Resource recommendation method and device, equipment, storage medium and program product

Abstract

The application discloses a resource recommendation method and device, equipment, a storage medium and a program product. The method comprises the following steps: acquiring prediction data, the prediction data being used for predicting the interaction probability of a first account and a candidate resource, the prediction data comprising attribute feature data and scene feature data, the attribute feature data being used for indicating the attribute features of the first account and the candidate resource, and the scene feature data being used for indicating the scene features of the interaction between the first account and the candidate resource; performing feature cross processing on the attribute feature data through at least two feature cross dimensions, to obtain at least two feature cross results; predicting the interaction probability of the first account and the candidate resource based on the at least two feature cross results and the scene feature data; and recommending the candidate resource to the first account based on the interaction probability in the case that the interaction probability meets the recommendation requirements, so that the probability prediction accuracy and the recommendation efficiency can be improved.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a resource recommendation method, apparatus, device, storage medium, and program product. Background Technology

[0002] In the gaming platform ecosystem, coupon recommendations are a crucial step in boosting user activity and driving business conversions. The scenarios in which users claim coupons are highly diverse, including various games' micro-community lifestyle benefit zones, where users can access these zones through multiple games to claim coupons.

[0003] In related technologies, high-order nonlinear relationships between features are mined through feature cross-networks to improve the accuracy of the model's final click-through rate prediction. Multi-scenario joint modeling is employed for various prediction scenarios, that is, data from multiple game scenarios are processed uniformly through a deep model including specific feature cross-networks.

[0004] However, the above-mentioned unified modeling and processing method is limited by a single feature crossover pattern, resulting in insufficient feature utilization, low prediction accuracy for differentiated scenarios, and thus low recommendation efficiency. Summary of the Invention

[0005] This application provides a resource recommendation method, apparatus, device, storage medium, and program product. The technical solution is as follows.

[0006] On the one hand, a resource recommendation method is provided, the method comprising:

[0007] Acquire prediction data, which is used to predict the probability of interaction between the first account and the candidate resource. The prediction data includes attribute feature data and scene feature data. The attribute feature data is used to indicate the attribute features of the first account and the candidate resource. The scene feature data is used to indicate the scene features of the interaction between the first account and the candidate resource.

[0008] The attribute feature data are subjected to feature cross processing by at least two feature cross dimensions to obtain at least two feature cross results;

[0009] Based on the at least two feature cross results and the scene feature data, predict the interaction probability between the first account and the candidate resource;

[0010] If the interaction probability meets the recommendation requirements, the candidate resource is recommended to the first account based on the interaction probability.

[0011] On the other hand, a resource recommendation device is provided, the device comprising:

[0012] An acquisition module is used to acquire prediction data, which is used to predict the interaction probability between a first account and a candidate resource. The prediction data includes attribute feature data and scene feature data. The attribute feature data is used to indicate the attribute features of the first account and the candidate resource, and the scene feature data is used to indicate the scene features of the interaction between the first account and the candidate resource.

[0013] The processing module is used to perform feature cross processing on the attribute feature data through at least two feature cross dimensions to obtain at least two feature cross results;

[0014] The processing module is further configured to predict the interaction probability between the first account and the candidate resource based on the at least two feature cross results and the scene feature data;

[0015] The processing module is further configured to recommend the candidate resources to the first account based on the interaction probability if the interaction probability meets the recommendation requirements.

[0016] In some embodiments, the acquisition module is further configured to acquire network input vectors corresponding to at least two feature cross networks based on the attribute feature data, wherein the at least two feature cross networks include a low-order feature cross network and a high-order feature cross network, wherein the low-order feature cross network corresponds to a low-order input vector and the high-order feature cross network corresponds to a high-order input vector;

[0017] The processing module is further configured to perform low-order feature crossing processing on the low-order input vector through a low-order feature crossing network to obtain a low-order feature crossing result.

[0018] The processing module is further configured to perform high-order feature crossing processing on the high-order input vector through a high-order feature crossing network to obtain the high-order feature crossing result.

[0019] In some embodiments, the processing module is further configured to:

[0020] The attribute feature data is vectorized to obtain attribute feature vectors, which include at least one of dense feature vectors, sparse feature vectors, and temporal feature vectors. The dense feature vectors are used to indicate continuous numerical features in the attribute feature data, the sparse feature vectors are used to indicate categorical features in the attribute feature data, and the temporal feature vectors are used to indicate temporal features in the attribute feature data.

[0021] The low-order input vector corresponding to the low-order feature cross network is obtained based on at least one of the dense feature vector and the sparse feature vector.

[0022] By concatenating the dense feature vector, the sparse feature vector, and the temporal feature vector, the higher-order input vector corresponding to the higher-order feature crossover network is obtained.

[0023] In some embodiments, the low-order feature cross network includes a first-order feature cross network and a second-order feature cross network;

[0024] The acquisition module is also used for:

[0025] Based on the dense feature vector and the sparse feature vector, obtain the first input vector corresponding to the first-order feature cross network;

[0026] The second input vector corresponding to the second-order feature cross network is obtained based on at least two dense feature vectors.

[0027] In some embodiments, the processing module is further configured to:

[0028] The first input vector is processed by the first-order feature crossing network to obtain the first-order feature crossing result.

[0029] The second input vector is subjected to feature crossing processing through the second-order feature crossing network to obtain the second-order feature crossing result;

[0030] The low-order feature cross result is obtained by fusing the first-order feature cross result and the second-order feature cross result.

[0031] In some embodiments, the processing module is further configured to:

[0032] The first input vector is expanded into a one-dimensional input vector by dimensional reshaping. The first input vector is a two-dimensional feature vector obtained based on the dense feature vector and the sparse feature vector.

[0033] The first-order feature crossing network is used to perform feature crossing processing on the one-dimensional input vector to obtain the first-order feature crossing result.

[0034] In some embodiments, the processing module is further configured to:

[0035] The network weight matrix in the second-order feature cross network is decomposed into a first low-dimensional matrix and a second low-dimensional matrix, wherein the number of columns of the first low-dimensional matrix is ​​equal to the number of rows of the second low-dimensional matrix, and the dimensions of the first low-dimensional matrix and the second low-dimensional matrix are lower than the dimension of the weight matrix.

[0036] The intermediate feature vector is obtained by performing a linear transformation on the second input vector using the first low-dimensional matrix.

[0037] The intermediate feature vector is linearly transformed using the second low-dimensional matrix to obtain the transformed input vector.

[0038] The transformed input vector and the second input vector are subjected to feature crossing processing by the second-order feature crossing network to obtain the second-order feature crossing result.

[0039] In some embodiments, the high-order feature cross-network includes a multi-expert cross-network;

[0040] The processing module is further configured to:

[0041] A general feature representation is extracted based on the higher-order input vector through a shared expert subnetwork in the multi-expert cross-network.

[0042] Task feature representations are extracted based on the higher-order input vector through the task expert subnetwork in the multi-expert cross network;

[0043] The higher-order feature cross result is obtained by fusing the general feature representation and the task feature representation.

[0044] In some embodiments, the processing module is further configured to:

[0045] The routing weight is determined based on the scene feature data, and the routing weight is used to indicate the degree of influence of the attribute feature data on the interaction probability.

[0046] Based on the routing weights, the at least two feature cross results are weighted and fused to obtain fused features;

[0047] The interaction probability is obtained by predicting the fused features through the prediction network in the probability prediction model, which is a pre-trained machine learning model for predicting the probability of a specified event occurring.

[0048] In some embodiments, the acquisition module is further configured to acquire sample prediction data, which is used to predict the sample interaction probability of a sample account with a sample resource. The sample prediction data includes sample attribute data and sample scenario data, and the sample prediction data corresponds to a reference label.

[0049] The processing module is further configured to perform feature cross processing on the sample prediction data through at least two feature cross networks in the candidate probability prediction model to obtain at least two sample feature cross results;

[0050] The processing module is further configured to predict the sample interaction probability of the sample account for the sample resource based on the cross-result of the at least two sample features and the sample scene data through the prediction network in the candidate probability prediction model;

[0051] The processing module is further configured to train the candidate prediction model based on the difference between the sample interaction probability and the reference label until the training requirements are met, thereby obtaining the probability prediction model.

[0052] In some embodiments, the acquisition module is further configured to:

[0053] Account attribute data, resource attribute data, and context data are obtained as the attribute feature data. The account attribute data is used to indicate the account data of the first account, the resource attribute data is used to indicate the resource attributes of the candidate resources, and the context data is used to indicate the environmental state of the first account.

[0054] The program feature data and time feature data are obtained as the scene feature data. The program feature data is used to indicate the type of program that recommends the candidate resource to the first account, and the time feature data is used to indicate the time when the candidate resource is recommended to the first account.

[0055] On the other hand, a computer device is provided, the computer device including a processor and a memory, the memory storing at least one instruction, at least one program, code set or instruction set, the at least one instruction, the at least one program, the code set or instruction set being loaded and executed by the processor to implement the resource recommendation method as described in any of the embodiments of this application above.

[0056] On the other hand, a computer-readable storage medium is provided, wherein at least one instruction, at least one program, code set, or instruction set is stored therein, wherein the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by a processor to implement the resource recommendation method as described in any of the embodiments of this application above.

[0057] On the other hand, a computer program product or computer program is provided, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform any of the resource recommendation methods described in the above embodiments.

[0058] The beneficial effects of the technical solutions provided in this application include at least the following:

[0059] By performing feature cross processing with at least two feature cross dimensions, at least two feature cross results are obtained based on attribute feature data. This allows for the full exploration of high-order nonlinear relationships between features from multiple dimensions, taking into account both the independent and joint effects of various features on interaction probabilities. As a result, using at least two feature cross results to participate in the prediction of interaction probabilities is more accurate, thereby improving the recommendation efficiency of resource recommendation based on account and resource interaction probabilities. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 This is a schematic diagram of a computer system provided in an exemplary embodiment of this application;

[0062] Figure 2 This is a flowchart of a resource recommendation method provided in an exemplary embodiment of this application;

[0063] Figure 3 This is a schematic diagram of probability prediction based on multi-dimensional feature intersection provided in an exemplary embodiment of this application;

[0064] Figure 4 This is a schematic diagram illustrating multi-scenario resource recommendation provided by an exemplary embodiment of this application;

[0065] Figure 5 This is a flowchart of a feature cross-processing method provided in an exemplary embodiment of this application;

[0066] Figure 6 This is a schematic diagram of a network structure provided in an exemplary embodiment of this application;

[0067] Figure 7 This is a schematic diagram of the structure of a second-order feature cross network provided in an exemplary embodiment of this application;

[0068] Figure 8 This is a schematic diagram of the structure of a multi-expert cross-network provided in an exemplary embodiment of this application;

[0069] Figure 9 This is a flowchart of a probability prediction method based on route weight provided in an exemplary embodiment of this application;

[0070] Figure 10 This is a schematic diagram of the model structure of a probabilistic prediction model based on dynamic routing provided in an exemplary embodiment of this application;

[0071] Figure 11 This is a schematic diagram of a single-layer structure of a multi-expert cross-network provided in an exemplary embodiment of this application;

[0072] Figure 12 This is a schematic diagram of dimensional reshaping provided in an exemplary embodiment of this application;

[0073] Figure 13 This is a schematic diagram of a network task provided in an exemplary embodiment of this application;

[0074] Figure 14 This is a flowchart of a training method for a probability prediction model provided in an exemplary embodiment of this application;

[0075] Figure 15 This is a structural block diagram of a resource recommendation apparatus provided in an exemplary embodiment of this application;

[0076] Figure 16 This is a structural block diagram of a terminal provided in an exemplary embodiment of this application. Detailed Implementation

[0077] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0078] It should be understood that although the terms first, second, etc., may be used in this disclosure to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, a first parameter may also be referred to as a second parameter without departing from the scope of this disclosure, and similarly, a second parameter may also be referred to as a first parameter. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0079] In the gaming platform ecosystem, coupon recommendations are a crucial step in enhancing user activity and monetization. The scenarios in which users claim coupons are highly diverse, including various game-specific micro-community welfare zones where users can access and claim coupons through different games. Related technologies utilize feature cross-networks to uncover higher-order nonlinear relationships between features, improving the model's accuracy in predicting final click-through rates. Multi-scenario joint modeling is employed for various prediction scenarios, using deep models including specific feature cross-networks to uniformly process data from multiple game scenarios. However, this unified modeling approach is limited by a single feature cross-mode, resulting in insufficient feature utilization and lower prediction accuracy for differentiated scenarios, leading to low recommendation efficiency.

