S302, performing discrete feature query processing based on the semantic feature data to obtain discrete feature data.
[0125] Discrete feature data may include recommendation sources, recommendation types, and search application identifiers. The recommendation source refers to the recommended position of the promotional information corresponding to the candidate recommendation word. The recommendation type refers to the recommended type of the promotional information corresponding to the candidate recommendation word. Taking the promotional information as a search advertisement in a search engine as an example, multiple advertisement display positions can be preset on the search page, each of which can be used to display a search advertisement for advertisement recommendation, and each advertisement display position can be represented by a different text identifier. The recommendation source is the advertisement position where the search advertisement corresponding to the candidate recommendation word is located; the search advertisement may include two types of recommendation types of advertisements, one is a search advertisement determined by an advertisement recommendation system, and the other is a search advertisement that can be determined directly through a search engine without going through an advertisement recommendation system, and the recommendation type is any one of the above two recommendation types. The search application identifier refers to the identifier of the search application used by the user to access the search engine website. For example, for the search engine website b of search engine A, the user can access the search engine website b through a variety of search applications.
[0126] In some embodiments, executing S302 may specifically include: obtaining a mapping relationship between information recommendation sources, querying the recommendation source corresponding to the candidate recommendation word and the recommendation type corresponding to the candidate recommendation word based on the mapping relationship between the information recommendation sources; obtaining a search application identifier corresponding to the user search text; and generating discrete feature data including the recommendation source, recommendation type, and search application identifier.
[0127] The information recommendation source mapping relationship may include a corresponding relationship between at least one reference candidate recommendation word, a reference recommendation source corresponding to the reference candidate recommendation word, and a reference recommendation type corresponding to the reference candidate recommendation word. The reference recommendation source refers to the recommended position of the promotion information corresponding to the reference candidate recommendation word, and the reference recommendation type refers to the recommended type of the promotion information corresponding to the reference candidate recommendation word.
[0128] The information recommendation source mapping relationship can be pre-configured and stored in a specified file, so that the information recommendation source mapping relationship can be directly obtained from the specified file later.
[0129] Specifically, in the information recommendation source mapping relationship, you can query when the reference candidate recommendation word is a candidate recommendation word, its corresponding reference recommendation source and its corresponding reference recommendation type. The reference recommendation source is the recommendation source corresponding to the candidate recommendation word, and the reference recommendation type is the recommendation type corresponding to the candidate recommendation word.
[0130] The search application identifier corresponding to the user search text is obtained by parsing the search request sent by the user terminal to obtain the search application identifier. The search request may carry the search application identifier of the search engine website visited by the user.
[0131] Furthermore, discrete feature data including recommendation source, recommendation type, and search application identifier is generated.
[0132] S303: Perform feature calculation processing based on the semantic feature data to obtain statistical feature data.
[0133] The statistical feature data may include similarity data and normalized click-through rate data. The similarity data includes the similarity between the search term and the candidate recommendation term. The normalized click-through rate data refers to the normalized click-through rate data obtained by normalizing the historical click-through rates corresponding to the candidate recommendation terms.
[0134] In some embodiments, executing S303 may specifically include: performing similarity calculation based on the search term and the candidate recommendation term to obtain similarity data; determining the historical click-through rate and normalized click count corresponding to the candidate recommendation term within a preset historical period, calculating the product of the normalized click count and the historical click-through rate, and obtaining the normalized click-through rate data corresponding to the candidate recommendation term; generating statistical feature data including similarity data and normalized click data.
[0135] Specifically, the cosine similarity between the search term and each candidate recommendation term may be calculated respectively to obtain similarity data.
[0136] Among them, the preset historical period can be the most recent preset number of days. The historical click-through rate refers to the click-through rate of the candidate recommendation words within the preset historical period. The normalized number of clicks is the ratio of the number of clicks of the candidate recommendation words within the preset historical period to the reference number of clicks, and the reference number of clicks is the difference between the maximum number of clicks and the minimum number of clicks among the clicks of all candidate recommendation words within the preset historical period. It is necessary to calculate the normalized number of clicks because the normalized click-through rate data corresponding to the recommendation words with a large number of historical clicks will approach the true posterior click-through rate after normalization, and the normalized click-through rate data corresponding to the recommendation words with a small number of historical clicks will approach 0 after normalization. Only the candidate recommendation words with high exposure and high click counts are statistically significant. Therefore, first calculating the normalized number of clicks, and then calculating the normalized click-through rate data through the normalized number of clicks can screen out candidate recommendation words with high exposure and high click counts.
[0137] After obtaining the above similarity data and normalized click rate data, statistical feature data including the similarity data and the normalized click rate data may be generated.
[0138] S304, using a click rate prediction model to perform click rate prediction processing on the semantic feature data, discrete feature data, and statistical feature data for the candidate recommendation words, and obtain the click rate of the reference promotion information corresponding to the candidate recommendation words.
