Resource recommendation method and device, electronic equipment and storage medium

Abstract

The present disclosure relates to a resource recommendation method and device, electronic equipment and storage medium, and relates to the field of artificial intelligence. The method comprises: in response to a resource recommendation request for an account, obtaining resource content features of candidate resources, and obtaining object features for the account and the candidate resources, including interaction features of the account, account attribute features of the account, and interaction features of the candidate resources, and resource attribute features of the candidate resources; inputting the resource content features and part of the object features into a feature cross network, performing feature cross processing, and obtaining cross features corresponding to the candidate resources; predicting the probability of interaction between the account and the candidate resources based on the object features and the cross features, and obtaining interaction prediction data of the candidate resources; and determining recommended resources recommended to the account from the multiple candidate resources based on the interaction prediction data of each candidate resource. The technical solutions provided by the embodiments of the present disclosure can effectively improve the accuracy of the recommendation results.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, and in particular to resource recommendation methods, apparatus, electronic devices, and storage media. Background Technology

[0002] Deep learning techniques based on neural networks and feature representations have been widely used in recommender systems, significantly improving the performance metrics and user experience of recommender systems, and have gradually become one of the experimental benchmarks in the field of recommender systems.

[0003] In related technologies, interaction prediction based on the interaction features corresponding to accounts and objects can yield predictive data characterizing the probability of interaction between accounts and objects. This predictive data can then be used to determine whether to recommend an object to an account. However, the accuracy of the recommendation results still needs improvement. Summary of the Invention

[0004] This disclosure provides a resource recommendation method, apparatus, electronic device, and storage medium, introducing resource content features to improve recommendation effectiveness. The technical solution of this disclosure is as follows: According to a first aspect of the present disclosure, a resource recommendation method is provided, comprising: In response to a resource recommendation request for an account, the resource content features of the candidate resources are obtained, and the object features for the account and the candidate resources are obtained. The object features include the interaction features of the account, the account attribute features of the account, the interaction features of the candidate resources, and the resource attribute features of the candidate resources. The resource content features and some of the object features are input into a feature cross-network for feature cross-processing to obtain the cross features corresponding to the candidate resources. Based on the object features and the cross features, the probability of the account interacting with the candidate resource is predicted to obtain the interaction prediction data of the candidate resource. Based on the interaction prediction data of each of the candidate resources, recommended resources to be recommended to the account are determined from the multiple candidate resources.

[0005] Optionally, the step of inputting the resource content features and part of the object features into a feature cross-network for feature cross-processing to obtain the cross features corresponding to the candidate resources includes: The resource content features and some of the object features are input into a feature cross-network and horizontally concatenated to obtain the concatenated features corresponding to the candidate resources. If the variance of the splicing feature meets the preset variance screening conditions, the splicing feature is used as the cross feature corresponding to the candidate resource.

[0006] Optionally, the step of predicting the probability of the account interacting with the candidate resource based on the object features and the cross features, to obtain interaction prediction data for the candidate resource, includes: Obtain the candidate sequence features corresponding to the candidate resources, wherein the candidate sequence features characterize the order of the candidate resources among the multiple candidate resources; Based on the object features, the cross features, and the candidate sequence features, the probability of the account interacting with the candidate resource is predicted, and the interaction prediction data of the candidate resource is obtained.

[0007] Optionally, the acquisition of resource content features of candidate resources includes: Obtain the resource content information of the candidate resources, wherein the resource content information includes modal information corresponding to at least one modality; Each modal information is input into a feature extraction network corresponding to the modality of the modal information, and the feature extraction network performs feature extraction processing on the modal information under the modality to obtain at least one resource modal feature; The resource content features of the candidate resource are obtained by combining the at least one resource modal features.

[0008] Optionally, the method further includes: Obtain the sample resource content information of the first sample resource, and obtain the sample information for the first sample resource and the sample account; the sample information includes the sample interaction information and sample account attribute information of the sample account, as well as the sample interaction information and sample resource attribute information of the first sample resource. Obtain the first sample resource content feature corresponding to the first sample resource content information, and obtain the first sample object feature and the second sample object feature of the sample information. The first sample object feature is obtained by feature extraction of all the sample information, and the second sample object feature is obtained by feature extraction of a portion of the sample information. The first sample resource content features and the second sample object features are input into the initial feature cross network and feature cross processing is performed to obtain the first sample cross features corresponding to the first sample resource. Based on the first sample object features and the first sample cross features, the probability of the sample account interacting with the first sample resource is predicted to obtain the first sample interaction prediction data of the first sample resource. Based on the first sample interaction prediction data and the historical interaction data for the first sample resource and the sample account, the first loss information is determined, and the initial feature cross network is adjusted based on the first loss information to obtain the trained feature cross network.

[0009] Optionally, the step of predicting the probability of the sample account interacting with the first sample resource based on the first sample object features and the first sample cross features, to obtain the first sample interaction prediction data for the first sample resource, includes: The first sample object features and the first sample cross features are input into the initial interaction prediction network to predict the probability of the sample account interacting with the first sample resource, thereby obtaining the first sample interaction prediction data of the first sample resource.

[0010] Optionally, the method further includes: Based on the first sample interaction prediction data and the historical interaction data, the second loss information is determined; Based on the second loss information, the initial feature cross network and the initial interaction prediction network are jointly adjusted to obtain the trained feature cross network and the trained interaction prediction network. The interaction prediction network is used to predict the probability of the account interacting with the candidate resource based on the object features and the cross features to obtain the interaction prediction data of the candidate resource.

[0011] According to a second aspect of the present disclosure, a resource recommendation apparatus is provided, comprising: The acquisition module is configured to perform an action in response to a resource recommendation request for an account, acquire the resource content features of candidate resources, and acquire object features for the account and the candidate resources. The object features include the interaction features of the account, the account attribute features of the account, and the interaction features and resource attribute features of the candidate resources. The feature cross module is configured to input the resource content features and part of the object features into the feature cross network, perform feature cross processing, and obtain the cross features corresponding to the candidate resource. The interaction prediction module is configured to predict the probability of the account interacting with the candidate resource based on the object features and the cross features, and obtain the interaction prediction data of the candidate resource. The resource recommendation module is configured to perform interactive prediction data based on each of the candidate resources to determine recommended resources to be recommended to the account from a plurality of candidate resources.