[0080] The resource recommendation method provided in this application embodiment, through feature cross processing of at least two feature cross dimensions, obtains at least two feature cross results based on attribute feature data. It can fully explore the high-order nonlinear relationship between features from multiple dimensions, taking into account the independent and joint effects of various features on the interaction probability. Therefore, the interaction probability obtained by using at least two feature cross results to participate in the prediction is more accurate, thereby improving the recommendation efficiency of resource recommendation based on account and resource interaction probability.

[0081] First, the computer system described in this application will be introduced. Please refer to... Figure 1 The illustration shows a schematic diagram of a computer system provided in an exemplary embodiment of this application, which includes a terminal 10 and a server 20.

[0082] Terminal 10 has a first application installed. The first application has a resource recommendation function. The candidate resources recommended in the first application can be determined by server 20. Server 20 is the backend server of the first application and is used to determine the candidate resources recommended in the first application.

[0083] Optionally, the first application can be any application with resource recommendation function, such as a virtual game program, social application, online shopping software, multimedia playback platform, e-reading software, online teaching platform, virtual reality (VR) application, augmented reality (AR) application, mixed reality (MR) application, etc. This application embodiment does not limit this.

[0084] In some embodiments, server 20 is a backend server for at least two first applications, and server 20 is used to coordinate and determine recommended candidate resources in at least two first applications.

[0085] Taking the example of terminal 10 logging into at least two first applications through a first account, server 20 needs to recommend candidate resources to the first account in a targeted manner based on the account attributes of the first account and the program type of the first application. That is, the candidate resources recommended to the first account in at least two first applications may be the same or different.

[0086] Indicatively, server 20 acquires prediction data, which is used to predict the interaction probability between the first account and candidate resources. The prediction data includes attribute feature data and scene feature data. The attribute feature data indicates the attribute features of the first account and candidate resources, and the scene feature data indicates the scene features of the interaction between the first account and candidate resources. The attribute feature data is processed by feature cross-processing through at least two feature cross-dimensions to obtain at least two feature cross-results. The interaction probability between the first account and candidate resources is predicted based on the at least two feature cross-results and the scene feature data. If the interaction probability meets the recommendation requirements, candidate resources are recommended to the first account based on the interaction probability.

[0087] Among them, scene feature data can be used to indicate the program type of the first application that recommends candidate resources to the first account.

[0088] The aforementioned terminal is optional and can be a desktop computer, laptop computer, mobile phone, tablet computer, e-book reader, Moving Picture Experts Group Audio Layer III (MP3) player, Moving Picture Experts Group Audio Layer IV (MP4) player, smart TV, smart vehicle, and other types of terminal devices. This application embodiment does not limit the specific terminal device to these types.

[0089] It is worth noting that the aforementioned servers can be independent physical servers, server clusters or distributed systems composed of multiple physical servers, or cloud servers that provide basic cloud computing services such as cloud services, cloud security, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0090] Cloud technology refers to a hosting technology that unifies a series of resources such as hardware, software, and networks within a wide area network or local area network to achieve data computing, storage, processing, and sharing.

[0091] In some embodiments, the server described above can also be implemented as a node in a blockchain system.

[0092] It should be noted that all information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, stored data, displayed data, etc.), and signals involved in this application have been authorized by the user or fully authorized by all parties, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant regions. For example, the operational data and account information involved in this application were obtained with full authorization.

[0093] To further clarify, this application may display a prompt interface, pop-up window, or output voice prompts before and during the collection of user-related data (e.g., account information, historical operation data, and real-time operation data involved in this application). These prompt interfaces, pop-ups, or voice prompts are used to inform the user that their relevant data is being collected. This ensures that the application only begins the steps for collecting user-related data after receiving confirmation from the user regarding the prompt interface or pop-up window; otherwise (i.e., without receiving confirmation from the user), the steps for collecting user-related data end, meaning no user-related data is collected. In other words, all user data collected in this application is collected with the user's consent and authorization, and the collection, use, and processing of relevant user data must comply with the relevant laws, regulations, and standards of the relevant regions.

[0094] This is illustrative; please refer to it. Figure 2 This document illustrates a flowchart of a resource recommendation method provided in an exemplary embodiment of this application. This method can be executed by a terminal, a server, or both simultaneously. This embodiment uses server execution as an example for illustration. Figure 2 As shown, the method includes the following steps:

[0095] Step 210: Obtain the prediction data.

[0096] The prediction data is used to predict the probability of interaction between the first account and the candidate resources. The prediction data includes attribute feature data and scene feature data. The attribute feature data is used to indicate the attribute features of the first account and the candidate resources, and the scene feature data is used to indicate the scene features of the interaction between the first account and the candidate resources.

[0097] Taking the recommendation of candidate resources to the first account in the first application as an example, the candidate resources can be any resources authorized and recommended by the first application, such as virtual game resources, virtual vouchers, multimedia playback permissions, e-reading permissions, social interaction permissions, etc. The resources authorized and recommended by the first application include, but are not limited to, the candidate resources provided in the first application, or at least one of the candidate resources recommended by the second application authorized and accessed by the first application.

[0098] The second application is an application authorized by the first application, which can access the first application and recommend candidate resources within the first application.

[0099] The interaction between the first account and the candidate resources includes, but is not limited to, any interaction method such as the first account viewing the candidate resources, the first account clicking on the candidate resources, the first account claiming the candidate resources, and the first account using the candidate resources. This application embodiment does not limit this.

[0100] Taking virtual vouchers as an example of candidate resources, the interaction probability can be the probability of the first account clicking on the virtual voucher, the probability of the first account receiving the virtual voucher, or the probability of the first account redeeming the virtual voucher, etc.

[0101] The scene features indicated by the scene feature data include, but are not limited to, at least one of the following: the program type of the first application that recommends candidate resources to the first account, or the recommendation time of the candidate resources to the first account.

[0102] In some embodiments, obtaining prediction data includes obtaining attribute feature data and scene feature data. Account attribute data, resource attribute data, and context data are obtained as attribute feature data. Account attribute data is used to indicate the account data of the first account, resource attribute data is used to indicate the resource attributes of the candidate resources, and context data is used to indicate the environmental state of the first account. Program feature data and time feature data are obtained as scene feature data. Program feature data is used to indicate the program type of the candidate resources recommended to the first account, and time feature data is used to indicate the recommendation time of the candidate resources recommended to the first account.

[0103] To acquire attribute feature data, account attribute data, resource attribute data, and context data can be collected in real time from the business database.

[0104] For illustrative purposes only, account attribute data includes, but is not limited to, user age, gender, account level, and historical behavior sequence; resource attribute data includes, but is not limited to, coupon identifier, brand category, discount level, and click-through rate; contextual data includes, but is not limited to, the device type, geographical location, and network environment corresponding to the first account.

[0105] Scene feature data includes program features and time features. Program features include, but are not limited to, the program category of the first application and the channel identifier for recommending candidate resources to the first account. Time features include, but are not limited to, the date, holiday type, and time period for recommending candidate resources to the first account.

[0106] In some embodiments, feature preprocessing includes: standardizing continuous numerical features (such as click-through rate), performing embedding mapping on categorical features (such as game ID), and encoding behavioral sequence features (such as recent click records) through an attention mechanism.

[0107] Step 220: Perform feature cross processing on the attribute feature data using at least two feature cross dimensions to obtain at least two feature cross results.

[0108] Feature cross processing of attribute feature data through at least two feature cross dimensions refers to capturing the interaction relationship between features from different levels or methods.

[0109] Optionally, at least two feature cross-dimensions include, but are not limited to, feature cross-processing of different orders, feature cross-processing of different interaction forms, and feature cross-processing of different feature types.

[0110] The order indicates the complexity of feature crossing, such as first-order feature crossing, second-order feature crossing, and higher-order feature crossing. The interaction forms include explicit crossing (such as inner product) or implicit crossing (such as automatic learning by neural networks). Feature crossing processing for different feature types includes differentiated processing for dense features, sparse features, and temporal features.

[0111] In some embodiments, the attribute feature data needs to be characterized first. This process includes feature vectorization of the attribute feature data to obtain attribute feature vectors.

[0112] The attribute feature vector includes at least one of dense feature vector, sparse feature vector and temporal feature vector. Dense feature vector is used to indicate continuous numerical features in attribute feature data, sparse feature vector is used to indicate categorical features in attribute feature data, and temporal feature vector is used to indicate temporal features in attribute feature data.

[0113] Indicatively, dense feature vectors can be used to indicate the click-through rate, redemption rate, user age, and number of days logged into the game for a certain type of voucher on a given day; sparse feature vectors can be used to indicate user gender, location, and the type of voucher brand; and time-series feature vectors can be used to indicate the historical sequence of vouchers received by a user and the historical sequence of brands exposed to the user.

[0114] By vectorizing attribute feature data into dense, sparse, and temporal feature vectors, the characteristics of attribute feature data in continuous numerical, categorical, and temporal dimensions can be fully reflected. This provides a foundation for implementing at least two feature cross-processing methods, enabling at least two feature cross-processing networks to fully learn the complex relationships between attribute feature data based on three-dimensional feature vectors, thereby improving the accuracy of predicting interaction probabilities based on attribute feature data. Multimodal feature support is achieved by being compatible with numerical, categorical, and temporal data, and the problem of feature heterogeneity is solved by vectorization to achieve unified data representation.

[0115] Optionally, at least two feature cross dimensions include a low-order feature cross dimension and a high-order feature cross dimension.

[0116] The low-order feature cross dimension is used to indicate the feature cross processing for at least one of the dense and sparse feature vectors, while the high-order feature cross dimension is used to indicate the feature cross processing for dense, sparse, and temporal feature vectors.

[0117] In some embodiments, the low-order feature cross dimension includes a one-dimensional feature cross dimension and a two-dimensional feature cross dimension.

[0118] One-dimensional feature cross dimension is used to indicate the degree of independent influence of a feature on the interaction probability, while two-dimensional feature cross dimension is used to indicate the degree of joint influence of at least two features on the interaction probability.

[0119] In some embodiments, in order to accurately predict interaction probabilities based on attribute feature data, at least two feature cross-dimensions can be implemented through at least two feature cross-networks.

[0120] Optionally, at least two feature cross-networks correspond to routing weights, which are used to indicate the output weights of the feature cross-networks. These routing weights can be determined based on scene feature data.

[0121] In a schematic manner, network input vectors corresponding to at least two feature cross-networks are obtained based on attribute feature data. The at least two feature cross-networks include a low-order feature cross-network and a high-order feature cross-network. The low-order feature cross-network corresponds to the low-order input vector, and the high-order feature cross-network corresponds to the high-order input vector. The low-order feature cross-network performs low-order feature cross-processing on the low-order input vector to obtain the low-order feature cross-processing result. The high-order feature cross-network performs high-order feature cross-processing on the high-order input vector to obtain the high-order feature cross-processing result.

