Data processing method and device for product promotion information and server

Through the multi-head self-attention network model, users' behavior and product information are processed, and the accurate matching and push of financial product promotion information is achieved, which solves the problem of poor promotion results in the existing technology and improves the user experience.

CN119048179APending Publication Date: 2024-11-29CHINA CONSTRUCTION BANK +1
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411185273.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-27
Publication Date
2024-11-29

AI Technical Summary

Technical Problem

When promoting financial products in financial business scenarios, existing technologies often use blind mass promotion information, resulting in the information that users are interested in being ignored, the promotion effect is poor, affecting the user experience, and lacking effective solutions.

Method used

Using a product promotion information recommendation model based on a multi-head self-attention network, the target user's historical behavior data and product information are obtained, and multi-dimensional representation data of content, form and product are generated, so as to predict the matching between users and products, and accurately push interest. promotion information.

Benefits of technology

It improves the accuracy and effectiveness of product promotion, reduces the possibility of pushing uninterested information to users, and improves the user experience.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119048179A_ABST
    Figure CN119048179A_ABST
Patent Text Reader

Abstract

The invention provides a data processing method and device for product promotion information and a server, and can be applied to the field of artificial intelligence. The method comprises the following steps: obtaining target promotion information of a target product and a user identifier of a target user; obtaining historical behavior data of a target user in a current time period, target content data used for describing a target product, target form data used for displaying the target product, and target product data about the target product; processing the data by using a promotion information coding module, a user behavior coding module and a prediction module in a preset product promotion information recommendation model, and determining and outputting a corresponding target processing result; judging whether the target promotion information is matched with the target user or not according to the target processing result; and pushing the target promotion information to the target user under the condition of determining that the target promotion information is matched with the target user. Therefore, the product promotion information in which the target user is interested can be accurately and automatically determined, and targeted pushing is carried out on the target user.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This specification belongs to the field of artificial intelligence technology, and in particular to a data processing method, device and server for product promotion information. Background Art

[0002] In financial business scenarios, trading platforms or institutions often push promotional information about financial products to users in order to promote the proposed financial products.

[0003] However, based on the existing methods, most of them blindly send a large number of promotional information involving different financial products to users, resulting in users often receiving a lot of promotional information that they are not interested in at all, or even unwilling to click on and browse. Based on the above methods, the promotion effect is often not ideal, and it is easy to cause user disgust and affect user experience.

[0004] To address the above problems, no effective solution has been proposed yet. Summary of the invention

[0005] The present specification provides a product promotion information data processing method, device and server, which can automatically and accurately determine the product promotion information that a target user is interested in, and push it to the target user in a targeted manner.

[0006] This specification provides a data processing method for product promotion information, including:

[0007] Obtaining targeted promotion information about the target product and the user ID of the target user;

[0008] According to the user ID of the target user, obtain the historical behavior data of the target user in the current time period;

[0009] According to the target promotion information, the product identification of the target product is extracted, which is used to describe the target content data of the target product, and the target form data for displaying the target product; at the same time, according to the product identification of the target product, the target product data of the target product is queried and obtained;

[0010] The promotion information encoding module in the preset product promotion information recommendation model is used to process the target content data, the target form data, and the target product data respectively to obtain the corresponding content representation data, form representation data, and product representation data; the user behavior encoding module in the preset product promotion information recommendation model is used to process the historical behavior data of the target user in the current time period to obtain the corresponding behavior representation data of the target user; wherein the preset product promotion information recommendation model is a model obtained by training based on a multi-head self-attention network;

[0011] Using the prediction module in the preset product promotion information recommendation model to process the content representation data, form representation data, product representation data, and target user behavior representation data, to obtain the corresponding target processing results;

[0012] According to the target processing results, determine whether the target promotion information matches the target user;

[0013] When it is determined that the target promotion information matches the target user, the target promotion information is pushed to the target user.

[0014] In one embodiment, the promotion information encoding module includes at least: a word embedding layer, a word-level multi-head self-attention network, and an accumulated word attention network;

[0015] Accordingly, the promotion information encoding module in the preset product promotion information recommendation model is used to process the target content data to obtain corresponding content representation data, including:

[0016] The target content data is processed using the word embedding layer to extract the corresponding content features; and a numerical target content embedding vector is generated based on the content features;

[0017] Using multiple attention heads in a word-level multi-head self-attention network to process the target content embedding vector respectively, learning and utilizing the interaction relationship between different content features, generating intermediate representation data of the content features; and combining the intermediate representation data of the content features of multiple attention heads to obtain multi-head representation data of the content features;

[0018] The multi-head representation data of content features are processed by using an accumulated word attention network to determine the attention weight of the content features; and the multi-head representation data of corresponding content features and the attention weight of the content features are jointly used to generate content representation data corresponding to the target content data.

[0019] In one embodiment, the generalized information encoding module further comprises at least: a UniLM sub-model, a CNN network, and a weight network layer; wherein the UniLM sub-model comprises a plurality of Transformer modules;

[0020] Accordingly, the promotion information encoding module in the preset product promotion information recommendation model is used to process the target content data to obtain corresponding content representation data, including:

[0021] The target content data is processed using the UniLM sub-model, and the corresponding content vector is obtained by generating and using the corresponding segment vector, word vector and position vector;

[0022] Processing the content vector using a plurality of Transformer modules, and generating a first content deep semantic vector and a second content deep semantic vector having deep bidirectional semantics by extracting and using context information;

[0023] Processing the first content depth semantic vector and the second content depth semantic vector using a CNN network, extracting and using local features to correct the first content depth semantic vector and the second content depth semantic vector, to obtain a corrected first content depth semantic vector and a corrected second content depth semantic vector;

[0024] The weighted network layer is used to obtain content representation data corresponding to the target content data by performing weighted operations based on a word-level attention mechanism and an item-level attention mechanism according to the revised first content deep semantic vector and the revised second content deep semantic vector.

[0025] In one embodiment, the promotion information encoding module further includes: a classifier, an initial encoder, and a multi-layer perceptron;

[0026] Accordingly, the method further includes:

[0027] Processing the target content data, the target form data, and the target product data using a classifier to determine a promotion information type and a promotion information subtype of the target promotion information;

[0028] Using the initial encoder, one-hot encoding is performed according to the promotion information type and the promotion information subtype of the target promotion information, respectively, to obtain a corresponding promotion information type encoding vector and promotion information subtype encoding vector;

[0029] Using a multi-layer perceptron to process the promotion information type encoding vector and the promotion information sub-type encoding vector, to obtain corresponding promotion information type features and promotion information sub-type features;

[0030] The weighted network layer is used to obtain content representation data corresponding to the target content data by performing weighted operations based on the word-level attention mechanism and the item-level attention mechanism according to the corrected first content deep semantic vector, the corrected second content deep semantic vector, the promotion information type features and the promotion information subtype features.

[0031] In one embodiment, the user behavior encoding module includes at least: a feature extraction layer, an information-level multi-head self-attention network, and an additional information attention network;

[0032] Accordingly, the user behavior encoding module in the preset product promotion information recommendation model is used to process the historical behavior data of the target user in the current time period to obtain the corresponding behavior representation data of the target user, including:

[0033] Using the feature extraction layer to process the historical behavior data of the target user in the current time period, extracting the user features related to the target user's behavior operations on the historical product promotion information as key features; and combining the key features to generate a numerical user feature embedding vector;

[0034] The user feature embedding vectors are processed by using multiple attention heads in the information-level multi-head self-attention network, and the interaction relationship between different key features is learned and utilized to enhance the key features to obtain enhanced key features; and the enhanced key features of multiple attention heads are combined to obtain multi-head representation data of the key features; wherein the multi-head representation data of a key feature corresponds to at least one historical product promotion information;

[0035] The additional information attention network is used to process the multi-head representation data of key features, and the influence weight of other historical product promotion information on the target user's behavioral operations regarding the historical product promotion information is learned and utilized; and the corresponding multi-head representation data of key features and the influence weight are jointly used to generate the target user's behavioral representation data corresponding to the target user's historical behavioral data.

[0036] In one embodiment, the prediction module includes at least a first operation structure, a second operation structure, a third operation structure connected in parallel, and a fourth operation structure connected in series with the first operation structure, the second operation structure, and the third operation structure;

[0037] Accordingly, the prediction module in the preset product promotion information recommendation model is used to process the content representation data, form representation data, product representation data, and target user behavior representation data to obtain corresponding target processing results, including:

[0038] The first operation structure is used to process the content characterization data and the behavior characterization data of the target user to obtain a corresponding first operation result; the second operation structure is used to process the form characterization data and the behavior characterization data of the target user to obtain a corresponding second operation result; the third operation structure is used to process the product characterization data and the behavior characterization data of the target user to obtain a corresponding third operation result; the fourth operation structure is used to generate a corresponding target processing result based on the first operation result, the second operation result, and the third operation result.

[0039] In one embodiment, the target processing result also carries a first operation result, a second operation result, and a third operation result;

[0040] Accordingly, after determining whether the target promotion information matches the target user according to the target processing result, the method further includes:

[0041] In the case where it is determined that the target promotion information does not match the target user, detecting whether the first operation result, the second operation result, and the third operation result meet the preset requirements;

[0042] When it is determined that the first operation result and the third operation result meet the preset requirements, and the second operation result does not meet the preset requirements, obtaining the behavior representation data of the target user;

[0043] According to the behavior representation data of the target user, a matching target information form template is determined from a preset information form template set;

[0044] Using the target information form template, modify the target promotion information in a form to obtain the modified target promotion information;

[0045] Push the modified target promotion information to target users.

[0046] In one embodiment, after acquiring historical behavior data of the target user in the current time period according to the user identification of the target user, the method further includes:

[0047] Check whether the amount of data contained in the historical behavior data of the target user in the current time period is greater than or equal to a preset data amount threshold;

[0048] When it is determined that the amount of data contained in the historical behavior data of the target user in the current time period is less than a preset data amount threshold, determining target neighboring users of the target user;

[0049] Obtain historical behavior data of target neighboring users in the current time period;

[0050] The historical behavior data of the target user in the current time period and the historical behavior data of the target neighboring users in the current time period are combined to serve as the expanded historical behavior data of the target user in the current time period.

[0051] In one embodiment, the method further comprises:

[0052] Obtain sample behavior records, sample promotion information, product data of sample products of sample users, as well as user historical behavior data, historical product promotion information, and product data of historical products in a specified time period;

[0053] A first training set is constructed based on sample behavior records of sample users, sample promotion information, and product data of sample products; a second training set is constructed based on user historical behavior data, historical product promotion information, and product data of historical products in a specified time period;

[0054] Constructing an initial model based on a multi-head self-attention network; wherein the initial model at least includes an initial promotion information encoding module, an initial user behavior encoding module, and an initial prediction module;

[0055] Use Spark to call the first training set, train the initial model, and obtain an intermediate model that meets the requirements;

[0056] Use Flink to call the second training set, train the intermediate model, and obtain a preset product promotion information recommendation model that meets the requirements.

[0057] In one embodiment, an initial model based on a multi-head self-attention network is constructed, including:

[0058] According to the sample promotion information and the product data of the sample product, the number of attention heads in the word-level multi-head self-attention network is determined; and the initial promotion information encoding module is configured according to the number of attention heads;

[0059] According to the sample behavior records of the sample users, the number of attention heads in the information-level multi-head self-attention network is determined; and the initial user behavior encoding module is configured according to the number of attention heads;

[0060] An initial prediction module is configured; and an initial promotion information encoding module, an initial user behavior encoding module, and an initial prediction module are combined to obtain an initial model.

[0061] In one embodiment, after obtaining the user's historical behavior data, historical product promotion information, and historical product data for a specified time period, the method further includes:

[0062] Use Flink to preprocess user historical behavior data, historical product promotion information, and historical product data for a specified time period.

[0063] In one embodiment, after obtaining a preset product promotion information recommendation model that meets the requirements, the method further includes:

[0064] At every specified time period, query the platform database to obtain historical product promotion information and historical promotion records for the current specified time period;

[0065] According to the historical promotion records of the current specified time period, obtain the user historical behavior data of the current specified time period; according to the historical product promotion information of the current specified time period, obtain the product data of the historical products of the current specified time period;

[0066] Flink is used to call the user historical behavior data of the current specified time period, the historical product promotion information of the current specified time period, and the product data of the historical products of the current specified time period to train the preset product promotion information recommendation model, and obtain the updated preset product promotion information recommendation model.