[0012] Optionally, the feature intersection module includes: The feature splicing unit is configured to input the resource content features and part of the object features into the feature cross network and perform horizontal splicing to obtain the spliced ​​features corresponding to the candidate resource. The feature filtering unit is configured to use the spliced ​​features as the cross features corresponding to the candidate resources when the variance of the spliced ​​features meets the preset variance filtering conditions.

[0013] Optionally, the interactive prediction module may include: The candidate sequence feature acquisition unit is configured to acquire the candidate sequence features corresponding to the candidate resource, wherein the candidate sequence features characterize the order of the candidate resource among multiple candidate resources; The second interaction prediction unit is configured to predict the probability of the account interacting with the candidate resource based on the object features, the cross features, and the candidate sequence features, and obtain the interaction prediction data of the candidate resource.

[0014] Optionally, the acquisition module includes: The resource content information acquisition unit is configured to acquire the resource content information of the candidate resource, wherein the resource content information includes modal information corresponding to at least one modality; The feature extraction unit is configured to execute a feature extraction network that inputs each modality information into a feature extraction network corresponding to the modality of the modality information, and the feature extraction network performs feature extraction processing on the modality information under the modality to obtain at least one resource modality feature; The feature combination unit is configured to combine the at least one resource modality feature to obtain the resource content feature of the candidate resource.

[0015] Optionally, the device includes: The first information acquisition unit is configured to acquire sample resource content information of the first sample resource and acquire sample information for the first sample resource and the sample account; the sample information includes sample interaction information and sample account attribute information of the sample account, as well as sample interaction information and sample resource attribute information of the first sample resource. The first feature acquisition unit is configured to acquire the first sample resource content feature corresponding to the first sample resource content information, and acquire the first sample object feature and the second sample object feature of the sample information. The first sample object feature is obtained by feature extraction of all the sample information, and the second sample object feature is obtained by feature extraction of a portion of the sample information. The feature crossing unit is configured to input the first sample resource content features and the second sample object features into the initial feature crossing network, perform feature crossing processing, and obtain the first sample crossing features corresponding to the first sample resource. The first interaction prediction unit is configured to perform a prediction of the probability that the sample account interacts with the first sample resource based on the first sample object features and the first sample cross features, so as to obtain the first sample interaction prediction data of the first sample resource. The first training unit is configured to perform a process based on the first sample interaction prediction data and historical interaction data for the first sample resource and the sample account to determine first loss information, and adjust the initial feature cross network based on the first loss information to obtain the trained feature cross network.

[0016] Optionally, the first interaction prediction unit includes: The interaction prediction subunit is configured to input the first sample object features and the first sample cross features into the initial interaction prediction network, predict the probability of the sample account interacting with the first sample resource, and obtain the first sample interaction prediction data of the first sample resource.

[0017] Optionally, the device further includes: The second loss calculation unit is configured to determine second loss information based on the first sample interaction prediction data and the historical interaction data. The second training unit is configured to perform joint adjustment of the initial feature cross network and the initial interaction prediction network based on the second loss information to obtain the trained feature cross network and the trained interaction prediction network. The interaction prediction network is used to predict the probability of the account interacting with the candidate resource based on the object features and the cross features to obtain the interaction prediction data of the candidate resource.

[0018] According to a third aspect of the present disclosure, an electronic device is provided, comprising: a processor; and a memory for storing processor-executable instructions; wherein the processor is configured to execute the instructions to implement the resource recommendation method described in any one of the first aspects of the present disclosure.

[0019] According to a fourth aspect of the present disclosure, a computer-readable storage medium is provided that, when instructions in the computer-readable storage medium are executed by a processor of an electronic device, enables the electronic device to perform a resource recommendation method as described in any one of the first aspects of the present disclosure.

[0020] According to a fifth aspect of the present disclosure, a computer program product is provided, including computer instructions that, when executed by a processor, implement the resource recommendation method as described in any one of the first aspects of the present disclosure.

[0021] The technical solutions provided by the embodiments of this disclosure have at least the following beneficial effects: The technical solution provided by the embodiments of this disclosure introduces resource content features of candidate resources in addition to account interaction features, account attribute features, candidate resource interaction features, and candidate resource resource attribute features, thus enriching the feature representation dimensions of candidate resources. Furthermore, by performing feature cross-referencing on resource content features and some object features, the resource content features of the content domain are mapped to cross-features of the interaction behavior domain, solving the problem of cross-domain features. Therefore, in the process of predicting the probability of interaction between an account and candidate resources based on object features and cross-features, the accuracy of interaction prediction data can be improved by leveraging cross-features. By combining the interaction prediction data corresponding to each candidate resource, recommended resources can be determined from multiple candidate resources to recommend to the account, effectively improving the accuracy of recommendation results, thereby increasing the recommendation-interaction conversion rate and optimizing the recommendation effect.

[0022] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0023] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure, and are not intended to unduly limit this disclosure.

[0024] Figure 1 This is a schematic diagram illustrating the application environment of a resource recommendation method according to an exemplary embodiment; Figure 2 This is a flowchart illustrating a resource recommendation method according to an exemplary embodiment; Figure 3 This is a flowchart illustrating a feature extraction method according to an exemplary embodiment; Figure 4 This is a schematic diagram of a model according to an exemplary embodiment; Figure 5 This is a schematic diagram of a network architecture for implementing a resource recommendation method, according to an exemplary embodiment. Figure 6 This is a schematic diagram illustrating the training of a network model according to an exemplary embodiment; Figure 7 This is a schematic diagram illustrating the training of another network model according to an exemplary embodiment; Figure 8 This is a block diagram illustrating a resource recommendation apparatus according to an exemplary embodiment; Figure 9 This is a block diagram of a mobile device for implementing a resource recommendation method, according to an exemplary embodiment. Figure 10 This is a service device block diagram illustrating a resource recommendation method according to an exemplary embodiment. Detailed Implementation

[0025] To enable those skilled in the art to better understand the technical solutions of this disclosure, the technical solutions in the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings.