[0139] In some embodiments, executing S304 may specifically include: performing semantic cross processing on semantic feature data to obtain a semantic cross feature vector; performing embedding processing on discrete feature data to obtain a discrete feature vector; performing vector splicing processing based on the semantic cross feature vector and the discrete feature vector to obtain a comprehensive feature vector; and performing click rate prediction processing on the comprehensive feature vector and the statistical feature data for the candidate recommendation words through a click rate prediction model to obtain the click rate of the reference promotion information corresponding to the candidate recommendation words.
[0140] The semantic feature data is subjected to semantic cross processing to obtain a semantic cross feature vector, which can be specifically: the semantic feature data is mapped into a target feature vector through a pre-trained language model, the target feature vector is a continuous high-dimensional feature vector, and the feature vector is subjected to feature compression processing and feature cross combination processing using a fully connected layer to obtain a semantic cross feature vector. Semantic feature data include multiple features such as search terms, candidate recommendation terms, historical click search term sets, and target promotion information titles. Pre-trained language models, such as Bidirectional Encoder Representations from Transformers (BERT), Generative Pre-trained Transformer (GPT), and other models, can be used to map search terms into a first feature vector, map candidate recommendation terms into a second feature vector, map historical click search term sets into a third feature vector, and map target promotion information titles into a fourth feature vector to obtain a target feature vector including the first feature vector, the second feature vector, the third feature vector, and the fourth feature vector; further, a fully connected layer is used to perform feature compression processing and feature cross-combination processing on the first feature vector, the second feature vector, the third feature vector, and the fourth feature vector to obtain a semantic cross-feature vector.
[0141] The discrete feature data are embedded to obtain discrete feature vectors, specifically: the recommendation source is word embedded to obtain the fifth feature vector, the recommendation type is word embedded to obtain the sixth feature vector, the search application identifier is word embedded to obtain the seventh feature vector, and a discrete feature vector including the fifth feature vector, the sixth feature vector, and the seventh feature vector is obtained.
[0142] The comprehensive feature vector may include a semantic cross-feature vector and a discrete feature vector. For example, the comprehensive feature vector is an m-dimensional vector, the values of the 1st to nth dimensions in the comprehensive feature vector are the values of the semantic cross-feature vector, and the values of the (n+1)th to mth dimensions in the comprehensive feature vector are the values of the discrete feature vector, m is greater than n, and both m and n are positive integers.
[0143] The CTR prediction model is used to process the comprehensive feature vector and statistical feature data for the candidate recommendation words to obtain the CTR of the reference promotion information corresponding to the candidate recommendation words. Specifically, the comprehensive feature vector is subjected to feature compression processing to obtain a compressed feature vector, and the compressed feature vector and statistical feature data are processed for the candidate recommendation words through the CTR prediction network in the CTR prediction model to obtain the CTR of the reference promotion information corresponding to the candidate recommendation words. The CTR prediction model may include a CTR prediction network and a conversion rate prediction network. The CTR prediction network is used to predict the click rate, and the conversion rate prediction network is used to predict the conversion rate.
[0144] S305, using a click rate prediction model to perform conversion rate prediction processing on the semantic feature data and the discrete feature data for the candidate recommendation words, and obtain the promotion information conversion rate corresponding to the candidate recommendation words.
[0145] In some embodiments, the conversion rate of the candidate recommendation words is estimated by the click-through rate estimation model on the comprehensive concatenated vector to obtain the promotion information conversion rate corresponding to the candidate recommendation words. Specifically, the comprehensive feature vector is subjected to feature compression to obtain a compressed feature vector, and the conversion rate estimation network in the click-through rate estimation model is used to estimate the conversion rate of the candidate recommendation words on the compressed feature vector to obtain the promotion information conversion rate corresponding to the candidate recommendation words.
[0146] S306, performing click-through rate calibration processing on the reference promotion information click-through rate and the promotion information conversion rate through a click-through rate prediction model to obtain the promotion information click-through rate corresponding to the candidate recommendation word.
[0147] In some embodiments, the product of the reference promotion information click rate and the promotion information conversion rate is calculated by the click rate prediction model to obtain the promotion information click rate corresponding to the candidate recommendation word. Specifically, in the click rate prediction model, the reference promotion information click rate and the promotion information conversion rate can be multiplied to obtain the promotion information click rate corresponding to the candidate recommendation word.
[0148] In some other embodiments, before executing step S301, the following steps can also be executed: creating an initial click-through rate prediction model for the promotion information click-through rate prediction scenario; performing feature extraction processing based on sample user search text to obtain sample semantic feature data, sample discrete feature data, and sample statistical feature data, and marking the sample semantic feature data, sample discrete feature data, and sample statistical feature data with sample promotion information click-through rate labels corresponding to sample candidate recommendation words; using the sample semantic feature data, sample discrete feature data, and sample statistical feature data to perform at least one round of model training on the initial click-through rate prediction model, and during the model training process, using the initial click-through rate prediction model to perform click-through rate prediction processing on the sample semantic feature data, sample discrete feature data, and sample statistical feature data for sample candidate recommendation words to obtain the predicted sample promotion information click-through rate corresponding to the sample candidate recommendation words, and calculating the model loss value based on the sample promotion information click-through rate label and the predicted sample promotion information click-through rate, and using the model loss value to perform model parameter adjustment processing on the initial click-through rate prediction model to obtain the click-through rate prediction model after training.