[0067] This manual provides a data processing device for product promotion information, including:

[0068] A first acquisition module is used to acquire target promotion information about a target product and a user ID of a target user;

[0069] The second acquisition module is used to acquire the historical behavior data of the target user in the current time period according to the user identification of the target user;

[0070] The third acquisition module is used to extract the product identification of the target product, the target content data for describing the target product, and the target form data for displaying the target product according to the target promotion information; at the same time, according to the product identification of the target product, query and obtain the target product data of the target product;

[0071] The first processing module is used to use the promotion information encoding module in the preset product promotion information recommendation model to process the target content data, the target form data, and the target product data respectively to obtain the corresponding content representation data, form representation data, and product representation data; use the user behavior encoding module in the preset product promotion information recommendation model to process the historical behavior data of the target user in the current time period to obtain the corresponding behavior representation data of the target user; wherein the preset product promotion information recommendation model is a model obtained by training based on a multi-head self-attention network;

[0072] The second processing module is used to process the content representation data, the form representation data, the product representation data, and the behavior representation data of the target user by using the prediction module in the preset product promotion information recommendation model to obtain the corresponding target processing result;

[0073] A determination module is used to determine whether the target promotion information matches the target user according to the target processing result;

[0074] The push module is used to push the target promotion information to the target user when it is determined that the target promotion information matches the target user.

[0075] This specification provides a server, including a processor and a memory for storing processor executable instructions, and the processor implements the relevant steps of the data processing method of product promotion information when executing the instructions.

[0076] This specification provides a computer-readable storage medium on which computer instructions are stored. When the instructions are executed by a processor, the steps of the data processing method for product promotion information are implemented.

[0077] This specification provides a computer program product, which includes a computer program, and when the computer program is executed by a processor, the steps of the data processing method for product promotion information are implemented.

[0078] Based on the data processing method, device and server for product promotion information provided in this specification, before specific implementation, a multi-head self-attention network can be used to construct and train to obtain information representation data about the product promotion information that is relatively comprehensive and rich based on multiple dimensions such as the content, form, and product of the product promotion information, as well as behavior representation data about the user based on time series, and then a preset product promotion information recommendation model can be used to accurately predict the matching between the user and the product promotion information based on the above information representation data and behavior representation data. In specific implementation, the target promotion information of the target product and the user identifier of the target user can be obtained first; according to the user identifier of the target user, the historical behavior data of the target user in the current time period can be obtained; according to the target promotion information, the product identifier of the target product, the target content data used to describe the target product in the target promotion information, and the target form data used to display the target product in the target promotion information are extracted; at the same time, according to the product identifier of the target product, the target product data of the target product is queried; the promotion information encoding module in the preset product promotion information recommendation model is used to process the target content data, the target form data, and the target product data respectively to obtain the corresponding content representation data, form representation data, and product representation data; then the user behavior encoding module in the preset product promotion information recommendation model is used to process the historical behavior data of the target user in the current time period to obtain the corresponding behavior representation data of the target user; the prediction module in the preset product promotion information recommendation model is used to process the content representation data, form representation data, product representation data, and the behavior representation data of the target user to obtain the corresponding target processing result; then, according to the target processing result, it is judged whether the target promotion information matches the target user; and if a match is determined, the target promotion information is pushed to the target user. In this way, it is possible to accurately and automatically determine the product promotion information that the target user is interested in and willing to click on, and then push targeted and accurate information to the target user. In this way, on the one hand, it can effectively improve the promotion effect of the product, and on the other hand, it can avoid pushing promotion information of products that the target user is not interested in to the target user, which affects the user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] In order to more clearly illustrate the embodiments of this specification, the drawings required for use in the embodiments will be briefly introduced below. The drawings described below are only some embodiments recorded in this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0080] Figure 1 It is a flowchart of a data processing method for product promotion information provided by an embodiment of this specification;

[0081] Figure 2 It is a schematic diagram of an embodiment of a data processing method for product promotion information provided by an embodiment of this specification, in a scenario example;

[0082] Figure 3 It is a schematic diagram of an embodiment of a data processing method for product promotion information provided by an embodiment of this specification, in a scenario example;

[0083] Figure 4 It is a schematic diagram of an embodiment of a data processing method for product promotion information provided by an embodiment of this specification, in a scenario example;

[0084] Figure 5 It is a schematic diagram of an embodiment of a data processing method for product promotion information provided by an embodiment of this specification, in a scenario example;

[0085] Figure 6 It is a schematic diagram of an embodiment of a data processing method for product promotion information provided by an embodiment of this specification, in a scenario example;

[0086] Figure 7 It is a schematic diagram of an embodiment of a data processing method for product promotion information provided by an embodiment of this specification, in a scenario example;

[0087] Figure 8 It is a schematic diagram of the structure of a server provided by an embodiment of this specification;

[0088] Fig. 9 It is a schematic diagram of the structure of a data processing device for product promotion information provided by an embodiment of this specification;

[0089] Fig.10 It is a schematic diagram of an embodiment of a data processing method for product promotion information provided by an embodiment of this specification, in a scenario example. DETAILED DESCRIPTION

[0090] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this specification.

[0091] It should be noted that the user-related information and data involved in the embodiments of this specification are all information and data authorized by the user or fully authorized by relevant parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of the relevant data comply with relevant laws, regulations and standards, take necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation entrances for users or relevant parties to choose to authorize or refuse.

[0092] It should also be noted that in the embodiments of this specification, certain software, components, models and other existing solutions in the industry may be mentioned, which should be regarded as exemplary. Their purpose is only to illustrate the feasibility of implementing the technical solution of this application, but it does not mean that the applicant has or will necessarily use the solution.

[0093] See also Figure 1 As shown, the embodiment of this specification provides a data processing method for product promotion information. Among them, the method is specifically applied to the server side. When implemented specifically, the method may include the following contents:

[0094] S101: Obtaining target promotion information about a target product and a user ID of a target user;

[0095] S102: Acquire historical behavior data of the target user in the current time period according to the user ID of the target user;

[0096] S103: extracting the product identification of the target product according to the target promotion information, target content data for describing the target product, and target form data for displaying the target product; at the same time, querying and obtaining target product data of the target product according to the product identification of the target product;

[0097] S104: using the promotion information encoding module in the preset product promotion information recommendation model to process the target content data, the target form data, and the target product data respectively, and obtain corresponding content representation data, form representation data, and product representation data; using the user behavior encoding module in the preset product promotion information recommendation model to process the historical behavior data of the target user in the current time period, and obtain the corresponding behavior representation data of the target user; wherein the preset product promotion information recommendation model is a model obtained by training based on a multi-head self-attention network;

[0098] S105: using a prediction module in a preset product promotion information recommendation model to process the content representation data, the form representation data, the product representation data, and the behavior representation data of the target user to obtain a corresponding target processing result;

[0099] S106: Determine whether the target promotion information matches the target user according to the target processing result;

[0100] S107: When it is determined that the target promotion information matches the target user, the target promotion information is pushed to the target user.

[0101] The target promotion information may be specifically understood as data information used to introduce the target product and to encourage users to pay attention to or purchase the target product. Specifically, for example, the target promotion information may be an advertisement about the target product.

[0102] The above-mentioned target promotion information may specifically be a video promotion information, a picture promotion information, a text promotion information, a voice promotion information, etc.

[0103] The target product may specifically be a financial product, such as stocks, funds, insurance, etc. The target product may also be an office supply, such as pencils, rulers, pencil sharpeners, etc. The target product may also be a related service, such as cleaning service, security service, designated driver service, etc.

[0104] It should be noted that the target products listed above are only schematic illustrations. In specific implementation, the target products may also include other types of products according to specific application scenarios and processing requirements. This specification does not limit this.

[0105] In this specification, the target product is mainly a financial product in a financial business scenario. For other types of target products, please refer to the embodiments of financial products in this specification. This specification does not elaborate on this.

[0106] The above-mentioned target users can be specifically understood as users to whom promotional information of related products is to be pushed.

[0107] The above-mentioned target content data can be specifically understood as content data in the target promotion information that is directly related to the target product, for example, text content about the price of the target product, text content about the purchase rules of the target product, or text content about the introduction of the provider of the target product, etc.

[0108] The above-mentioned target form data can be specifically understood as information data in the target promotion information that is not directly related to the target product, for example, the text of the target promotion information used to attract the user's attention, the main color used in the target promotion information, the layout rules of text and pictures in the target promotion information, etc.

[0109] The above-mentioned target product data can be specifically understood as content data directly related to the target product that is not included in the target product promotion information and is obtained through other media other than the target promotion information, such as evaluation texts about the target product collected through multiple channels on relevant platforms, introductory information of the target product published on the official website and not included in the target promotion information, historical revenue records of the target product, etc.

[0110] The above-mentioned historical behavior data of the target user in the current time period may specifically be the historical operation records of the target user in the most recent time period (e.g., the most recent month, or the most recent six months, etc.) regarding the relevant product promotion information. For example, the user clicks on and browses the product promotion information pushed to him, and the user further clicks on the product link in the product promotion information and purchases the relevant product. For another example, the user directly swipes away the product promotion information pushed to him. For another example, the content data, form data, product data, etc. of the product promotion information that the user clicks on and browses.

[0111] Furthermore, in order to more accurately and comprehensively describe the relevant characteristic information of the target user, the historical behavior data of the target user in the current time period may also carry the attribute information of the target user in the current time period. The attribute information may specifically include one or more of the following: the user's gender, the user's age, the user's occupation, the user's monthly income, the user's risk assessment results, the user's credit data, etc.

[0112] The above-mentioned preset product promotion information recommendation model can be specifically understood as a neural network model that is constructed and trained in advance according to preset training rules using a multi-head self-attention network. It can automatically determine whether the user matches the product promotion information based on the input product promotion information-related data and user behavior-related data, based on multiple different dimensions, combined with time series analysis, and through corresponding processing operations.

[0113] In some embodiments, during specific implementation, the user identification of the target user (for example, the user name, user number, or registered mobile phone number of the target user) can be obtained and the platform history database can be queried to obtain the operation record of the target user for the relevant product promotion information in the current time period; at the same time, the platform user database can also be queried based on the user identification of the target user to obtain the attribute information of the target user; and the above operation records and attribute information are combined as the historical behavior data of the target user in the current time period.

[0114] In specific implementation, the target product targeted by the target promotion information can be determined by performing OCR recognition and / or semantic recognition on the target promotion information, and the product identification of the target product can be obtained.

[0115] The product type of the target product can be queried and determined based on the product identification of the target product; a matching product content extraction model can be determined based on the product type of the target product; and the target promotion information can be processed using the content extraction model to automatically extract the target content data used to describe the target product in the target promotion information.

[0116] According to the preset form detection rules, the target promotion information can be detected, the speech text can be extracted, and the promotion information features such as color tone, layout, font, etc. can be determined; then the above speech text and promotion information features are combined to obtain the corresponding target form data.

[0117] The target product data can be obtained according to the product identification of the target product by querying the platform product information database and / or crawling data related to the product identification of the target product from other trading platforms and websites with which there is a cooperative relationship.

[0118] In some embodiments, see Figure 2 As shown, the above-mentioned preset product promotion information recommendation model at least includes: a promotion information encoding module, a user behavior encoding module, and a prediction module.

[0119] The promotion information encoding module is connected in parallel with the user behavior encoding module, and the prediction module is connected in series with the promotion information encoding module and the user behavior encoding module respectively.

[0120] The above-mentioned promotion information encoding module is used to process the input related data of the promotion information including content data, form data, product data, etc., analyze and extract corresponding features, generate and output representation data including content representation data, form representation data, product representation data, etc., which can comprehensively reflect the characteristics of the impact of the promotion information on users from multiple dimensions such as content, form, and product.

[0121] The above-mentioned user behavior encoding module is used to process input data related to user behavior, including user historical behavior data, extract and utilize corresponding features through time series analysis, generate and output representation data including user behavior representation data that can reflect the characteristics of the user's operational behavior on promotional information.

[0122] The above-mentioned promotion information encoding module can specifically be a network module constructed and trained based on a multi-head self-attention network. In this way, the above-mentioned promotion information encoding module can use the multi-head self-attention network to simultaneously associate multiple different dimensions to generate more accurate and comprehensive content representation data, form representation data, and product representation data about the promotion information.

[0123] The user behavior encoding module can be a network module constructed and trained based on a multi-head self-attention network. In this way, the user behavior encoding module can use the multi-head self-attention network to perform temporal analysis on the user's behavior within a time period, and at the same time combine the user's attribute information to generate more detailed and reliable behavior representation data about the user's behavior.

[0124] The prediction module is used to jointly process the representation data output by the promotion information encoding module and the user behavior encoding module, and through corresponding analysis and calculation, predict whether the user will be interested in the promotion information and output the corresponding processing result.

[0125] In some embodiments, see Figure 2 As shown, the promotion information encoding module may include at least: a word embedding layer, a word-level multi-head self-attention network, and an accumulated word attention network;

[0126] Accordingly, see Figure 3 As shown, the above-mentioned promotion information encoding module in the preset product promotion information recommendation model is used to process the target content data to obtain the corresponding content representation data. When implemented specifically, the following contents may be included:

[0127] S1: Use the word embedding layer to process the target content data and extract the corresponding content features; and generate a numerical target content embedding vector based on the content features;

[0128] S2: Use multiple attention heads in the word-level multi-head self-attention network to process the target content embedding vector respectively, learn and use the interaction relationship between different content features, and generate intermediate representation data of the content features; and combine the intermediate representation data of the content features of multiple attention heads to obtain the multi-head representation data of the content features;

[0129] S3: Use the cumulative word attention network to process the multi-head representation data of the content features to determine the attention weight of the content features; and jointly use the corresponding multi-head representation data of the content features and the attention weight of the content features to generate content representation data corresponding to the target content data.

