Resource recommendation model training method, resource pushing method, device and medium

Abstract

This disclosure provides a training method, resource recommendation method, apparatus, and medium for a resource recommendation model, relating to the fields of artificial intelligence, digital marketing, and intelligent retrieval. The training method includes: acquiring a sample set related to a sample object; processing a first basic resource feature and a second basic resource feature using an initial resource recommendation model to obtain a recommendation result for the sample; evaluating the interaction resource features of negative samples based on the first interaction resource feature to obtain a second interaction resource feature for the negative samples; and training the initial resource recommendation model based on the sample recommendation result and the tag recommendation result determined based on the second interaction resource feature using a knowledge transfer mechanism to obtain a trained resource recommendation model. The resource recommendation model is used to process basic resource features related to a target object to obtain a target recommendation result.

Description

Technical Field

[0001] This disclosure relates to the fields of artificial intelligence, digital marketing, and intelligent retrieval, and more specifically, to a training method for a resource recommendation model, a resource push method, an apparatus, a medium, an electronic device, and a program product. Background Technology

[0002] With the rapid development of internet technology, users can conveniently browse news, videos, and other resources through smartphones and other terminal devices. Related internet platforms can also recommend resources to users based on their needs, thereby improving the convenience of accessing resources and enhancing the user experience.

[0003] In realizing the present invention, the inventors discovered at least the following problems in the related technology: the matching degree between the resources pushed to users and the users' actual needs is low, making it difficult to accurately meet the users' resource acquisition needs. Furthermore, the deep learning model used to determine recommended resources based on the user's resource-related characteristics requires a large amount of training data and training time during training, which increases the computational overhead of the computing devices used to perform model training and raises the power consumption level required for model training. Summary of the Invention

[0004] In view of this, this disclosure provides a training method, a resource recommendation method, an apparatus, a medium, an electronic device, and a program product for a resource recommendation model.

[0005] One aspect of this disclosure provides a method for training a resource recommendation model, comprising: acquiring a sample set related to a sample object, wherein positive samples in the sample set include a first basic resource feature and a first interactive resource feature, the first basic resource feature being related to the resource features of the positive sample resource in a historical period, the historical period being the period before the positive sample resource is interacted with by the sample object, the first interactive resource feature being determined based on the interaction operation performed by the sample object on the positive sample resource, and negative samples in the sample set including a second basic resource feature of the negative sample resource; processing the first basic resource feature and the second basic resource feature using an initial resource recommendation model to obtain a sample recommendation result; evaluating the interactive resource feature of the negative sample based on the first interactive resource feature to obtain a second interactive resource feature of the negative sample; and training the initial resource recommendation model based on the sample recommendation result and the tag recommendation result determined based on the second interactive resource feature using a knowledge transfer mechanism to obtain a trained resource recommendation model, wherein the resource recommendation model is used to process basic resource features related to a target object to obtain a target recommendation result.

[0006] According to embodiments of this disclosure, evaluating the interaction resource features of the negative samples based on the first interaction resource features to obtain the second interaction resource features of the negative samples includes: performing a causal inference operation based on the first interaction resource features of the positive samples and preset interaction operation prior information to obtain feature distribution data of the target preset interaction features in the sample set, wherein the interaction operation prior information characterizes the similarity between the negative samples and at least one of the positive samples; determining a target distribution vector based on the feature distribution data of each of the multiple target preset interaction features; performing a solution operation on the target optimization function based on the optimization conditions determined by the target distribution vector according to the optimal transmission mechanism to obtain an optimized target transmission matrix, wherein the matrix elements of the target transmission matrix include the target feature distribution data of the negative samples, and the target feature distribution data characterizes the target distribution probability of the negative samples in the target preset interaction resource features; and determining the second interaction resource features of the negative samples based on the target feature distribution data.

[0007] According to embodiments of this disclosure, multiple positive samples correspond to multiple prior information of interaction operations; wherein, performing causal inference based on the first interaction resource feature of the positive sample and the preset prior information of interaction operations includes: determining intermediate distribution data of at least one target preset interaction feature in the positive sample based on the ratio between the first interaction resource feature of the positive sample and the prior information of interaction operations associated with the positive sample; and fusing the intermediate distribution data corresponding to each of the multiple positive samples in the sample set to obtain feature distribution data of at least one target preset interaction feature in the sample set.

[0008] According to embodiments of this disclosure, the prior information of the interaction operation is determined based on the following operations: processing the first interaction resource feature and the second interaction resource feature using a pre-trained reference recommendation index prediction network to obtain the reference recommendation index of each of the positive sample and the negative sample; and the prior information of the interaction operation of each of the positive sample and the negative sample based on the reference recommendation index of each of the positive sample and the negative sample.

[0009] According to an embodiment of this disclosure, determining the second interactive resource feature of the negative sample based on the target feature distribution data includes: for the target negative sample in the negative sample, determining the fusion weight corresponding to the target preset interactive resource feature based on the target feature distribution data corresponding to the target preset interactive resource feature; and performing weighted fusion of multiple target preset interactive resource features based on the fusion weight to obtain the second interactive resource feature of the target negative sample.

[0010] According to an embodiment of this disclosure, the objective optimization function is constructed based on a cost matrix and a transfer matrix to be optimized. The cost matrix is ​​determined by the following operations: obtaining hidden features of sample resources, which are obtained by feature extraction of the first basic resource features and the second basic resource features; predicting interaction resource features based on the hidden features of sample resources to obtain the cost matrix, wherein the matrix elements of the cost matrix represent the cost feature distribution data of the feature type of the objective preset interaction feature in each sample in the sample set, and the samples include the positive samples or the negative samples.

[0011] According to embodiments of this disclosure, the first interactive resource feature or the second interactive resource feature mentioned above includes at least one of the following: comment operation feature, purchase operation feature, resource browsing duration feature, and resource exposure location feature.

[0012] According to embodiments of this disclosure, the above-mentioned tag recommendation result is determined based on the following operations: fusing the above-mentioned second interactive resource features and sample resource hidden features to obtain sample fusion features, wherein the above-mentioned sample resource hidden features are obtained by feature extraction of the above-mentioned first basic resource features and the above-mentioned second basic resource features; and evaluating the recommendation result based on the above-mentioned sample fusion features to obtain the above-mentioned tag recommendation result.

[0013] Another aspect of this disclosure provides a resource recommendation method, comprising: acquiring target basic resource features related to a target object; processing the target basic resource features using a resource recommendation model trained according to the above training method to obtain a target recommendation result; and pushing the target resource corresponding to the target recommendation result to the target object.

[0014] Another aspect of this disclosure provides a training apparatus for a resource recommendation model, comprising: a first acquisition module, configured to acquire a sample set related to a sample object, wherein positive samples in the sample set include a first basic resource feature and a first interactive resource feature, the first basic resource feature being related to the resource features of the positive sample resource in a historical period, the historical period being the period before the positive sample resource is interactively operated on by the sample object, the first interactive resource feature being determined based on the interaction operation performed by the sample object on the positive sample resource, and negative samples in the sample set including a second basic resource feature of the negative sample resource; a sample recommendation result acquisition module, configured to process the first basic resource feature and the second basic resource feature using an initial resource recommendation model to obtain a sample recommendation result; a second interactive resource feature acquisition module, configured to evaluate the interactive resource features of the negative sample based on the first interactive resource feature to obtain the second interactive resource feature of the negative sample; and a training module, configured to train the initial resource recommendation model based on a knowledge transfer mechanism, according to the sample recommendation result and the tag recommendation result determined based on the second interactive resource feature, to obtain a trained resource recommendation model, wherein the resource recommendation model is used to process the basic resource features related to a target object to obtain a target recommendation result.

[0015] Another aspect of this disclosure provides a resource push device, comprising: a second acquisition module for acquiring target basic resource features related to a target object; a target recommendation result acquisition module for processing the target basic resource features using a resource recommendation model trained according to the above training method to obtain a target recommendation result; and a push module for pushing target resources corresponding to the target recommendation result to the target object.

[0016] Another aspect of this disclosure provides an electronic device, including: one or more processors; and a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to perform the method described above.