[0149] Among them, the initial click rate prediction model is a double-tower model, and the initial click rate prediction model includes a click rate prediction tower and a conversion rate prediction tower.
[0150] Based on the sample user search text, feature extraction processing is performed to obtain sample semantic feature data, sample discrete feature data, and sample statistical feature data. Specifically, it can be: based on the sample user search text, search word extraction processing is performed to obtain sample search words, based on the sample search words, recommendation word matching processing is performed to obtain sample candidate recommendation words, and a sample historical click search word set corresponding to the sample candidate recommendation words is obtained; the sample promotion information title corresponding to the sample candidate recommendation word is determined, and sample semantic feature data including the sample search word, the sample candidate recommendation word, the sample historical click search word set, and the sample promotion information title is generated; the sample recommendation source and sample recommendation type corresponding to the sample candidate recommendation word are obtained, and the sample search application identifier corresponding to the sample user search text is obtained, and sample discrete feature data including the sample recommendation source, the sample recommendation type, and the sample search application identifier are generated; based on the sample search word and the sample candidate recommendation word, similarity calculation processing is performed to obtain sample similarity data, the sample historical click rate and the sample normalized click number corresponding to the sample candidate recommendation word within a preset historical period are determined, and the product of the sample normalized click number and the sample historical click rate is calculated to obtain the sample normalized click rate data corresponding to the sample candidate recommendation word, and sample statistical feature data including the sample similarity data and the sample normalized click data are generated.
[0151] The initial click-through rate estimation model is used to perform click-through rate estimation processing on the sample semantic feature data, the sample discrete feature data, and the sample statistical feature data for the sample candidate recommendation words to obtain the predicted sample promotion information click-through rate corresponding to the sample candidate recommendation words. Specifically, it can be: the click-through rate estimation tower in the initial click-through rate estimation model is used to perform click-through rate estimation processing on the sample semantic feature data, the sample discrete feature data, and the sample statistical feature data for the sample candidate recommendation words to obtain the reference sample promotion information click-through rate corresponding to the sample candidate recommendation words; the conversion rate estimation tower in the initial click-through rate estimation model is used to perform promotion information conversion rate estimation processing on the sample semantic feature data and the sample discrete feature data for the sample candidate recommendation words to obtain the predicted sample promotion information conversion rate corresponding to the sample candidate recommendation words; the initial click-through rate estimation model is used to perform click-through rate calibration processing on the reference sample promotion information click-through rate and the predicted sample promotion information conversion rate to obtain the predicted promotion information click-through rate corresponding to the sample candidate recommendation words.
[0152] The model loss value is calculated based on the sample promotion information click rate label and the predicted sample promotion information click rate. Specifically, the loss calculation formula is used to calculate the sample promotion information click rate label and the predicted sample promotion information click rate to obtain the model loss value. The loss calculation formula satisfies the following formula:
[0153]
[0154] L(θ cvr ,θ ctr ) represents the model loss value, l() represents the cross loss function, y i represents the click-through rate label of the sample promotion information, f(x i θ ctr ) represents the click-through rate of the reference sample promotion information, f(x i θ ctr ×f(x i θ cvr )) represents the predicted click-through rate of sample promotion information, f(x i θ cvr ) represents the predicted sample promotion information conversion rate, represents the logical AND operation result of the sample promotion information click rate label and the sample promotion information conversion rate label, θ ctr and θ cvr Represents the network parameters of the initial CTR prediction model.
[0155] It can be understood that the sample promotion information click rate label includes the actual click rate of the promotion information corresponding to the sample candidate recommendation word; the sample promotion information conversion rate label includes the actual conversion rate of the promotion information corresponding to the sample candidate recommendation word.
[0156] The model loss value is used to adjust the model parameters of the initial click rate prediction model to obtain the click rate prediction model after training. The click rate prediction model after training needs to meet the model training end conditions, which may include, for example, the value of the loss function is less than or equal to the preset loss function threshold, the number of iterations reaches the preset number threshold, etc. The specific model training end conditions can be determined based on the actual situation and are not specifically limited here.