[0026] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0027] To facilitate understanding of the technical solutions and their effects described in the embodiments of this application, the relevant technical terms are explained in the embodiments of this application: End-to-end learning, also known as end-to-end training, refers to training without dividing the learning process into modules or stages, directly optimizing the overall goal of the task. In end-to-end learning, it is generally not necessary to explicitly define the functions of different modules or stages, and the intermediate process requires no human intervention. Training data for end-to-end learning is in the form of "input-output" pairs, requiring no additional information.

[0028] Multilayer Perceptron (MLP), also known as Artificial Neural Network (ANN), has multiple hidden layers in addition to the input and output layers. The simplest MLP contains only one hidden layer.

[0029] Feature crosses (also known as feature combinations) are combinations of features of different types or dimensions.

[0030] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for display, data used for analysis, etc.) involved in this disclosure are all information and data authorized by the user or fully authorized by all parties.

[0031] Please see Figure 1The diagram illustrates an application environment for a resource recommendation method according to an exemplary embodiment. The application environment may include a terminal 110 and a server 120, which can be connected via a wired network or a wireless network.

[0032] Terminal 110 may be, but is not limited to, electronic devices such as smartphones, desktop computers, tablets, laptops, smart speakers, digital assistants, augmented reality (AR) / virtual reality (VR) devices, and smart wearable devices. Terminal 110 may have an application (App) installed. This application can be a standalone application or a subroutine within a standalone application. Users of Terminal 110 can log in to the application using pre-registered user information, which may include an account and password. Optionally, the operating system running on Terminal 110 may include, but is not limited to, Android, iOS, Linux, and Windows. Server 120 may be a server providing backend services for the application in Terminal 110, or other servers connected and communicating with the application's backend server. It can be a single server, a server cluster or distributed system composed of multiple servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms.

[0033] In the embodiments of this disclosure, terminal 110 may respond to a resource recommendation request sent by an account to server 120, where the account is the user's login application account. Server 120 receives the resource recommendation request, determines recommended resources from multiple candidate resources in response to the request, and pushes the resource display information of the recommended resources to terminal 110. Terminal 110 then displays the resource display information of the recommended resources to the user.

[0034] Specifically, server 120 performs interaction prediction processing based on object features and cross features to obtain interaction prediction data for each candidate resource among multiple candidate resources. Then, server 120 can determine recommended resources from among the multiple candidate resources based on the interaction prediction data for each candidate resource. The object features include account interaction features and account attribute features, as well as the interaction features and resource attribute features of the candidate resources. By introducing resource content features of candidate resources into the interaction prediction, the resource recommendation effect can be effectively improved.

[0035] In addition, it should be noted that, Figure 1 The example shown is merely one application environment provided by this disclosure. In practical applications, other application environments may also be included, such as more terminals. This disclosure does not limit this.

[0036] Figure 2 This is a flowchart illustrating a resource recommendation method according to an exemplary embodiment, such as... Figure 2 As shown, the resource recommendation method is used in a recommendation system, and the method may include the following steps: In step S210, in response to a resource recommendation request for an account, the resource content characteristics of the candidate resources are obtained, and the object characteristics for the account and the candidate resources are obtained.

[0037] In this embodiment, the candidate resources can be categorized as articles, videos, music, advertisements, images, web pages, or social media updates. The recommendation system, by recommending suitable resources to accounts, can provide users associated with those accounts with information of interest, thus satisfying their information needs. To achieve accuracy, personalization, diversity, and high conversion rates in recommendations, it is necessary to select some or all of the candidate resources from a pool of candidates as recommended resources.

[0038] In this embodiment of the disclosure, the object features include the account's interaction features, the account's account attribute features, the candidate resource's interaction features, and the candidate resource's resource attribute features. The account's interaction features describe the user's interactive behavior information based on the account, and may include, but are not limited to, the frequency and timing of various interactive behaviors such as viewing, clicking, forwarding, or purchasing, and the sequence features of resources associated with each interactive behavior. The account's attribute features describe the attribute information of the account itself and the associated users, and may include, but are not limited to, account identifier, user age, user gender, and user preferences. The candidate resource's interaction features describe the interactive behavior information generated for the candidate resource, and may include, but are not limited to, the sequence features of accounts that have performed various interactive behaviors for the candidate resource, and the frequency of each interactive behavior for the candidate resource. The candidate resource's resource attribute features describe the attribute information of the candidate resource itself, and may include, but are not limited to, resource identifier and resource category. Furthermore, the object features may also include the account's context features, which may include the device identifier, device type, system version, and application version of the logged-in device.

[0039] In this embodiment of the disclosure, the resource content feature of the candidate resource is a feature representation of the content information contained in the candidate resource. The resource information contained in the candidate resource may include, but is not limited to, text information, image information, voice information, web page link address information, etc.

[0040] In one embodiment of this disclosure, the above-mentioned features are feature representations extracted based on corresponding information. For example... Figure 3 As shown, step S210 may specifically include: In step S211, object information for the account and candidate resources and resource content information for the candidate resources are obtained.

[0041] The object information includes account interaction information, account attribute information, candidate resource interaction information, and candidate resource resource attribute information. Account interaction information describes user interaction behaviors based on the account, and may include, but is not limited to, behavioral metrics such as frequency and timing of various interaction behaviors (viewing, clicking, forwarding, or purchasing) and the sequence characteristics of resources associated with each interaction behavior. Account attribute information describes the attributes of the account itself and associated users, and may include, but is not limited to, account identifier, user age, user gender, and user preferences. Candidate resource interaction information describes interaction behaviors performed on candidate resources, and may include, but is not limited to, the sequence characteristics of accounts that have performed various interaction behaviors on candidate resources and behavioral metrics such as the frequency of each interaction behavior on candidate resources. Candidate resource resource attribute information describes the attribute information of the candidate resource itself, and may include, but is not limited to, resource identifier and resource category.