[0017] Another aspect of this disclosure provides a computer-readable storage medium storing computer-executable instructions, which, when executed, are used to implement the method described above.

[0018] Another aspect of this disclosure provides a computer program product including computer-executable instructions that, when executed, are used to implement the method described above.

[0019] According to embodiments of this disclosure, by acquiring a sample set containing positive and negative samples, and processing the first and second basic resource features using an initial resource recommendation model, sample recommendation results related to the sample object are obtained. This enables the initial resource recommendation model to initially predict recommendation results based on resource features of the sample object before performing an interaction operation on the sample resource. By using the first interaction resource feature to evaluate the interaction resource features of negative samples, it is possible to predict the second interaction resource feature for negative resources for which the sample object did not perform an interaction operation in historical periods. This allows for the estimation of interaction resource features that might result after the sample object performs an interaction operation on the negative resource. Consequently, the tag recommendation result determined based on the second interaction resource feature can predict the tag recommendation result corresponding to each sample in the sample set, provided that both positive and negative samples contain interaction operation resource features. This makes the tag recommendation result more closely match the actual interaction operation performed on the positive and negative resources. This knowledge transfer mechanism, which trains the initial resource recommendation model based on sample and tag recommendation results, allows for a relatively accurate characterization of the feature types generated after negative sample resources are interacted with by sample objects, even when negative samples have not undergone any interaction. This enables the resource recommendation model to learn the resource features of both positive and negative samples more fully, avoiding the problem of reduced prediction accuracy caused by negative knowledge transfer during the initial resource recommendation model training process. This improves the accuracy and robustness of the target recommendation results generated by the resource recommendation model, thereby enhancing the convenience and accuracy of resource acquisition for target objects and improving the user experience. Attached Figure Description

[0020] The above and other objects, features and advantages of this disclosure will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0021] Figure 1 This illustration schematically shows an exemplary system architecture of a training method and apparatus for applying a resource recommendation model according to embodiments of the present disclosure;

[0022] Figure 2 A flowchart illustrating a method for training a resource recommendation model according to an embodiment of the present disclosure is shown schematically.

[0023] Figure 3 The schematic diagram illustrates the principle of a training method for a resource recommendation model provided according to an embodiment of the present disclosure;

[0024] Figure 4 The schematic diagram illustrates the principle of a training method for an interactive resource feature evaluation layer according to an embodiment of the present disclosure;

[0025] Figure 5A flowchart illustrating a resource push method according to an embodiment of the present disclosure is shown schematically.

[0026] Figure 6 A block diagram of a training apparatus for a resource recommendation model according to an embodiment of the present disclosure is shown schematically.

[0027] Figure 7 A block diagram schematically illustrates a resource push device according to embodiments of the present disclosure; and

[0028] Figure 8 A block diagram of an electronic device suitable for implementing a training method for a resource recommendation model and a resource push method according to embodiments of the present disclosure is shown schematically. Detailed Implementation

[0029] The embodiments of the present disclosure will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the disclosure. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the present disclosure for ease of explanation. However, it will be apparent that one or more embodiments may be practiced without these specific details. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concepts of the present disclosure.

[0030] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit this disclosure. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0031] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0032] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0033] In the embodiments disclosed herein, the collection, updating, analysis, processing, use, transmission, provision, disclosure, and storage of data (e.g., including but not limited to user personal information) comply with relevant laws and regulations, are used for legitimate purposes, and do not violate public order and good morals. In particular, necessary measures have been taken to prevent unauthorized access to user personal information data and to safeguard user personal information security, network security, and national security.

[0034] In the embodiments disclosed herein, user authorization or consent is obtained before acquiring or collecting user personal information.

[0035] The inventors discovered that resource recommendation models used to predict resource recommendation results such as click-through rates and conversion rates have low prediction accuracy, making it difficult to accurately recommend resources that meet user needs. This may be due to imbalanced sample data and negative transfer during the training of the resource recommendation model. Low training accuracy of the resource recommendation model may lead to longer training times and higher computational costs on the computing devices performing the training tasks.

[0036] This disclosure provides a method for training a resource recommendation model, a resource push method, an apparatus, a medium, an electronic device, and a program product. The method includes: acquiring a sample set related to a sample object, wherein positive samples in the sample set include a first basic resource feature and a first interactive resource feature, the first basic resource feature being related to the resource features of the positive sample resource in a historical period, the historical period being the period before the positive sample resource was interactively operated on by the sample object, the first interactive resource feature being determined based on the sample object performing an interactive operation on the positive sample resource, and negative samples in the sample set including a second basic resource feature of the negative sample resource; processing the first basic resource feature and the second basic resource feature using an initial resource recommendation model to obtain a sample recommendation result; evaluating the interactive resource feature of the negative sample based on the first interactive resource feature to obtain the second interactive resource feature of the negative sample; and training the initial resource recommendation model based on the sample recommendation result and the tag recommendation result determined based on the second interactive resource feature using a knowledge transfer mechanism to obtain a trained resource recommendation model, the resource recommendation model being used to process the basic resource features related to the target object to obtain a target recommendation result.

[0037] According to embodiments of this disclosure, by acquiring a sample set containing positive and negative samples, and processing the first and second basic resource features using an initial resource recommendation model, sample recommendation results related to the sample object are obtained. This enables the initial resource recommendation model to initially predict recommendation results based on resource features of the sample object before performing an interaction operation on the sample resource. By using the first interaction resource feature to evaluate the interaction resource features of negative samples, it is possible to predict the second interaction resource feature for negative resources for which the sample object did not perform an interaction operation in historical periods. This allows for the estimation of interaction resource features that might result after the sample object performs an interaction operation on the negative resource. Consequently, the tag recommendation result determined based on the second interaction resource feature can predict the tag recommendation result corresponding to each sample in the sample set, provided that both positive and negative samples contain interaction operation resource features. This makes the tag recommendation result more closely match the actual interaction operation performed on the positive and negative resources. This knowledge transfer mechanism, which trains the initial resource recommendation model based on sample and tag recommendation results, allows for a relatively accurate characterization of the feature types generated after negative sample resources are interacted with by sample objects, even when negative samples have not undergone any interaction. This enables the resource recommendation model to learn the resource features of both positive and negative samples more fully, avoiding the problem of reduced prediction accuracy caused by negative knowledge transfer during the initial resource recommendation model training process. This improves the accuracy and robustness of the target recommendation results generated by the resource recommendation model, thereby enhancing the convenience and accuracy of resource acquisition for target objects and improving the user experience.

[0038] Figure 1 This illustration schematically depicts an exemplary system architecture for training methods and apparatus for resource recommendation models applicable according to embodiments of this disclosure. It should be noted that... Figure 1 The examples shown are merely examples of system architectures that can be applied to the embodiments of this disclosure, in order to help those skilled in the art understand the technical content of this disclosure, but do not mean that the embodiments of this disclosure cannot be used in other devices, systems, environments or scenarios.

[0039] like Figure 1 As shown, the system architecture 100 according to this embodiment may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 serves as a medium for providing communication links between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. The network 104 may include various connection types, such as wired and / or wireless communication links, etc.

[0040] Users can use the first terminal device 101, the second terminal device 102, and the third terminal device 103 to interact with the server 105 via the network 104 to receive or send messages, etc. Various communication client applications can be installed on the first terminal device 101, the second terminal device 102, and the third terminal device 103, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, and / or social media platform software, etc. (for example only).

[0041] The first terminal device 101, the second terminal device 102, and the third terminal device 103 can be various electronic devices with displays and support web browsing, including but not limited to smartphones, tablets, laptops, and desktop computers.

[0042] Server 105 can be a server that provides various services, such as a backend management server that supports websites browsed by users using the first terminal device 101, the second terminal device 102, and the third terminal device 103 (this is just an example). The backend management server can analyze and process data such as received user requests, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices.

[0043] It should be noted that the training method for the resource recommendation model provided in this embodiment can generally be executed by server 105. Correspondingly, the training device for the resource recommendation model provided in this embodiment can generally be located in server 105. The training method for the resource recommendation model provided in this embodiment can also be executed by a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105. Correspondingly, the training device for the resource recommendation model provided in this embodiment can also be located in a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105.