[0130] According to the above method, by utilizing the relevant structure in the promotion information encoding module to process the target content data, corresponding content representation data can be generated, which contains refined feature information related to the content of the target promotion information and is convenient for further processing by subsequent models.

[0131] In the specific implementation, according to a similar method as described above, the target form data is first processed by using the word embedding layer in the promotion information encoding module to generate a corresponding target form embedding vector; then the target form embedding vector is processed separately by multiple attention heads in the word-level multi-head self-attention network to generate multi-head representation data of each content feature; finally, the multi-head representation data of the content features is processed by the cumulative word attention network to generate form representation data corresponding to the target form data.

[0132] In a similar manner to that described above, the target product data is first processed using the word embedding layer in the promotion information encoding module to generate the corresponding target product embedding vector; the target product embedding vector is then processed separately using multiple attention heads in the word-level multi-head self-attention network to generate multi-head representation data of each content feature; finally, the multi-head representation data of the content features is processed using the cumulative word attention network to generate product representation data corresponding to the target product data.

[0133] Based on the above embodiment, the promotion information encoding module in the preset product promotion information recommendation model can be used to simultaneously associate with multiple different dimensions of the content, form, and product of the target promotion information to generate more accurate and comprehensive content representation data, form representation data, and product representation data for the target promotion information.

[0134] For details, see Figure 2 As shown, in the promotion information encoding module, the word embedding layer is connected in series with the word-level multi-head self-attention network, and the word-level multi-head self-attention network is connected in series with the cumulative word attention network. Among them, the above-mentioned word-level multi-head self-attention network is a multi-head self-attention network. The above-mentioned word-level multi-head self-attention network can specifically include multiple attention heads (or self-attention heads).

[0135] In specific implementation, the word embedding layer can be used to first obtain a content feature template that matches the content data; based on the content feature template, the target content data is processed to extract content features such as the name of the financial product, risk level, investment type, etc.; and the above content features are converted into corresponding numerical data; and then the above numerical data are combined in sequence to construct the corresponding content representation data.

[0136] Specifically, for example, the following form of content representation data can be constructed: [e1, e2, ..., e M ]. Among them, e M Indicates the numerical data corresponding to the content feature numbered M, where M is the total number of content features.

[0137] In specific implementation, each of the multiple attention heads in the word-level multi-head self-attention network can be used to process the target content embedding vector based on its own perspective and the knowledge it has learned and mastered, and learn the interaction relationship parameters between different content features in the target content embedding vector; and use the interaction relationship parameters between different content features to process the numerical data of each content feature to obtain the processed content features that can be associated with other related content features as the intermediate representation data of the content feature; and then combine the intermediate representation data of the same content feature from multiple attention heads to obtain multi-head representation data for each content feature.

[0138] Specifically, for example, the interaction relationship parameter of the attention head numbered k with respect to the content feature numbered i can be calculated according to the following formula:

[0139]

[0140] in, is the interaction relation parameter of the attention head numbered k with respect to the content feature numbered j on the content feature numbered i, is the first projection parameter of the attention head numbered k.

[0141] Furthermore, the content feature processed by the attention head numbered k with respect to the content feature numbered i can be calculated according to the following formula as the intermediate representation data of the attention head numbered k with respect to the content feature numbered i:

[0142]

[0143] in, is the intermediate representation data of the attention head numbered k about the content feature numbered i, is the second projection parameter of the attention head numbered k.

[0144] Then, the intermediate representation data of the content feature numbered i from multiple attention heads are combined and used in the following manner to construct the multi-head representation data of the content feature numbered i: Where h is the total number of attention heads.

[0145] In specific implementation, the cumulative word attention network can be used to first use relevant projection parameters and query vectors to separately process the multi-head representation data of each content feature to obtain the initial weight of each content feature; then the initial weights of multiple content features are used jointly to determine the attention weight of each content feature; finally, the multi-head representation data of the corresponding content features and the attention weights of the content features are used jointly through weighted processing to obtain the content representation data corresponding to the target content data.

[0146] Specifically, for example, the initial weight of the content feature numbered i may be calculated according to the following formula:

[0147]

[0148] in, is the initial weight of the content feature numbered i, is the multi-head representation data of the content feature numbered i, q w is the query vector, V m is the first relevant projection parameter, v m is the second related projection parameter.

[0149] Then calculate the attention weight of the content feature numbered i according to the following formula:

[0150]

[0151] in, is the attention weight of the content feature numbered i, is the initial weight of the content feature numbered i, is the initial weight of the content feature numbered j.

[0152] Finally, the corresponding content representation data is constructed in the following way:

[0153] In some embodiments, see Figure 2 As shown, the word-level multi-head self-attention network can specifically include a first multi-head self-attention sub-network, a second multi-head self-attention sub-network, and a third multi-head self-attention sub-network connected in parallel.

[0154] The first multi-head self-attention sub-network is used to specifically process content embedding vectors, the second multi-head self-attention sub-network is used to specifically process form embedding vectors, and the third multi-head self-attention sub-network is used to specifically process product embedding vectors. The first multi-head self-attention sub-network, the second multi-head self-attention sub-network, and the third multi-head self-attention sub-network may each include multiple attention heads.

[0155] In this way, the word-level multi-head self-attention network can be used to more accurately and specifically process the corresponding content embedding vector, form embedding vector, and product embedding vector to achieve relatively better processing results.

[0156] In some embodiments, after the intermediate representation data of the content feature of multiple attention heads are combined to obtain the multi-head representation data of the content feature, the method may further include the following when implemented:

[0157] A word-level multi-head self-attention network is used to process multi-head representation data of content features, multi-head representation data of form features, and multi-head representation data of product features, and the interaction between the multi-head representation data of form features and the multi-head representation data of product features and the multi-head representation data of content features is learned and used to adjust the multi-head representation data of content features to obtain adjusted multi-head representation data of content features; wherein the multi-head representation data of form features is generated by processing a target form embedding vector using multiple attention heads in a word-level multi-head self-attention network; and the multi-head representation data of product features is generated by processing a target product embedding vector using multiple attention heads in a word-level multi-head self-attention network.

[0158] According to the above method, the word-level multi-head self-attention network can be used to process the multi-head representation data of content features, the multi-head representation data of formal features, and the multi-head representation data of product features, learn and use the interaction between the multi-head representation data of content features and the multi-head representation data of product features and the multi-head representation data of formal features, adjust the multi-head representation data of formal features, and obtain the adjusted multi-head representation data of formal features.

[0159] Furthermore, a word-level multi-head self-attention network can be used to process multi-head representation data of content features, multi-head representation data of formal features, and multi-head representation data of product features, learn and use the interaction between the multi-head representation data of content features and the multi-head representation data of formal features and the multi-head representation data of product features, adjust the multi-head representation data of product features, and obtain the adjusted multi-head representation data of product features.

[0160] Based on the above embodiments, considering that for the same promotional information, different types of form features, content features, and product features often have certain interactions. Therefore, by introducing the interaction relationship of other types of features for correction, multi-head representation data of content features with relatively higher accuracy and relatively better effect can be obtained, and the adjusted multi-head representation data of content features can be obtained to further improve the overall processing accuracy and reduce errors.

[0161] In some embodiments, another promotion information encoding module may be used. The promotion information encoding module at least includes: a UniLM sub-model, a CNN network, a weight network layer and other structures; wherein the UniLM sub-model includes multiple Transformer modules;

[0162] Accordingly, the above-mentioned promotion information encoding module in the preset product promotion information recommendation model is used to process the target content data to obtain the corresponding content representation data. When implemented specifically, the following contents may be included:

[0163] S1: Use the UniLM sub-model to process the target content data, and obtain the corresponding content vector by generating and using the corresponding segment vector, word vector and position vector;

[0164] S2: Processing the content vector using multiple Transformer modules, extracting and using context information to generate a first content deep semantic vector and a second content deep semantic vector having deep bidirectional semantics;

[0165] S3: Processing the first content depth semantic vector and the second content depth semantic vector by using a CNN network, extracting and modifying the first content depth semantic vector and the second content depth semantic vector by using local features, and obtaining a modified first content depth semantic vector and a modified second content depth semantic vector;

[0166] S4: Using the weight network layer, according to the corrected first content deep semantic vector and the corrected second content deep semantic vector, by performing weighted operations based on the word-level attention mechanism and the item-level attention mechanism, content representation data corresponding to the target content data is obtained.

[0167] Among them, UniLM (Unified Language Model) can be specifically understood as a model structure that can obtain dynamic word embedding of the components of input data, and by using dynamic word embedding, the output vector has semantic information at different levels of context. The above UniLM sub-model is a model obtained by pre-training with relevant sample data based on the UniLM model structure.

[0168] CNN (Convolutional Neural Networks) network, specifically refers to a convolutional neural network, which is a feedforward neural network that includes convolution calculations and has a deep structure.

[0169] Specifically, UniLM has a similar encoder structure to BERT in terms of network structure; but in pre-training tasks, UniLM can not only use the context of masked tokens for training like the autoencoder language model, but also train from left to right like the autoregressive language model, and can even encode the input text first and then generate a sequence from left to right like the Encoder-Decoder architecture model, thereby obtaining relatively good feature processing effects.

[0170] Correspondingly, using the above-mentioned UniLM sub-model to process the target content data, different from the conventional static word embedding method, can more fully and comprehensively consider the contextual semantic information of the input data components, thereby obtaining a semantic vector with higher accuracy and better effect.

[0171] The above-mentioned UniLM sub-model may specifically include multiple Transformer modules. In specific implementation, the UniLM sub-model may first process and transform the input target content data to obtain a content vector of uniform length containing corresponding different component vectors (e.g., segment vectors, word vectors, and position vectors, etc.); the UniLM sub-model then uses multiple Transformer modules to process the content vector, extract and use contextual information, and perform further feature processing to obtain a first content deep semantic vector and a second content deep semantic vector with deep bidirectional semantics.

[0172] Specifically, the first content depth semantic vector can be expressed as: The above second content depth semantic vector can be expressed as:

[0173] Considering that the relevant sample data used in the previous pre-trained UniLM sub-model may contain noisy data, which in turn affects the model accuracy of the UniLM sub-model, the first content depth semantic vector and the second content depth semantic vector directly obtained based on the UniLM sub-model will still have noise errors.

[0174] In order to further improve the subsequent processing accuracy, the CNN network can be used to process the first content depth semantic vector and the second content depth semantic vector to extract and use local features to correct the first content depth semantic vector and the second content depth semantic vector, so as to obtain the corrected first content depth semantic vector and the corrected second content depth semantic vector with higher accuracy and eliminated noise errors.

[0175] Specifically, the local features extracted by the CNN network can be expressed as follows:

[0176] t″ i =ReLU(w c T i:i+h-1 +b c )

[0177] Among them, w c represents the convolution kernel of size h, T i:i+h-1 It means taking out the submatrix consisting of rows i to i+h-1 from vector T, b c represents the local bias parameter.

[0178] By using the above local features to correct the vector, the corrected first content depth semantic vector can be specifically expressed as: Similarly, the modified second content depth semantic vector can be specifically expressed as:

[0179] Next, considering that different component vectors (word vectors) in product promotion information differ in importance and information volume, we can first use the weighted network layer based on the word-level attention mechanism to process the revised first content depth semantic vector and the revised second content depth semantic vector respectively, and determine the first weight value of each component vector; then use the corresponding first weight value to obtain the corresponding weighted first content depth semantic vector and the weighted second content depth semantic vector through a weighted operation.

[0180] Specifically, for example, the following formula may be referred to to obtain the first content depth semantic vector after a weighting:

[0181]

[0182] Among them, r a is the first content depth semantic vector after weighting, W a is the first type of training parameter, q i is the attention query vector, b a represents the first bias parameter, n a Represents the number of component vectors in the modified first content depth semantic vector.

[0183] In a similar way, the second content depth semantic vector after weighting can be obtained: b .

[0184] Furthermore, the weight network layer can be used again based on the item-level attention mechanism to determine the second weight value for the first content depth semantic vector after the first weighting and the second content depth semantic vector after the first weighting from the overall dimension, and then through a secondary weighted operation, the corresponding content representation data can be finally obtained.

[0185] In some embodiments, the promotion information encoding module may further include: a classifier, an initial encoder, a multi-layer perceptron and other structures;

[0186] Accordingly, when the method is implemented specifically, it may also include the following contents:

[0187] S1: using a classifier to process target content data, target form data, and target product data to determine a promotion information type and a promotion information subtype of the target promotion information;

[0188] S2: using the initial encoder to perform one-hot encoding according to the promotion information type and promotion information subtype of the target promotion information, respectively, to obtain the corresponding promotion information type encoding vector and promotion information subtype encoding vector;

[0189] S3: using a multi-layer perceptron to process the promotion information type encoding vector and the promotion information sub-type encoding vector to obtain corresponding promotion information type features and promotion information sub-type features;

[0190] S4: Using the weighted network layer, according to the corrected first content deep semantic vector, the corrected second content deep semantic vector, the promotion information type feature and the promotion information subtype feature, by performing weighted operations based on the word-level attention mechanism and the item-level attention mechanism, content representation data corresponding to the target content data is obtained.