[0157] The click-through rate prediction method provided in the embodiment of the present application obtains semantic feature data by performing semantic feature extraction processing on user search text, obtains discrete feature data by performing discrete feature query processing on the semantic feature data, obtains statistical feature data by performing feature calculation processing on the semantic feature data, and then performs click-through rate prediction processing on the semantic feature data, discrete feature data, and statistical feature data for candidate recommendation words through a click-through rate prediction model to obtain the click-through rate of reference promotion information corresponding to the candidate recommendation words, and then performs conversion rate prediction processing on the semantic feature data and discrete feature data for the candidate recommendation words through the click-through rate prediction model to obtain the promotion information conversion rate corresponding to the candidate recommendation words, and then performs click-through rate calibration processing on the reference promotion information click-through rate and the promotion information conversion rate through the click-through rate prediction model to obtain the promotion information click-through rate corresponding to the candidate recommendation words. In this way, the embodiment of the present application utilizes the natural law that promotional information is first exposed, then clicked, and then converted, and combines the promotional information conversion rate (such as the search ad conversion rate) to estimate and calibrate the promotional information click-through rate (such as the search ad click-through rate). Compared with the traditional click-through rate estimation task, which overestimates the click-through rate due to the influence of noise data (such as misclicks), the embodiment of the present application uses the promotional information conversion rate to calibrate the estimated promotional information click-through rate, solve the problem of overestimating the promotional information click-through rate in related technologies, and thus obtain a more accurate promotional information click-through rate. In addition, the promotional information click-through rate can determine the recommended ranking order of its corresponding candidate recommendation words among all candidate recommendation words. Therefore, the embodiment of the present application can improve the accuracy of the recommended ranking order of candidate recommendation words under the semantics of the current user search text.
[0158] The following will be combined Figure 4 , the click rate prediction device provided in the embodiment of the present application is introduced in detail. It should be noted that, Figure 4 The click rate estimation device shown is used to execute this application Figure 1 to Figure 3 For the convenience of explanation, only the part related to the embodiment of the present application is shown. For the specific technical details not disclosed, please refer to the present application. Figure 1 to Figure 3 The embodiment shown.
[0159] See also Figure 4, which shows a schematic diagram of the structure of the click rate prediction device of an embodiment of the present application. The click rate prediction device 1 can be implemented as all or part of the device through software, hardware or a combination of both. According to some embodiments, the click rate prediction device 1 includes a feature extraction module 11 and a model prediction module 12, which are specifically used to:
[0160] A feature extraction module 11 is used to obtain a user search text input by a user, and perform feature extraction processing based on the user search text to obtain semantic feature data, discrete feature data, and statistical feature data;
[0161] The model estimation module 12 is used to estimate the click rate of the promotion information for the candidate recommendation words by using a click rate estimation model based on the semantic feature data, the discrete feature data and the statistical feature data, so as to obtain the click rate of the promotion information corresponding to the candidate recommendation words.
[0162] Optionally, the feature extraction module 11 includes:
[0163] A first extraction unit, configured to perform semantic feature extraction processing based on the user search text to obtain semantic feature data;
[0164] A second extraction unit, configured to perform discrete feature query processing based on the semantic feature data to obtain discrete feature data;
[0165] The third extraction unit is used to perform feature calculation processing based on the semantic feature data to obtain statistical feature data.
[0166] Optionally, the first extraction unit is specifically used to:
[0167] Performing a search word extraction process based on the user search text to obtain a search word, and performing a recommendation word matching process based on the search word to obtain a candidate recommendation word;
[0168] Obtain a set of historical click search terms corresponding to the candidate recommendation terms;
[0169] Determine the target promotion information title corresponding to the candidate recommendation word;
[0170] Generate semantic feature data including the search term, the candidate recommendation term, the historical click search term set, and the target promotion information title.
[0171] Optionally, the second extraction unit is specifically used to:
[0172] Acquire an information recommendation source mapping relationship, and query the recommendation source corresponding to the candidate recommendation word and the recommendation type corresponding to the candidate recommendation word based on the information recommendation source mapping relationship;
[0173] Obtaining a search application identifier corresponding to the user search text;
[0174] Generate discrete feature data including the recommendation source, the recommendation type, and the search application identifier.
[0175] Optionally, the third extraction unit is specifically used to:
[0176] Performing similarity calculation based on the search term and the candidate recommendation term to obtain similarity data;
[0177] Determine the historical click rate and normalized click count corresponding to the candidate recommendation word within a preset historical period, calculate the product of the normalized click count and the historical click rate, and obtain the normalized click rate data corresponding to the candidate recommendation word;
[0178] Statistical feature data including the similarity data and the normalized click data is generated.
[0179] Optionally, see Figure 5 , is a schematic diagram of the structure of a model estimation module provided in an embodiment of the present application. Figure 5 As shown, the model estimation module 12 includes a first estimation unit 121, a second estimation unit 122 and a third estimation unit 123, which are specifically used for:
[0180] The first estimation unit 121 is used to estimate the click rate of the promotion information for the candidate recommendation word by using a click rate estimation model on the semantic feature data, the discrete feature data, and the statistical feature data to obtain a click rate of the reference promotion information corresponding to the candidate recommendation word;
[0181] The second estimation unit 122 is used to estimate the promotion information conversion rate of the candidate recommendation word on the semantic feature data and the discrete feature data through the click rate estimation model to obtain the promotion information conversion rate corresponding to the candidate recommendation word;
[0182] The third estimation unit 123 is used to perform click rate calibration processing on the reference promotion information click rate and the estimated promotion information conversion rate through the click rate estimation model to obtain the promotion information click rate corresponding to the candidate recommendation word.