[0044] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0045] Figure 2 A flowchart illustrating a training method for a resource recommendation model according to an embodiment of the present disclosure is shown.

[0046] like Figure 2 As shown, the training method for this resource recommendation model includes operations S210~S240.

[0047] In operation S210, obtain the sample set related to the sample object.

[0048] According to embodiments of this disclosure, the positive samples in the sample set include a first basic resource feature and a first interactive resource feature. The first basic resource feature is related to the resource feature of the positive sample resource in a historical period. The historical period is the period before the positive sample resource is interactively operated on by the sample object. The first interactive resource feature is determined based on the interactive operation performed by the sample object on the positive sample resource. The negative samples in the sample set include the second basic resource feature of the negative sample resource.

[0049] In operation S220, the first and second basic resource features are processed using the initial resource recommendation model to obtain the sample recommendation results.

[0050] In operation S230, the interaction resource features of the negative sample are evaluated based on the first interaction resource features to obtain the second interaction resource features of the negative sample.

[0051] In operation S230, based on the knowledge transfer mechanism, an initial resource recommendation model is trained according to the sample recommendation results and the label recommendation results determined based on the second interactive resource features, thus obtaining the trained resource recommendation model.

[0052] According to embodiments of this disclosure, negative sample resources can be resources for which the sample object did not perform any interactive operations during a historical period. Since negative samples have not been interacted with by the sample object, the obtained sample set does not include the interactive resource features related to the negative sample resources and the sample object.

[0053] It should be noted that the sample set may include multiple sample resources, which can be positive or negative sample resources. The first and second basic resource features can be the basic resource features of the sample resources in the sample set. The basic resource features can be pre-click features related to the sample resources. Pre-click features can represent the feature data related to the sample resource or sample object before the sample object performs an interaction operation on the positive or negative sample resource.

[0054] According to embodiments of this disclosure, sample resources may include any type of resources such as advertising resources, news resources, video resources, and product detail page resources. Embodiments of this disclosure do not limit the specific type of sample resources.

[0055] In one example, basic resource features may include resource text information, resource image information, resource classification information, etc., of positive or negative sample resources. Basic resource features may also include user object attribute information, user object historical interaction information, etc., related to positive or negative sample resources.

[0056] It should be noted that a user object can be a user who has performed interactive operations on either positive or negative sample resources during a historical period. The user object and the sample object can be different users. For example, a user object may have performed interactive operations on a negative sample resource, but a sample object may not have performed any interactive operations on a negative sample resource.

[0057] According to embodiments of this disclosure, the first interactive resource feature may include resource features generated based on interactive operations of sample objects on positive sample resources during historical time periods. The first interactive resource feature may include resource-related feature data such as the dwell time and page exposure position of the positive sample resource. Alternatively, the first interactive resource feature may also include subsequent interactive behavior feature data from historical time periods, such as comment data and purchase behavior data of sample objects on the positive sample resource.

[0058] According to embodiments of this disclosure, the interactive operation performed by the sample object on the positive sample resource during the historical operation period can be any type of interactive operation such as a click operation.

[0059] According to embodiments of this disclosure, the initial resource recommendation model can be constructed based on any type of deep learning algorithm. For example, it can be constructed based on a convolutional neural network algorithm, but it is not limited to this. It can also be constructed based on other types of network algorithms such as long short-term memory networks. The embodiments of this disclosure do not limit the specific algorithm type of the initial resource recommendation model.

[0060] According to embodiments of this disclosure, the sample recommendation results may include recommendation evaluation metrics for recommending sample resources within a sample set. For example, the sample recommendation results may include at least one of sample click-through rate and sample conversion rate.

[0061] According to embodiments of this disclosure, evaluating the interaction resource features of negative samples based on first interaction resource features to obtain second interaction resource features of negative samples may include using a pre-trained deep learning model to process the first interaction resource features and predict the second interaction resource features of negative samples.

[0062] In one example, a pre-trained deep learning model can be used to process the first interaction resource features and the hidden features of the sample resources to obtain the second interaction resource features of the negative sample. The hidden features of the sample resources can be obtained by feature extraction from the first and second basic resource features. The pre-trained deep learning model can be based on an attention network algorithm to learn the similarity between the first interaction resource features and the hidden features of the sample resources based on the attention mechanism, thereby predicting the second interaction resource features of the negative sample and improving the matching degree between the second interaction resource features and the interaction behavior of the sample object.

[0063] According to embodiments of this disclosure, the label recommendation result can be obtained by evaluating the recommendation results of the second interactive resource features. For example, the second interactive resource features can be processed based on a multilayer perceptron to obtain the label recommendation result of the negative sample.

[0064] In one example, the recommendation results can be evaluated based on the second and first interactive resource features to obtain the label recommendation results for each positive and negative sample. Based on a knowledge transfer mechanism, the initial resource recommendation model is trained using the label recommendation results for each positive and negative sample, as well as the sample recommendation results for each positive and negative sample.

[0065] According to embodiments of this disclosure, training an initial resource recommendation model based on a knowledge transfer mechanism, using sample recommendation results and tag recommendation results determined based on second interactive resource features, may include determining a loss value based on the difference between the sample recommendation results and the tag recommendation results, adjusting the model parameters of the initial resource recommendation model based on the loss value, until the training rounds reach a preset round threshold or the loss value converges, thereby obtaining the trained resource recommendation model.

[0066] According to embodiments of this disclosure, a resource recommendation model is used to process basic resource features related to a target object to obtain a target recommendation result. At least one target resource related to the basic resource features can be pushed to the target object based on the target recommendation result.

[0067] According to embodiments of this disclosure, by evaluating the interaction resource features of negative samples based on the first interaction resource features, the evaluated second interaction resource features can more accurately characterize the resource-related features and object interaction behavior features that the sample object may generate after performing interaction operations on the negative sample resources. This allows for the evaluation of recommendation results for sample resources in the sample set based on the second interaction resource features of negative samples. This enables the label recommendation results to be evaluated under the condition of sample imbalance where negative samples lack interaction resource features, making the label recommendation results more accurately characterize the matching degree between negative and positive sample resources and the resource needs of the sample object. By using a knowledge transfer mechanism to train the initial resource recommendation model using the differences between the sample recommendation results and the label recommendation results, the label recommendation results can be used as pseudo-labels to perform knowledge transfer on the initial resource recommendation model. This at least overcomes the situation where the imbalance of sample data in the sample set causes negative knowledge transfer interference in the initial resource recommendation model, and avoids the resource recommendation model from over-relying on positive samples during training, thereby improving the prediction accuracy of the resource recommendation model for the target object.

[0068] According to embodiments of this disclosure, the sample recommendation result includes at least one of the following: sample click-through rate (CTR) and sample conversion rate. The CTR represents the probability that a sample object performs an interactive operation on a sample resource. The sample object can perform subsequent operations such as commenting, searching, favorites, liking, adding to cart, and purchasing by interacting with the sample resource. The conversion rate represents the probability that, among the subsequent interactive operations performed by the sample object after interacting with the sample resource, any type of preset target operation, such as purchasing or commenting, is involved.

[0069] It should be noted that the interactive operations involved in the embodiments of this disclosure can be understood as click operations performed by an object on a sample resource or a target resource. The object can view the details page of the sample resource or target resource by performing the interactive operation. The subsequent interactive operations or follow-up operations designed in the embodiments of this disclosure can be understood as operations performed by the object after performing the interactive operation for viewing the sample resource or target resource. The follow-up interactive operations can be related to the characteristics of the interactive resource. For example, the follow-up operation performed by the sample object on a positive sample resource can be a comment operation, and the first interactive resource characteristic of the positive sample can include the comment content entered in the comment operation.

[0070] In one example, the initial resource recommendation model can be a multi-objective task recommendation model. Sample recommendation results can include sample click-through rate (CTR) and sample conversion rate, while tag recommendation results can include tag CTR and tag conversion rate. A first loss value is determined based on the difference between sample CTR and tag CTR; a second loss value is determined based on the difference between sample conversion rate and tag conversion rate; and a target loss value is determined based on the sum of the first and second loss values. This target loss value is then used to adjust the model parameters of the initial resource recommendation model until the target loss value converges or reaches a preset number of training epochs, resulting in the trained resource recommendation model.