[0042] Part of the object information can be part of the account's interaction information or account attribute information, or part of the candidate resource's interaction information or resource attribute information, or a combination of part of the account's interaction information or account attribute information and part of the candidate resource's interaction information or resource attribute information, such as a combination of the account's identification information and the candidate resource's identification information.

[0043] The resource content information of candidate resources may include, but is not limited to, text information, image information, audio information, web page link address information, etc.

[0044] In step S213, the object information is input into the first feature representation network for feature extraction processing to obtain object features for accounts and candidate resources.

[0045] Feasibly, the object information is input into the first feature representation network to extract features, and the extracted features are encoded or embedded to obtain the corresponding object features.

[0046] In step S215, the resource content information is input into the second feature representation network for feature extraction processing to obtain the resource content features of the candidate resources.

[0047] Feasibly, the resource content information is input into the second feature representation to extract the resource content features, and the extracted features are embedded to obtain the corresponding resource content features.

[0048] Optionally, the second feature indicates that the network can be a convolutional neural network, such as a lightweight convolutional neural network like MobileNet or ResNet18.

[0049] In step S217, partial object information is input into the third feature representation network for feature extraction processing to obtain partial object features.

[0050] Feasibly, partial object information can be input into a third feature representation network to extract features, and the extracted features can be encoded or embedded to obtain the corresponding partial object features.

[0051] Optionally, the third feature representation network can be a lightweight feature representation network that is compressed from the first feature representation network, thereby effectively improving the efficiency of feature extraction.

[0052] In the above embodiments, the object information, partial object information, and resource content information are extracted and represented respectively to obtain object features, partial object features, and resource content features, providing effective feature data for subsequent feature cross and interaction prediction.

[0053] Furthermore, the resource content information of the candidate resources includes modality information corresponding to at least one modality, and the second feature representation network includes a feature extraction network corresponding to at least one modality. Step S215 can be implemented as follows: In step S2151, the resource content information of the candidate resource is obtained, and the resource content information includes modal information corresponding to at least one modality.

[0054] Modalities can be text, images, voice, links, etc., and modal information corresponds to text information, image information, voice information, link information, etc.

[0055] In step S2153, each modal information is input into a feature extraction network corresponding to the modality of the modal information. The feature extraction network performs feature extraction processing on the modal information under the modality to obtain at least one resource modal feature.

[0056] like Figure 4As shown, the resource content information of candidate resources can include image information, text information, and speech information. Image information is input into an image feature representation network to obtain image features, text information is input into a text feature representation network to obtain text features, and speech information is input into a speech feature representation network to obtain speech features. That is, the second feature representation network includes at least one modality-corresponding feature extraction network, and each feature extraction network is used to extract features from the modality information of one modality.

[0057] In step S2153, at least one resource modal feature is combined to obtain the resource content features of the candidate resource.

[0058] It is feasible to horizontally concatenate at least one resource modality feature to obtain resource content features.

[0059] In the above embodiments, the resource content information contained in the candidate resources is modally subdivided, and features are extracted from each modality information according to the modality. This can effectively extract resource modality features of different modalities and combine them to obtain multimodal resource content features, thus enriching the feature dimensions of the resource content.

[0060] In step S230, the resource content features and some object features are input into the feature cross-network for feature cross-processing to obtain the cross features corresponding to the candidate resources.

[0061] In this embodiment, based on data analysis, mining, and a deep understanding of recommendation services, it is believed that whether an account interacts with a candidate resource is related to the resource content characteristics of the candidate resource. However, related technologies that rely solely on object characteristics (excluding resource content characteristics) for interaction prediction in this embodiment ignore the impact of resource content on the interaction. Therefore, by adding resource content characteristics to the candidate resource side, this embodiment can effectively improve recommendation performance and interaction conversion rates.

[0062] It is understandable that resource content features are feature representations in the content space, while object features are feature representations in the behavior space. The content space and behavior space belong to feature spaces of different domains. In current recommendation systems, which predict the probability of an account interacting with candidate resources using features from the behavior space, directly inputting resource content features and object features into the interaction prediction network for interaction prediction processing will cause the network to focus more on object features and ignore resource content features. This makes it difficult to achieve the goal of improving recommendation performance using resource content features even with the addition of this feature dimension. Therefore, in this embodiment, feature cross-referencing is performed on resource content features and a portion of object features. That is, resource content features are mapped to the behavior space using a portion of the object features within the behavior space. The resulting cross-features can represent resource content features aligned with the interaction behavior space, thereby solving the problem of cross-domain features.

[0063] In this embodiment of the disclosure, the partial object features include some features among the object features. The partial object features can be a portion of the account's interaction features or account attribute features, a portion of the candidate resource's interaction features or resource attribute features, or a combination of a portion of the account's interaction features or account attribute features and a portion of the candidate resource's interaction features or resource attribute features, such as a combination of the account's identifier features and the candidate resource's identifier features. Since partial object features are part of the object features and also belong to the feature representation of the behavior space, they can help solve the problem of cross-domain features. Compared to the full set of object features, using partial object features reduces the number of introduced features, thereby reducing the workload of feature representation and feature intersection, and improving overall efficiency.

[0064] In one embodiment of this disclosure, step S230 may include: In step S231, the resource content features and some object features are input into the feature cross-network and horizontally concatenated to obtain the concatenated features corresponding to the candidate resources.

[0065] For example, the resource content features and some object features are input into a feature cross-network, and the identification features of candidate resources in the resource content features, the identification features of accounts in the partial object features, and the identification features of accounts in the partial object features are horizontally concatenated to obtain concatenated features.

[0066] Some object features use the identification features of candidate resources and the identification features of accounts. Optionally, current context features may also be included, which are used to characterize the current interaction environment of the account.

[0067] In step S233, if the variance of the splicing feature meets the preset variance screening conditions, the splicing feature is used as the cross feature corresponding to the candidate resource.

[0068] The preset variance condition limits the variance threshold of the features. If the variance of the concatenated features is too low, the concatenated features carry less information and have little effect on interactive prediction. Therefore, the effectiveness of the features can be improved by retaining or deleting the concatenated features according to the variance screening condition.