[0183] Optionally, the first estimating unit 121 is specifically configured to:
[0184] Performing semantic cross processing on the semantic feature data to obtain a semantic cross feature vector, and performing embedding processing on the discrete feature data to obtain a discrete feature vector;
[0185] Performing vector concatenation processing based on the semantic cross feature vector and the discrete feature vector to obtain a comprehensive feature vector;
[0186] The click rate prediction model is used to predict the click rate of the promotion information for the candidate recommendation word by processing the comprehensive feature vector and the statistical feature data to obtain the click rate of the reference promotion information corresponding to the candidate recommendation word.
[0187] Optionally, the second estimating unit 122 is specifically configured to:
[0188] The promotion information conversion rate of the candidate recommendation word is estimated by the comprehensive splicing vector through a click-through rate estimation model to obtain the promotion information conversion rate corresponding to the candidate recommendation word.
[0189] Optionally, the third estimating unit 123 is specifically configured to:
[0190] The product of the reference promotion information click-through rate and the estimated promotion information conversion rate is calculated by the click-through rate prediction model to obtain the promotion information click-through rate corresponding to the candidate recommendation word.
[0191] Optionally, the click rate estimation device 1 further includes:
[0192] The first training module is used to create an initial click rate prediction model for the promotion information click rate prediction scenario;
[0193] The second training module is used to perform feature extraction processing based on sample user search text to obtain sample semantic feature data, sample discrete feature data, and sample statistical feature data, and to label the sample semantic feature data, the sample discrete feature data, and the sample statistical feature data with a sample promotion information click-through rate label corresponding to the sample candidate recommendation word;
[0194] The third training module is used to use the sample semantic feature data, the sample discrete feature data, and the sample statistical feature data to perform at least one round of model training on the initial click-through rate prediction model. During the model training process, the initial click-through rate prediction model is used to perform promotion information click-through rate estimation processing on the sample semantic feature data, the sample discrete feature data, and the sample statistical feature data for the sample candidate recommendation words to obtain the predicted sample promotion information click-through rate corresponding to the sample candidate recommendation words, and the model loss value is calculated based on the sample promotion information click-through rate label and the predicted sample promotion information click-through rate. The model loss value is used to adjust the model parameters of the initial click-through rate prediction model to obtain the click-through rate prediction model after training.
[0195] Optionally, the third training module is specifically used to:
[0196] The sample promotion information click rate label and the predicted sample promotion information click rate are calculated using a loss calculation formula to obtain a model loss value;
[0197] The loss calculation formula satisfies the following formula:
[0198]
[0199] L(θ cvr ,θ ctr ) represents the model loss value, l() represents the cross loss function, y i Indicates the click-through rate label of the sample promotion information, f(x i θ ctr ) represents the click-through rate of the reference sample promotion information, f(x i θ ctr ×f(x i θ cvr )) represents the predicted click rate of the sample promotion information, f(x i θ cvr ) represents the predicted sample promotion information conversion rate, represents the logical AND operation result of the sample promotion information click rate label and the sample promotion information conversion rate label, θ ctr and θ cvr Represents the network parameters of the initial click rate prediction model.
[0200] Please refer to Figure 6 , Figure 6 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. Exemplarily, the electronic device in the embodiment of the present application may be a server, and the electronic device may include one or more of the following components: a processor 110, a memory 120, an input device 130, an output device 140, and a bus 150. The processor 110, the memory 120, the input device 130, and the output device 140 may be connected via the bus 150.
[0201] The processor 110 may include one or more processing cores. The processor 110 uses various interfaces and lines to connect various parts of the entire electronic device, and executes various functions of the electronic device and processes data by running or executing instructions, programs, code sets or instruction sets stored in the memory 120, and calling data stored in the memory 120. Optionally, the processor 110 can be implemented in at least one hardware form of digital signal processing (DSP), field-programmable gate array (FPGA), and programmable logic array (PLA). The processor 110 can integrate one or a combination of a central processing unit (CPU), a graphics processing unit (GPU), and a modem. Among them, the CPU mainly processes the operating system, user interface, and application programs; the GPU is responsible for rendering and drawing display content; and the modem is used to process wireless communications. It can be understood that the above-mentioned modem may not be integrated into the processor 110, but may be implemented separately through a communication chip.
[0202] The memory 120 may include a random access memory (RAM) or a read-only memory (ROM). Optionally, the memory 120 includes a non-transitory computer-readable storage medium. The memory 120 may be used to store instructions, programs, codes, code sets or instruction sets. The memory 120 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for implementing at least one function (e.g., a touch function, a sound playback function, an image playback function, etc.), instructions for implementing the following various method embodiments, etc. The operating system may be an Android system, including a system deeply developed based on the Android system, an IOS system developed by Apple, including a system deeply developed based on the IOS system or other systems.