[0191] In specific implementation, the classifier can be used to automatically identify and determine the promotion information type and promotion information subtype of the target promotion information by processing the target content data, target form data, and target product data. The above promotion information subtype can be specifically understood as the next level of subdivision of the promotion information type.

[0192] Considering that the promotion information type and promotion information subtype contain fewer text characters and no context, the initial encoder can be used to directly one-hot encode the above promotion information type and promotion information subtype, respectively, to obtain the promotion information type encoding vector and promotion information subtype encoding vector corresponding to the same length, which are respectively denoted as: T′ c , T′ sc .

[0193] Among them, one-hot encoding can be specifically understood as a commonly used feature encoding method, which is mainly used to convert the values ​​of categorical variables into binary vector form.

[0194] In order to extract more type feature information, the promotion information type encoding vector and the promotion information subtype encoding vector can be respectively input into the multi-layer perceptron for further feature processing to extract the corresponding promotion information type features and promotion information subtype features, which are respectively denoted as: r c 、r sc .

[0195] Then, the weighted network layer can be used to process the corrected first content depth semantic vector and the corrected second content depth semantic vector based on the word-level attention mechanism, and a weighted processing based on the word-level attention mechanism can be performed to obtain a weighted first content depth semantic vector r a , and the weighted second content depth semantic vector r b .

[0196] Then, the weighted network layer is used to jointly process the first content deep semantic vector after weighting once, the second content deep semantic vector after weighting once, the promotion information type features and the promotion information sub-type features based on the item-level attention mechanism, and the second weight values ​​for the first content deep semantic vector after weighting once and the second content deep semantic vector after weighting once are determined from the overall dimension respectively; at the same time, the weighted network layer is used to determine the weight values ​​of the promotion information type features and the promotion information sub-type features.

[0197] Specifically, for example, the second weight value of the first content depth semantic vector after weighting may be determined according to the following formula:

[0198] r′ a =tanh(W r r a +b r )

[0199]

[0200] Among them, α a Represents the second weight value, r′ a Represents the intermediate weight value, W r represents the second type of training parameters, b r Represents the second bias parameter.

[0201] In a similar manner, the second weight value of the second content depth semantic vector after the first weighting, the weight value of the promotion information type feature, and the weight value of the promotion information subtype feature can be further calculated and recorded as: α b , αc , α sc .

[0202] Finally, the weighted network layer is used to perform a weighted operation by jointly using the first content depth semantic vector after weighting, the second content depth semantic vector after weighting, the promotion information type feature, the promotion information subtype feature, and the corresponding weight value to obtain and output the content representation data corresponding to the target content data: x = α a r a +α b r b +α c r c +α sc r sc .

[0203] In a similar manner as described above, the promotion information encoding module may be used to process and obtain corresponding form representation data and product representation data.

[0204] In some embodiments, see Figure 2 The above-mentioned user behavior encoding module may at least include: a feature extraction layer, an information-level multi-head self-attention network, an additional information attention network and other structures.

[0205] Among them, the feature extraction layer is connected in series with the information-level multi-head self-attention network, and the information-level multi-head self-attention network is connected in series with the additional information attention network.

[0206] Accordingly, see Figure 4 As shown, the user behavior encoding module in the preset product promotion information recommendation model is used to process the historical behavior data of the target user in the current time period to obtain the corresponding behavior representation data of the target user. When implemented specifically, the following contents may be included:

[0207] S1: Use the feature extraction layer to process the historical behavior data of the target user in the current time period, extract the user features related to the target user's behavior operation on the historical product promotion information as key features; and combine the key features to generate a numerical user feature embedding vector;

[0208] S2: Use multiple attention heads in the information-level multi-head self-attention network to process the user feature embedding vectors respectively, learn and use the interaction relationship between different key features, enhance the key features, and obtain enhanced key features; and combine the enhanced key features of multiple attention heads to obtain multi-head representation data of the key features; wherein, the multi-head representation data of a key feature corresponds to at least one historical product promotion information;

[0209] S3: Use the additional information attention network to process the multi-head representation data of key features, learn and use the influence weight of other historical product promotion information on the target user's behavioral operations regarding the historical product promotion information; and jointly use the corresponding multi-head representation data of key features and the influence weight to generate the target user's behavioral representation data corresponding to the target user's historical behavioral data.

[0210] Based on the above embodiment, the user behavior encoding module in the preset product promotion information recommendation model can be used to perform a temporal analysis on the user's behavior within a time period, and at the same time, combined with the user's attribute information, to generate more detailed and reliable behavior representation data about the user's behavior.

[0211] In specific implementation, the feature extraction layer can be used to first process the historical behavior data of the target user in the current time period, and filter and extract the required key features. Among them, the above key features at least include the target user's operation features for product promotion information; further, the above key features can also include the attribute features of the target user, and the features of the product promotion information accepted and browsed by the target user (including: product promotion information content data, form data, product data); then convert the above key features into corresponding numerical data, and combine the above numerical data in sequence to obtain the corresponding user feature embedding vector. For example, it can be expressed as [r1, r2, ..., r M ]. Among them, r M Indicates the numerical data corresponding to the key feature numbered M, where M is the total number of key features.

[0212] In the specific implementation, considering that the product promotion information that the target user is willing to click and browse in the current time period often has a certain correlation, each of the multiple attention heads in the information-level multi-head self-attention network can be used to learn the interaction relationship between different key features respectively, and determine the interaction relationship parameters between different key features; then use the interaction relationship parameters between different key features to process the numerical data of each key feature, and obtain enhanced key features that can be associated with key features of other product promotion information; then combine the enhanced key features of multiple attention heads to obtain multi-head representation data of the key features.

[0213] Specifically, for example, the interaction relationship parameter of the attention head numbered k with respect to the key feature numbered i can be calculated according to the following formula:

[0214]

[0215] in, is the interaction relationship parameter of the attention head numbered k with respect to the key feature numbered j with respect to the key feature numbered i, is the first projection parameter of the attention head numbered k.

[0216] Furthermore, the enhanced key features of the attention head numbered k with respect to the key features numbered i can be calculated according to the following formula:

[0217]

[0218] in, is the enhanced key feature of the attention head numbered k about the key feature numbered i, is the second projection parameter of the attention head numbered k.

[0219] Then, the enhanced key features of the key feature numbered i are combined and used by multiple attention heads in the following manner to construct multi-head representation data of the key feature numbered i: Where h is the total number of attention heads.

[0220] In specific implementation, the additional information attention network can be used to first use the corresponding network parameters and operation vectors to separately process the multi-head representation data of each key feature to obtain the initial weight of each key feature; then the initial weights of multiple key features are used jointly to determine the influence weight of each key feature; finally, the multi-head representation data of the corresponding key features and the influence weights of the key features are used jointly through weighted processing to obtain the behavior representation data corresponding to the behavior data of the target user.

[0221] Specifically, for example, the initial weight of the key feature numbered i can be calculated according to the following formula:

[0222]

[0223] in, is the initial weight of the key feature numbered i, is the multi-head characterization data of the key feature numbered i, q n is the operation vector, V n is the first network parameter, v m is the second network parameter.

[0224] Then calculate the influence weight of the key feature numbered i according to the following formula:

[0225]

[0226] in, is the influence weight of the key feature numbered i, is the initial weight of the key feature numbered i, is the initial weight of the key feature numbered j.

[0227] Finally, the corresponding behavior representation data is constructed in the following way:

[0228] In some embodiments, see Figure 2 As shown, the prediction module at least includes a first operation structure, a second operation structure, a third operation structure connected in parallel, and a fourth operation structure connected in series with the first operation structure, the second operation structure, and the third operation structure.

[0229] Accordingly, the above-mentioned prediction module in the preset product promotion information recommendation model is used to process the content representation data, the form representation data, the product representation data, and the behavior representation data of the target user to obtain the corresponding target processing result. When implemented specifically, it may include:

[0230] The first operation structure is used to process the content characterization data and the behavior characterization data of the target user to obtain a corresponding first operation result; the second operation structure is used to process the form characterization data and the behavior characterization data of the target user to obtain a corresponding second operation result; the third operation structure is used to process the product characterization data and the behavior characterization data of the target user to obtain a corresponding third operation result; the fourth operation structure is used to generate a corresponding target processing result based on the first operation result, the second operation result, and the third operation result.

[0231] Based on the above embodiment, the prediction module in the preset product promotion information recommendation model can be used to better access and process the representation data output by the promotion information encoding module and the user behavior encoding module to obtain a target processing result with a higher reference machine value.

[0232] In a specific implementation, the first operation structure, the second operation structure, and the third operation structure can respectively perform inner product operations on two input representation data to obtain corresponding operation results.

[0233] For example, the first operation structure may perform an inner product operation on the content representation data and the behavior representation data according to the following formula to obtain a corresponding first operation result:

[0234] y1=u T r c

[0235] Among them, y1 identifies the first operation result, u is the behavior representation data, and r is the content representation data.

[0236] In specific implementation, the fourth operation structure can distinguish the influence of different operation results according to the learned knowledge, and determine the weight coefficients for different operation results; and then perform weighted operations according to the corresponding weight coefficients and operation results to obtain the final target processing result.

[0237] Generally, when the data value of the target processing result is larger, it means that the target user is more interested in the target promotion information, the more it matches the target promotion information, and the probability of clicking on the target promotion information is relatively higher. On the contrary, when the data value of the target processing result is smaller, it means that the target user is less interested in the target promotion information, the less it matches the target promotion information, and the probability of clicking on the target promotion information is relatively lower.

[0238] In some embodiments, during specific implementation, it can be detected whether the target processing result is greater than a preset matching threshold; when it is determined to be greater than the preset matching threshold, it is determined that the target promotion information matches the target user.

[0239] On the contrary, when it is determined that the value is less than or equal to the preset matching threshold, it is determined that the target promotion information does not match the target user.

[0240] The preset matching threshold is determined by statistics and learning of relevant data of a large number of test users.

[0241] In some embodiments, the target processing result may also carry a first operation result, a second operation result, and a third operation result;

[0242] Accordingly, after determining whether the target promotion information matches the target user based on the target processing result, refer to Figure 5 As shown, when the method is implemented specifically, it may also include the following contents:

[0243] S1: When it is determined that the target promotion information does not match the target user, detecting whether the first operation result, the second operation result, and the third operation result meet the preset requirements;

[0244] S2: when it is determined that the first operation result and the third operation result meet the preset requirements, and the second operation result does not meet the preset requirements, obtaining the behavior representation data of the target user;

[0245] S3: According to the behavior representation data of the target user, a matching target information form template is determined from a preset information form template set;

[0246] S4: using the target information form template, modifying the form of the target promotion information to obtain the modified target promotion information;

[0247] S5: Push the modified target promotion information to the target users.

[0248] Based on the above embodiment, when it is determined that the target promotion information does not match the target user, the specific cause of the mismatch can be further determined from three levels, namely, the content, form, and product of the target promotion information, based on the first operation result, the second operation result, and the third operation result carried by the target processing result; and targeted processing can then be performed.

[0249] In specific implementation, when it is determined that the first operation result and the third operation result meet the preset requirements, and the second operation result does not meet the preset requirements, it can be judged that the target user is interested in the content of the target promotion information and the target product itself, but because the form of the current target promotion information does not match the personal preference of the target user, the target user is unwilling to click to browse the target promotion information.

[0250] At this time, according to the target user's behavior characterization data, a matching Yisuhu information form template set can be determined from the preset information form template set as the target information form template; then, according to the target form template, the form of the target promotion information can be modified in a targeted manner, for example, the font, color tone, and wording text of the target promotion information can be modified to obtain the modified target promotion information that matches the target user's personal preference. Then, the modified target promotion information is pushed to the target user, and the target user is more likely to be willing to click and browse.

[0251] The preset information form template set includes a plurality of preset information form templates, each of which corresponds to a user behavior representation data range.

[0252] Before specific implementation, a large number of information forms that are of interest to different users, as well as user behavior representation data, can be collected; and the information form and behavior representation data of the same user can be combined into a data body; multiple data bodies are then clustered to obtain multiple data groups; wherein one data group corresponds to one type of user; common features in the information forms contained in the data body in each data group are extracted and combined to obtain a preset information form template corresponding to the data group; at the same time, based on the user behavior representation data contained in the data body in the data group, the behavior representation data range corresponding to this type of user is determined by calculating and utilizing the corresponding data statistical distribution; and the user behavior representation data range corresponding to the same data group is then associated with the preset information form template to obtain a preset information form template set.