[0069] In addition, the spliced ​​features can be filtered by missing rate, constant value ratio, etc., to obtain cross features.

[0070] In one feasible implementation of this disclosure, when the resource content features include multiple resource modal features, the resource modal features corresponding to low weights in the splicing features are filtered according to the weight parameters of each modality to obtain cross features. In another feasible implementation, the order of the multiple resource modal features is determined according to the weight parameters of each modality, and the multiple resource modal features are re-spliced ​​with a portion of the object features according to the order to obtain cross features.

[0071] Optionally, the feature crossover network can be a multilayer perceptron, which performs feature crossover and filtering on resource content features and some object features.

[0072] In the above embodiments, feature cross-referencing is performed on resource content features and partial object features. This involves mapping resource content features from the content space to the behavior space using features within the behavior space, thereby resolving the cross-domain feature issue and enhancing the competitiveness of resource content features in interactive prediction processing. This achieves the goal of improving recommendation performance by leveraging resource content features. Furthermore, partial object features are some interactive features within object features, also belonging to the feature representation of the behavior space. This not only helps solve the cross-domain feature problem but also reduces the number of features introduced compared to using all object features, thus reducing the workload of feature cross-referencing and improving overall efficiency.

[0073] In step S250, based on object features and cross features, the probability of an account interacting with candidate resources is predicted to obtain interaction prediction data for candidate resources.

[0074] In this embodiment of the disclosure, the interaction prediction data represents the probability of an account interacting with a candidate resource. The interaction prediction data may include the probability of at least one interaction category. The interaction category may include, but is not limited to, interactive behaviors such as watching, clicking, purchasing, forwarding, commenting, liking, or blocking. The probability of each interaction category indicates the likelihood of an account and a candidate resource generating the corresponding category of interactive behavior.

[0075] In one embodiment of this disclosure, step S250 may include: In step S251, the object features and cross features are input into the interactive prediction network to perform feature fusion and obtain the target fused features.

[0076] It is feasible to add the cross feature as a feature of one dimension of the object feature corresponding to the candidate resource to the object feature to obtain the fused feature.

[0077] In step S253, interactive prediction processing is performed based on the fusion features to obtain interactive prediction data.

[0078] Interaction prediction networks are built upon arbitrary recommendation algorithm models. The network's input is features, and its output is interaction prediction data. When predictions are required for multiple interaction categories, interaction prediction networks can be built based on MMoE (Multi-gate Mixture-of-Experts) and MLP (Multilayer Perceptron) models.

[0079] In the above embodiments, by introducing cross features in the interaction prediction process, where cross features represent resource content features of the alignment behavior space, the feature dimensions used are enriched, the accuracy of the interaction prediction data is improved, and thus the effect of object recommendation can be effectively improved.

[0080] In one embodiment of this disclosure, step S250 may further include: In step S255, the candidate sequence features corresponding to the candidate resources are obtained, and the candidate sequence features represent the order of the candidate resources among multiple candidate resources; In step S257, based on object features, cross features, and candidate sequence features, the probability of an account interacting with a candidate resource is predicted to obtain interaction prediction data for the candidate resource.

[0081] In the above embodiments, considering that in addition to pushing a single candidate resource to an account, a sequence of candidate resources can also be pushed, and the order of the candidate resources in the candidate resource sequence will also affect the interaction, in addition to object features and cross features, candidate sequence features can also be introduced to further improve the accuracy of interaction prediction.

[0082] In step S270, based on the interactive prediction data of each candidate resource, recommended resources to be recommended to the account are determined from multiple candidate resources.

[0083] In one embodiment of this disclosure, recommended resources that can be pushed to an account can be selected from multiple candidate resources based on the interaction prediction data of each candidate resource and preset interaction prediction conditions. The recommended resources can be one or more candidate resources, and the interaction prediction data of each candidate resource in the recommended resources meets the preset interaction prediction conditions.

[0084] When the interaction prediction data includes the predicted probabilities corresponding to at least one interaction category, the interaction prediction condition can be limited to a threshold for the predicted probability corresponding to a single interaction category, or it can be limited to a threshold for the average predicted probabilities corresponding to multiple interaction categories. For example, the target interaction data for each candidate resource can be obtained by weighting and summing the predicted probabilities corresponding to each interaction category included in the interaction prediction data, based on the weights of each interaction category. The target interaction data can be used to measure the overall recommendation effect when the candidate resource is used as a recommendation resource.

[0085] The interactive prediction condition can also be limited to the number of candidate resources in the recommended resources. Based on the interactive prediction data corresponding to each candidate resource, each candidate resource is sorted, and the top N candidate resources are selected as recommended resources. The top N candidate resources constitute the recommendation sequence.

[0086] Feasibly, after the recommended resources are determined, the server sends the display information corresponding to the recommended resources to the terminal logged into the account, and the terminal displays the display information.

[0087] Figure 5 This is a schematic diagram of a network architecture for implementing a resource recommendation method, according to an exemplary embodiment. Figure 5 As shown, in this embodiment, a cross-domain subnet module is constructed in addition to the first feature representation network and the interaction prediction network. The cross-domain subnet module includes a second feature representation network, a third feature representation network, and a feature cross-network. The second feature representation network is used to extract and represent features of the resource content information of candidate resources to obtain resource content features. The third feature representation network is used to extract and represent features of some object information to obtain some object features. The feature cross-network is used to perform feature cross-combination of resource content features and some object features to obtain cross features, which can characterize resource content features mapped to the behavior space. The interaction prediction network can perform interaction prediction processing based on the newly introduced cross features and the object features output by the original first feature representation network to obtain interaction prediction data for accounts and candidate resources. The interaction prediction data can be used to determine whether to use the candidate resource as the recommended resource corresponding to the account. The specific processes of feature representation, feature cross-combination, and interaction prediction processing of the above information can be referred to the foregoing embodiments and will not be repeated here.

[0088] Figure 6This is a schematic diagram illustrating the training of a network model according to an exemplary embodiment, such as... Figure 6 As shown, the method may further include: In step S310, the sample resource content information of the first sample resource is obtained, and the sample information for the first sample resource and the sample account is obtained.