[0071] According to embodiments of this disclosure, the first interactive resource feature or the second interactive resource feature includes at least one of the following: comment operation feature, purchase operation feature, resource browsing duration feature, and resource exposure location feature.

[0072] According to embodiments of this disclosure, comment operation features can represent the text content of a sample object's comment on a sample resource, comment preference attributes (such as likes and other comment-related operations), etc.

[0073] According to embodiments of this disclosure, purchase operation features can represent relevant feature data of a sample object performing a purchase operation on goods or services of a sample resource, such as the quantity and category of the purchased goods or services.

[0074] According to embodiments of this disclosure, the resource browsing duration feature can represent the browsing duration of any jump from the sample resource's details page, resource introduction page, etc., to a page related to the sample resource after the sample object performs an interactive operation on the sample resource.

[0075] According to embodiments of this disclosure, resource exposure location features can represent feature data characterizing the exposure location of a sample resource after the sample object performs an interactive operation on the sample resource.

[0076] According to embodiments of this disclosure, sample resources may include positive sample resources and negative sample resources. A first interaction resource feature may be a resource feature related to the positive sample resource generated after a sample object performs an interaction operation on the positive sample resource. A second interaction resource feature may represent a resource feature related to the negative sample resource generated after a sample object may perform an interaction operation on the negative sample resource. The second interaction resource feature can be understood as a resource feature related to potential subsequent interaction operations of the sample object on the negative sample resource. For example, potential comment operation features, potential purchase operation features, potential browsing duration features, and potential exposure location features of the negative sample resource.

[0077] It should be noted that the information acquisition involved in the embodiments of this disclosure, including but not limited to the first basic resource characteristics, the second basic resource characteristics, and the first interactive resource characteristics, is all obtained after obtaining authorization from the relevant users or organizations. Before acquiring information, the relevant users or organizations are informed in advance of the purpose of the information, and after acquiring the information, necessary encryption or de-identification measures are adopted to protect the information from being disclosed. The execution process of the methods provided in the embodiments of this disclosure complies with the provisions of relevant laws and regulations and does not violate public order and good morals.

[0078] According to embodiments of this disclosure, the tag recommendation result is determined based on the following operations: fusing the second interactive resource features and the hidden features of the sample resources to obtain sample fusion features; and evaluating the recommendation result based on the sample fusion features to obtain the tag recommendation result.

[0079] According to embodiments of this disclosure, the sample resource hiding features are obtained by feature extraction from a first basic resource feature and a second basic resource feature. The sample resource hiding features may include multiple features, and each of the multiple sample resource hiding features may correspond to a multiple sample.

[0080] In one example, the first basic resource features of each of the N positive samples can be respectively associated with the N first sample resource hidden features, and the second basic resource features of each of the M negative samples can be respectively associated with the M second sample resource hidden features.

[0081] According to embodiments of this disclosure, feature fusion of the second interactive resource features and the hidden features of the sample resources may include processing the second interactive resource features and the hidden features of the sample resources based on a fusion function to obtain sample fusion features. The fusion function may include any type of function such as a multiplication function, a concatenation function, or an accumulation function, and embodiments of this disclosure do not limit the specific type of fusion function.

[0082] In one example, feature fusion of the second interactive resource features and the hidden sample resource features may further include processing the second interactive resource features and the hidden sample resource features based on a neural network algorithm to obtain fused sample features. For example, an attention network algorithm may be used to process the second interactive resource features and the hidden sample resource features to obtain fused sample features. However, this is not limited to this; other types of neural network algorithms may also be used to process the second interactive resource features and the hidden sample resource features to obtain fused sample features. The embodiments of this disclosure do not limit the specific algorithm type for fusing the second interactive resource features and the hidden sample resource features.

[0083] According to embodiments of this disclosure, feature fusion of the second interactive resource features and the hidden features of the sample resources may further include feature fusion of the first interactive resource features, the second interactive resource features, and the hidden features of the sample resources to obtain sample fusion features. These sample fusion features can include resource features of the sample object before it performs an interactive operation with positive and negative sample resources, resource features generated after the sample object performs an interactive operation with positive sample resources, and potential resource features generated after the sample object performs an interactive operation with negative sample resources. This at least partially solves the problem of imbalanced or missing training data caused by the lack of post-click features of negative samples, improving the matching degree between subsequent tag recommendation results and the resource needs of the sample object, and improving the accuracy of tag recommendation results. Therefore, relatively accurate tag recommendation results can be used as pseudo-tags to achieve knowledge distillation of the initial resource recommendation model based on a knowledge transfer mechanism, thereby improving the prediction accuracy of the recommendation results of the trained resource recommendation model.

[0084] Figure 3 The schematic diagram illustrates the principle of a training method for a resource recommendation model provided according to an embodiment of the present disclosure.

[0085] like Figure 3As shown, the initial recommendation model 310, serving as a student model, can include a backbone network 311 and a first recommendation result output layer 312. The backbone network 311 can be constructed based on convolutional neural network algorithms or attention network algorithms. For example, it can be constructed based on neural network algorithms such as Feature Pyramid Networks (FPN) or Transformer networks. The backbone network 311 can perform feature extraction and feature fusion on the basic resource features F311 of multiple sample resources in the sample set, obtaining multiple hidden features F312 corresponding to each sample resource. The first recommendation result output layer 312 can be constructed based on a multilayer perceptron algorithm. The first recommendation result output layer 312 processes the multiple hidden features F312 of the sample resources to obtain the sample recommendation result 301.

[0086] The teacher model 320 may include an interactive resource feature evaluation layer 321 and a second recommendation result output layer 322. The interactive resource feature evaluation layer 321 can be constructed based on any type of algorithm; for example, it can be constructed based on an attention network algorithm. Multiple first interactive resource features F3211, multiple prior information on interactive operations 301, and multiple hidden features of sample resources F312 are input into the resource feature evaluation layer 321 to obtain the second interactive resource features of negative samples. The multiple prior information on interactive operations 301 can correspond to multiple positive samples, and the prior information on interactive operations can represent the similarity between positive and negative sample resources in the sample set.

[0087] The second interactive resource features of the negative samples and multiple first interactive resource features F3211 can be used as multiple interactive resource features F322 of the sample set. Multiple interactive resource features F322 (e.g., multiple first interactive resource features F3211 and the second interactive resource features of the negative samples) are concatenated with multiple sample resource hidden features F312 to obtain sample fusion features. The sample fusion features are input into the second recommendation result output layer 322 to obtain the label recommendation result 302. The sample recommendation result 301 and the label recommendation result 302 are processed using a loss function to obtain a loss value. The model parameters of the initial resource recommendation model 310 are adjusted based on the loss value until the loss value converges, resulting in the trained resource recommendation model.

[0088] According to embodiments of this disclosure, evaluating the interaction resource features of negative samples based on first interaction resource features to obtain second interaction resource features of negative samples includes the following operations.

[0089] For the target preset interactive resource features, a causal inference operation is performed based on the first interactive resource features of positive samples and the preset prior information of interactive operations to obtain the feature distribution data of the target preset interactive features in the sample set; the target distribution vector is determined based on the feature distribution data of multiple target preset interactive features; according to the optimal transmission mechanism, the target optimization function is solved based on the optimization conditions determined by the target distribution vector to obtain the optimized target transmission matrix; the second interactive resource features of negative samples are determined based on the target feature distribution data.

[0090] According to embodiments of this disclosure, the target preset interactive resource feature is a preset interactive resource feature in a preset interactive resource feature library. The feature type of the first resource interactive feature or the feature type of the second resource interactive feature may correspond to the preset feature type of at least one preset interactive resource feature.

[0091] For example, multiple preset interactive resource features in the preset interactive resource feature library may each have feature types A, B, and C. The feature types of the first interactive resource feature can be type A, type B, and type C, and the feature types of the second interactive resource feature can be type A and type B. The feature types of each preset interactive resource feature may include, for example, comment operation type, purchase operation type, resource browsing duration type, and resource exposure location type.