[0122] Low-order feature cross results are representation vectors generated by explicitly modeling the basic interaction relationships between features (such as first-order linear terms and second-order combination terms).

[0123] The low-order feature cross results are generated based on dense and sparse features in the attribute feature data, including first-order linear terms, second-order combination terms, and higher-order cross terms. The first-order linear terms are used to indicate the degree of independent influence of a single feature on the interaction probability, the second-order combination terms are used to indicate the pairwise latent vector product of sparse features, and the higher-order cross terms are used to indicate explicit element-wise crosses while retaining the original features.

[0124] The high-order feature cross result is a deep representation vector generated by implicitly learning the complex nonlinear relationships (third-order and above combinations) between features.

[0125] The high-order feature cross result is generated by feature cross processing based on the full concatenation result of dense features, sparse features and temporal features in the attribute feature data.

[0126] Illustratively, low-order feature cross networks can include Factorization Machines (FM) and Deep & Cross Networks (DCN), such as Deep & Cross Networks version 2 (DCNv2), while high-order cross networks can include Shared and Task-specific Embeddings (STEM).

[0127] FM can be used to implement low-order explicit crosses of sparse features, such as handling the impact of basic attribute features like account identifiers and resource identifiers on interaction probabilities; DCNv2 is used to implement high-order explicit crosses of dense or sparse features, modeling complex feature combinations, such as time series × category; STEM is used to implement multi-task dynamic routing and high-order implicit crosses, adapting to multi-task objectives such as click-through rate / redeem rate.

[0128] Schematic, the FM network includes first-order and second-order terms. The first-order term calculates the linear effect, such as user age multiplied by discount level; the second-order term calculates the latent vector inner product of user identifier and voucher identifier. The DCNv2 network multiplies the original features element-wise with the output of the cross layer, generating a high-order feature combination through three layers of cross processing. The STEM network extracts general representations of user interests through shared experts, predicts short-term behavioral characteristics of experts based on click-through rates among task experts, predicts long-term consumption preferences based on redemption rates among task experts, and dynamically mixes expert outputs according to the current scenario through a gating network. After feature cross processing, the prediction network concatenates the low-order / high-order cross results, and the interaction probability is output through a two-layer fully connected network.

[0129] It is worth noting that the structure and data processing of the above probability prediction model are merely illustrative examples, and the embodiments of this application do not limit them.

[0130] Step 230: Based on at least two feature cross results and scene feature data, predict the interaction probability between the first account and the candidate resources.

[0131] In recommendation systems, potential association features between accounts and candidate resources are extracted through at least two feature cross-dimensional methods. In this embodiment, at least two feature cross-results are obtained. For example, low-order feature cross-results (such as FM network output) capture the explicit combination relationship between user basic attributes (such as age and gender) and voucher attributes (such as brand and discount level); high-order feature cross-results (such as DCNv2 network output) mine the deep non-linear association between user operation sequences (such as recent clicks and redemption records) and voucher statistical features (such as exposure rate and redemption rate).

[0132] Scene feature data is used to describe the context in which recommendations occur, including but not limited to application scenarios, time features, and channel features.

[0133] Application scenario is used to indicate the interaction scenario between the first account and the candidate resources, such as the program type of the first application that recommends candidate resources to the first account; time feature is used to indicate the recommendation time period when recommending candidate resources to the first account, such as weekdays / weekends, morning and evening peak hours, etc.; channel feature is used to indicate the entry source of the first account to enter the resource recommendation scenario, such as the activity page in the first application, etc.

[0134] In some embodiments, the interaction probability between the first account and the candidate resource is predicted by dynamic fusion based on the cross-results of at least two features and scene feature data.

[0135] The dynamic fusion method includes scenario weighting of low-order feature cross results and expert routing of high-order feature cross results.

[0136] As an illustration, for scene weighting of low-order feature cross-results, scene weights can be generated using the Sigmoid function to adjust the feature importance of the FM output. For example, a candidate resource of the food coupon type may have a higher weight in application A than in application B. For expert routing of high-order feature cross-results, in the STEM network, scene features are concatenated with expert network inputs, and the weights of different experts are dynamically assigned through Softmax gating to adapt to the differences in data distribution across multiple scenarios.

[0137] In some embodiments, the weighted low-order cross-tabulation results, high-order cross-tabulation results, and scene features are concatenated to form the final feature representation. Based on this feature representation, an interaction probability value between 0 and 1 is output through a fully connected layer and a sigmoid activation function, representing the interaction probability between the first account and the candidate resource, such as the probability of a user clicking / consuming / redeeming a voucher.

[0138] This is illustrative; please refer to it. Figure 3 , Figure 3 This is a schematic diagram of probability prediction based on multi-dimensional feature intersection provided in an exemplary embodiment of this application, such as... Figure 3 As shown, prediction data 310 is obtained, which includes attribute feature data 311 and scene feature data 312. The attribute feature data 311 is subjected to feature cross processing through at least two feature cross dimensions to obtain at least two feature cross results 320, such as feature cross results 1 to n, where n is a positive integer. Based on the at least two feature cross results 320 and scene feature data 312, the interaction probability 330 between the first account and the candidate resource is predicted.

[0139] Step 240: If the interaction probability meets the recommendation requirements, recommend candidate resources to the first account based on the interaction probability.

[0140] Optionally, candidate vouchers can be sorted by predicted probability, or candidate resources with interaction probabilities sorted from high to low within a preset ranking range can be pushed to the user interface.

[0141] In some embodiments, a first candidate resource whose interaction probability meets the recommendation requirements is obtained from the candidate resources; the first candidate resource is recommended to the first account in descending order of interaction probability, wherein the interaction probability corresponding to the first candidate resource is positively correlated with the recommendation display priority of the first candidate resource in the client where the first account is located, and the recommendation display priority is used to indicate the display order and visual prominence of the first candidate resource in the client where the first account is located.

[0142] For illustrative purposes, the recommendation requirement includes an interaction probability of 70%. The first candidate resource with an interaction probability of 70% is selected from the candidate resources and recommended to the first account. When recommending the first candidate resource to the first account, the server can send the first candidate resource to the client of the first account. The first candidate resource is associated with a recommendation display priority, which is positively correlated with the interaction probability. The client of the first account can prioritize rendering the first candidate resource with the higher recommendation display priority during the interface loading process, and display the first candidate resource with higher visual prominence according to the recommendation display priority. For example, according to the interface rendering template of the resource recommendation interface in the first application, the first candidate resource with the higher recommendation display priority is prioritized and displayed in a prominent position.

[0143] This is illustrative; please refer to it. Figure 4 , Figure 4 This is a schematic diagram illustrating multi-scenario resource recommendation provided by an exemplary embodiment of this application, such as... Figure 4As shown, a server is jointly recommended for both game program A and game program B. During resource recommendation for game program A, the game interface 410 of game program A is displayed. In response to a trigger operation on the first control 411, a first resource recommendation interface 420 is displayed in game program A. The first resource recommendation interface 420 includes candidate resources 421 recommended to the first account. The candidate resources 421 are recommended based on the interaction probability determined by the attribute feature data and scene feature data of the first account's interaction with candidate resources in game program A. During resource recommendation for game program B, in response to a trigger operation on the second control 431 in the game interface 430 of game program B, a second resource recommendation interface 440 is displayed in game program B. The second resource recommendation interface 440 includes candidate resources 441 recommended to the first account. The candidate resources 441 are recommended based on the interaction probability determined by the attribute feature data and scene feature data of the first account's interaction with candidate resources in game program B.

[0144] In summary, the method provided in this application, through feature cross processing of at least two feature cross dimensions, obtains at least two feature cross results based on attribute feature data. This method can fully explore the high-order nonlinear relationships between features from multiple dimensions, taking into account both the independent and joint effects of various features on the interaction probability. As a result, the interaction probability obtained by using at least two feature cross results in the prediction is more accurate, thereby improving the recommendation efficiency of resource recommendation based on account and resource interaction probability.

[0145] In some embodiments, feature crossing is performed using a feature crossing network. Please refer to [reference needed]. Figure 5 , Figure 5 This is a flowchart of a feature cross-processing method provided in an exemplary embodiment of this application. This method can be executed by a terminal, a server, or both simultaneously. This embodiment illustrates the method as being executed by a server. Figure 5 As shown, step 220 above includes the following steps:

[0146] Step 221: Obtain the network input vectors corresponding to at least two feature cross-networks based on the attribute feature data.

[0147] At least two feature cross-networks include a low-order feature cross-network and a high-order feature cross-network, with the low-order feature cross-network corresponding to the low-order input vector and the high-order feature cross-network corresponding to the high-order input vector.

[0148] This application embodiment takes at least two feature cross-networks, including a low-order feature cross-network and a high-order feature cross-network, as an example. The low-order feature cross-network is used to explicitly model the basic interaction relationship between attribute features, and the high-order feature cross-network is used to implicitly capture the deep nonlinear interaction pattern between attribute features.

[0149] To illustrate, low-order information features are extracted from attribute feature data through a low-order feature cross network to obtain low-order feature cross results; high-order information features are extracted from attribute feature data through a high-order feature cross network to obtain high-order feature cross results.

[0150] This application embodiment takes a low-order feature cross network, including a low-order cross sub-network (FM+SENet) and a high-order cross sub-network (DCNv2), and a high-order feature cross network, including a multi-expert cross network, as an example. The low-order cross sub-network is used to capture explicit low-order interaction relationships between attribute features, such as first-order linear relationships and second-order combination relationships, mainly solving the basic correlation problem under sparse features; the high-order cross sub-network is used to model implicit high-order nonlinear relationships (third-order and above complex combinations) between features and to mine deep feature patterns.

[0151] Optionally, the low-order cross subnetwork includes a first-order feature network and a second-order feature network. The first-order feature network is used to quantify the independent impact of a single feature, such as the linear contribution of user age to click-through rate. The second-order feature network is used to model pairwise feature interactions through latent vector inner product, such as the combined effect of user identifier × product identifier.

[0152] In some embodiments, information features are extracted based on dense and sparse feature vectors through a low-order cross sub-network to obtain a low-order cross sub-result. The low-order cross sub-network includes a first-order feature cross network and a second-order feature cross network. The first-order feature cross network is used to quantify the degree of independent influence of dense and sparse feature vectors on the interaction probability, and the second-order feature cross network is used to quantify the degree of joint influence of at least two sparse feature vectors on the interaction probability. At least two of the attribute feature vectors are concatenated to obtain high-order cross input features. The high-order cross input features are then processed layer by layer through the high-order cross sub-network to obtain a high-order cross sub-result. Finally, the low-order cross sub-result and the high-order cross sub-result are fused to obtain a low-order feature cross result.

[0153] In some embodiments, the attribute feature data needs to be characterized first. This process includes feature vectorization of the attribute feature data to obtain attribute feature vectors.

[0154] The attribute feature vector includes at least one of dense feature vector, sparse feature vector and temporal feature vector. Dense feature vector is used to indicate continuous numerical features in attribute feature data, sparse feature vector is used to indicate categorical features in attribute feature data, and temporal feature vector is used to indicate temporal features in attribute feature data.

[0155] Indicatively, dense feature vectors can be used to indicate the click-through rate, redemption rate, user age, and number of days logged into the game for a certain type of voucher on a given day; sparse feature vectors can be used to indicate user gender, location, and the type of voucher brand; and time-series feature vectors can be used to indicate the historical sequence of vouchers received by a user and the historical sequence of brands exposed to the user.

[0156] Taking the consumption voucher recommendation scenario as an example, user attribute feature data is classified and processed, and then assigned to different levels of feature cross-networks.