[0089] The sample information includes sample interaction information and sample account attribute information for the sample account, as well as sample interaction information and sample resource attribute information for the first sample resource. The sample information and sample resource content information can be referenced from the object information and resource content information, and will not be elaborated upon here.

[0090] In step S311, the first sample resource content feature corresponding to the first sample resource content information is obtained, and the first sample object feature and the second sample object feature of the sample information are obtained. The first sample object feature is obtained by feature extraction of all sample information, and the second sample object feature is obtained by feature extraction of a portion of the sample information.

[0091] In other words, the features of the second sample object are part of the features of the first sample object.

[0092] like Figure 6 As shown, the trained second feature representation network is used to extract features from the sample resource content information to obtain the corresponding first sample resource content features. The trained third feature representation network is then used to extract features from a portion of the sample information to obtain the second sample object features. Furthermore, the trained first feature representation network is used to extract features from the sample information to obtain the first sample object features.

[0093] In step S312, the first sample resource content features and the second sample object features are input into the initial feature cross network for feature cross processing to obtain the first sample cross features corresponding to the first sample resource.

[0094] Feature crossover processing can be referred to in the aforementioned embodiments, and will not be repeated here.

[0095] In step S313, based on the first sample object features and the first sample cross features, the probability of the sample account interacting with the first sample resource is predicted to obtain the first sample interaction prediction data of the first sample resource.

[0096] Feasibly, the first sample cross-features can be input as follows: Figure 6The trained interaction prediction network shown performs interaction prediction processing to obtain the first sample interaction prediction data. The interaction prediction network can be an algorithmic model that predicts one or more specific interactive behaviors. The obtained first sample interaction prediction data indicates the probability that the sample account and the first sample resource will produce that interactive behavior, such as the click probability.

[0097] In step S314, based on the first sample interaction prediction data and the historical interaction data for the first sample resources and sample accounts, the first loss information is determined, and the initial feature cross network is adjusted based on the first loss information to obtain the trained feature cross network.

[0098] It should be noted that the first sample resource is the historical recommendation resource pushed to the sample account, and there is corresponding historical interaction data. The historical interaction data represents the historical interaction information between the sample account and the first sample resource. It can be the tag information corresponding to a specific type of interaction behavior. The above tag information represents whether this type of interaction behavior has occurred, such as 1 indicating that an interaction has occurred and 0 indicating that no interaction has occurred.

[0099] When the interaction prediction processing during the training phase is only for predicting a specific type of interaction, the required interaction prediction network is compared to... Figure 5 The interactive prediction network shown is simpler and requires less data to process during the training phase, which can speed up training and improve training efficiency.

[0100] Feasibly, based on the first loss information, the network parameters of the initial feature cross network are adjusted to obtain the trained feature cross network.

[0101] The trained feature cross-network is used to perform feature cross-referencing between resource content features and some object features.

[0102] The specific processes of feature extraction, feature crossing, and interactive prediction processing of the above information can be referred to the aforementioned embodiments, and will not be repeated here.

[0103] In the above embodiments, for Figure 5 The cross-domain subnet module shown can be trained independently for its feature cross-networks. This allows the trained feature cross-networks to better map resource content features to the behavior space, thereby enhancing the competitiveness of resource content-dimensional features in interaction prediction processing and improving the accuracy of interaction prediction data. Furthermore, the first feature representation network, third feature representation network, and feature cross-network involved in the cross-domain subnet module can be jointly trained end-to-end. This allows the trained first feature representation network to more effectively mine and extract resource content features related to the interaction.

[0104] Figure 7This is a training diagram of another network model illustrated according to an exemplary embodiment, such as... Figure 7 As shown, the method may further include: In step S410, the sample resource content information of the first sample resource is obtained, and the sample information for the first sample resource and the sample account is obtained.

[0105] The sample information includes sample interaction information and sample account attribute information of the sample account, as well as sample interaction information and sample resource attribute information of the first sample resource.

[0106] In step S411, the first sample resource content feature corresponding to the first sample resource content information is obtained, and the first sample object feature and the second sample object feature of the sample information are obtained. The first sample object feature is obtained by feature extraction of all sample information, and the second sample object feature is obtained by feature extraction of a portion of the sample information.

[0107] In step S412, the first sample resource content features and the second sample object features are input into the initial feature cross network for feature cross processing to obtain the first sample cross features corresponding to the first sample resource.

[0108] Steps S410 to S412 are the same as steps S310 to S312, and will not be repeated here.

[0109] In step S413, the first sample object features and the first sample cross features are input into the initial interaction prediction network to predict the probability of the sample account interacting with the first sample resource, thereby obtaining the first sample interaction prediction data of the first sample resource.

[0110] The initial interaction prediction network is Figure 5 The interaction prediction network shown is in an incomplete training state, that is, in the embodiments of this disclosure, the main network of the recommendation system - the initial interaction prediction network and the cross-domain network are jointly trained.

[0111] In step S414, the second loss information is determined based on the first sample interaction prediction data and historical interaction data.

[0112] It is feasible to calculate the second loss information based on the cross-entropy loss function, using the first sample interaction prediction data and historical interaction data.

[0113] In step S415, based on the second loss information, the initial feature cross network and the initial interaction prediction network are jointly adjusted to obtain the trained feature cross network and the trained interaction prediction network.

[0114] Interaction prediction networks are used to predict the probability of an account interacting with candidate resources based on object features and cross features, thus obtaining interaction prediction data for candidate resources.

[0115] In addition, the initial first feature representation network, the initial second feature representation network, and the initial third feature representation network can be jointly trained. The trained first feature representation network is used to obtain object features, the trained second feature representation network is used to obtain resource content features, and the trained third feature representation network is used to obtain partial object features.

[0116] Among them, the trained feature cross network is used to perform feature cross between resource content features and some object features, and the trained interaction prediction network is used to perform interaction prediction processing required by the business based on object features and cross features.

[0117] The specific processes of feature extraction, feature crossing, and interactive prediction processing of the above information can be referred to the aforementioned embodiments, and will not be repeated here.