[0253] In specific implementation, if it is determined that the third operation result does not meet the preset requirements, it can be determined that the target user is not interested in the target product. At this time, if the target promotion information is pushed to the target user, on the one hand, the user has only a small probability of clicking to browse; on the other hand, even if the user clicks to browse, he will not pay attention to or buy the relevant target product, and may even make the user feel bored because of being pushed products that he is not interested in, affecting the user experience.

[0254] Therefore, when it is determined that the third operation result does not meet the preset requirements, it is possible to directly give up pushing the target promotion information to the target user, detect whether there is product promotion information matching the target user in the remaining other product promotion information, and push it to the target user.

[0255] In specific implementation, when it is determined that the first operation result does not meet the preset requirements, and the third operation result meets the preset requirements, it can be judged that the target user itself is interested in the target product, but the content in the target promotion information may prevent the target user from better understanding the target product, or may cause annoyance. At this time, it is possible to obtain and determine whether the target user is a high-value user based on the attribute information of the target user (for example, membership level information, user tags, etc.); when it is determined that the target user is a high-value user, the historical behavior data, target promotion information, and target product data of the target user in the current time period are sent to the promotion message generation end, so that the promotion message generation end generates updated target promotion information about the target product that matches the target user based on the above data; and then pushes the updated target promotion information to the target user. In this way, the order conversion rate can be effectively improved, the loss of high-value users can be avoided, and a relatively good push effect can be obtained.

[0256] In some embodiments, after obtaining the historical behavior data of the target user in the current time period according to the user identification of the target user, referring to Figure 6 As shown, when the method is implemented specifically, it may also include the following contents:

[0257] S1: Detect whether the amount of data contained in the historical behavior data of the target user in the current time period is greater than or equal to a preset data amount threshold;

[0258] S2: when it is determined that the amount of data contained in the historical behavior data of the target user in the current time period is less than a preset data amount threshold, determining target neighboring users of the target user;

[0259] S3: Obtain historical behavior data of target neighboring users in the current time period;

[0260] S4: combining the historical behavior data of the target user in the current time period and the historical behavior data of the target neighboring users in the current time period to serve as the expanded historical behavior data of the target user in the current time period.

[0261] The target neighboring users may be specifically understood as other users whose attribute information has a similarity with the target user greater than a preset similarity threshold and have a high reference value for the behavior prediction of the target user.

[0262] In specific implementation, when it is detected that the amount of data contained in the historical behavior data of the target user in the current time period is less than the preset data amount threshold, it can be judged that the historical behavior data of the target user in the current time period currently used for prediction is too sparse, and it is impossible to obtain high-precision behavior characterization data based on the above historical behavior data.

[0263] At this time, the server can query the platform user database based on the attribute information of the target user, and determine other users who are similar to the target user in attribute information as target neighboring users; then obtain the historical behavior data of the target neighboring users in the current time period; and by combining the historical behavior data of the target user in the current time period and the historical behavior data of the target neighboring users in the current time period, effectively expand the historical behavior data of the target user in the current time period, and obtain the expanded historical behavior data of the target user in the current time period with a richer and more substantial data volume. Subsequently, based on the above-mentioned expanded historical behavior data of the target user in the current time period, more accurate behavior representation data of the target user with higher reference value can be obtained.

[0264] Based on the above embodiments, the problem that the amount of historical behavior data of some target users in the current time period is small, resulting in sparse data and affecting the accuracy of subsequent prediction processing can be effectively solved.

[0265] In some embodiments, when the query platform database fails to find the target neighboring user that meets the requirements, the method further includes: generating and sending a data expansion assistance request to the partner platform with which the cooperation relationship exists according to the preset cooperation agreement; wherein the data expansion assistance request carries at least the attribute information of the target user, but does not carry the user identification of the target user. The partner platform receives and responds to the data expansion assistance request, queries the user database of the platform according to the attribute information of the target user, determines the target neighboring user that meets the requirements; and obtains the behavior data of the target neighboring user in the current time period; and then feeds back the behavior data of the target neighboring user in the current time period without the user identification of the target neighboring user to the server for data expansion.

[0266] In some embodiments, see Figure 7 As shown, when the method is implemented specifically, it may also include the following contents:

[0267] S1: Obtain sample behavior records, sample promotion information, product data of sample products of sample users, as well as user historical behavior data, historical product promotion information, and product data of historical products in a specified time period;

[0268] S2: construct a first training set based on sample user's sample behavior records, sample promotion information, and product data of sample products; construct a second training set based on user's historical behavior data, historical product promotion information, and product data of historical products in a specified time period;

[0269] S3: construct an initial model based on a multi-head self-attention network; wherein the initial model at least includes an initial promotion information encoding module, an initial user behavior encoding module, and an initial prediction module;

[0270] S4: Use Spark to call the first training set, train the initial model, and obtain an intermediate model that meets the requirements;

[0271] S5: Use Flink to call the second training set, train the intermediate model, and obtain a preset product promotion information recommendation model that meets the requirements.

[0272] The spark mentioned above is specifically a batch processing-based framework, which is suitable for batch processing of offline data sets and has high processing efficiency.

[0273] The above-mentioned Flink is specifically an event-driven stream processing framework, which is suitable for real-time processing of data streams, supports stateful computing, and has good real-time responsiveness.

[0274] The above-mentioned specified time period may be the most recent month, or the most recent year, etc.

[0275] The first training set mentioned above can be specifically understood as a relatively large amount of stock data obtained based on historical data and test data.

[0276] The second training set mentioned above can be specifically understood as incremental data with a relatively small amount of data accessed and processed by the platform in a recent specified time period.

[0277] In specific implementation, by using spark to call the first training set with a large amount of data to train the initial model, it is possible to effectively use the advantages and characteristics of spark in batch processing of stock data, and more efficiently complete the first stage of training to obtain an intermediate model. Then, by using Flink to call the second training set with a smaller amount of data, further model training is performed on the basis of the initial model, and it is possible to effectively use the advantages and characteristics of Flink in real-time processing of incremental data, and more accurately complete the second stage of training, and obtain a model that is more compatible with the latest actual business scenarios and has a higher reference value, as a preset product promotion information recommendation model that meets the requirements.

[0278] Based on the above embodiment, by acquiring and using two different training data, the first training set and the second training set, and introducing and utilizing two different processing frameworks, spark and Flink, it is possible to fully and effectively utilize the advantages of different training data and different processing frameworks, and efficiently train to obtain a preset product promotion information recommendation model with higher accuracy and better effect.

[0279] In specific implementation, after using Spark to call the first training set, train the initial model, and obtain an intermediate model that meets the requirements, the intermediate model can be stored in a shared storage area based on Kafka. In this way, Flink can be used to access the shared storage area and efficiently call the intermediate model for further model training.

[0280] Among them, the above-mentioned Kafka is an open source stream processing platform, which can be specifically a high-throughput distributed publish-subscribe messaging system.

[0281] In some embodiments, the above-mentioned construction of the initial model based on the multi-head self-attention network may include the following contents during specific implementation:

[0282] S1: Determine the number of attention heads in the word-level multi-head self-attention network according to the sample promotion information and product data of the sample product; and configure the initial promotion information encoding module according to the number of attention heads;

[0283] S2: Determine the number of attention heads in the information-level multi-head self-attention network according to the sample behavior records of the sample users; and configure the initial user behavior encoding module according to the number of attention heads;

[0284] S3: configure an initial prediction module; and combine an initial promotion information encoding module, an initial user behavior encoding module, and an initial prediction module to obtain an initial model.

[0285] Based on the above embodiments, an initial model that is more compatible with the actual business scenario and has better effects can be constructed.

[0286] In some embodiments, after obtaining the user's historical behavior data, historical product promotion information, and historical product data for a specified time period, the method may further include the following when implemented:

[0287] Use Flink to preprocess user historical behavior data, historical product promotion information, and historical product data for a specified time period.

[0288] In specific implementation, Flink can be used to perform pre-processing such as data cleaning and format unification on historical user behavior data, historical product promotion information, and historical product data for a specified time period.

[0289] Based on the above embodiment, preprocessed data with higher accuracy and less noise can be obtained, and then the preprocessed data can be used as a second training set to train a preset product promotion information recommendation model with better effect.

[0290] In some embodiments, after obtaining a preset product promotion information recommendation model that meets the requirements, the method may further include the following when implemented:

[0291] S1: query the platform database at every specified time interval to obtain the historical product promotion information and historical promotion records of the current specified time period;

[0292] S2: Based on the historical promotion records of the current specified time period, obtain the user historical behavior data of the current specified time period; based on the historical product promotion information of the current specified time period, obtain the product data of the historical products of the current specified time period;

[0293] S3: Use Flink to call the user historical behavior data of the current specified time period, the historical product promotion information of the current specified time period, and the product data of the historical products of the current specified time period to train the preset product promotion information recommendation model to obtain an updated preset product promotion information recommendation model.

[0294] Based on the above embodiment, Flink can be used to call the user historical behavior data of the current specified time period, the historical product promotion information of the current specified time period, and the product data of the historical products of the current specified time period at every specified time period, and the currently used preset product promotion information recommendation model can be trained and updated online, so that the preset product promotion information recommendation model used by the server can always better fit the latest actual business scenarios, and then the preset product promotion information recommendation model can be used to continuously recommend to users product promotion information that users are interested in and willing to click on and browse, thereby improving user experience.

[0295] As can be seen from the above, the data processing method for product promotion information provided in the embodiments of this specification can, before being implemented, use a multi-head self-attention network to construct and train to obtain information representation data about the product promotion information that is relatively comprehensive and rich based on multiple dimensions such as the content, form, and product of the product promotion information, as well as behavior representation data about the user based on time series, and then be able to accurately predict the matching between the user and the product promotion information based on the above information representation data and behavior representation data. Preset product promotion information recommendation model.

[0296] In specific implementation, the target promotion information of the target product and the user identifier of the target user can be obtained first; according to the user identifier of the target user, the historical behavior data of the target user in the current time period can be obtained; according to the target promotion information, the product identifier of the target product is extracted to describe the target content data of the target product, and the target form data for displaying the target product; at the same time, according to the product identifier of the target product, the target product data of the target product is queried; the promotion information encoding module in the preset product promotion information recommendation model is used to process the target content data, the target form data, and the target product data respectively to obtain the corresponding content representation data, form representation data, and product representation data; the user behavior encoding module in the preset product promotion information recommendation model is used to process the historical behavior data of the target user in the current time period to obtain the corresponding behavior representation data of the target user; the prediction module in the preset product promotion information recommendation model is used to process the content representation data, form representation data, product representation data, and the behavior representation data of the target user to obtain the corresponding target processing result; then, according to the target processing result, it is judged whether the target promotion information matches the target user; and when a match is determined, the target promotion information is pushed to the target user.

[0297] In this way, the product promotion information that the target users are interested in can be automatically determined more accurately and pushed to the target users in a targeted manner. On the one hand, this can effectively improve the promotion effect of the product, and on the other hand, it can avoid pushing product promotion information that the target users are not interested in to affect the user experience.

[0298] This specification embodiment provides a server, referring to Figure 8 The server includes a network communication port 801, a processor 802 and a memory 803, and the above structures are connected through internal cables so that each structure can perform specific data interaction.

[0299] The network communication port 801 may be used to obtain target promotion information about the target product and the user ID of the target user;

[0300] The processor 802 can be specifically used to obtain the historical behavior data of the target user in the current time period according to the user identifier of the target user; extract the product identifier of the target product according to the target promotion information, which is used to describe the target content data of the target product, and the target form data for displaying the target product; at the same time, query and obtain the target product data of the target product according to the product identifier of the target product; use the promotion information encoding module in the preset product promotion information recommendation model to process the target content data, the target form data, and the target product data respectively to obtain the corresponding content representation data, the form representation data, and the product representation data; use the user behavior encoding module in the preset product promotion information recommendation model to process the historical behavior data of the target user in the current time period to obtain the corresponding behavior representation data of the target user; wherein the preset product promotion information recommendation model is a model obtained by training based on a multi-head self-attention network; use the prediction module in the preset product promotion information recommendation model to process the content representation data, the form representation data, the product representation data, and the behavior representation data of the target user to obtain the corresponding target processing result; determine whether the target promotion information matches the target user according to the target processing result; and push the target promotion information to the target user when it is determined that the target promotion information matches the target user.

[0301] The memory 803 may be specifically used to store corresponding instruction programs, as well as intermediate data such as historical behavior data of the target user in the current time period, target content data, target form data, target product data, etc.

[0302] Based on the above method, the relevant structural performance of the server can be effectively utilized, the data processing speed of the electronic device can be improved, and the data processing of product promotion information can be efficiently realized.

[0303] In this embodiment, the network communication port 801 can be a virtual port that is bound to different communication protocols so that different data can be sent or received. For example, the network communication port can be a port responsible for web data communication, a port responsible for FTP data communication, or a port responsible for email data communication. In addition, the network communication port can also be a physical communication interface or communication chip. For example, it can be a wireless mobile network communication chip, such as GSM, CDMA, etc.; it can also be a Wifi chip; it can also be a Bluetooth chip.