[0157] To illustrate, for low-order input vectors, dense features (such as user age, coupon click-through rate) can be concatenated with sparse features (such as coupon ID, brand category) to form the first input vector, which is then fed into a first-order feature cross-network (FM+SENet). Only dense features (such as the redemption rate in the past 7 days) are used to form the second input vector, which is then fed into a second-order feature cross-network (DCNv2).

[0158] For the high-order input vector, all features (dense + sparse + temporal) are concatenated and then used as the high-order input vector to input the multi-expert cross-network (STEM).

[0159] Efficient information extraction is achieved by matching feature types with network capabilities.

[0160] Step 222: Perform low-order feature crossing processing on the low-order input vector through a low-order feature crossing network to obtain the low-order feature crossing result.

[0161] The low-order feature crossing network is used to perform low-order feature crossing processing of low-order feature crossing dimensions. In the embodiments of this application, the low-order feature crossing network is used to perform feature crossing processing based on at least one feature vector among dense feature vectors and sparse feature vectors.

[0162] Optionally, the low-order feature crossing network includes a first-order feature crossing network and a second-order feature crossing network. Before the low-order feature crossing process, it is necessary to obtain the low-order input vectors corresponding to the first-order feature crossing network and the second-order feature crossing network, respectively.

[0163] In some embodiments, a first input vector corresponding to a first-order feature cross-network is obtained based on dense feature vectors and sparse feature vectors; and a second input vector corresponding to a second-order feature cross-network is obtained based on at least two dense feature vectors.

[0164] In a schematic way, the first input vector is obtained by concatenating the sparse feature vector and the dense feature vector, which is used as the network input of FM+SENet. The second input vector is obtained based on the dense feature vector, which is used as the network input of DCNv2.

[0165] After obtaining the first input vector and the second input vector, step 222 includes: performing feature crossing processing on the first input vector through a first-order feature crossing network to obtain a first-order feature crossing result; performing feature crossing processing on the second input vector through a second-order feature crossing network to obtain a second-order feature crossing result; and fusing the first-order feature crossing result and the second-order feature crossing result to obtain a low-order feature crossing result.

[0166] Indicatively, in the first-order cross-network processing, the first input vector (dense + sparse features) is input to FM+SENet. Feature importance weights are generated through SENet to adjust the input features. The independent influence of weighted features is calculated through first-order summation, such as the contribution of "young users" to the probability of receiving the coupon. The combined effect of features is captured through second-order inner product, such as the joint influence of "food coupon A + female users". The first-order feature cross-processing result is output (concatenating independent and joint influence terms).

[0167] In some embodiments, the first input vector is expanded to obtain a one-dimensional input vector by dimensional reshaping. The first input vector is a two-dimensional feature vector obtained based on dense feature vectors and sparse feature vectors. The one-dimensional input vector is then subjected to feature crossing processing through a first-order feature crossing network to obtain a first-order feature crossing result.

[0168] In the second-order cross-network processing, the second input vector (dense features only) is input to DCNv2; feature cross-processing is performed through explicit cross-processing, and high-order interaction of features is constructed through the DCNv2 network with Low-rank decomposition, such as "high click-through rate × low redemption rate"; the original features are preserved through residual connections to avoid information loss; and the second-order feature cross-processing result is output.

[0169] The first-order feature cross results are concatenated with the second-order feature cross results to form a low-order feature cross result, which covers feature relationships from simple to medium complexity.

[0170] Indicatively, a first-order feature cross-network (i.e., a low-order cross-subnetwork) includes the FM network. The core idea of ​​the FM network is to capture the interaction relationships between features through factorization. Please refer to Equation 1 below for its model principle.

[0171]

[0172] Where, x i It is the input feature, v iIt is the i-th input x i The embedding vector, v j It is the j-th input x j The embedding vector, w i w0 and w0 are the weights and biases, respectively. This formula uses two distinct features x... i With x j The inner product of the corresponding Embedding vectors is used as the second-order feature x. i x j The calculated weights explicitly capture the interaction relationships between features.

[0173] In some embodiments, a squeeze-and-excitation network (SENet) is also used in the first-order feature cross-network. For an illustrative example, please refer to [reference needed]. Figure 6 , Figure 6 This is a schematic diagram of a network structure provided in an exemplary embodiment of this application, such as... Figure 6 As shown, the network structure 600 includes the SENet network, corresponding to the Squeeze stage 610, the Excitation stage 620, and the Scale stage 630.

[0174] SENet primarily performs a squeeze operation on the input to generate weights, which are then combined with the original input element-wise. For an input SENet tensor of shape (C, H, W), max-pooling yields a weight tensor of dimension (1, 1, C). This weight tensor is then passed through two dense layers with ReLU and Sigmoid activation functions, and used as weights along dimension C for element-wise multiplication with the original input tensor. Essentially, SENet automatically filters the input tensor along dimension C.

[0175] In some embodiments, the first input vector is expanded to obtain a one-dimensional input vector by dimensional reshaping. The first input vector is a two-dimensional feature vector obtained based on the dense feature vector and the sparse feature vector. The one-dimensional input vector is then subjected to feature crossing processing by a first-order feature crossing network to obtain a first-order feature crossing result.

[0176] Second-order feature cross networks (i.e., higher-order cross sub-networks) include the DCNv2 network; please refer to [link / reference]. Figure 7 , Figure 7 This is a schematic diagram of the structure of a second-order feature cross network provided in an exemplary embodiment of this application, as shown below. Figure 7As shown, the second-order feature cross network 700 includes a stacked structure 710 and a parallel structure 720. In the stacked structure 710, the cross layer and the depth layer are stacked in sequence to form a deep network. In the parallel structure 720, the cross layer and the depth layer process the input features in parallel and then merge the results.

[0177] The input layer of DCNv2 is used to input raw features, which typically include numerical and categorical features. Categorical features are usually transformed into dense vectors through an embedding layer.

[0178] DCNv2's Cross Layer is used to explicitly capture high-order interactions between features. The processing logic of the Cross Layer is shown in Equation 2 below.

[0179] x l+1 =x0⊙(W l x l +b l )+x l Formula 2,

[0180] Where x0 is the input feature, x l It is the output of the l-th layer, W l and b l These are the weights and biases, respectively. This formula modifies the input feature x0 by comparing it with the output x0 of the current layer. l Perform cross-products to explicitly capture the interactions between features.

[0181] DCNv2's deep layers consist of multiple fully connected layers for automatically learning feature representations. By introducing non-linear activation functions and a multi-layered structure, deep layers are able to capture complex patterns in the data.

[0182] In some embodiments, the network weight matrix in the second-order feature cross network is decomposed into a first low-dimensional matrix and a second low-dimensional matrix, wherein the number of columns in the first low-dimensional matrix is ​​equal to the number of rows in the second low-dimensional matrix, and the dimensions of the first and second low-dimensional matrices are lower than the dimension of the weight matrix; the second input vector is linearly transformed using the first low-dimensional matrix to obtain an intermediate feature vector; the intermediate feature vector is linearly transformed using the second low-dimensional matrix to obtain a transformed input vector; the transformed input vector and the second input vector are subjected to feature cross processing using the second-order feature cross network to obtain the second-order feature cross result.

[0183] The method provided in this application obtains the first input vector corresponding to the first-order feature cross-network based on dense and sparse feature vectors, enabling the first-order feature cross-network to effectively analyze the degree of independent influence of dense and sparse feature vectors on the interaction probability through explicit feature interaction. It obtains the second input vector corresponding to the second-order feature cross-network based on at least two dense feature vectors, enabling the second-order feature cross-network to effectively analyze the degree of joint influence of at least two dense feature vectors on the interaction probability through implicit feature interaction. It selects different types of feature cross-networks to perform feature cross-processing of different dimensions according to different types of features, realizing the refinement of low-order networks. It can adapt to the differential design of different features in multiple scenarios such as sparse features (such as brand ID) and dense features (such as click rate) in consumer voucher recommendation.

[0184] The method provided in this application combines first-order results (independent influence) and second-order results (joint influence) to cover more comprehensive feature relationships, achieve information complementarity, and explicitly retain the two types of cross results, which facilitates the analysis of key influencing factors.

[0185] Step 223: Perform high-order feature crossing processing on the high-order input vector through a high-order feature crossing network to obtain the high-order feature crossing result.

[0186] The high-order feature cross network is used to perform high-order feature cross processing of high-order feature cross dimensions. In the embodiments of this application, the high-order feature cross network is used to perform feature cross processing based on the full concatenation vector of dense feature vectors, sparse feature vectors and temporal feature vectors.

[0187] In some embodiments, the high-order feature cross network includes a multi-expert cross network. A general feature representation is extracted based on the high-order input vector through a shared expert subnetwork in the multi-expert cross network; a task feature representation is extracted based on the high-order input vector through a task expert subnetwork in the multi-expert cross network; and the general feature representation and the task feature representation are fused to obtain the high-order feature cross result.

[0188] To illustrate, dense feature vectors, sparse feature vectors, and temporal feature vectors are concatenated to obtain a high-order input vector, which serves as the input feature for the expert network.

[0189] In the multi-expert cross-network of high-order feature cross-network, the shared expert sub-network is used to extract general features across scenarios, such as "the general preference for catering vouchers"; the task expert sub-network is used to learn specific scenario features, such as "the unique interest of users in milk tea vouchers in application A".

[0190] In some embodiments, the high-order feature cross-network adopts dynamic routing to adapt to multi-scenario prediction. In this process, scene-aware gating is introduced into the multi-expert cross-network, for example, expert weights are dynamically allocated based on scene features such as game ID and time. The weighted aggregation of general and task feature representations generates high-order feature cross-results that adapt to multi-scenario data distributions.

[0191] Higher-order feature cross-networks include STEM networks; please refer to [link / reference]. Figure 8 , Figure 8 This is a schematic diagram of the structure of a multi-expert cross-network provided in an exemplary embodiment of this application, such as... Figure 8 As shown, the high-order feature cross-network 800 includes an embedding layer 810, an expert layer 820, a gating layer 830, and an output layer 840.

[0192] The embedding level of STEM is used to input the original features, which typically include numerical and categorical features. These features are then transformed into dense vectors through the embedding level. Here, a different set of dense vectors is generated for different tasks and shared experts.

[0193] The STEM Expert Level is used to capture the interaction relationships between features and includes multiple fully connected layers. Experts are further divided into task-specific experts and shared experts. Each expert module consists of multiple fully connected layers and can capture different data features respectively.

[0194] The gate level in STEM is determined by the number of tasks and is responsible for aggregating the outputs of multiple different experts. Its output is a weighted sum of the outputs of multiple different experts, and the weights are often obtained by passing the input embedding dense vector through a gated network and a softmax activation function.

[0195] The output layer (Tower Level) of STEM is determined by the number of tasks and includes multiple fully connected layers, each used to output the model's prediction y for task A. A And the prediction of task B y B .

[0196] The method provided in this application embodiment achieves high-order feature cross-processing through a multi-expert cross-network. It can learn general features across scenarios by sharing experts and focus on scene-specific features by task experts, thus achieving scene adaptation. At the same time, it isolates the expert weights of different scenarios through a gating mechanism to avoid negative transfer problems and improve the efficiency of high-order feature cross-processing.

[0197] In summary, the method provided in this application embodiment achieves feature cross processing of at least two feature cross dimensions through at least two feature cross networks to obtain at least two feature cross results. This allows for a full exploration of the influence of features on interaction probability based on attribute feature data, thereby improving the accuracy of interaction probability prediction.

[0198] The method provided in this application transforms attribute feature data into attribute feature vectors, including dense feature vectors, sparse feature vectors, and temporal feature vectors. This enables the representation of attribute features from multiple dimensions. A low-order input vector is then constructed based on at least one of the dense and sparse feature vectors. A high-order input vector is obtained by concatenating the dense, sparse, and temporal feature vectors. This allows for the targeted use of network input vectors that match the dimensions of the feature cross-network to participate in feature cross-processing, improving the efficiency of feature cross-processing in each network. It effectively utilizes multi-dimensional features to evaluate the impact of features on interaction probabilities, enabling the low-order network to focus on dense and sparse features, while the high-order network integrates all features, balancing performance and complexity.