[0118] In the above embodiments, the resource recommendation method provided in this disclosure is trained end-to-end, which enables the overall performance of the network to meet the requirements of the final interaction prediction, thereby achieving the goal of improving the recommendation effect by utilizing the features of the resource content dimension.

[0119] As can be seen from the technical solutions provided by the embodiments of this disclosure above, the technical solutions provided by the embodiments of this disclosure, in addition to the interaction features of the account, the account attribute features of the account, the interaction features of the candidate resources, and the resource attribute features of the candidate resources, introduce the resource content features of the candidate resources, enriching the feature representation dimensions of the candidate resources; and by performing feature cross-features on the resource content features and some object features, the resource content features of the content domain are mapped to the cross-features of the interaction behavior domain, solving the problem of cross-domain features; thus, in the process of predicting the probability of interaction between the account and the candidate resources based on object features and cross-features, the accuracy of the interaction prediction data can be improved by leveraging the cross-features; and by combining the interaction prediction data corresponding to each candidate resource, the recommended resources to be recommended to the account can be determined from multiple candidate resources, which can effectively improve the accuracy of the recommendation results, thereby improving the recommendation-interaction conversion rate and optimizing the recommendation effect.

[0120] Figure 8 This is a block diagram illustrating an object recommendation device according to an exemplary embodiment. (Refer to...) Figure 8 The device includes: The acquisition module 810 is configured to perform a response to a resource recommendation request for an account, acquire the resource content features of candidate resources, and acquire object features for the account and the candidate resources. The object features include the interaction features of the account, the account attribute features of the account, and the interaction features and resource attribute features of the candidate resources. The feature cross module 820 is configured to input the resource content features and part of the object features into the feature cross network, perform feature cross processing, and obtain the cross features corresponding to the candidate resource. The interaction prediction module 830 is configured to predict the probability of the account interacting with the candidate resource based on the object features and the cross features, and obtain the interaction prediction data of the candidate resource. The resource recommendation module 840 is configured to perform interactive prediction data based on each of the candidate resources to determine recommended resources to be recommended to the account from a plurality of candidate resources.

[0121] Optionally, the feature intersection module 820 includes: The feature splicing unit is configured to input the resource content features and part of the object features into the feature cross network and perform horizontal splicing to obtain the spliced ​​features corresponding to the candidate resource. The feature filtering unit is configured to use the spliced ​​features as the cross features corresponding to the candidate resources when the variance of the spliced ​​features meets the preset variance filtering conditions.

[0122] Optionally, the interaction prediction module 830 may include: The candidate sequence feature acquisition unit is configured to acquire the candidate sequence features corresponding to the candidate resource, wherein the candidate sequence features characterize the order of the candidate resource among multiple candidate resources; The second interaction prediction unit is configured to predict the probability of the account interacting with the candidate resource based on the object features, the cross features, and the candidate sequence features, and obtain the interaction prediction data of the candidate resource.

[0123] Optionally, the acquisition module 810 includes: The resource content information acquisition unit is configured to acquire the resource content information of the candidate resource, wherein the resource content information includes modal information corresponding to at least one modality; The feature extraction unit is configured to execute a feature extraction network that inputs each modality information into a feature extraction network corresponding to the modality of the modality information, and the feature extraction network performs feature extraction processing on the modality information under the modality to obtain at least one resource modality feature; The feature combination unit is configured to combine the at least one resource modality feature to obtain the resource content feature of the candidate resource.

[0124] Optionally, the device includes: The first information acquisition unit is configured to acquire sample resource content information of the first sample resource and acquire sample information for the first sample resource and the sample account; the sample information includes sample interaction information and sample account attribute information of the sample account, as well as sample interaction information and sample resource attribute information of the first sample resource. The first feature acquisition unit is configured to acquire the first sample resource content feature corresponding to the first sample resource content information, and acquire the first sample object feature and the second sample object feature of the sample information. The first sample object feature is obtained by feature extraction of all the sample information, and the second sample object feature is obtained by feature extraction of a portion of the sample information. The feature crossing unit is configured to input the first sample resource content features and the second sample object features into the initial feature crossing network, perform feature crossing processing, and obtain the first sample crossing features corresponding to the first sample resource. The first interaction prediction unit is configured to perform a prediction of the probability that the sample account interacts with the first sample resource based on the first sample object features and the first sample cross features, so as to obtain the first sample interaction prediction data of the first sample resource. The first training unit is configured to perform a process based on the first sample interaction prediction data and historical interaction data for the first sample resource and the sample account to determine first loss information, and adjust the initial feature cross network based on the first loss information to obtain the trained feature cross network.

[0125] Optionally, the first interaction prediction unit includes: The interaction prediction subunit is configured to input the first sample object features and the first sample cross features into the initial interaction prediction network, predict the probability of the sample account interacting with the first sample resource, and obtain the first sample interaction prediction data of the first sample resource.

[0126] Optionally, the device further includes: The second loss calculation unit is configured to determine second loss information based on the first sample interaction prediction data and the historical interaction data. The second training unit is configured to perform joint adjustment of the initial feature cross network and the initial interaction prediction network based on the second loss information to obtain the trained feature cross network and the trained interaction prediction network. The interaction prediction network is used to predict the probability of the account interacting with the candidate resource based on the object features and the cross features to obtain the interaction prediction data of the candidate resource.

[0127] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0128] In an exemplary embodiment of this disclosure, an electronic device is also provided, including: a processor; and a memory for storing processor-executable instructions; wherein the processor is configured to execute the instructions to implement the resource recommendation method as described in the embodiments of this disclosure.

[0129] Figure 9 This is a block diagram illustrating an electronic device for implementing a resource recommendation method according to an exemplary embodiment. The electronic device may be a terminal, and its internal structure diagram may be as follows: Figure 9 As shown, the electronic device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a resource recommendation method. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the device's casing, or an external keyboard, touchpad, or mouse.

[0130] Figure 10 This is a block diagram illustrating an electronic device for implementing a resource recommendation method according to an exemplary embodiment. The electronic device may be a server, and its internal structure diagram may be as follows: Figure 10As shown, this electronic device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a resource recommendation method.