[0304] In this embodiment, the processor 802 may be implemented in any appropriate manner. For example, the processor may take the form of a microprocessor or processor and a computer-readable medium storing a computer-readable program code (such as software or firmware) executable by the (micro)processor, a logic gate, a switch, an application specific integrated circuit (ASIC), a programmable logic controller, and an embedded microcontroller, etc. This specification does not limit this.

[0305] In this embodiment, the memory 803 may include multiple levels. In a digital system, anything that can store binary data can be a memory; in an integrated circuit, a circuit with a storage function that has no physical form is also called a memory, such as RAM, FIFO, etc.; in a system, a storage device with a physical form is also called a memory, such as a memory stick, TF card, etc.

[0306] The embodiment of the present specification also provides a computer-readable storage medium based on the data processing method of the above-mentioned product promotion information, wherein the computer-readable storage medium stores computer program instructions, and when the computer program instructions are executed, the following steps are implemented: obtaining target promotion information about the target product and the user identification of the target user; obtaining the historical behavior data of the target user in the current time period according to the user identification of the target user; extracting the product identification of the target product according to the target promotion information, which is used to describe the target content data of the target product, and the target form data for displaying the target product; at the same time, according to the product identification of the target product, querying and obtaining the target product data of the target product; using the promotion information encoding module in the preset product promotion information recommendation model to process the target content data, target form data and the target content data, respectively. data, and target product data to obtain corresponding content representation data, form representation data, and product representation data; use the user behavior encoding module in the preset product promotion information recommendation model to process the historical behavior data of the target user in the current time period to obtain the corresponding behavior representation data of the target user; wherein the preset product promotion information recommendation model is a model obtained based on multi-head self-attention network training; use the prediction module in the preset product promotion information recommendation model to process the content representation data, form representation data, product representation data, and the behavior representation data of the target user to obtain the corresponding target processing result; according to the target processing result, determine whether the target promotion information matches the target user; when it is determined that the target promotion information matches the target user, push the target promotion information to the target user.

[0307] In this embodiment, the storage medium includes, but is not limited to, a random access memory (RAM), a read-only memory (ROM), a cache, a hard disk (HDD), or a memory card. The memory may be used to store computer program instructions. The network communication unit may be an interface for network connection communication set in accordance with the standard specified by the communication protocol.

[0308] In this embodiment, the functions and effects specifically implemented by the program instructions stored in the computer-readable storage medium can be explained in comparison with other implementations and will not be described in detail here.

[0309] The present specification also provides a computer program product, which at least includes a computer program. When the computer program is executed by a processor, the following method steps are implemented: obtaining target promotion information about a target product and a user identifier of a target user; obtaining historical behavior data of the target user in the current time period according to the user identifier of the target user; extracting a product identifier of the target product according to the target promotion information, target content data for describing the target product, and target form data for displaying the target product; at the same time, according to the product identifier of the target product, querying and obtaining target product data of the target product; using a promotion information encoding module in a preset product promotion information recommendation model to process the target content data, target form data, and target product data respectively, and obtaining Corresponding content representation data, form representation data, and product representation data; using the user behavior encoding module in the preset product promotion information recommendation model to process the historical behavior data of the target user in the current time period, and obtain the corresponding behavior representation data of the target user; wherein the preset product promotion information recommendation model is a model obtained based on multi-head self-attention network training; using the prediction module in the preset product promotion information recommendation model to process the content representation data, form representation data, product representation data, and the behavior representation data of the target user, and obtain the corresponding target processing result; according to the target processing result, determine whether the target promotion information matches the target user; when it is determined that the target promotion information matches the target user, push the target promotion information to the target user.

[0310] See also Fig. 9 As shown, the embodiment of this specification also provides a data processing device for product promotion information, which may specifically include the following structural modules:

[0311] The first acquisition module 901 may be specifically used to acquire target promotion information about a target product and a user ID of a target user;

[0312] The second acquisition module 902 may be specifically used to acquire the historical behavior data of the target user in the current time period according to the user identification of the target user;

[0313] The third acquisition module 903 may be specifically used to extract the product identification of the target product, target content data for describing the target product, and target form data for displaying the target product according to the target promotion information; at the same time, according to the product identification of the target product, query and obtain the target product data of the target product;

[0314] The first processing module 904 may be specifically used to process the target content data, the target form data, and the target product data respectively using the promotion information encoding module in the preset product promotion information recommendation model to obtain corresponding content representation data, form representation data, and product representation data; and process the historical behavior data of the target user in the current time period using the user behavior encoding module in the preset product promotion information recommendation model to obtain corresponding behavior representation data of the target user; wherein the preset product promotion information recommendation model is a model obtained by training based on a multi-head self-attention network;

[0315] The second processing module 905 may be specifically used to process the content representation data, the form representation data, the product representation data, and the behavior representation data of the target user by using the prediction module in the preset product promotion information recommendation model to obtain the corresponding target processing result;

[0316] The determination module 906 may be specifically used to determine whether the target promotion information matches the target user according to the target processing result;

[0317] The push module 907 may be specifically configured to push the target promotion information to the target user when it is determined that the target promotion information matches the target user.

[0318] In some embodiments, the promotion information encoding module may include at least: a word embedding layer, a word-level multi-head self-attention network, an accumulated word attention network and other structures;

[0319] Accordingly, when the above-mentioned first processing module 904 is implemented, the promotion information encoding module in the preset product promotion information recommendation model can be used to process the target content data in the following manner to obtain corresponding content representation data: use the word embedding layer to process the target content data and extract the corresponding content features; and generate a numerical target content embedding vector based on the content features; use multiple attention heads in the word-level multi-head self-attention network to process the target content embedding vector respectively, learn and use the interaction relationship between different content features, and generate intermediate representation data of the content features; and combine the intermediate representation data of the content features of multiple attention heads to obtain multi-head representation data of the content features; use the cumulative word attention network to process the multi-head representation data of the content features and determine the attention weight of the content features; and jointly use the corresponding multi-head representation data of the content features and the attention weight of the content features to generate content representation data corresponding to the target content data.

[0320] In some embodiments, after the intermediate representation data of the content features of multiple attention heads are combined to obtain the multi-head representation data of the content features, the device, when implemented, can also be used to: use a word-level multi-head self-attention network to process the multi-head representation data of content features, the multi-head representation data of form features, and the multi-head representation data of product features, learn and use the interaction between the multi-head representation data of form features and the multi-head representation data of product features and the multi-head representation data of content features, adjust the multi-head representation data of content features, and obtain the adjusted multi-head representation data of content features; wherein the multi-head representation data of form features is generated by processing the target form embedding vector using multiple attention heads in the word-level multi-head self-attention network; the multi-head representation data of product features is generated by processing the target product embedding vector using multiple attention heads in the word-level multi-head self-attention network.

[0321] In some embodiments, the user behavior encoding module may include at least: a feature extraction layer, an information-level multi-head self-attention network, an additional information attention network and other structures;

[0322] Accordingly, when the first processing module 904 is specifically implemented, the user behavior encoding module in the preset product promotion information recommendation model can be used to process the historical behavior data of the target user in the current time period in the following manner to obtain the corresponding behavior representation data of the target user: the historical behavior data of the target user in the current time period is processed by the feature extraction layer, and the user features related to the target user's behavior operation on the historical product promotion information are extracted as key features; and the key features are combined to generate a numerical user feature embedding vector; the user feature embedding vector is processed by multiple attention heads in the information-level multi-head self-attention network respectively, the interaction relationship between different key features is learned and utilized, and the key features are enhanced to obtain the enhanced key features; and the enhanced key features of multiple attention heads are combined to obtain the multi-head representation data of the key features; wherein the multi-head representation data of a key feature corresponds to at least one historical product promotion information; the multi-head representation data of the key feature is processed by the additional information attention network, and the influence weight of other historical product promotion information on the target user's behavior operation on the historical product promotion information is learned and utilized; and the corresponding multi-head representation data and influence weight of the key feature are jointly used to generate the behavior representation data of the target user corresponding to the historical behavior data of the target user.

[0323] In some embodiments, the promotion information encoding module may at least include: a UniLM sub-model, a CNN network, a weight network layer and other structures; wherein the UniLM sub-model includes a plurality of Transformer modules;

[0324] Accordingly, when the above-mentioned first processing module 904 is implemented, the promotion information encoding module in the preset product promotion information recommendation model can be used to process the target content data in the following manner to obtain corresponding content representation data: the UniLM sub-model is used to process the target content data, and the corresponding content vector is obtained by generating and using the corresponding segment vector, word vector and position vector; the content vector is processed by multiple Transformer modules, and the first content deep semantic vector and the second content deep semantic vector with deep bidirectional semantics are generated by extracting and using context information; the first content deep semantic vector and the second content deep semantic vector are processed by the CNN network, and the local features are extracted and used to correct the first content deep semantic vector and the second content deep semantic vector to obtain the corrected first content deep semantic vector and the corrected second content deep semantic vector; the weight network layer is used to perform weighted operations based on the word-level attention mechanism and the item-level attention mechanism according to the corrected first content deep semantic vector and the corrected second content deep semantic vector to obtain content representation data corresponding to the target content data.

[0325] In some embodiments, the promotion information encoding module may further include: a classifier, an initial encoder, and a multi-layer perceptron; accordingly, when the above-mentioned first processing module 904 is implemented, the classifier may also be used to process the target content data, the target form data, and the target product data to determine the promotion information type and the promotion information subtype of the target promotion information; the initial encoder may be used to perform one-hot encoding according to the promotion information type and the promotion information subtype of the target promotion information, respectively, to obtain the corresponding promotion information type encoding vector and the promotion information subtype encoding vector; the multi-layer perceptron may be used to process the promotion information type encoding vector and the promotion information subtype encoding vector to obtain the corresponding promotion information type features and the promotion information subtype features; the weight network layer may be used to perform weighted operations based on the word-level attention mechanism and the item-level attention mechanism according to the corrected first content deep semantic vector, the corrected second content deep semantic vector, the promotion information type features, and the promotion information subtype features to obtain content representation data corresponding to the target content data.

[0326] In some embodiments, the prediction module may include at least a first operation structure, a second operation structure, a third operation structure connected in parallel, and a fourth operation structure connected in series with the first operation structure, the second operation structure, and the third operation structure;

[0327] Accordingly, when the above-mentioned second processing module 905 is implemented, the prediction module in the preset product promotion information recommendation model can be used to process the content representation data, the form representation data, the product representation data, and the behavior representation data of the target user in the following manner to obtain the corresponding target processing result: the first operation structure is used to process the content representation data and the behavior representation data of the target user to obtain the corresponding first operation result; the second operation structure is used to process the form representation data and the behavior representation data of the target user to obtain the corresponding second operation result; the third operation structure is used to process the product representation data and the behavior representation data of the target user to obtain the corresponding third operation result; the fourth operation structure is used to generate the corresponding target processing result according to the first operation result, the second operation result, and the third operation result.

[0328] In some embodiments, the target processing result also carries a first operation result, a second operation result, and a third operation result;

[0329] Correspondingly, after determining whether the target promotion information matches the target user based on the target processing result, the device, when implemented specifically, can also be used to: when it is determined that the target promotion information does not match the target user, detect whether the first operation result, the second operation result, and the third operation result meet the preset requirements; when it is determined that the first operation result and the third operation result meet the preset requirements, and the second operation result does not meet the preset requirements, obtain the behavior characterization data of the target user; determine a matching target information form template from a preset information form template set based on the behavior characterization data of the target user; use the target information form template to modify the form of the target promotion information to obtain the modified target promotion information; and push the modified target promotion information to the target user.

[0330] In some embodiments, after obtaining the historical behavior data of the target user in the current time period based on the user identification of the target user, the device, when implemented, can also be used to: detect whether the amount of data contained in the historical behavior data of the target user in the current time period is greater than or equal to a preset data amount threshold; when it is determined that the amount of data contained in the historical behavior data of the target user in the current time period is less than the preset data amount threshold, determine the target neighboring users of the target user; obtain the historical behavior data of the target neighboring users in the current time period; and combine the historical behavior data of the target user in the current time period and the historical behavior data of the target neighboring users in the current time period as the expanded historical behavior data of the target user in the current time period.

[0331] In some embodiments, when the device is implemented, it can also be used to: obtain sample behavior records, sample promotion information, product data of sample products, and user historical behavior data, historical product promotion information, and product data of historical products in a specified time period; construct a first training set based on the sample behavior records, sample promotion information, and product data of sample products in a specified time period; construct a second training set based on user historical behavior data, historical product promotion information, and product data of historical products in a specified time period; construct an initial model based on a multi-head self-attention network; wherein the initial model includes at least an initial promotion information encoding module, an initial user behavior encoding module, and an initial prediction module; use spark to call the first training set, train the initial model, and obtain an intermediate model that meets the requirements; use Flink to call the second training set, train the intermediate model, and obtain a preset product promotion information recommendation model that meets the requirements.