[0092] According to embodiments of this disclosure, prior information about interactive operations characterizes the similarity between a negative sample and at least one positive sample. Feature distribution data can represent the probability or value of a target preset interactive feature with the same feature type among multiple preset interactive resource features in a preset interactive resource feature library within a sample set. Multiple feature distribution data in the target distribution vector can represent the probability or value of each of the multiple target preset interactive features of different feature types within the overall sample set.

[0093] According to embodiments of this disclosure, performing causal inference based on the first interactive resource features of positive samples and preset prior information of interactive operations may include using the first basic resource features of multiple positive samples and the second basic resource features of multiple negative samples as basic resource features X, using the preset prior information of interactive operations as the causal factor Y for the sample object to potentially perform interactive operations on the negative sample resources, and using the potential second interactive resource features of the negative samples and the first interactive resource features of the positive samples as interactive resource features Z. A causal logical relationship exists between the causal factor Y and the interactive resource features Z. Causal inference is performed on the basic resource features X, the causal factor Y, and the interactive resource features Z based on the do operator (or do-calculus) to obtain the feature distribution data of the result information Z in the sample set.

[0094] According to embodiments of this disclosure, the matrix elements of the target transmission matrix include target feature distribution data of negative samples, the target feature distribution data corresponds to target preset interactive resource features, and the target feature distribution data characterizes the target distribution probability of negative samples in the target preset interactive resource features.

[0095] According to embodiments of this disclosure, the optimization conditions may include a target distribution vector and a preset source distribution vector. The source distribution vector may include a uniform distribution of feature types of multiple target preset interactive resource features in the sample set. The target optimization function may be constructed based on the transfer matrix to be optimized. The matrix elements of the transfer matrix to be optimized may be defined as the distribution values ​​or distribution probabilities of the feature types of the target preset interactive resource features in the sample set. By solving the target optimization function based on the optimal transfer mechanism and optimization conditions, the problem of predicting the second interactive feature of negative samples can be transformed into an optimization problem for the transfer matrix to be optimized. Thus, the matrix elements of the obtained target transfer matrix can represent the target distribution probabilities of the negative samples in multiple target preset interactive resource features. Furthermore, based on one or more target distribution probabilities of the negative samples, the feature types of the potential target preset interactive resource features of the negative samples can be determined, thereby determining the potential second interactive resource features of the negative samples.

[0096] According to embodiments of this disclosure, the objective optimization function is constructed based on the cost matrix and the transfer matrix to be optimized.

[0097] In one example, the objective function and the optimization conditions can be expressed based on formulas (1) and (2), respectively.

[0098] (1);

[0099] (2)

[0100] in, Let π represent the cost matrix for optimal transmission, a represent the source distribution vector (a discrete uniform distribution vector), b represent the target distribution vector, and δ represent the optimization condition. h(π) is the entropy regularization term, and ε represents the entropy regularization coefficient. The objective optimization function after adding the entropy regularization term can be solved quickly using the Sinkhorn iterative algorithm. The obtained π... ε This represents the probability distribution of positive or negative samples corresponding to the target preset interactive resource features, which can be understood as π. ε This represents the distribution data of the target features.

[0101] In one example, the second interactive resource feature of a negative sample can be determined based on the following formula (3).

[0102] (3);

[0103] Where E represents the embedding vector table of the target preset interactive resource features, E stores the embedding vector corresponding to the feature category of each target preset interactive resource feature, π ε’ h represents the distribution data of the target features compared to the negative samples. t This indicates the characteristics of the second interactive resource.

[0104] According to embodiments of this disclosure, determining the second interactive resource feature of a negative sample based on target feature distribution data may further include: for a target negative sample in the negative sample, determining a fusion weight corresponding to the target preset interactive resource feature based on target feature distribution data corresponding to the target preset interactive resource feature; and performing weighted fusion of multiple target preset interactive resource features based on the fusion weight to obtain the second interactive resource feature of the target negative sample.

[0105] According to embodiments of this disclosure, when the matrix elements of the target transmission matrix represent target negative samples and multiple target preset interactive resource features corresponding to target feature distribution data, the multiple target feature distribution data can be normalized to obtain multiple fusion weights. Thus, the potential logical relationship between the negative sample resource and each target preset interactive resource feature can be represented based on the fusion weights. Weighted fusion of multiple target preset interactive resource features based on the fusion weights allows the obtained second interactive resource feature to more accurately represent the potential post-click features of the negative sample resource, thereby improving the accuracy of the tag recommendation results.

[0106] According to embodiments of this disclosure, the cost matrix is ​​determined based on the following operations.

[0107] Obtain hidden features of sample resources; predict interactive resource features based on the hidden features of sample resources to obtain the cost matrix.

[0108] The elements of the cost matrix represent the cost feature distribution data of the target preset interaction feature for each sample in the sample set, including positive and negative samples. The cost feature distribution data can represent the probability distribution or distribution value of the target preset interaction feature's feature type corresponding to positive or negative samples.

[0109] In one example, the cost matrix can be determined based on the following formula (4).

[0110] (4)

[0111] Among them, h s Let f() represent the hidden features of the sample resources, and let f() represent the cost matrix prediction sublayer, which can be constructed based on any type of neural network algorithm. P represents the cost matrix element features corresponding to the elements of the cost matrix. Q= This represents the cost matrix.

[0112] In one example, the cost matrix prediction sub-layer can be semi-supervised based on the following formula (5), and the hidden features of sample resources can be processed based on the trained cost matrix prediction sub-layer to obtain the cost matrix.

[0113] (5)

[0114] Where N represents the number of positive samples in a batch of sample sets. This represents the true post-click feature (or first interactive resource feature) of the m-th positive sample in a batch of samples. This represents the loss function. Using the actual click-after features of positive samples to generate auxiliary semi-supervised signals to train the cost matrix prediction sublayer allows it to more accurately predict the distribution of the target preset interactive resource features in the sample set.

[0115] In one example, the cost matrix, target distribution vector, and target transfer matrix can be represented based on Tables 1 to 3 below.

[0116] Table 1

[0117]

[0118] Table 2

[0119]

[0120] Table 3

[0121]

[0122] Referring to Tables 1 to 3, the sample set includes samples 1 to 4, and the feature types of the target preset interactive resource features can include feature type 1 to feature type 4. In the case that sample 1 is a negative sample, the target feature distribution data of negative sample 1 can include target feature distribution data 11, target feature distribution data 12, target feature distribution data 13, and target feature distribution data 14.

[0123] In one example, causal inference can be performed based on the calculation process shown in formulas (6) and (7) below.

[0124] (6);

[0125] (7);

[0126] Where do() represents the do operator, This represents the feature distribution data of multiple negative samples in the target distribution vector. X, Y, and Z represent the basic resource features, interoperability prior information, and interaction resource features, respectively. This can be represented based on the feature characteristics of the cost matrix elements obtained from the cost matrix prediction sublayer. However, since subsequent interactive operations (or post-click actions) only occur after the interactive operation, the cost matrix prediction sublayer lacks the ability to capture the causal relationship between the second basic resource feature and the second interactive resource feature of negative samples in the sample set. Therefore, the distribution probability of the interactive resource feature (target preset interactive resource feature) in the entire sample set can be estimated based on the above formulas (6) and (7) to solve the counterfactual scenario. For negative samples that have not undergone interactive operations, a counterfactual problem can be simulated based on the above formulas (6) and (7): that is, if the sample object clicks on the negative sample resource, how will the sample object change after clicking (performing an interactive operation)? This can be understood as the feature distribution data of the target preset interaction features in the sample set.

[0127] In one example, multiple positive samples correspond to multiple prior information for interaction operations. The prior information for interaction operations is determined based on the following operations.

[0128] The network is predicted to process the first and second interaction resource features using pre-trained reference recommendation metrics to obtain the reference recommendation metrics for positive and negative samples respectively; and the prior information of the interaction operations of positive and negative samples is obtained based on the reference recommendation metrics for positive and negative samples respectively.

[0129] According to embodiments of this disclosure, the pre-trained reference recommendation metric prediction network can be constructed based on any type of deep learning algorithm, and embodiments of this disclosure do not limit the specific algorithm type of the reference recommendation metric prediction network.