[0199] In some embodiments, routing weights are determined using scene feature data to achieve differentiated recommendations based on the recommendation scenario. Please refer to [reference needed]. Figure 9 , Figure 9 This is a flowchart of a probability prediction method based on route weights provided in an exemplary embodiment of this application. This method can be executed by a terminal, a server, or both simultaneously. This embodiment illustrates the method as being executed by a server. Figure 9 As shown, step 230 above includes the following steps:

[0200] Step 231: Determine the routing weights based on the scene feature data.

[0201] Routing weights are used to indicate the degree of influence of attribute feature data on the probability of interaction.

[0202] In recommendation scenarios, users' preferences for resources vary significantly across different contexts. Therefore, it is necessary to consider the specific scenarios in which resources are recommended to users in order to achieve differentiated recommendations.

[0203] To illustrate, for example, players of virtual game A might prefer restaurant vouchers, while players of virtual game B might be more interested in discounts on digital products; similarly, for example, users might prefer fast food vouchers during weekday lunchtimes, while they might prefer movie theater discounts on weekend evenings.

[0204] In some embodiments, processing all scenarios with a single model leads to limited recommendation effects. The method provided in the embodiments of the present application enables the model to automatically adjust the importance of different features according to scenario features through a dynamic routing weight mechanism, so as to improve recommendation accuracy.

[0205] Specifically, the routing weight is determined based on the scenario feature data, so that the routing weight can determine the degree of influence of the attribute feature data on the interaction probability based on the scenario features, thereby realizing scenario awareness, and the recommendation results obtained by the same user in different scenarios are differentiated, for example, coffee coupons are recommended on workdays, and restaurant packages are recommended on weekends; on the other hand, feature-level dynamic adjustment can also be realized, for example, during game promotions, the weight of the "discount intensity" feature is automatically increased; and by generating weights, only a small number of parameters need to be added to the probability prediction model, without retraining a new model for each different scenario, which improves the recommendation efficiency.

[0206] In some embodiments, the routing weight can be obtained by adding a gating mechanism in the probability prediction model. When training the probability prediction model, by assigning reference labels to the training data, the model can learn the degree of influence of different features in the training data on the prediction results in different scenarios while learning the probability prediction capability, so that the trained probability prediction model can automatically generate the routing weight according to the scenario feature data based on the gating mechanism, thereby realizing an automatic dynamic routing mechanism.

[0207] In some embodiments, when fusing the first-order feature interaction result and the second-order feature interaction result, the network output weights of the two feature interaction networks can also be adjusted based on the routing weight, for example, weighted calculation is performed on the first-order feature interaction result and the second-order feature interaction result based on the routing weight to obtain a low-order feature interaction result.

[0208] Step 232: Perform weighted fusion on at least two feature interaction results based on the routing weight to obtain a fused feature.

[0209] In some embodiments, the i-th network output result is determined based on the i-th feature interaction result among the at least two feature interaction results and the routing weight corresponding to the i-th feature interaction network, where i is a positive integer.

[0210] In some embodiments, the low-order network output result is determined based on the low-order feature interaction result and the low-order routing weight corresponding to the low-order feature interaction network; the high-order network output result is determined based on the high-order feature interaction result and the high-order routing weight corresponding to the high-order feature interaction network.

[0211] The low-order routing weight is determined based on the network parameters of the low-order feature interaction network and the scenario feature data; the high-order routing weight is determined based on the network parameters of the high-order feature interaction network and the scenario feature data.

[0212] In some embodiments, low-order routing weights are determined based on network parameters and scene feature data of low-order feature cross-networks; high-order routing weights are determined based on network parameters, scene feature data, and attribute feature data of high-order feature cross-networks.

[0213] Step 233: The interaction probability is obtained by predicting based on the fusion features through the prediction network in the probability prediction model.

[0214] A probabilistic prediction model is a pre-trained machine learning model used to predict the probability of a specified event occurring.

[0215] In some embodiments, the network outputs corresponding to at least two feature cross-results are concatenated to obtain fused features, and the interaction probability is predicted by a prediction network based on the fused features.

[0216] This is illustrative; please refer to it. Figure 10 , Figure 10 This is a schematic diagram of the model structure of a probabilistic prediction model based on dynamic routing provided in an exemplary embodiment of this application, as shown below. Figure 10 As shown, the probability prediction model 1000 includes a low-order feature cross-network 1010, a high-order feature cross-network 1020, and a prediction network 1030.

[0217] The resource recommendation method provided in this application, in determining the interaction probability between a first account and candidate resources, employs, as follows: Figure 10 The probabilistic prediction model 1000 shown executes steps 1 to 6, including step 1, information collection; step 2, feature vectorization; step 3, vector concatenation; step 4, information extraction; step 5, network task; and step 6, task prediction.

[0218] Regarding the information collection process, collect data that may be relevant to conversion rate prediction.

[0219] It is illustrative to collect basic information about the brands associated with the recommended vouchers, basic information about the vouchers themselves, statistical information such as the vouchers' recent click-through rate and conversion rate, as well as basic information about the users and their behavioral sequences.

[0220] For the feature vectorization process, feature data is mapped to a high-dimensional space.

[0221] Intuitively, the information collected in the first step is divided into three categories: dense features, sparse features, and sequence features. Dense features are continuous numerical features, such as the click-through rate and redemption rate of a certain type of coupon on a given day, the user's age, and the number of days they logged into the game. Sparse features are categorical features, such as user gender, location, and the brand category of the coupon. Sequence features are time-related data, such as the historical sequence of coupons a user has received or the historical sequence of brands a user has been exposed to. These features are vectorized and transformed into input tensors for the model. Both sparse and dense features use lookup embedding, representing each feature with a 16-dimensional feature tensor. Since the numerical values ​​of dense features have specific meanings, the vectorized dense features are multiplied by their numerical values ​​to obtain the final result. Sequence features, due to their temporal relationship, utilize an attention mechanism to extract deep features. In the attention mechanism, query is brand_id, and key is the sequence of brand_ids for user exposure, claiming, consumption, and subscription. The final sequence tensor representation can be calculated using the formula: Attention(query,key,value)=softmax(MLP(query,key))*value.

[0222] The sequence information contains users' historical behavior regarding in-game voucher services, reflecting their overall interests. Therefore, the feature module uses user behavior sequences to represent shared information. When using the Attention mechanism for information extraction, the query will differ depending on the game the user is playing. The query information includes not only the brand information to be recommended but also information about the current game. This allows for the full extraction of shared interest features from the user sequence and the differences between different games.

[0223] For the vector concatenation process, vectors are concatenated into a matrix vector according to a unified dimension.

[0224] To illustrate, based on actual business needs, vector concatenation mainly consists of three parts: using dense feature vectors as input to the SENet & DCNv2 network; concatenating sparse feature vectors and dense feature vectors to form a new vector as input to the SENet & FM network; and concatenating sparse feature, dense feature, and sequence vectors to form a new vector as input to the STEM network.

[0225] For the information extraction process, low-order feature cross-networks and high-order feature cross-networks are used to extract information respectively.

[0226] Indicatively, the system fully extracts information from the input features, mainly divided into three parts: the first-order feature crossover (FM) module, the second-order feature crossover (DCNv2) module, and the high-order feature crossover (Deep) module. At the same time, the SENet module is used for automatic feature selection, and two scene information gating modules are used to perceive the differences in scene distribution.

[0227] First, the high-order feature cross-network provided in the embodiments of this application will be further introduced.

[0228] The basic structure of the STEM network has been introduced previously. In this embodiment, a two-layer STEM network is used to form the Deep Layer. The input of the STEM network is a concatenation of Dense Feature, Sparse Feature, and Sequence Feature, and it is divided into two parts in the Embedding dimension: the Task part x T and Shared part x S Please refer to Figure 11 , Figure 11 This is a schematic diagram of a single-layer structure of a multi-expert cross-network provided in an exemplary embodiment of this application, as shown below. Figure 11 As shown, each layer of the multi-expert cross-network model structure shares an expert module 1110 and a task expert module 1120. The shared expert module 1110 includes at least two shared expert sub-networks, and the task expert module 1120 includes at least two task expert sub-networks. The shared expert module 1110 corresponds to the shared gating network 1101, and the task expert module 1120 corresponds to the task gating network 1102.

[0229] Where, x Scene Representing scene characteristics, including scene-related contextual features such as channel ID, game ID, day of the week, and hour. The input represents the input from the level l task expert. The input represents the l-th layer shared expert. Both Task Expert and Shared Expert consist of two fully connected layers and have the same structure.

[0230] G represents a gating network, whose input is a concatenation of the expert network's input and scene features, and whose output is the aggregated weights of each expert network. and It is the weighted sum of all expert outputs. Please refer to Formulas 3 to 9 below for details.

[0231]

[0232]

[0233] in, For the output features of the i-th task expert, Task_Expert i For the i-th task expert subnetwork, The input features (task-specific features) for the l-th layer task expert. Shared_Expert is the output feature of the i-th shared expert. i For the i-th shared expert subnetwork, The input features (cross-task general features) are shared by the experts at layer l.

[0234] W1 and W2 are the two-layer weight matrices of the expert network. d represents the feature dimension, and b1 and b2 are the bias terms of the expert network. This is the ReLU activation function.

[0235] x Scene For scene feature data, W T This is the task gating weight matrix. d scene For scene feature dimensions, b T For the bias of task experts, Φ(x) is the Softmax activation function, ω T To output weights, This involves assigning weights to n task experts and n shared experts, where n is the number of expert networks. W S To share the gate weight matrix, b S For the bias of shared experts, ω S To share gating weights, That is, to allocate shared expert weights.

[0236] Let i be the weight of the expert for the i-th task. Let be the contribution weight of the i-th shared expert to the task output. This is the input for the task expert at level l+1. Let be the weight of the i-th shared expert; Reality and Identical items can be merged; This is the input for the shared expert at level l+1.

[0237] In the final layer, the output of the experts in the Task section will be... As the final output x of the STEM network STEM .

[0238] The first-order feature cross-network provided in the embodiments of this application will be further described.

[0239] In some embodiments, the first input vector is expanded to obtain a one-dimensional input vector by dimensional reshaping. The first input vector is a two-dimensional feature vector obtained based on dense feature vectors and sparse feature vectors. The one-dimensional input vector is then subjected to feature crossing processing through a first-order feature crossing network to obtain a first-order feature crossing result.

[0240] The basic structures of SENet and FM networks have been introduced above. In this embodiment, the standard SENet is improved by using a reshape method instead of the original max-pooling dimensionality reduction method, which reduces information loss.

[0241] Please refer to Figure 12 , Figure 12 This is a schematic diagram of dimensional reshaping provided in an exemplary embodiment of this application, such as... Figure 12 As shown, x i Given a tensor of 1201 as input to SENet, x re 1202 is the flattened one-dimensional tensor, 1203 is the MLP, ω is the weight of the MLP output, 1204 is ⊙, which represents element-wise multiplication, and x o The final SENet output is 1205.

[0242] If we represent the two-dimensional tensor as a column, please refer to the following formula for the representation of the input feature tensor and the dynamic weight vector.

[0243] Equations 10 and 11.

[0244]

[0245] ω=(ω 1 ,ω 2 ,…,ω f ) Formula 11,

[0246] Where, x i The input is a two-dimensional feature tensor with dimensions f×d, where f is the number of features, d is the embedding dimension of each feature (e.g., 16-dimensional), and ω is the feature importance weight vector generated by SENet, with the same dimension as the number of features. j The weight of the j-th feature is between 0 and 1, and is generated by SENet through the Sigmoid activation output.