[0131] Those skilled in the art will understand that Figure 9 and Figure 10 The structure shown is merely a block diagram of a portion of the structure related to the present disclosure and does not constitute a limitation on the electronic device to which the present disclosure is applied. A specific electronic device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0132] In exemplary embodiments of this disclosure, a computer-readable storage medium including instructions is also provided, which, when executed by a processor of an electronic device, enable the electronic device to perform the resource recommendation method in the embodiments of this disclosure.

[0133] In an exemplary embodiment of this disclosure, a computer program product is also provided, including computer instructions that, when executed by a processor, implement the resource recommendation method in the embodiments of this disclosure.

[0134] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0135] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the following claims.

[0136] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.

Claims

1. A resource recommendation method, characterized in that, The method includes: In response to a resource recommendation request for an account, the resource content features of candidate resources are obtained, and object features for the account and the candidate resources are obtained. The object features include the interaction features of the account, the account attribute features of the account, the interaction features of the candidate resources, and the resource attribute features of the candidate resources. The resource content features are feature representations of modal information corresponding to at least one modality contained in the candidate resources. The resource content features and a portion of the object features are input into a feature cross-network for feature cross-processing to obtain the cross features corresponding to the candidate resource; the cross features represent the features after mapping the resource content features to the interaction behavior space using a portion of the object features. Based on the object features and the cross features, the probability of the account interacting with the candidate resource is predicted to obtain the interaction prediction data of the candidate resource. Based on the interaction prediction data of each of the candidate resources, recommended resources to be recommended to the account are determined from the multiple candidate resources.

2. The method according to claim 1, characterized in that, The step of inputting the resource content features and part of the object features into a feature cross-network for feature cross-processing to obtain the cross features corresponding to the candidate resources includes: The resource content features and some of the object features are input into a feature cross-network and horizontally concatenated to obtain the concatenated features corresponding to the candidate resources. If the variance of the splicing feature meets the preset variance screening conditions, the splicing feature is used as the cross feature corresponding to the candidate resource.

3. The method according to claim 1, characterized in that, The step of predicting the probability of the account interacting with the candidate resource based on the object features and the cross features, to obtain interaction prediction data for the candidate resource, includes: Obtain the candidate sequence features corresponding to the candidate resources, wherein the candidate sequence features characterize the order of the candidate resources among the multiple candidate resources; Based on the object features, the cross features, and the candidate sequence features, the probability of the account interacting with the candidate resource is predicted, and the interaction prediction data of the candidate resource is obtained.

4. The method according to claim 1, characterized in that, The resource content features for obtaining candidate resources include: Obtain the resource content information of the candidate resources, wherein the resource content information includes modal information corresponding to at least one modality; Each modal information is input into a feature extraction network corresponding to the modality of the modal information, and the feature extraction network performs feature extraction processing on the modal information under the modality to obtain at least one resource modal feature; The resource content features of the candidate resource are obtained by combining the at least one resource modal features.

5. The method according to claim 1, characterized in that, The method further includes: Obtain the sample resource content information of the first sample resource, and obtain the sample information for the first sample resource and the sample account; the sample information includes the sample interaction information and sample account attribute information of the sample account, as well as the sample interaction information and sample resource attribute information of the first sample resource. Obtain the first sample resource content feature corresponding to the first sample resource content information, and obtain the first sample object feature and the second sample object feature of the sample information. The first sample object feature is obtained by feature extraction of all the sample information, and the second sample object feature is obtained by feature extraction of a portion of the sample information. The first sample resource content features and the second sample object features are input into the initial feature cross network and feature cross processing is performed to obtain the first sample cross features corresponding to the first sample resource. Based on the first sample object features and the first sample cross features, the probability of the sample account interacting with the first sample resource is predicted to obtain the first sample interaction prediction data of the first sample resource. Based on the first sample interaction prediction data and the historical interaction data for the first sample resource and the sample account, the first loss information is determined, and the initial feature cross network is adjusted based on the first loss information to obtain the trained feature cross network.

6. The method according to claim 5, characterized in that, The step of predicting the probability of interaction between the sample account and the first sample resource based on the first sample object features and the first sample cross features, to obtain the first sample interaction prediction data for the first sample resource, includes: The first sample object features and the first sample cross features are input into the initial interaction prediction network to predict the probability of the sample account interacting with the first sample resource, thereby obtaining the first sample interaction prediction data of the first sample resource.

7. The method according to claim 6, characterized in that, The method further includes: Based on the first sample interaction prediction data and the historical interaction data, the second loss information is determined; Based on the second loss information, the initial feature cross network and the initial interaction prediction network are jointly adjusted to obtain the trained feature cross network and the trained interaction prediction network. The interaction prediction network is used to predict the probability of the account interacting with the candidate resource based on the object features and the cross features to obtain the interaction prediction data of the candidate resource.

8. A resource recommendation device, characterized in that, The device includes: The acquisition module is configured to perform an action in response to a resource recommendation request for an account, acquire resource content features of candidate resources, and acquire object features for the account and the candidate resources. The object features include interaction features of the account, account attribute features of the account, interaction features of the candidate resources, and resource attribute features of the candidate resources. The resource content features are feature representations of modal information corresponding to at least one modality contained in the candidate resources. The feature cross module is configured to input the resource content features and a portion of the object features into the feature cross network, perform feature cross processing, and obtain the cross features corresponding to the candidate resource; the cross features represent the features after mapping the resource content features to the interaction behavior space using a portion of the object features. The interaction prediction module is configured to predict the probability of the account interacting with the candidate resource based on the object features and the cross features, and obtain the interaction prediction data of the candidate resource. The resource recommendation module is configured to perform interactive prediction data based on each of the candidate resources to determine recommended resources to be recommended to the account from a plurality of candidate resources.

9. An electronic device, characterized in that, include: processor; Memory used to store the processor's executable instructions; The processor is configured to execute the instructions to implement the resource recommendation method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, When the instructions in the computer-readable storage medium are executed by the processor of the electronic device, the electronic device is able to perform the resource recommendation method as described in any one of claims 1 to 7.

Images