[0203] In order for the operating system to distinguish the specific application scenarios of third-party applications, it is necessary to open up data communication between third-party applications and the operating system so that the operating system can obtain the current scenario information of third-party applications at any time, and then perform targeted system resource adaptation based on the current scenario.
[0204] The input device 130 is used to receive input commands or data, and includes but is not limited to a keyboard, a mouse, a camera, a microphone, or a touch device. The output device 140 is used to output commands or data, and includes but is not limited to a display device and a speaker. In one example, the input device 130 and the output device 140 can be combined, and the input device 130 and the output device 140 are touch screen displays.
[0205] The touch display screen can be designed as a full screen, a curved screen or a special-shaped screen. The touch display screen can also be designed as a combination of a full screen and a curved screen, or a combination of a special-shaped screen and a curved screen, which is not limited in the embodiments of the present application.
[0206] In addition, those skilled in the art will appreciate that the structure of the electronic device shown in the above drawings does not constitute a limitation on the electronic device, and the electronic device may include more or fewer components than shown, or combine certain components, or arrange the components differently. For example, the electronic device also includes a radio frequency circuit, an input unit, a sensor, an audio circuit, a wireless fidelity (WiFi) module, a power supply, a Bluetooth module and other components, which will not be described in detail here.
[0207] exist Figure 6 In the electronic device shown, the processor 110 can be used to call the program of the click rate prediction method stored in the memory 120, and specifically perform the following operations:
[0208] Acquire a user search text input by a user, and perform feature extraction processing based on the user search text to obtain semantic feature data, discrete feature data, and statistical feature data;
[0209] Based on the semantic feature data, the discrete feature data, and the statistical feature data, a click-through rate prediction model is used to perform click-through rate prediction processing on the promotion information for the candidate recommendation words to obtain the click-through rate of the promotion information corresponding to the candidate recommendation words.
[0210] In one embodiment, when the processor 110 performs the step of performing feature extraction processing based on the user search text to obtain semantic feature data, discrete feature data, and statistical feature data, the following operations are specifically performed:
[0211] Performing semantic feature extraction processing based on the user search text to obtain semantic feature data;
[0212] Performing discrete feature query processing based on the semantic feature data to obtain discrete feature data;
[0213] Feature calculation processing is performed based on the semantic feature data to obtain statistical feature data.
[0214] In one embodiment, when the processor 110 performs the step of performing semantic feature extraction processing based on the user search text to obtain semantic feature data, the processor 110 specifically performs the following operations:
[0215] Performing a search word extraction process based on the user search text to obtain a search word, and performing a recommendation word matching process based on the search word to obtain a candidate recommendation word;
[0216] Obtain a set of historical click search terms corresponding to the candidate recommendation terms;
[0217] Determine the target promotion information title corresponding to the candidate recommendation word;
[0218] Generate semantic feature data including the search term, the candidate recommendation term, the historical click search term set, and the target promotion information title.
[0219] In one embodiment, when the processor 110 executes the step of performing discrete feature query processing based on the semantic feature data to obtain discrete feature data, the processor 110 specifically performs the following operations:
[0220] Acquire an information recommendation source mapping relationship, and query the recommendation source corresponding to the candidate recommendation word and the recommendation type corresponding to the candidate recommendation word based on the information recommendation source mapping relationship;
[0221] Obtaining a search application identifier corresponding to the user search text;
[0222] Generate discrete feature data including the recommendation source, the recommendation type, and the search application identifier.
[0223] In one embodiment, when the processor 110 performs the step of performing feature calculation processing based on the semantic feature data to obtain statistical feature data, the processor 110 specifically performs the following operations:
[0224] Performing similarity calculation based on the search term and the candidate recommendation term to obtain similarity data;
[0225] Determine the historical click rate and normalized click count corresponding to the candidate recommendation word within a preset historical period, calculate the product of the normalized click count and the historical click rate, and obtain the normalized click rate data corresponding to the candidate recommendation word;
[0226] Statistical feature data including the similarity data and the normalized click data is generated.
[0227] In one embodiment, when the processor 110 performs the step of estimating the click rate of the promotion information for the candidate recommendation word using the click rate estimation model based on the semantic feature data, the discrete feature data, and the statistical feature data to obtain the click rate of the promotion information corresponding to the candidate recommendation word, the following operations are specifically performed:
[0228] The semantic feature data, the discrete feature data, and the statistical feature data are processed by a click-through rate prediction model to predict the click-through rate of the promotion information for the candidate recommendation word, so as to obtain the click-through rate of the reference promotion information corresponding to the candidate recommendation word;
[0229] The click rate prediction model is used to perform promotion information conversion rate prediction processing on the semantic feature data and the discrete feature data for the candidate recommendation words to obtain the promotion information conversion rate corresponding to the candidate recommendation words;
[0230] The click rate prediction model is used to perform click rate calibration processing on the reference promotion information click rate and the estimated promotion information conversion rate to obtain the promotion information click rate corresponding to the candidate recommendation word.