[0247] The weighted feature output is based on formulas 10 and 11. Please refer to formula 12 below.

[0248]

[0249] Where, x o The weighted feature tensor still has dimensions f×d. This is to assign importance weights to the j-th eigenvector.

[0250] make The final output x is then obtained through a standard FM network. FM Please refer to Formulas 13 to 15 below.

[0251]

[0252] x FM =Concat(c1,c2) ​​Formula 15,

[0253] Where c1 is a first-order interaction term, used to represent the independent influence result of concatenating all features, and Concat() is the concatenation function. Represents the weighted eigenvector of the j-th element. The summation of all elements, c2 is a second-order cross term, i.e., pairwise tensors. Inner product splicing, Used to indicate the vector inner product operation, x FM This is the first-order feature crossover result, which is the final output of the FM network. It retains both the independent and interactive effects of the features for use by subsequent prediction networks.

[0254] It is worth noting that in actual business applications, the output of the FM module fluctuates significantly. Therefore, to increase the stability of training, the input x of the SENet&FM network is... i Before, for the initial input x I A module to prevent gradient backpropagation has been added; please refer to Formula 16 below.

[0255] x i =Stop_gradient(x I ) Equation 16,

[0256] Where, x i x is the network input of a first-order feature cross network. I The original input features are unweighted two-dimensional tensors. Stop_gradient() is used to indicate the operation to prevent gradient backpropagation, fix the input gradient of SENet, avoid the fluctuation of the FM module from affecting the weight update of SENet, and improve training stability.

[0257] The method provided in this application replaces pooling with dimension reshaping, which reduces the computational complexity of first-order feature cross processing while preserving the complete information of sparse features, achieving lossless information processing and improving the efficiency of first-order feature cross processing. On the other hand, by preventing gradient backpropagation, the interference of the FM module with SENet's weight learning can be avoided, improving the stability of the model during training.

[0258] The second-order feature cross-network provided in the embodiments of this application will be further described.

[0259] In some embodiments, the network weight matrix in the second-order feature cross network is decomposed into a first low-dimensional matrix and a second low-dimensional matrix; the second input vector is linearly transformed by the first low-dimensional matrix to obtain an intermediate feature vector; the intermediate feature vector is linearly transformed by the second low-dimensional matrix to obtain a transformed input vector; the transformed input vector and the second input vector are subjected to feature cross processing by the second-order feature cross network to obtain the second-order feature cross result.

[0260] In this matrix, the number of columns in the first low-dimensional matrix is ​​equal to the number of rows in the second low-dimensional matrix, and the dimensions of the first and second low-dimensional matrices are lower than the dimension of the weight matrix.

[0261] For the second-order feature cross-network, SENet, with the same structure as in FM, is adopted. Since there are many dense features in actual business applications, a two-layer DCNv2 network can be used to extract second-order combined features only from the dense features.

[0262] If we let the input of the DCNv2 network be x D The output is x DCN For the feature crossover process, please refer to Formulas 17 and 18 below.

[0263]

[0264] Here, ⊙ is used to indicate element-wise multiplication of tensors. For the ReLU activation function, x d1 For the first-level cross-output, retain residual connections +x D x D For the input dense feature matrix, This is the first layer cross-weight matrix. This is the first-level bias term. DCN This is the final output of the DCNv2 network. This is the second layer cross-weight matrix. This is the second-level bias term.

[0265] However, in actual business operations, due to the input x DThe dimension is D = f * d, where f is the number of features and d is the embedding dimension, which is a large number, leading to difficulties in... and The computational complexity is D. 2 This causes the model to consume a lot of resources during training and deployment, leading to a surge in latency. Therefore, a Low-rank approach can be used to reduce latency. and To break it down into two low-dimensional matrices, please refer to Formulas 19 and 20 below.

[0266]

[0267] in, This is the first layer cross-weight matrix. This is the second layer cross-weight matrix. and It is a left decomposition matrix. and It is a right decomposition matrix. k is a user-defined constant, k << D; in the actual model, k = 16. Therefore, and The computational complexity was reduced to kD, while the model performance remained almost unchanged.

[0268] Ultimately, x STEm x FM and x DCN The data is then spliced ​​together as the output of the information extraction module.

[0269] The method provided in this application embodiment, through second-order cross Low-rank decomposition, reduces the number of parameters from O(D) to O(D). 2 The computational complexity is reduced to O(KD), which accelerates the computation in the second-order feature crossing process, reduces the computational complexity, and improves the efficiency of the second-order feature crossing process.

[0270] For network tasks, please refer to Figure 13 , Figure 13 This is a schematic diagram of a network task provided in an exemplary embodiment of this application, such as... Figure 13 As shown, input data 1310 is received through the input layer, and the input data 1310 includes a one-dimensional or multi-dimensional tensor; the input data 1310 is processed by the fully connected layer and the activation function to obtain the mapped data 1320; the output layer outputs the output data 1330 based on the mapped data 1320, and the output data 1330 is one-dimensional data.

[0271] Specifically, after extracting information from the model input through different modules, the resulting tensors are concatenated and then input into the network task module. The network task module contains multiple fully connected layers, which essentially reduce the dimensionality of the extracted tensors layer by layer, ultimately obtaining a 1-dimensional output, denoted as logit, which serves as the output of the network task.

[0272] By concatenating the outputs of multiple networks and then using them for prediction, it is possible to integrate low-order and high-order cross-results, thereby improving the model's expressive power.

[0273] After obtaining the logit output, the Sigmoid activation function transforms it into a value between 0 and 1, which serves as the model's prediction of the coupon's click-through rate. From this, the binary cross-entropy loss function can be calculated; the specific calculation method is shown in Formula 21 below.

[0274]

[0275] Where y represents the real label. To predict the interaction probabilities, the model parameters are continuously updated during training through gradient backpropagation.

[0276] In summary, the method provided in this application generates low-order routing weights using scene feature data for low-order feature cross networks and generates high-order routing weights using concatenated features of scene feature data and attribute feature data for high-order feature cross networks. This enables the paths generated for different feature cross networks to match the feature cross order in the feature cross networks, thereby improving the accuracy of dynamic routing.

[0277] In some embodiments, the resource recommendation method provided in this application further includes a training process for a probabilistic prediction model. For details on the method provided in this application, please refer to... Figure 14 This document illustrates a flowchart of a training method for a probabilistic prediction model provided in an exemplary embodiment of this application. This method can be executed by a terminal, a server, or both simultaneously. This embodiment uses server execution as an example for illustration. Figure 14 As shown, the method includes the following steps:

[0278] Step 1410: Obtain sample prediction data.

[0279] The sample prediction data is used to predict the probability of sample account interaction with sample resources. The sample prediction data includes sample attribute data and sample scenario data, and the sample prediction data has corresponding reference labels.

[0280] Taking the prediction of user interaction probability with resources in at least two first applications as an example, sample prediction data can be obtained from the business logs of at least two first applications. The sample prediction data may include historical business data and real-time business data of at least two first applications.

[0281] For illustrative purposes, the sample attribute data includes dense features (such as click-through rate, user level, etc.), sparse features (such as user identifier, product identifier, etc.), and time-series features (including behavior sequences). The sample scenario data includes the recommendation channel identifier and recommendation time period of the recommended resources. The reference label can be used to indicate whether the user clicks on the recommended resource, and the value is 0 or 1.

[0282] Step 1420: Perform feature cross processing on the sample prediction data through at least two feature cross networks in the candidate probability prediction model to obtain at least two sample feature cross results.

[0283] The prediction process includes a feature cross processing process and a probability prediction process. The feature cross processing process includes a low-order cross processing process and a high-order cross processing process. For example, the candidate probability prediction model includes the FM network, DCNv2 network, and STEM network to be trained. The low-order cross processing process processes first-order and second-order feature crosses through the FM network and third-order feature crosses through DCNv2.

[0284] Step 1430: The prediction network in the candidate probability prediction model predicts the sample interaction probability of the sample account for the sample resource based on the cross-results of at least two sample features and the sample scenario data.

[0285] In the probability prediction process, the network outputs of the FM network and the DCNv2 network are weighted and fused according to the sample routing weights; in the higher-order cross processing process, the STEM network performs feature cross processing based on the task network and the expert network respectively, and dynamically weights the network outputs of the task network and the expert network according to the sample routing weights.

[0286] Step 1440: Train the candidate prediction model based on the difference between the sample interaction probability and the reference label until it meets the training requirements, and obtain the probability prediction model.

[0287] During training, using `stop_gradient` stops gradient backpropagation, preventing routing weights from affecting the adjustment of model parameters in the feature cross model.

[0288] As an illustration, you can refer to the above formula to calculate the predicted loss value based on the difference between the sample interaction probability and the reference label, and train the candidate probability prediction model based on the predicted loss value. This will not be elaborated further here.

[0289] Optionally, the training requirements include, but are not limited to, at least one of the following: convergence of the predicted loss value, or reaching a preset threshold number of training iterations. This application embodiment does not limit these requirements.

[0290] In summary, the method provided in this application embodiment realizes an end-to-end training optimization model training scheme for routing weights and model parameters, which can ensure the scene perception capability of the model and enable the trained probabilistic prediction model to improve prediction accuracy based on scene features.

[0291] The following section provides a more detailed description of the application scenarios corresponding to the resource recommendation method provided in this application embodiment. The resource recommendation method provided in this application embodiment can be applied to any application scenario involving resource recommendation, such as online social scenarios, online shopping scenarios, virtual game scenarios, AR scenarios, and VR scenarios; this application embodiment does not limit it in this regard.

[0292] I. Applied to virtual game scenarios.

[0293] Taking virtual game scenarios as an example, a game platform can connect to at least two game applications and recommend resources to players in at least two game applications through a recommendation system.

[0294] Players can log in to different game applications using a unified game account on the gaming platform. However, for different game applications, or even different time periods when a player logs into a game application, different recommendation scenarios exist. Even for the same player, preferences may differ. Therefore, it is necessary to fully consider scenario features in determining the interaction probability between players and resources, and then recommend corresponding resources based on the interaction probability to improve recommendation efficiency. Correspondingly, in multiple scenarios, it is also necessary to consider the cross-processing of features from multiple dimensions to fully utilize the degree of influence of features on the interaction probability in different scenarios and improve prediction accuracy.

[0295] This application example illustrates the use of recommending consumer vouchers provided by brands integrated into game applications to players, with the interaction probability including the probability of players clicking on the consumer vouchers.

[0296] Acquire prediction data, which is used to predict the click probability of a player account and candidate resources. The prediction data includes attribute feature data and scene feature data. The attribute feature data indicates the attribute characteristics of the player account and candidate resources, and the scene feature data indicates the scene characteristics of the interaction between the player account and candidate resources. Perform feature cross processing on the attribute feature data through at least two feature cross dimensions to obtain at least two feature cross results. Based on the at least two feature cross results and the scene feature data, predict the click probability of the player account and candidate resources. If the click probability meets the recommendation requirements, recommend candidate resources to the player account based on the click probability.

[0297] In some embodiments, during the multi-dimensional feature cross processing, network input vectors corresponding to at least two feature cross networks are obtained based on attribute feature data. The at least two feature cross networks include a low-order feature cross network and a high-order feature cross network. The low-order feature cross network corresponds to the low-order input vector, and the high-order feature cross network corresponds to the high-order input vector. The low-order feature cross network is used to perform low-order feature cross processing on the low-order input vector to obtain the low-order feature cross result. The high-order feature cross network is used to perform high-order feature cross processing on the high-order input vector to obtain the high-order feature cross result.