[0332] In some embodiments, when the device is implemented, the initial model based on the multi-head self-attention network can be constructed in the following manner: according to the sample promotion information and the product data of the sample product, the number of attention heads in the word-level multi-head self-attention network is determined; and the initial promotion information encoding module is configured according to the number of attention heads; according to the sample behavior records of the sample users, the number of attention heads in the information-level multi-head self-attention network is determined; and the initial user behavior encoding module is configured according to the number of attention heads; the initial prediction module is configured; and the initial promotion information encoding module, the initial user behavior encoding module, and the initial prediction module are combined to obtain the initial model.

[0333] In some embodiments, after obtaining the user's historical behavior data, historical product promotion information, and product data of historical products for a specified time period, the device can also be used, when implemented, to pre-process the user's historical behavior data, historical product promotion information, and product data of historical products for a specified time period using Flink.

[0334] In some embodiments, after obtaining a preset product promotion information recommendation model that meets the requirements, the device can also be used when it is implemented to: query the platform database at every specified time period to obtain historical product promotion information for the current specified time period, and historical promotion records for the current specified time period; obtain user historical behavior data for the current specified time period based on historical promotion records for the current specified time period; obtain product data of historical products for the current specified time period based on historical product promotion information for the current specified time period; use Flink to call user historical behavior data for the current specified time period, historical product promotion information for the current specified time period, and product data of historical products for the current specified time period to train the preset product promotion information recommendation model, and obtain an updated preset product promotion information recommendation model.

[0335] It should be noted that the units, devices or modules described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. For the convenience of description, the above devices are described separately by functions divided into various modules. Of course, when implementing this specification, the functions of each module can be implemented in the same or more software and / or hardware, or the modules that implement the same function can be implemented by a combination of multiple sub-modules or sub-units. The device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0336] As can be seen from the above, the data processing device for product promotion information provided in the embodiment of this specification can automatically and accurately determine the product promotion information that the target user is interested in, and push it to the target user in a targeted manner. On the one hand, it can effectively improve the promotion effect of the product, and on the other hand, it can avoid pushing product promotion information that the target user is not interested in to the target user, affecting the user experience.

[0337] In a specific scenario example, the data processing method for product promotion information provided in this specification can be applied to recommend financial product advertisements (a type of product promotion information) based on multi-head attention. Fig.10 As shown, the following contents may be included.

[0338] In this scenario example, considering that most of the financial services of many banks (platforms) are delivered online, multiple online channels are used to reach users to achieve functions such as message notifications, advertising push, and customer service replies. In the process of user interaction with the channel, the user's historical behavior is recorded, and the relevance of the user's interests and financial products is analyzed to achieve precise advertising marketing for thousands of people.

[0339] However, existing methods often have the following pain points: (1) Sensitivity to data sparsity. When there are relatively few common interactions between users and financial products, it is difficult for the target user to find neighboring users, and it is difficult to extract meaningful information from a small number of interactions. (2) Limited expressive ability. The ability to capture some duplicate features and relationships is relatively weak. The multi-head attention mechanism can learn more complex feature representations through deep learning models, thereby providing more accurate recommendations. (3) Poor flexibility and scalability. Existing technologies are not convenient to integrate and combine with other models and technologies to meet different recommendation needs. (4) Poor adaptability to new data. When facing new users, existing technologies need a certain amount of time to accumulate enough interaction data to provide accurate recommendations. The multi-head attention mechanism can adapt to new data faster by utilizing other relevant information or training models.

[0340] To address the above pain points, an improved financial product advertising recommendation method based on multi-head attention is proposed, which includes at least the following points:

[0341] (1) Construct a financial product advertisement encoder (e.g., a promotion information encoding module in a preset product promotion information recommendation model). Based on this encoder, key features such as the name of a financial product, historical performance, investment period, issuer reputation, and risk level can be converted into numerical embedding vectors; then input into a multi-head attention network, and add additional layers and optimize the loss function; finally, a vector representation of the product advertisement (e.g., content representation data, form representation data, and product representation data related to the target promotion information) is obtained through offline evaluation.

[0342] (2) Construct a user encoder (for example, a user behavior encoding module in a preset product promotion information recommendation model). Based on this encoder, the interactive feature data of users such as click-through rate, dwell time, conversion rate, product holding time, and rating of product advertisements can be converted into an embedding vector. The RNN sequence model is used to process the data, and a multi-head attention mechanism (such as investment preference and active time period) is introduced to output a comprehensive user representation vector (for example, the behavior representation data of the target user).

[0343] (3) Build a mechanism for jointly processing new and old data. Use Spark to process existing data, Flink to process new data, and train related encoder models for prediction.

[0344] When implementing it, first refer to Fig.10 As shown in the figure, a financial product advertising recommendation process framework is built.

[0345] (1) The bank's clients include mobile banking, personal online banking, bank WeChat public accounts, bank WeChat mini-programs, etc. When users use the banking services provided by these channels, after authorization, the bank system will automatically collect the user's behavior data, such as browsing time, click-through rate, search keywords, purchased products, product ratings, etc.

[0346] (2) The collected user behavior data and financial product advertisements in the bank are used as data sets. The existing data is input into the defined multi-head attention mechanism neural network using Spark for distributed training to obtain a model that can be used initially and stored in the shared storage system Kafka.

[0347] (3) Use Flink to build a predictor for the newly added data, read the preliminary model and new data for training, and optimize the model.

[0348] (4) The target user is input into the multi-head attention network financial product advertising recommendation algorithm module, and the model will generate an advertising recommendation list for the user, thereby achieving personalized financial product advertising recommendations for each individual.

[0349] Next, a financial product advertising recommendation algorithm based on the multi-head attention mechanism can be introduced to complete specific predictions.

[0350] (1) Construct a financial product advertisement encoder module. It has three layers: word embedding layer, word-level multi-attention self-attention network (e.g., word-level multi-head self-attention network), and cumulative word attention network.

[0351] In the word embedding layer, key features such as financial product name, historical performance, risk level, and issuing institution reputation are converted into numerical embedding vectors, and M represents the number of features of financial product advertisements, that is, [e1, e2, ..., e M ].

[0352] In the word-level multi-attention self-attention network, the i-th feature learned by the k-th attention head is calculated as:

[0353]

[0354] in, is the interaction relation parameter of the attention head numbered k with respect to the content feature numbered j on the content feature numbered i, is the first projection parameter of the attention head numbered k. V is the intermediate representation data of the attention head numbered k about the content feature numbered i, k w is the second projection parameter of the attention head numbered k.

[0355] Further, the multi-head representation data of the content feature numbered i: Where h is the total number of attention heads.

[0356] In the cumulative word attention network, different features in the same advertisement may have different importance in representing the advertisement. The weight α of the i-th feature in the advertisement is i w Calculated as:

[0357]

[0358] in, is the initial weight of the content feature numbered i, is the multi-head representation data of the content feature numbered i, q w is the query vector, V m is the first relevant projection parameter, v m is the second related projection parameter. is the attention weight of the content feature numbered i, is the initial weight of the content feature numbered j.

[0359] The final representation of an ad is the weighted sum of the contextual feature representations (content representation data), expressed as:

[0360] When constructing a financial product advertisement encoder module, the following scheme can also be adopted.

[0361] Considering that text, color, and numbers are important components of advertisements, advertisements contain various types of information such as people, objects, text, color, numbers, etc. By analyzing this information, the model can understand the detailed content, related topics, and background of the advertisement. This information helps to learn the deep semantics of the advertisement. Therefore, we can first build a text encoder to capture the semantics of the advertisement, and then use the color-object encoder to capture the impression theme of the advertisement. The features of the text, objects, colors, and people in the advertisement are extracted respectively, and then fused to obtain the financial advertisement representation.

[0362] Specifically, word embedding is the process of converting advertising components into low-dimensional continuous vectors. Traditional word embedding methods use one-hot encoding or TF-IDF to represent components as a sparse vector. However, these static word embedding methods usually ignore the contextual information between components, so that a component is represented by the same vector in any context, which obviously loses useful information. In order to address the shortcomings of traditional word embedding, a pre-trained language model (for example, the UniLM sub-model) can be used to obtain dynamic embedding of components so that the embedded word vector has semantic information at different levels of context.

[0363] Specifically, the component data (e.g., target content data) containing the text, color, object, and person of the advertisement can be input into the UniLM pre-trained model. UniLM adds the input information to the segment vector, word vector, and position vector and converts it into a vector of uniform length (e.g., content vector). After multiple Transformer blocks extract context information, the advertisement theme and the text, color, object, and person of the advertisement are converted into vectors with deep bidirectional semantics. and

[0364] Considering that UniLM uses a lot of corpus information that is not targeted at ad recommendations during pre-training, this may introduce noise. In order to make word vectors more suitable for ad recommendations, we further input the word vectors of ad text, objects, and colors into convolutional neural networks (CNN) to extract local features of ads:

[0365] t″ i =ReLU(w c T i:i+j-1 +bc )

[0366] Among them, w c is the convolution kernel of size h, T i:i+h-1 It means taking the submatrix consisting of rows i to i+h-1 from T. The text, object, and color of the advertisement are represented by extracting local features through CNN. and

[0367] Considering that different words in the advertising text have different importance and information content, for example, in the advertising text "CCB Cloud Tax Loan, helping small and micro enterprises to start a new chapter", the word "Cloud Tax Loan" has a higher importance than the word "small and micro enterprises" because it indicates that the advertisement is related to Cloud Tax Loan. Therefore, this chapter's model designs a word-level attention mechanism to assign different weights to different words, so as to better learn the global vector representation of different text types. Taking financial advertising text as an example, the vector representation of financial advertising text is obtained:

[0368]

[0369] Among them, W a represents the trainable parameters, b a represents the bias, q i Represents the attention query vector. Similarly, the vector representation r of the financial advertisement text can be obtained b .

[0370] Since advertising categories and subcategories generally contain only a few words, one-hot encoding is used to obtain a uniform length vector representation T′ of advertising categories and subcategories. c and T′ sc , in order to extract more additional information, they are input into a multi-layer perceptron to extract advertising category and subcategory features r c and r sc .

[0371] Considering that advertising text, color, objects, and characters have different importance for learning advertising representation, an item-level attention mechanism is introduced at the end of the financial advertising encoder to make the model focus on important information and calculate attention weights for different types of advertising information:

[0372] r′ a =tanh(W r r a +b r )

[0373]

[0374] Among them, W r represents the trainable parameters, b rIndicates bias.

[0375] Similarly, the attention weights α for ad text, category, and subcategory can be calculated b , α c and α sc , perform weighted summation to obtain advertising representation data: x = α a r a +α b r b +α c r c +α sc r sc .

[0376] (2) Construct a user encoder module, which consists of two layers: a financial product advertising-level multi-head self-attention network and an additional advertising attention network.

[0377] In the multi-head self-attention network at the financial product advertising level. Usually, product advertisements that the same user follows may have certain correlations. The representation of advertisements is enhanced by capturing the interaction between product advertisements. The i-th feature learned by the k-th attention head is expressed as:

[0378]

[0379] in, is the interaction relationship parameter of the attention head numbered k with respect to the key feature numbered j with respect to the key feature numbered i, is the first projection parameter of the attention head numbered k. is the enhanced key feature of the attention head numbered k about the key feature numbered i, is the second projection parameter of the attention head numbered k.

[0380] Further, the multi-head characterization data of the key feature numbered i: Where h is the total number of attention heads.

[0381] In the attached ad attention network, different ads may have different amounts of information in representing the user. The weight of the i-th ad is calculated as:

[0382]

[0383] in, is the initial weight of the key feature numbered i, is the multi-head characterization data of the key feature numbered i, q n is the operation vector, V n is the first network parameter, v m is the second network parameter. is the influence weight of the key feature numbered i, is the initial weight of the key feature numbered j.

[0384] The final user representation is the weighted sum of the user's interaction representation with the ad (behavior representation data), and its representation is:

[0385] (3) Click predictor (prediction module): The click predictor module is used to predict the probability of a user clicking on a candidate ad. c Represented as r c Next, the click probability score y (target processing result) is calculated by the inner product of the user representation vector and the advertisement representation vector, that is, y = u T r c .

[0386] Through the above scenario examples, it is verified that the data processing method for product promotion information provided in this manual has the following advantages: when processing sparse data, the multi-attention network mechanism can better fill in missing values ​​through learned features; it has strong flexibility and scalability. Through the multi-head attention mechanism, multiple attention vectors can be generated to provide richer feature representations. Since the number of heads is configurable, the appropriate number of attention heads can be flexibly selected according to the complexity of the task and computing resources, and it is also more convenient for modular design; the user vector is extracted using a time series model, which can more accurately capture user interests and predict click-through rates; it has strong adaptability to new data. Use Flink to preprocess real-time data and input it into the model for real-time prediction. According to the real-time data and prediction results, the network model is regularly updated and the parameters are fine-tuned.