[0231] In one embodiment, when the processor 110 performs the step of performing click-through rate estimation processing on the semantic feature data, the discrete feature data, and the statistical feature data for the candidate recommendation word through the click-through rate estimation model to obtain the click-through rate of the reference promotion information corresponding to the candidate recommendation word, the following operations are specifically performed:
[0232] Performing semantic cross processing on the semantic feature data to obtain a semantic cross feature vector, and performing embedding processing on the discrete feature data to obtain a discrete feature vector;
[0233] Performing vector concatenation processing based on the semantic cross feature vector and the discrete feature vector to obtain a comprehensive feature vector;
[0234] The click rate prediction model is used to predict the click rate of the promotion information for the candidate recommendation word by processing the comprehensive feature vector and the statistical feature data to obtain the click rate of the reference promotion information corresponding to the candidate recommendation word.
[0235] In one embodiment, when the processor 110 performs the step of performing the promotion information conversion rate estimation processing on the semantic feature data and the discrete feature data for the candidate recommendation word through the click rate estimation model to obtain the promotion information conversion rate corresponding to the candidate recommendation word, the following operations are specifically performed:
[0236] The promotion information conversion rate of the candidate recommendation word is estimated by the comprehensive splicing vector through a click-through rate estimation model to obtain the promotion information conversion rate corresponding to the candidate recommendation word.
[0237] In one embodiment, when the processor 110 performs the step of performing click rate calibration processing on the reference promotion information click rate and the estimated promotion information conversion rate through the click rate prediction model to obtain the promotion information click rate corresponding to the candidate recommendation word, the processor 110 specifically performs the following operations:
[0238] The product of the reference promotion information click-through rate and the estimated promotion information conversion rate is calculated by the click-through rate prediction model to obtain the promotion information click-through rate corresponding to the candidate recommendation word.
[0239] In one embodiment, the processor 110 further performs the following operations:
[0240] Create an initial click rate prediction model for promotion information click rate prediction scenarios;
[0241] Perform feature extraction processing based on sample user search text to obtain sample semantic feature data, sample discrete feature data, and sample statistical feature data, and annotate the sample semantic feature data, the sample discrete feature data, and the sample statistical feature data with a sample promotion information click-through rate label corresponding to the sample candidate recommendation word;
[0242] The sample semantic feature data, the sample discrete feature data, and the sample statistical feature data are used to perform at least one round of model training on the initial click-through rate prediction model. During the model training process, the sample semantic feature data, the sample discrete feature data, and the sample statistical feature data are used to perform promotion information click-through rate estimation processing for sample candidate recommendation words to obtain the predicted sample promotion information click-through rate corresponding to the sample candidate recommendation words, and a model loss value is calculated based on the sample promotion information click-through rate label and the predicted sample promotion information click-through rate. The model loss value is used to adjust the model parameters of the initial click-through rate prediction model to obtain a click-through rate prediction model after training.
[0243] In one embodiment, when the processor 110 executes the step of calculating the model loss value based on the sample promotion information click-through rate label and the predicted sample promotion information click-through rate, the processor 110 specifically performs the following operations:
[0244] The sample promotion information click rate label and the predicted sample promotion information click rate are calculated using a loss calculation formula to obtain a model loss value;
[0245] The loss calculation formula satisfies the following formula:
[0246]
[0247] L(θ cvr,θ ctr ) represents the model loss value, l() represents the cross loss function, y i Indicates the click-through rate label of the sample promotion information, f(x i θ ctr ) represents the click-through rate of the reference sample promotion information, f(x i θ ctr ×f(x i θ cvr )) represents the predicted click rate of the sample promotion information, f(x i θ cvr ) represents the predicted sample promotion information conversion rate, represents the logical AND operation result of the sample promotion information click rate label and the sample promotion information conversion rate label, θ ctr and θ cvr Represents the network parameters of the initial click rate prediction model.
[0248] An embodiment of the present application further provides a computer-readable storage medium, which stores at least one instruction, and the at least one instruction is used to be executed by a processor to implement the click rate estimation method as described in the above embodiments.
[0249] An embodiment of the present application further provides a computer program product, which stores at least one instruction, and the at least one instruction is loaded and executed by the processor to implement the click rate prediction method described in the above embodiments.
[0250] Those skilled in the art should be aware that in one or more of the above examples, the functions described in the embodiments of the present application can be implemented with hardware, software, firmware, or any combination thereof. When implemented using software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or codes on a computer-readable medium. Computer-readable media include computer storage media and communication media, wherein the communication media include any media that facilitates the transmission of a computer program from one place to another. The storage medium can be any available medium that a general or special-purpose computer can access.
[0251] The above description is only an optional embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.
Claims
1. A method for estimating click rate. It is characterized in that The method comprises: Acquire a user search text input by a user, and perform feature extraction processing based on the user search text to obtain semantic feature data, discrete feature data, and statistical feature data; Based on the semantic feature data, the discrete feature data, and the statistical feature data, a click-through rate prediction model is used to perform click-through rate prediction processing on the promotion information for the candidate recommendation words to obtain the click-through rate of the promotion information corresponding to the candidate recommendation words.