[0298] For the process of obtaining the network input vector of at least two feature cross-networks, the attribute feature data is vectorized to obtain attribute feature vectors. The attribute feature vectors include at least one of dense feature vectors, sparse feature vectors, and temporal feature vectors. Dense feature vectors are used to indicate continuous numerical features in the attribute feature data, sparse feature vectors are used to indicate categorical features in the attribute feature data, and temporal feature vectors are used to indicate temporal features in the attribute feature data. Based on at least one of dense feature vectors and sparse feature vectors, the low-order input vector corresponding to the low-order feature cross-network is obtained. The dense feature vectors, sparse feature vectors, and temporal feature vectors are concatenated to obtain the high-order input vector corresponding to the high-order feature cross-network.

[0299] Optionally, the low-order feature cross network includes a first-order feature cross network and a second-order feature cross network.

[0300] In some embodiments, a first input vector corresponding to a first-order feature cross-network is obtained based on dense feature vectors and sparse feature vectors; and a second input vector corresponding to a second-order feature cross-network is obtained based on at least two dense feature vectors.

[0301] The low-order feature crossing process includes: performing feature crossing processing on the first input vector through a first-order feature crossing network to obtain a first-order feature crossing result; performing feature crossing processing on the second input vector through a second-order feature crossing network to obtain a second-order feature crossing result; and fusing the first-order feature crossing result and the second-order feature crossing result to obtain the low-order feature crossing result.

[0302] In some embodiments, the first-order feature cross processing process includes expanding the first input vector into a one-dimensional input vector by dimensional reshaping, wherein the first input vector is a two-dimensional feature vector obtained based on dense feature vectors and sparse feature vectors; and performing feature cross processing on the one-dimensional input vector through a first-order feature cross network to obtain a first-order feature cross result.

[0303] In some embodiments, the second-order feature cross processing includes: decomposing the network weight matrix in the second-order feature cross network into a first low-dimensional matrix and a second low-dimensional matrix, wherein the number of columns in the first low-dimensional matrix is ​​equal to the number of rows in the second low-dimensional matrix, and the dimensions of the first and second low-dimensional matrices are lower than the dimension of the weight matrix; performing a linear transformation on the second input vector through the first low-dimensional matrix to obtain an intermediate feature vector; performing a linear transformation on the intermediate feature vector through the second low-dimensional matrix to obtain a transformed input vector; and performing feature cross processing on the transformed input vector and the second input vector through the second-order feature cross network to obtain the second-order feature cross result.

[0304] In some embodiments, the high-order feature cross network includes a multi-expert cross network.

[0305] The high-order feature cross-processing process includes: extracting general feature representations based on high-order input vectors through the shared expert subnetwork in the multi-expert cross-processing network; extracting task feature representations based on high-order input vectors through the task expert subnetwork in the multi-expert cross-processing network; and fusing the general feature representations and task feature representations to obtain the high-order feature cross-processing result.

[0306] In some embodiments, a multi-scene routing function is also included. Routing weights are determined based on scene feature data, whereby the routing weights indicate the degree of influence of attribute feature data on the click probability. The results of at least two feature interactions are weighted and fused based on the routing weights to obtain fused features. The click probability is then predicted based on the fused features using a prediction network in a probability prediction model, where the probability prediction model is a pre-trained machine learning model used to predict the probability of a specified event occurring.

[0307] In some embodiments, the process further includes training a probabilistic prediction model, acquiring sample prediction data, which is used to predict the sample click probability of a sample account on a sample resource, the sample prediction data including sample attribute data and sample scene data, and the sample prediction data corresponding to a reference label; performing feature cross processing on the sample prediction data through at least two feature cross networks in the candidate probabilistic prediction model to obtain at least two sample feature cross results; predicting the sample click probability of a sample account on a sample resource based on at least two sample feature cross results and sample scene data through the prediction network in the candidate probabilistic prediction model; training the candidate prediction model based on the difference between the sample click probability and the reference label until the training requirements are met, thus obtaining the probabilistic prediction model.

[0308] In some embodiments, the process of acquiring prediction data includes: acquiring account attribute data, resource attribute data, and context data as attribute feature data, wherein the account attribute data is used to indicate the account data of the player account, the resource attribute data is used to indicate the resource attributes of the candidate resources, and the context data is used to indicate the environmental state of the player account; and acquiring program feature data and time feature data as scene feature data, wherein the program feature data is used to indicate the program type of the candidate resources recommended to the player account, and the time feature data is used to indicate the recommendation time of the candidate resources recommended to the player account.

[0309] It is worth noting that the above-mentioned click probabilities are merely illustrative examples. The resource recommendation method provided in this application embodiment can also recommend candidate resources based on at least one type of interaction probability, such as viewing probability, usage probability, and redemption probability. This application embodiment does not limit this.

[0310] II. Applied to multimedia playback scenarios.

[0311] Taking a multimedia playback scenario as an example, a multimedia playback platform can connect to at least two multimedia playback applications and recommend resources to viewers in at least two multimedia playback applications through a recommendation system.

[0312] Viewers can log in to different multimedia playback applications using a unified viewer account on the playback platform. However, depending on the multimedia playback application or the time period during which a viewer logs in, different recommendation scenarios may exist. Even for the same viewer, preferences for recommended resources may differ. Therefore, it is necessary to fully consider scenario features in determining the interaction probability between the viewer and resources, and then recommend corresponding resources based on the interaction probability to improve recommendation efficiency. Furthermore, it is necessary to fully consider multi-dimensional feature cross-patterns across multiple scenarios to fully utilize attribute feature data in predicting interaction probabilities and improve the accuracy of probability prediction.

[0313] This application example illustrates the use of recommending multimedia content from a multimedia playback application to a viewer, where the interaction probability includes the probability of the viewer viewing the multimedia content.

[0314] In the resource recommendation process, prediction data is acquired to predict the viewing probability of audience accounts and candidate resources. The prediction data includes attribute feature data and scene feature data. Attribute feature data indicates the attribute features of audience accounts and candidate resources, while scene feature data indicates the scene features of audience accounts interacting with candidate resources. Feature cross processing is performed on the attribute feature data through at least two feature cross dimensions to obtain at least two feature cross results. Based on the at least two feature cross results and scene feature data, the viewing probability of audience accounts and candidate resources is predicted. If the viewing probability meets the recommendation requirements, candidate resources are recommended to audience accounts based on the viewing probability.

[0315] It is worth noting that the above application scenarios are merely illustrative examples, and the embodiments of this application do not limit them.

[0316] Figure 15 This is a structural block diagram of a resource recommendation apparatus provided in an exemplary embodiment of this application, such as... Figure 15 As shown, the device includes the following parts:

[0317] The acquisition module 1510 is used to acquire prediction data, which is used to predict the interaction probability between the first account and the candidate resource. The prediction data includes attribute feature data and scene feature data. The attribute feature data is used to indicate the attribute features of the first account and the candidate resource, and the scene feature data is used to indicate the scene features of the interaction between the first account and the candidate resource.

[0318] Processing module 1520 is used to perform feature cross processing on the attribute feature data through at least two feature cross dimensions to obtain at least two feature cross results;

[0319] The processing module 1520 is further configured to predict the interaction probability between the first account and the candidate resource based on the at least two feature cross results and the scene feature data;

[0320] The processing module 1520 is further configured to recommend the candidate resources to the first account based on the interaction probability when the interaction probability meets the recommendation requirements.

[0321] In some embodiments, the acquisition module 1510 is further configured to acquire network input vectors corresponding to at least two feature cross networks based on the attribute feature data, wherein the at least two feature cross networks include a low-order feature cross network and a high-order feature cross network, the low-order feature cross network corresponds to a low-order input vector, and the high-order feature cross network corresponds to a high-order input vector.

[0322] The processing module 1520 is further configured to perform low-order feature crossing processing on the low-order input vector through a low-order feature crossing network to obtain a low-order feature crossing result.

[0323] The processing module 1520 is further configured to perform high-order feature crossing processing on the high-order input vector through a high-order feature crossing network to obtain a high-order feature crossing result.

[0324] In some embodiments, the processing module 1520 is further configured to:

[0325] The attribute feature data is vectorized to obtain attribute feature vectors, which include at least one of dense feature vectors, sparse feature vectors, and temporal feature vectors. The dense feature vectors are used to indicate continuous numerical features in the attribute feature data, the sparse feature vectors are used to indicate categorical features in the attribute feature data, and the temporal feature vectors are used to indicate temporal features in the attribute feature data.

[0326] The low-order input vector corresponding to the low-order feature cross network is obtained based on at least one of the dense feature vector and the sparse feature vector.

[0327] By concatenating the dense feature vector, the sparse feature vector, and the temporal feature vector, the higher-order input vector corresponding to the higher-order feature crossover network is obtained.

[0328] In some embodiments, the low-order feature cross network includes a first-order feature cross network and a second-order feature cross network;

[0329] The acquisition module 1510 is further configured to:

[0330] Based on the dense feature vector and the sparse feature vector, obtain the first input vector corresponding to the first-order feature cross network;

[0331] The second input vector corresponding to the second-order feature cross network is obtained based on at least two dense feature vectors.

[0332] In some embodiments, the processing module 1520 is further configured to:

[0333] The first input vector is processed by the first-order feature crossing network to obtain the first-order feature crossing result.

[0334] The second input vector is subjected to feature crossing processing through the second-order feature crossing network to obtain the second-order feature crossing result;

[0335] The low-order feature cross result is obtained by fusing the first-order feature cross result and the second-order feature cross result.

[0336] In some embodiments, the processing module 1520 is further configured to:

[0337] The first input vector is expanded into a one-dimensional input vector by dimensional reshaping. The first input vector is a two-dimensional feature vector obtained based on the dense feature vector and the sparse feature vector.

[0338] The first-order feature crossing network is used to perform feature crossing processing on the one-dimensional input vector to obtain the first-order feature crossing result.

[0339] In some embodiments, the processing module 1520 is further configured to:

[0340] The network weight matrix in the second-order feature cross network is decomposed into a first low-dimensional matrix and a second low-dimensional matrix, wherein the number of columns of the first low-dimensional matrix is ​​equal to the number of rows of the second low-dimensional matrix, and the dimensions of the first low-dimensional matrix and the second low-dimensional matrix are lower than the dimension of the weight matrix.

[0341] The intermediate feature vector is obtained by performing a linear transformation on the second input vector using the first low-dimensional matrix.

[0342] The intermediate feature vector is linearly transformed using the second low-dimensional matrix to obtain the transformed input vector.

[0343] The transformed input vector and the second input vector are subjected to feature crossing processing by the second-order feature crossing network to obtain the second-order feature crossing result.

[0344] In some embodiments, the high-order feature cross-network includes a multi-expert cross-network;

[0345] The processing module 1520 is also used for:

[0346] A general feature representation is extracted based on the higher-order input vector through a shared expert subnetwork in the multi-expert cross-network.

[0347] Task feature representations are extracted based on the higher-order input vector through the task expert subnetwork in the multi-expert cross network;

[0348] The higher-order feature cross result is obtained by fusing the general feature representation and the task feature representation.

[0349] In some embodiments, the processing module 1520 is further configured to:

[0350] The routing weight is determined based on the scene feature data, and the routing weight is used to indicate the degree of influence of the attribute feature data on the interaction probability.

[0351] Based on the routing weights, the at least two feature cross results are weighted and fused to obtain fused features;

[0352] The interaction probability is obtained by predicting the fused features through the prediction network in the probability prediction model, which is a pre-trained machine learning model for predicting the probability of a specified event occurring.

[0353] In some embodiments, the acquisition module 1510 is further configured to acquire sample prediction data, which is used to predict the sample interaction probability of a sample account to a sample resource. The sample prediction data includes sample attribute data and sample scenario data, and the sample prediction data corresponds to a reference label.