[0387] Although the present specification provides method operation steps as described in the embodiments or flow charts, more or less operation steps may be included based on conventional or non-creative means. The order of steps listed in the embodiments is only one way of executing the order of many steps, and does not represent a unique execution order. When the device or client product in practice is executed, it can be executed in sequence or in parallel according to the method shown in the embodiments or the drawings (for example, a parallel processor or a multi-threaded processing environment, or even a distributed data processing environment). The term "include", "include" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, product or device including a series of elements includes not only those elements, but also includes other elements that are not explicitly listed, or also includes elements inherent to such a process, method, product or device. In the absence of more restrictions, it is not excluded that there are other identical or equivalent elements in the process, method, product or device including the elements. The first, second, etc. words are used to represent the name, and do not represent any particular order.

[0388] Those skilled in the art also know that, in addition to implementing the controller in a purely computer-readable program code, the controller can be made to implement the same function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, such a controller can be considered as a hardware component, and the devices for implementing various functions included therein can also be considered as structures within the hardware component. Or even, the devices for implementing various functions can be considered as both software modules for implementing the method and structures within the hardware component.

[0389] This specification may be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, classes, etc. that perform specific tasks or implement specific abstract data types. This specification may also be practiced in distributed computing environments where tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules may be located in local and remote computer-readable storage media including storage devices.

[0390] Through the description of the above embodiments, it can be known that those skilled in the art can clearly understand that the present specification can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present specification can essentially be embodied in the form of a software product, which can be stored in a storage medium such as ROM / RAM, a disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a mobile terminal, a server, or a network device, etc.) to execute the methods described in each embodiment of the present specification or some parts of the embodiments.

[0391] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. This specification can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld devices or portable devices, tablet devices, multi-processor systems, microprocessor-based systems, set-top boxes, programmable electronic devices, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, etc.

[0392] Although the present specification is described through embodiments, those skilled in the art will appreciate that there are many modifications and changes to the present specification without departing from the spirit of the present specification, and it is intended that the appended claims include these modifications and changes without departing from the spirit of the present specification.

Claims

1. A data processing method for product promotion information, It is characterized in that include: Obtaining targeted promotion information about the target product and the user ID of the target user; According to the user ID of the target user, obtain the historical behavior data of the target user in the current time period; According to the target promotion information, the product identification of the target product is extracted, which is used to describe the target content data of the target product, and the target form data for displaying the target product; at the same time, according to the product identification of the target product, the target product data of the target product is queried and obtained; The promotion information encoding module in the preset product promotion information recommendation model is used to process the target content data, the target form data, and the target product data respectively to obtain the corresponding content representation data, form representation data, and product representation data; the user behavior encoding module in the preset product promotion information recommendation model is used to process the historical behavior data of the target user in the current time period to obtain the corresponding behavior representation data of the target user; wherein the preset product promotion information recommendation model is a model obtained by training based on a multi-head self-attention network; Using the prediction module in the preset product promotion information recommendation model to process the content representation data, form representation data, product representation data, and target user behavior representation data, to obtain the corresponding target processing results; According to the target processing results, determine whether the target promotion information matches the target user; When it is determined that the target promotion information matches the target user, the target promotion information is pushed to the target user.

2. The method according to claim 1, It is characterized in that The promotion information encoding module at least includes: a word embedding layer, a word-level multi-head self-attention network, and an accumulated word attention network; Accordingly, the promotion information encoding module in the preset product promotion information recommendation model is used to process the target content data to obtain corresponding content representation data, including: The target content data is processed using the word embedding layer to extract the corresponding content features; and a numerical target content embedding vector is generated based on the content features; Using multiple attention heads in a word-level multi-head self-attention network to process the target content embedding vector respectively, learning and utilizing the interaction relationship between different content features, generating intermediate representation data of the content features; and combining the intermediate representation data of the content features of multiple attention heads to obtain multi-head representation data of the content features; The multi-head representation data of content features are processed by using an accumulated word attention network to determine the attention weight of the content features; and the multi-head representation data of corresponding content features and the attention weight of the content features are jointly used to generate content representation data corresponding to the target content data.

3. The method according to claim 1, It is characterized in that The promotion information encoding module at least includes: a UniLM sub-model, a CNN network, and a weight network layer; wherein the UniLM sub-model includes a plurality of Transformer modules; Accordingly, the promotion information encoding module in the preset product promotion information recommendation model is used to process the target content data to obtain corresponding content representation data, including: The target content data is processed using the UniLM sub-model, and the corresponding content vector is obtained by generating and using the corresponding segment vector, word vector and position vector; Processing the content vector using a plurality of Transformer modules, and generating a first content deep semantic vector and a second content deep semantic vector having deep bidirectional semantics by extracting and using context information; Processing the first content depth semantic vector and the second content depth semantic vector using a CNN network, extracting and using local features to correct the first content depth semantic vector and the second content depth semantic vector, to obtain a corrected first content depth semantic vector and a corrected second content depth semantic vector; The weighted network layer is used to obtain content representation data corresponding to the target content data by performing weighted operations based on a word-level attention mechanism and an item-level attention mechanism according to the revised first content deep semantic vector and the revised second content deep semantic vector.

4. The method according to claim 3, It is characterized in that The promotion information encoding module also includes: a classifier, an initial encoder, and a multi-layer perceptron; Accordingly, the method further includes: Processing the target content data, the target form data, and the target product data using a classifier to determine a promotion information type and a promotion information subtype of the target promotion information; Using the initial encoder, one-hot encoding is performed according to the promotion information type and the promotion information subtype of the target promotion information, respectively, to obtain a corresponding promotion information type encoding vector and promotion information subtype encoding vector; Using a multi-layer perceptron to process the promotion information type encoding vector and the promotion information sub-type encoding vector, to obtain corresponding promotion information type features and promotion information sub-type features; The weighted network layer is used to obtain content representation data corresponding to the target content data by performing weighted operations based on the word-level attention mechanism and the item-level attention mechanism according to the corrected first content deep semantic vector, the corrected second content deep semantic vector, the promotion information type features and the promotion information subtype features.

5. The method according to claim 1, It is characterized in that The user behavior encoding module at least includes: a feature extraction layer, an information-level multi-head self-attention network, and an additional information attention network; Accordingly, the user behavior encoding module in the preset product promotion information recommendation model is used to process the historical behavior data of the target user in the current time period to obtain the corresponding behavior representation data of the target user, including: Using the feature extraction layer to process the historical behavior data of the target user in the current time period, extracting the user features related to the target user's behavior operations on the historical product promotion information as key features; and combining the key features to generate a numerical user feature embedding vector; The user feature embedding vectors are processed by using multiple attention heads in the information-level multi-head self-attention network, and the interaction relationship between different key features is learned and utilized to enhance the key features to obtain enhanced key features; and the enhanced key features of multiple attention heads are combined to obtain multi-head representation data of the key features; wherein the multi-head representation data of a key feature corresponds to at least one historical product promotion information; The additional information attention network is used to process the multi-head representation data of key features, and the influence weight of other historical product promotion information on the target user's behavioral operations regarding the historical product promotion information is learned and utilized; and the corresponding multi-head representation data of key features and the influence weight are jointly used to generate the target user's behavioral representation data corresponding to the target user's historical behavioral data.

6. The method according to claim 1, It is characterized in that The prediction module at least includes a first operation structure, a second operation structure, a third operation structure connected in parallel, and a fourth operation structure connected in series with the first operation structure, the second operation structure, and the third operation structure; Accordingly, the prediction module in the preset product promotion information recommendation model is used to process the content representation data, form representation data, product representation data, and target user behavior representation data to obtain corresponding target processing results, including: The first operation structure is used to process the content characterization data and the behavior characterization data of the target user to obtain a corresponding first operation result; the second operation structure is used to process the form characterization data and the behavior characterization data of the target user to obtain a corresponding second operation result; the third operation structure is used to process the product characterization data and the behavior characterization data of the target user to obtain a corresponding third operation result; the fourth operation structure is used to generate a corresponding target processing result based on the first operation result, the second operation result, and the third operation result.

7. The method according to claim 6, It is characterized in that The target processing result also carries a first operation result, a second operation result, and a third operation result; Accordingly, after determining whether the target promotion information matches the target user according to the target processing result, the method further includes: In the case where it is determined that the target promotion information does not match the target user, detecting whether the first operation result, the second operation result, and the third operation result meet the preset requirements; When it is determined that the first operation result and the third operation result meet the preset requirements, and the second operation result does not meet the preset requirements, obtaining the behavior representation data of the target user; According to the behavior representation data of the target user, a matching target information form template is determined from a preset information form template set; Using the target information form template, modify the target promotion information in a form to obtain the modified target promotion information; Push the modified target promotion information to target users.

8. The method according to claim 1, It is characterized in that After acquiring historical behavior data of the target user in the current time period according to the user identifier of the target user, the method further includes: Check whether the amount of data contained in the historical behavior data of the target user in the current time period is greater than or equal to a preset data amount threshold; When it is determined that the amount of data contained in the historical behavior data of the target user in the current time period is less than a preset data amount threshold, determining target neighboring users of the target user; Obtain historical behavior data of target neighboring users in the current time period; The historical behavior data of the target user in the current time period and the historical behavior data of the target neighboring users in the current time period are combined to serve as the expanded historical behavior data of the target user in the current time period.

9. The method according to claim 1, It is characterized in that The method further comprises: Obtain sample behavior records, sample promotion information, product data of sample products of sample users, as well as user historical behavior data, historical product promotion information, and product data of historical products in a specified time period; A first training set is constructed based on sample behavior records of sample users, sample promotion information, and product data of sample products; a second training set is constructed based on user historical behavior data, historical product promotion information, and product data of historical products in a specified time period; Constructing an initial model based on a multi-head self-attention network; wherein the initial model at least includes an initial promotion information encoding module, an initial user behavior encoding module, and an initial prediction module; Use Spark to call the first training set, train the initial model, and obtain an intermediate model that meets the requirements; Use Flink to call the second training set, train the intermediate model, and obtain a preset product promotion information recommendation model that meets the requirements.

10. The method according to claim 9, It is characterized in that Build an initial model based on a multi-head self-attention network, including: According to the sample promotion information and the product data of the sample product, the number of attention heads in the word-level multi-head self-attention network is determined; and the initial promotion information encoding module is configured according to the number of attention heads; According to the sample behavior records of the sample users, the number of attention heads in the information-level multi-head self-attention network is determined; and the initial user behavior encoding module is configured according to the number of attention heads; An initial prediction module is configured; and an initial promotion information encoding module, an initial user behavior encoding module, and an initial prediction module are combined to obtain an initial model.

11. The method according to claim 9, It is characterized in that After obtaining the user's historical behavior data, historical product promotion information, and historical product data for a specified time period, the method further includes: Use Flink to preprocess user historical behavior data, historical product promotion information, and historical product data for a specified time period.

12. The method according to claim 9, It is characterized in that After obtaining a preset product promotion information recommendation model that meets the requirements, the method further includes: At every specified time period, query the platform database to obtain historical product promotion information and historical promotion records for the current specified time period; According to the historical promotion records of the current specified time period, obtain the user historical behavior data of the current specified time period; according to the historical product promotion information of the current specified time period, obtain the product data of the historical products of the current specified time period; Flink is used to call the user historical behavior data of the current specified time period, the historical product promotion information of the current specified time period, and the product data of the historical products of the current specified time period to train the preset product promotion information recommendation model, and obtain the updated preset product promotion information recommendation model.

13. A data processing device for product promotion information, It is characterized in that include: A first acquisition module is used to acquire target promotion information about a target product and a user ID of a target user; The second acquisition module is used to acquire the historical behavior data of the target user in the current time period according to the user identification of the target user; The third acquisition module is used to extract the product identification of the target product, the target content data for describing the target product, and the target form data for displaying the target product according to the target promotion information; at the same time, according to the product identification of the target product, query and obtain the target product data of the target product; The first processing module is used to use the promotion information encoding module in the preset product promotion information recommendation model to process the target content data, the target form data, and the target product data respectively to obtain the corresponding content representation data, form representation data, and product representation data; use the user behavior encoding module in the preset product promotion information recommendation model to process the historical behavior data of the target user in the current time period to obtain the corresponding behavior representation data of the target user; wherein the preset product promotion information recommendation model is a model obtained by training based on a multi-head self-attention network; The second processing module is used to process the content representation data, the form representation data, the product representation data, and the behavior representation data of the target user by using the prediction module in the preset product promotion information recommendation model to obtain the corresponding target processing result; A determination module is used to determine whether the target promotion information matches the target user according to the target processing result; The push module is used to push the target promotion information to the target user when it is determined that the target promotion information matches the target user.

14. A server, It is characterized in that The method comprises a processor and a memory for storing processor-executable instructions, wherein the processor implements the steps of the method according to any one of claims 1 to 12 when executing the instructions.

15. A computer-readable storage medium, It is characterized in that Computer instructions are stored thereon, and when the instructions are executed by a processor, the steps of the method according to any one of claims 1 to 12 are implemented.

16. A computer program product, It is characterized in that The invention comprises a computer program, which implements the steps of the method according to any one of claims 1 to 12 when being executed by a processor.

Citation Information

Cited By

  • Small program message pushing method

    CN121125827A