2. The method according to claim 1, It is characterized in that The feature extraction process based on the user search text to obtain semantic feature data, discrete feature data, and statistical feature data includes: Performing semantic feature extraction processing based on the user search text to obtain semantic feature data; Performing discrete feature query processing based on the semantic feature data to obtain discrete feature data; Feature calculation processing is performed based on the semantic feature data to obtain statistical feature data.
3. The method according to claim 2, It is characterized in that The step of performing semantic feature extraction processing based on the user search text to obtain semantic feature data includes: Performing a search word extraction process based on the user search text to obtain a search word, and performing a recommendation word matching process based on the search word to obtain a candidate recommendation word; Obtain a set of historical click search terms corresponding to the candidate recommendation terms; Determine the target promotion information title corresponding to the candidate recommendation word; Generate semantic feature data including the search term, the candidate recommendation term, the historical click search term set, and the target promotion information title.
4. The method according to claim 3, It is characterized in that The performing discrete feature query processing based on the semantic feature data to obtain discrete feature data includes: Acquire an information recommendation source mapping relationship, and query the recommendation source corresponding to the candidate recommendation word and the recommendation type corresponding to the candidate recommendation word based on the information recommendation source mapping relationship; Obtaining a search application identifier corresponding to the user search text; Generate discrete feature data including the recommendation source, the recommendation type, and the search application identifier.
5. The method according to claim 3, It is characterized in that The step of performing feature calculation based on the semantic feature data to obtain statistical feature data includes: Performing similarity calculation based on the search term and the candidate recommendation term to obtain similarity data; Determine the historical click rate and normalized click count corresponding to the candidate recommendation word within a preset historical period, calculate the product of the normalized click count and the historical click rate, and obtain the normalized click rate data corresponding to the candidate recommendation word; Statistical feature data including the similarity data and the normalized click data is generated.
6. The method according to claim 1, It is characterized in that The process of using a click rate prediction model to perform click rate prediction processing on the promotion information for the candidate recommendation word based on the semantic feature data, the discrete feature data, and the statistical feature data to obtain the click rate of the promotion information corresponding to the candidate recommendation word includes: The semantic feature data, the discrete feature data, and the statistical feature data are processed by a click-through rate prediction model to predict the click-through rate of the promotion information for the candidate recommendation word, so as to obtain the click-through rate of the reference promotion information corresponding to the candidate recommendation word; The click rate prediction model is used to perform promotion information conversion rate prediction processing on the semantic feature data and the discrete feature data for the candidate recommendation words to obtain the promotion information conversion rate corresponding to the candidate recommendation words; The click rate prediction model is used to perform click rate calibration processing on the reference promotion information click rate and the estimated promotion information conversion rate to obtain the promotion information click rate corresponding to the candidate recommendation word.
7. The method according to claim 6, It is characterized in that The process of performing click-through rate estimation processing on the semantic feature data, the discrete feature data, and the statistical feature data for the candidate recommendation word to obtain the click-through rate of the reference promotion information corresponding to the candidate recommendation word includes: Performing semantic cross processing on the semantic feature data to obtain a semantic cross feature vector, and performing embedding processing on the discrete feature data to obtain a discrete feature vector; Performing vector concatenation processing based on the semantic cross feature vector and the discrete feature vector to obtain a comprehensive feature vector; The click rate prediction model is used to predict the click rate of the promotion information for the candidate recommendation word by processing the comprehensive feature vector and the statistical feature data to obtain the click rate of the reference promotion information corresponding to the candidate recommendation word.
8. A click rate estimation device, It is characterized in that The device comprises: A feature extraction module is used to obtain a user search text input by a user, and perform feature extraction processing based on the user search text to obtain semantic feature data, discrete feature data, and statistical feature data; The model estimation module is used to use a click-through rate estimation model to perform click-through rate estimation processing on the promotion information of the candidate recommendation words based on the semantic feature data, the discrete feature data, and the statistical feature data to obtain the click-through rate of the promotion information corresponding to the candidate recommendation words.
9. A computer storage medium, It is characterized in that The computer storage medium stores a plurality of instructions, and the instructions are suitable for being loaded by a processor and executing the method according to any one of claims 1 to 7.
10. An electronic device, It is characterized in that include: A processor and a memory; wherein the memory stores a computer program, and the computer program is suitable for being loaded by the processor and executing the method according to any one of claims 1 to 7.
Citation Information
Patent Citations
Click rate estimation method, medium, system and computing device
CN109582862A
Search result recommendation method and system based on text click rate
CN112487274A
Estimated conversion rate calibration method and device, computing equipment and storage medium
CN115330426A
Click rate estimation model training method, click rate estimation method, device and equipment
CN116932864A
Risk premiums for conversion-based online advertisement bidding
US20100228636A1
Cited By
Real-time search click rate estimation method, system, equipment, medium and product
CN122470966A