[0130] According to embodiments of this disclosure, the reference recommendation metric can include any type of recommendation metric such as click-through rate, conversion rate, or click-to-conversion rate. The reference recommendation metric can be used to predict the probability that positive and negative samples will perform interactive operations.

[0131] According to embodiments of this disclosure, based on the respective reference recommendation indices of positive and negative samples, and the respective prior information of interaction operations of positive and negative samples, it may include using the reference recommendation indices as prior information of interaction operations, or it may also include normalizing the reference recommendation indices of each sample in the sample set to obtain the respective prior information of interaction operations of positive samples.

[0132] According to embodiments of this disclosure, performing a causal inference operation based on the first interactive resource features of a positive sample and preset prior information of interactive operations may include: determining intermediate distribution data of at least one target preset interactive feature in the positive sample based on the ratio between the first interactive resource features of the positive sample and the prior information of interactive operations associated with the positive sample; and fusing the intermediate distribution data corresponding to each of the multiple positive samples in the sample set to obtain feature distribution data of at least one target preset interactive feature in the sample set.

[0133] In one example, the intermediate distribution data of the positive sample can be represented based on the following formula (8).

[0134] (8);

[0135] in, Z represents the prior information of the interaction operation of the i-th positive sample in the sample set. i Let A be the first interactive resource feature of the i-th positive sample. Let A represent the intermediate distribution data of the i-th positive sample.

[0136] The sample set can be a batch of sample data, and the sample set can include N samples. The feature distribution data of the target preset interaction features in the sample set can be represented by formula (9).

[0137] (9);

[0138] in, This represents the feature distribution data of the target preset interaction features in the sample set.

[0139] Figure 4 The schematic diagram illustrates the principle of a training method for an interactive resource feature evaluation layer provided according to an embodiment of the present disclosure.

[0140] like Figure 4As shown, the interaction resource feature evaluation layer 400 includes a cost matrix prediction sublayer 410, a causal inferrer 420, and an optimal transport prediction sublayer 430. Multiple sample resource hidden features F410 can include the resource hidden features corresponding to the 1st sample, 2nd sample, ..., up to the nth sample in the sample set. The 2nd sample can be a negative sample. Among the multiple interaction resource features F420 corresponding to the 1st sample, 2nd sample, ..., up to the nth sample in the sample set, the interaction resource feature corresponding to the 2nd sample can be a preset initialization feature. The other interaction resource features in the multiple interaction resource features F420 besides the preset initialization feature corresponding to the 2nd sample can be the first interaction resource features corresponding to each of the multiple positive samples. Multiple interaction operation prior information T430 can each correspond to the 1st sample, 2nd sample, ..., up to the nth sample in the sample set. Inputting the multiple sample resource hidden features F410 into the cost matrix prediction sublayer 410 yields the cost matrix M401. Multiple interactive resource features F420 and multiple interactive operation prior information T430 are input into the causal inferrer 420, and the target distribution vector M402 is output. The causal inferrer 420 can be constructed based on formula (9). The cost matrix M401, the target distribution vector M402 and the source distribution vector M403 are input into the optimal transmission prediction sublayer 430, and the target transmission matrix M404 is output. The source distribution vector M403 can be a discrete uniform distribution vector representing the uniform distribution of multiple target preset interactive resource features in multiple samples of the sample set. The matrix elements of the target transmission matrix M404 include the target feature distribution data of the second sample as a negative sample in the multiple target preset interactive resource features. Based on the feature type of the target feature distribution data and the embedding vector corresponding to the feature type in the preset interactive resource feature library, the second interactive resource feature F432 of the second sample can be obtained.

[0141] Based on the training method of the resource recommendation model provided in the embodiments of this disclosure, the embodiments of this disclosure also provide a resource push method.

[0142] Figure 5 A flowchart illustrating a resource push method according to an embodiment of the present disclosure is shown schematically.

[0143] like Figure 5 As shown, the resource push method includes operations S510~S530.

[0144] In operation S510, the target basic resource characteristics related to the target object are obtained.

[0145] In operation S520, the resource recommendation model trained using the training method provided in this embodiment of the resource recommendation model processes the target basic resource features to obtain the target recommendation result; and

[0146] When operating S530, target resources corresponding to the target recommendation results are pushed to the target object.

[0147] According to embodiments of this disclosure, the target basic resource characteristics related to the target object may include the target basic resource characteristics of the resources to be recommended to the target object, or may also include interactive behavior characteristics related to the interactive operation behavior of the target object.

[0148] According to embodiments of this disclosure, the target recommendation result may include recommendation metrics such as the target click-through rate and target conversion rate of the resource to be recommended. The target resource may include a target resource determined from the resources to be recommended based on the target recommendation result. For example, the target resource may be a resource to be recommended that corresponds to the predicted highest click-through rate.

[0149] According to embodiments of this disclosure, pushing target resources corresponding to target recommendation results to a target object may include pushing messages related to the target resource to the target object's terminal device, or it may also include displaying the target resource at a specified location on the page the target object is browsing. Embodiments of this disclosure do not limit the specific method of pushing target resources to the target object, as long as the target object can obtain the target resource in a timely manner.

[0150] According to the resource push method provided in this disclosure, since the resource recommendation model trained based on the training method provided in this disclosure is used to process the target basic resource features, it can more accurately predict the matching degree between the resource to be recommended and the resource needs of the target object when the target object has not performed any interactive operations on the resource to be recommended. This avoids the problem of low accuracy of recommendation results caused by imbalanced samples during training. It is also suitable for more accurate cold start delivery of the resource to be recommended, and can more accurately meet the resource acquisition needs of the target object and the resource push needs of the resource provider.

[0151] Figure 6 A block diagram of a training apparatus for a resource recommendation model according to an embodiment of the present disclosure is shown schematically.

[0152] like Figure 6 As shown, the training device 600 for the resource recommendation model includes: a first acquisition module 610, a sample recommendation result acquisition module 620, a second interactive resource feature acquisition module 630, and a training module 640.

[0153] The first acquisition module 610 is used to acquire a sample set related to the sample object. The positive samples in the sample set include a first basic resource feature and a first interactive resource feature. The first basic resource feature is related to the resource feature of the positive sample resource in a historical period. The historical period is the period before the positive sample resource is interactively operated on by the sample object. The first interactive resource feature is determined based on the interactive operation performed by the sample object on the positive sample resource. The negative samples in the sample set include a second basic resource feature of the negative sample resource.

[0154] The sample recommendation result acquisition module 620 is used to process the first basic resource feature and the second basic resource feature using the initial resource recommendation model to obtain the sample recommendation result.

[0155] The second interactive resource feature acquisition module 630 is used to evaluate the interactive resource features of negative samples based on the first interactive resource features, and obtain the second interactive resource features of the negative samples.

[0156] Training module 640 is used to train an initial resource recommendation model based on the knowledge transfer mechanism, the sample recommendation results, and the label recommendation results determined based on the second interactive resource features, to obtain a trained resource recommendation model. The resource recommendation model is used to process the basic resource features related to the target object to obtain the target recommendation result.

[0157] According to embodiments of this disclosure, the second interactive resource feature acquisition module 630 includes: a feature distribution data acquisition unit, a target distribution vector determination unit, a target transmission matrix acquisition unit, and a second interactive resource feature acquisition unit.

[0158] The feature distribution data acquisition unit is used to perform causal inference operations based on the first interaction resource features of positive samples and the preset interaction operation prior information for the target preset interaction resource features, and obtain the feature distribution data of the target preset interaction features in the sample set. The interaction operation prior information represents the similarity between negative samples and at least one positive sample.

[0159] The target distribution vector determination unit is used to determine the target distribution vector based on the feature distribution data of each of the preset interactive features of multiple targets.

[0160] The target transmission matrix acquisition unit is used to perform a solution operation on the target optimization function based on the optimization conditions determined by the target distribution vector according to the optimal transmission mechanism, and obtain the optimized target transmission matrix. The matrix elements of the target transmission matrix include the target feature distribution data of the negative samples, which represent the target distribution probability of the negative samples in the target preset interactive resource features.

[0161] The second interactive resource feature acquisition unit is used to determine the second interactive resource features of negative samples based on the target feature distribution data.