[0354] The processing module 1520 is further configured to perform feature cross processing on the sample prediction data through at least two feature cross networks in the candidate probability prediction model to obtain at least two sample feature cross results.

[0355] The processing module 1520 is further configured to predict the sample interaction probability of the sample account for the sample resource based on the cross-result of the at least two sample features and the sample scene data through the prediction network in the candidate probability prediction model;

[0356] The processing module 1520 is further configured to train the candidate prediction model based on the difference between the sample interaction probability and the reference label until the training requirements are met, thereby obtaining the probability prediction model.

[0357] In some embodiments, the acquisition module 1510 is further configured to:

[0358] Account attribute data, resource attribute data, and context data are obtained as the attribute feature data. The account attribute data is used to indicate the account data of the first account, the resource attribute data is used to indicate the resource attributes of the candidate resources, and the context data is used to indicate the environmental state of the first account.

[0359] The program feature data and time feature data are obtained as the scene feature data. The program feature data is used to indicate the type of program that recommends the candidate resource to the first account, and the time feature data is used to indicate the time when the candidate resource is recommended to the first account.

[0360] In summary, the apparatus provided in this application, through feature cross processing of at least two feature cross dimensions, obtains at least two feature cross results based on attribute feature data. This fully explores the high-order nonlinear relationships between features from multiple dimensions, taking into account both the independent and joint effects of various features on the interaction probability. As a result, the interaction probability obtained by using at least two feature cross results in the prediction is more accurate, thereby improving the recommendation efficiency of resource recommendation based on account and resource interaction probability.

[0361] It should be noted that the resource recommendation device provided in the above embodiments is only an example of the division of the above functional modules. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above.

[0362] Figure 16 This illustration shows a structural block diagram of a terminal 1600 provided in an exemplary embodiment of this application. The terminal 1600 may be a smartphone, tablet computer, MP3 player, MP4 player, laptop computer, or desktop computer. The terminal 1600 may also be referred to as user equipment, portable terminal, laptop terminal, desktop terminal, or other names.

[0363] Typically, terminal 1600 includes a processor 1601 and a memory 1602.

[0364] Processor 1601 may include one or more processing cores, such as a quad-core processor, an octa-core processor, etc. Processor 1601 may be implemented using at least one hardware form selected from Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), and Programmable Logic Array (PLA). Processor 1601 may also include a main processor and a coprocessor. The main processor, also known as the Central Processing Unit (CPU), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, processor 1601 may integrate a Graphics Processing Unit (GPU), which is responsible for rendering and drawing the content to be displayed on the screen. In some embodiments, processor 1601 may also include an Artificial Intelligence (AI) processor, which is used to handle computational operations related to machine learning.

[0365] Memory 1602 may include one or more computer-readable storage media, which may be non-transitory. Memory 1602 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In some embodiments, the non-transitory computer-readable storage media in memory 1602 is used to store at least one instruction, which is executed by processor 1601 to implement the resource recommendation method provided in the method embodiments of this application.

[0366] In some embodiments, the terminal 1600 also includes other components 1603, as those skilled in the art will understand. Figure 16 The structure shown does not constitute a limitation on terminal 1600 and may include more or fewer components than shown, or combine certain components, or use different component arrangements.

[0367] Embodiments of this application also provide a computer device that can be implemented as follows: Figure 1 The terminal or server shown. The computer device includes a processor and a memory, the memory storing at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, at least one program, code set, or instruction set is loaded and executed by the processor to implement the resource recommendation method provided in the above-described method embodiments.

[0368] Embodiments of this application also provide a computer-readable storage medium storing at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, at least one program, code set, or instruction set is loaded and executed by a processor to implement the resource recommendation method provided in the above-described method embodiments.

[0369] Embodiments of this application also provide a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the resource recommendation method provided in the above-described method embodiments.

[0370] Optionally, the computer-readable storage medium may include: read-only memory (ROM), random access memory (RAM), solid-state drives (SSDs), or optical discs, etc. The random access memory may include resistive random access memory (ReRAM) and dynamic random access memory (DRAM). The sequence numbers of the embodiments in this application are merely descriptive and do not represent the superiority or inferiority of the embodiments.

[0371] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0372] The above description is merely an optional embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A resource recommendation method, characterized in that, The method includes: Acquire prediction data, which is used to predict the probability of interaction between the first account and the candidate resource. The prediction data includes attribute feature data and scene feature data. The attribute feature data is used to indicate the attribute features of the first account and the candidate resource. The scene feature data is used to indicate the scene features of the interaction between the first account and the candidate resource. The attribute feature data are subjected to feature cross processing by at least two feature cross dimensions to obtain at least two feature cross results; Based on the at least two feature cross results and the scene feature data, predict the interaction probability between the first account and the candidate resource; If the interaction probability meets the recommendation requirements, the candidate resource is recommended to the first account based on the interaction probability.

2. The method according to claim 1, characterized in that, The attribute feature data is subjected to feature cross processing through at least two feature cross dimensions to obtain at least two feature cross results, including: Based on the attribute feature data, at least two feature cross networks are obtained, each corresponding to a network input vector. The at least two feature cross networks include a low-order feature cross network and a high-order feature cross network. The low-order feature cross network corresponds to a low-order input vector, and the high-order feature cross network corresponds to a high-order input vector. The low-order input vector is processed by a low-order feature crossing network to obtain the low-order feature crossing result. The higher-order input vector is processed by a higher-order feature crossing network to obtain the higher-order feature crossing result.

3. The method according to claim 2, characterized in that, The step of obtaining network input vectors corresponding to at least two feature cross-networks based on the attribute feature data includes: The attribute feature data is vectorized to obtain attribute feature vectors, which include at least one of dense feature vectors, sparse feature vectors, and temporal feature vectors. The dense feature vectors are used to indicate continuous numerical features in the attribute feature data, the sparse feature vectors are used to indicate categorical features in the attribute feature data, and the temporal feature vectors are used to indicate temporal features in the attribute feature data. The low-order input vector corresponding to the low-order feature cross network is obtained based on at least one of the dense feature vector and the sparse feature vector. By concatenating the dense feature vector, the sparse feature vector, and the temporal feature vector, the higher-order input vector corresponding to the higher-order feature crossover network is obtained.

4. The method according to claim 3, characterized in that, The low-order feature cross network includes a first-order feature cross network and a second-order feature cross network; The step of obtaining the low-order input vector corresponding to the low-order feature cross network based on at least one of the dense feature vector and the sparse feature vector includes: Based on the dense feature vector and the sparse feature vector, obtain the first input vector corresponding to the first-order feature cross network; The second input vector corresponding to the second-order feature cross network is obtained based on at least two dense feature vectors.

5. The method according to claim 4, characterized in that, The step of performing low-order feature crossing processing on the low-order input vector through a low-order feature crossing network to obtain the low-order feature crossing result includes: The first input vector is processed by the first-order feature crossing network to obtain the first-order feature crossing result. The second input vector is subjected to feature crossing processing through the second-order feature crossing network to obtain the second-order feature crossing result; The low-order feature cross result is obtained by fusing the first-order feature cross result and the second-order feature cross result.

6. The method according to claim 5, characterized in that, The step of performing feature crossing processing on the first input vector through the first-order feature crossing network to obtain the first-order feature crossing result includes: The first input vector is expanded into a one-dimensional input vector by dimensional reshaping. The first input vector is a two-dimensional feature vector obtained based on the dense feature vector and the sparse feature vector. The first-order feature crossing network is used to perform feature crossing processing on the one-dimensional input vector to obtain the first-order feature crossing result.

7. The method according to claim 5, characterized in that, The step of performing feature crossing processing on the second input vector through the second-order feature crossing network to obtain the second-order feature crossing result includes: The network weight matrix in the second-order feature cross network is decomposed into a first low-dimensional matrix and a second low-dimensional matrix, wherein the number of columns of the first low-dimensional matrix is ​​equal to the number of rows of the second low-dimensional matrix, and the dimensions of the first low-dimensional matrix and the second low-dimensional matrix are lower than the dimension of the weight matrix. The intermediate feature vector is obtained by performing a linear transformation on the second input vector using the first low-dimensional matrix. The intermediate feature vector is linearly transformed using the second low-dimensional matrix to obtain the transformed input vector. The transformed input vector and the second input vector are subjected to feature crossing processing by the second-order feature crossing network to obtain the second-order feature crossing result.

8. The method according to claim 3, characterized in that, The higher-order feature cross-network includes a multi-expert cross-network; The step of performing high-order feature crossing processing on the high-order input vector through a high-order feature crossing network to obtain the high-order feature crossing result includes: A general feature representation is extracted based on the higher-order input vector through a shared expert subnetwork in the multi-expert cross-network. Task feature representations are extracted based on the higher-order input vector through the task expert subnetwork in the multi-expert cross network; The higher-order feature cross result is obtained by fusing the general feature representation and the task feature representation.

9. The method according to any one of claims 1 to 8, characterized in that, The prediction of the interaction probability between the first account and the candidate resource based on the at least two feature cross results and the scene feature data includes: The routing weight is determined based on the scene feature data, and the routing weight is used to indicate the degree of influence of the attribute feature data on the interaction probability. Based on the routing weights, the at least two feature cross results are weighted and fused to obtain fused features; The interaction probability is obtained by predicting the fused features through the prediction network in the probability prediction model, which is a pre-trained machine learning model for predicting the probability of a specified event occurring.

10. The method according to claim 9, characterized in that, The method further includes: Obtain sample prediction data, which is used to predict the probability of sample interaction between sample accounts and sample resources. The sample prediction data includes sample attribute data and sample scenario data, and the sample prediction data corresponds to reference tags. The sample prediction data is subjected to feature cross processing by at least two feature cross networks in the candidate probability prediction model to obtain at least two sample feature cross results. The prediction network in the candidate probability prediction model predicts the sample interaction probability of the sample account for the sample resource based on the cross-results of the features of at least two samples and the sample scenario data. The candidate prediction model is trained based on the difference between the sample interaction probability and the reference label until the training requirements are met, thus obtaining the probability prediction model.

11. The method according to any one of claims 1 to 8, characterized in that, The acquisition of prediction data includes: Account attribute data, resource attribute data, and context data are obtained as the attribute feature data. The account attribute data is used to indicate the account data of the first account, the resource attribute data is used to indicate the resource attributes of the candidate resources, and the context data is used to indicate the environmental state of the first account. The program feature data and time feature data are obtained as the scene feature data. The program feature data is used to indicate the type of program that recommends the candidate resource to the first account, and the time feature data is used to indicate the time when the candidate resource is recommended to the first account.

12. A resource recommendation device, characterized in that, The device includes: An acquisition module is used to acquire prediction data, which is used to predict the interaction probability between a first account and a candidate resource. The prediction data includes attribute feature data and scene feature data. The attribute feature data is used to indicate the attribute features of the first account and the candidate resource, and the scene feature data is used to indicate the scene features of the interaction between the first account and the candidate resource. The processing module is used to perform feature cross processing on the attribute feature data through at least two feature cross dimensions to obtain at least two feature cross results; The processing module is further configured to predict the interaction probability between the first account and the candidate resource based on the at least two feature cross results and the scene feature data; The processing module is further configured to recommend the candidate resources to the first account based on the interaction probability if the interaction probability meets the recommendation requirements.

13. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing at least one computer program, the at least one computer program being loaded and executed by the processor to implement the resource recommendation method as described in any one of claims 1 to 11.

14. A computer-readable storage medium, characterized in that, The storage medium stores at least one computer program, which is loaded and executed by a processor to implement the resource recommendation method as described in any one of claims 1 to 11.

15. A computer program product, characterized in that, It includes a computer program that, when executed by a processor, implements the resource recommendation method as described in any one of claims 1 to 11.

Images