[0162] According to embodiments of this disclosure, multiple positive samples correspond to multiple prior information of interactive operations.

[0163] According to embodiments of this disclosure, the feature distribution data acquisition unit includes: an intermediate distribution data acquisition subunit and a feature distribution data acquisition subunit.

[0164] The intermediate distribution data acquisition subunit is used to determine the intermediate distribution data of at least one target preset interaction feature in the positive sample based on the ratio between the first interaction resource feature of the positive sample and the prior information of the interaction operation associated with the positive sample.

[0165] The feature distribution data acquisition sub-unit is used to fuse the intermediate distribution data corresponding to multiple positive samples in the sample set to obtain the feature distribution data of at least one target preset interaction feature in the sample set.

[0166] According to embodiments of this disclosure, the prior information of the interaction operation is determined based on the following operations: using pre-trained reference recommendation indicators to predict the network processing of the first interaction resource feature and the second interaction resource feature to obtain the reference recommendation indicators for positive samples and negative samples respectively; and based on the reference recommendation indicators for positive samples and negative samples respectively, the prior information of the interaction operation for positive samples and negative samples respectively.

[0167] According to embodiments of this disclosure, the second interactive resource feature acquisition unit includes: a first determining subunit and a second interactive resource feature acquisition subunit.

[0168] The first determining subunit is used to determine the fusion weight corresponding to the target preset interactive resource features based on the target feature distribution data corresponding to the target preset interactive resource features for the target negative sample in the negative sample.

[0169] The second interactive resource feature acquisition subunit is used to perform weighted fusion of multiple target preset interactive resource features based on fusion weights to obtain the second interactive resource features of the target negative sample.

[0170] According to embodiments of this disclosure, the objective optimization function is constructed based on a cost matrix and a transfer matrix to be optimized. The cost matrix is ​​determined by the following operations: obtaining hidden features of sample resources, which are obtained by feature extraction of a first basic resource feature and a second basic resource feature; predicting interaction resource features based on the hidden features of sample resources to obtain the cost matrix, where the matrix elements of the cost matrix represent the cost feature distribution data of the feature type of the target preset interaction feature in each sample in the sample set, and the samples include positive samples or negative samples.

[0171] According to embodiments of this disclosure, the first interactive resource feature or the second interactive resource feature includes at least one of the following: comment operation feature, purchase operation feature, resource browsing duration feature, and resource exposure location feature.

[0172] According to embodiments of this disclosure, the tag recommendation result is determined based on the following operations: fusing the second interactive resource features and the sample resource hidden features to obtain sample fusion features, wherein the sample resource hidden features are obtained by feature extraction of the first basic resource features and the second basic resource features; and evaluating the recommendation result based on the sample fusion features to obtain the tag recommendation result.

[0173] Figure 7 A block diagram of a resource push device according to an embodiment of the present disclosure is shown schematically.

[0174] like Figure 7 As shown, the resource push device 700 includes: a second acquisition module 710, a target recommendation result acquisition module 720, and a push module 730.

[0175] The second acquisition module 710 is used to acquire the target basic resource characteristics related to the target object;

[0176] The target recommendation result acquisition module 720 is used to process the target basic resource features using the resource recommendation model trained by the training method of the resource recommendation model provided in the embodiments of this disclosure, and obtain the target recommendation result.

[0177] The push module 730 is used to push target resources corresponding to the target recommendation results to the target object.

[0178] Any one or more of the modules, units, and subunits according to the embodiments of this disclosure, or at least part of the functions of any one or more of them, can be implemented in one module. Any one or more of the modules, units, and subunits according to the embodiments of this disclosure can be implemented by dividing them into multiple modules. Any one or more of the modules, units, and subunits according to the embodiments of this disclosure can be at least partially implemented as hardware circuitry, such as a Field-Programmable Gate Array (FPGA), a Programmable Logic Array (PLA), a System-on-Chip, a System-on-a-Substrate, a System-on-Package, an Application-Specific Integrated Circuit (ASIC), or implemented in hardware or firmware by any other reasonable means of integrating or packaging the circuitry, or implemented in software, hardware, or firmware, or in any suitable combination of any of these three implementation methods. Alternatively, one or more of the modules, units, and subunits according to the embodiments of this disclosure can be at least partially implemented as a computer program module, which, when run, can perform corresponding functions.

[0179] For example, any multiple modules among the first acquisition module 610, sample recommendation result acquisition module 620, second interactive resource feature acquisition module 630, and training module 640, or the second acquisition module 710, target recommendation result acquisition module 720, and push module 730, can be combined into one module / unit / subunit, or any one of these modules / units / subunits can be split into multiple modules / units / subunits. Alternatively, at least some of the functions of one or more of these modules / units / subunits can be combined with at least some of the functions of other modules / units / subunits and implemented in one module / unit / subunit. According to embodiments of this disclosure, at least one of the first acquisition module 610, the sample recommendation result acquisition module 620, the second interactive resource feature acquisition module 630, and the training module 640, or the second acquisition module 710, the target recommendation result acquisition module 720, and the push module 730, can be at least partially implemented as a hardware circuit, such as a field-programmable gate array (FPGA), a programmable logic array (PLA), a system-on-a-chip, a system-on-a-substrate, a system-on-package, an application-specific integrated circuit (ASIC), or any other reasonable method of integrating or packaging the circuit, or implemented in software, hardware, or firmware, or in any appropriate combination of any of these three implementation methods. Alternatively, at least one of the first acquisition module 610, the sample recommendation result acquisition module 620, the second interactive resource feature acquisition module 630, and the training module 640, or the second acquisition module 710, the target recommendation result acquisition module 720, and the push module 730, can be at least partially implemented as a computer program module, which can perform corresponding functions when the computer program module is run.

[0180] It should be noted that the training device part of the resource recommendation model in the embodiments of this disclosure corresponds to the training method part of the resource recommendation model in the embodiments of this disclosure. For a detailed description of the training device part of the resource recommendation model, please refer to the training method part of the resource recommendation model, which will not be repeated here.

[0181] It should be noted that the resource push device part in the embodiments of this disclosure corresponds to the resource push method part in the embodiments of this disclosure. For a detailed description of the resource push device part, please refer to the resource push method part, which will not be repeated here.

[0182] Figure 8 A block diagram of an electronic device suitable for implementing a training method for a resource recommendation model and a resource push method according to embodiments of the present disclosure is shown schematically. Figure 8 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments disclosed herein.

[0183] like Figure 8 As shown, an electronic device 800 according to an embodiment of this disclosure includes a processor 801, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 802 or a program loaded from a storage portion 808 into a random access memory (RAM) 803. The processor 801 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 801 may also include onboard memory for caching purposes. The processor 801 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of this disclosure.

[0184] RAM 803 stores various programs and data required for the operation of electronic device 800. Processor 801, ROM 802, and RAM 803 are interconnected via bus 804. Processor 801 performs various operations of the method flow according to embodiments of the present disclosure by executing programs in ROM 802 and / or RAM 803. It should be noted that the programs may also be stored in one or more memories other than ROM 802 and RAM 803. Processor 801 may also perform various operations of the method flow according to embodiments of the present disclosure by executing programs stored in said one or more memories.

[0185] According to embodiments of this disclosure, the electronic device 800 may further include an input / output (I / O) interface 805, which is also connected to a bus 804. The electronic device 800 may also include one or more of the following components connected to the input / output (I / O) interface 805: an input section 806 including a keyboard, mouse, etc.; an output section 807 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 808 including a hard disk, etc.; and a communication section 809 including a network interface card such as a LAN card, modem, etc. The communication section 809 performs communication processing via a network such as the Internet. A drive 810 is also connected to the input / output (I / O) interface 805 as needed. A removable medium 811, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 810 as needed so that computer programs read from it can be installed into the storage section 808 as needed.

[0186] According to embodiments of this disclosure, the method flow according to embodiments of this disclosure can be implemented as a computer software program. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable storage medium, the computer program containing program code for performing the methods shown in the flowchart. In such embodiments, the computer program can be downloaded and installed from a network via communication section 809, and / or installed from removable medium 811. When the computer program is executed by processor 801, it performs the functions defined in the system of embodiments of this disclosure. According to embodiments of this disclosure, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0187] This disclosure also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs that, when executed, implement the method according to the embodiments of this disclosure.

[0188] According to embodiments of this disclosure, the computer-readable storage medium can be a non-volatile computer-readable storage medium. Examples include, but are not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this disclosure, the computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0189] For example, according to embodiments of this disclosure, a computer-readable storage medium may include the ROM 802 and / or RAM 803 described above and / or one or more memories other than ROM 802 and RAM 803.

[0190] Embodiments of this disclosure also include a computer program product, which includes a computer program containing program code for performing the methods provided in the embodiments of this disclosure. When the computer program product is run on an electronic device, the program code is used to enable the electronic device to implement the training method and resource push method of the resource recommendation model provided in the embodiments of this disclosure.

[0191] When the computer program is executed by the processor 801, it performs the functions defined in the system / apparatus of this disclosure embodiments. According to embodiments of this disclosure, the systems, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0192] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and may be downloaded and installed via the communication section 809, and / or installed from a removable medium 811. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0193] According to embodiments of this disclosure, program code for executing the computer programs provided in embodiments of this disclosure can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, languages ​​such as Java, C++, Python, "C", or similar programming languages. The program code can execute entirely on a user's computing device, partially on a user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0194] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions. Those skilled in the art will understand that the features described in the various embodiments of the present disclosure can be combined and / or combined in various ways, even if such combinations are not explicitly described in the present disclosure. In particular, the features described in the various embodiments of this disclosure may be combined and / or combined in various ways without departing from the spirit and teachings of this disclosure. All such combinations and / or combinations fall within the scope of this disclosure.

[0195] The embodiments of this disclosure have been described above. However, these embodiments are for illustrative purposes only and are not intended to limit the scope of this disclosure. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. Various substitutions and modifications can be made by those skilled in the art without departing from the scope of this disclosure, and all such substitutions and modifications should fall within the scope of this disclosure.

Claims

1. A method for training a resource recommendation model, comprising: Obtain a sample set related to the sample object, wherein the positive samples of the sample set include a first basic resource feature and a first interaction resource feature, the first basic resource feature is related to the resource feature of the positive sample resource in a historical period, the historical period is the period before the positive sample resource is interacted with by the sample object, the first interaction resource feature is determined based on the interaction operation performed by the sample object on the positive sample resource, and the negative samples of the sample set include a second basic resource feature of the negative sample resource; The first basic resource feature and the second basic resource feature are processed using the initial resource recommendation model to obtain the sample recommendation result; Based on the first interactive resource feature, the negative sample is evaluated for interactive resource features to obtain the second interactive resource feature of the negative sample. Based on the knowledge transfer mechanism, the initial resource recommendation model is trained according to the sample recommendation results and the label recommendation results determined based on the second interactive resource features to obtain the trained resource recommendation model. The resource recommendation model is used to process the basic resource features related to the target object to obtain the target recommendation result.

2. The method according to claim 1, wherein, Based on the first interactive resource feature, the negative sample is evaluated for interactive resource features, resulting in the second interactive resource feature of the negative sample, which includes: For the target preset interaction resource features, a causal inference operation is performed based on the first interaction resource features of the positive sample and the preset interaction operation prior information to obtain the feature distribution data of the target preset interaction features in the sample set, wherein the interaction operation prior information characterizes the similarity between the negative sample and at least one of the positive samples. Determine the target distribution vector based on the feature distribution data of each of the multiple target preset interactive features; According to the optimal transmission mechanism, the target optimization function is solved based on the optimization conditions determined by the target distribution vector to obtain the optimized target transmission matrix. The matrix elements of the target transmission matrix include the target feature distribution data of the negative sample, and the target feature distribution data characterizes the target distribution probability of the negative sample in the target preset interactive resource features. The second interactive resource feature of the negative sample is determined based on the target feature distribution data.

3. The method according to claim 2, wherein, Each of the positive samples corresponds to a priori information of the interaction operation; The causal inference operation based on the first interactive resource features of the positive samples and the preset prior information of the interactive operations includes: Based on the ratio between the first interactive resource feature of the positive sample and the prior information of the interactive operation associated with the positive sample, at least one of the target preset interactive features is determined as intermediate distribution data in the positive sample; and The intermediate distribution data corresponding to each of the multiple positive samples in the sample set are fused to obtain the feature distribution data of at least one target preset interaction feature in the sample set.

4. The method according to claim 2 or 3, wherein, The prior information for the interactive operation is determined based on the following operations: The reference recommendation index prediction network is used to process the first interaction resource features and the second interaction resource features to obtain the reference recommendation indexes for the positive samples and the negative samples respectively. as well as Based on the reference recommendation metrics of the positive samples and the negative samples, and the prior information of the interaction operations of the positive samples and the negative samples.

5. The method according to claim 2, wherein, Determining the second interactive resource features of the negative sample based on the target feature distribution data includes: For the target negative sample in the negative sample, based on the target feature distribution data corresponding to the target preset interactive resource features, determine the fusion weight corresponding to the target preset interactive resource features; and Based on the fusion weight, multiple target preset interactive resource features are weighted and fused to obtain the second interactive resource features of the target negative sample.

6. The method according to claim 2, wherein, The objective optimization function is constructed based on the cost matrix and the transfer matrix to be optimized. The cost matrix is ​​determined by the following operation: Obtain hidden features of sample resources, wherein the hidden features of sample resources are obtained by feature extraction from the first basic resource features and the second basic resource features; Based on the hidden features of the sample resources, the interaction resource features are predicted to obtain the cost matrix. The matrix elements of the cost matrix represent the cost feature distribution data of each sample in the sample set, where the feature type of the target preset interaction feature is included. The sample includes the positive sample or the negative sample.

7. The method according to claim 1, wherein, The first interactive resource feature or the second interactive resource feature includes at least one of the following: Features of commenting actions, purchase actions, resource browsing duration, and resource exposure location.

8. The method according to claim 1, wherein, The tag recommendation results are determined based on the following operations: The second interactive resource feature and the sample resource hidden feature are fused to obtain the sample fusion feature, wherein the sample resource hidden feature is obtained by feature extraction from the first basic resource feature and the second basic feature; as well as The recommendation results are evaluated based on the sample fusion features to obtain the label recommendation results.

9. A resource push method, comprising: Obtain the target's basic resource characteristics related to the target object; The target basic resource features are processed using the resource recommendation model trained according to any one of claims 1 to 8 to obtain the target recommendation result; as well as Push target resources corresponding to the target recommendation results to the target object.

10. A training device for a resource recommendation model, comprising: The first acquisition module is used to acquire a sample set related to the sample object. The positive samples of the sample set include a first basic resource feature and a first interaction resource feature. The first basic resource feature is related to the resource feature of the positive sample resource in a historical period. The historical period is the period before the positive sample resource is interacted with by the sample object. The first interaction resource feature is determined based on the interaction operation performed by the sample object on the positive sample resource. The negative samples of the sample set include a second basic resource feature of the negative sample resource. The sample recommendation result acquisition module is used to process the first basic resource feature and the second basic resource feature using the initial resource recommendation model to obtain the sample recommendation result. The second interactive resource feature acquisition module is used to evaluate the interactive resource features of the negative sample based on the first interactive resource features to obtain the second interactive resource features of the negative sample. The training module is used to train the initial resource recommendation model based on the sample recommendation results and the label recommendation results determined based on the second interactive resource features, using a knowledge transfer mechanism, to obtain the trained resource recommendation model. The resource recommendation model is used to process the basic resource features related to the target object to obtain the target recommendation result.

11. A resource delivery device, comprising: The second acquisition module is used to acquire the target basic resource characteristics related to the target object; The target recommendation result acquisition module is used to process the target basic resource features using the resource recommendation model trained according to any one of claims 1 to 8, and obtain the target recommendation result; as well as The push module is used to push target resources corresponding to the target recommendation results to the target object.

12. An electronic device, comprising: One or more processors; Memory, used to store one or more programs. Wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the method of any one of claims 1 to 9.

13. A computer-readable storage medium having stored thereon executable instructions that, when executed by a processor, cause the processor to perform the method of any one of claims 1 to 9.

14. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1 to 9.

Images