Resource processing method and device, equipment and storage medium

Abstract

The invention discloses a resource processing method and device, equipment and a storage medium. The method comprises the steps of obtaining object information of a target object and resource information of a target resource to be recommended to the target object; a multi-target prediction model is obtained, a plurality of prediction targets of the multi-target prediction model are determined based on the dependency relationship among a plurality of business targets, and the prediction target corresponding to any dependent business target comprises a resource feedback behavior indicated by the business target on which the corresponding business target depends after the resource is executed; the probability that the resource feedback behavior indicated by the corresponding service target is executed by the corresponding resource; and calling a multi-target prediction model, performing multi-target prediction processing on the target resource according to the object information of the target object and the resource information of the target resource, and performing recommendation decision processing on the target resource based on the prediction result of each prediction target. According to the invention, the accuracy of resource recommendation decisions can be improved.

Description

Technical Field

[0001] The present application relates to the field of Internet technology, specifically to the field of computer technology, and in particular to a resource processing method, apparatus, device, and storage medium. Background Art

[0002] Currently, resource recommendation services typically configure multiple business objectives, such as click-through and conversion targets. When recommending resources to a target user (any user), each business objective is typically directly used as a prediction target in a multi-objective prediction model. This model then uses the resource information of the resource to be recommended to predict multiple business objectives. Based on the predicted results for each business objective (such as the predicted click-through rate (CTR) and conversion rate), a decision is made on whether to recommend the corresponding resource to the target user. This approach can easily lead to low accuracy in resource recommendation decisions. Summary of the Invention

[0003] The embodiments of the present application provide a resource processing method, apparatus, device, and storage medium, which can improve the accuracy of multi-objective prediction, thereby improving the accuracy of resource recommendation decisions.

[0004] In one aspect, an embodiment of the present application provides a resource processing method, the method comprising:

[0005] Acquiring object information of a target object and acquiring resource information of a target resource to be recommended to the target object;

[0006] Obtaining a multi-objective prediction model, wherein multiple prediction objectives of the multi-objective prediction model are determined based on dependency relationships between multiple business objectives, where each business objective indicates a resource feedback behavior; the multiple prediction objectives include a prediction objective corresponding to each business objective, and the prediction objective corresponding to any dependent business objective includes: a probability that the corresponding resource will perform the resource feedback behavior indicated by the business objective on which the corresponding business objective depends, after the resource performs the resource feedback behavior indicated by the business objective on which the corresponding business objective depends;

[0007] Calling the multi-objective prediction model, performing multi-objective prediction processing on the target resource according to the object information of the target object and the resource information of the target resource, and obtaining a prediction result for each prediction target;

[0008] Based on the prediction result of each prediction target, a recommendation decision process is performed on the target resource.

[0009] On the other hand, an embodiment of the present application provides a resource processing device, the device comprising:

[0010] an acquiring unit, configured to acquire object information of a target object and resource information of a target resource to be recommended to the target object;

[0011] The acquisition unit is further configured to acquire a multi-objective prediction model, wherein multiple prediction targets of the multi-objective prediction model are determined based on dependency relationships between multiple business objectives, where each business objective indicates a resource feedback behavior; the multiple prediction targets include a prediction target corresponding to each business objective, and the prediction target corresponding to any dependent business objective includes: after the resource executes the resource feedback behavior indicated by the business objective on which the corresponding business objective depends, the probability that the corresponding resource will execute the resource feedback behavior indicated by the corresponding business objective;

[0012] a processing unit, configured to call the multi-objective prediction model, perform multi-objective prediction processing on the target resource according to the object information of the target object and the resource information of the target resource, and obtain a prediction result for each prediction target;

[0013] The processing unit is further configured to perform recommendation decision processing on the target resource based on the prediction result of each prediction target.

[0014] In a specific embodiment, the object information of the target object includes at least: a target behavior sequence of the target object under each business objective; any target behavior sequence includes: a resource identifier of at least one resource; wherein, among the multiple business objectives, a dependent business objective is represented as a first business objective; when the acquisition unit is used to acquire the target behavior sequence of the target object under the first business objective, it can be specifically used to:

[0015] Obtaining a first historical behavior sequence of the target object under the first business objective, where the first historical behavior sequence is a sequence consisting of resource identifiers of at least one resource for which the target object has historically executed a first resource feedback behavior; the first resource feedback behavior is the resource feedback behavior indicated by the first business objective;

[0016] If there are L missing values ​​in the obtained first historical behavior sequence, then obtain a second historical behavior sequence of the target object under the second business goal, where the second business goal refers to the business goal on which the first business goal depends; the second historical behavior sequence refers to a sequence consisting of resource identifiers of at least one resource for which the target object has historically executed a second resource feedback behavior; the second resource feedback behavior refers to the resource feedback behavior indicated by the second business goal; and L is a positive integer.

[0017] Based on the second historical behavior sequence, the first historical behavior sequence is dynamically supplemented with missing values; and the dynamically supplemented first historical behavior sequence is used as the target behavior sequence of the target object under the first business goal.

[0018] In another specific implementation, when the acquisition unit is used to dynamically supplement missing values ​​of the first historical behavior sequence based on the second historical behavior sequence, it can be specifically used to:

[0019] Taking the resources corresponding to the resource identifiers in the second historical behavior sequence as candidate resources; and counting the total number of times the first resource feedback behavior has been executed on each candidate resource in the past;

[0020] Filter out the first L candidate resources from the candidate resources in descending order of the total number of times; and dynamically add the resource identifiers of the first L candidate resources to the first historical behavior sequence.

[0021] In another specific implementation, when the acquisition unit is used to dynamically supplement missing values ​​of the first historical behavior sequence based on the second historical behavior sequence, it can be specifically used to:

[0022] Determine at least one reference object, where the reference object refers to an object that meets the following conditions: no missing values ​​exist in the historical behavior sequence under the first business objective and the historical behavior sequence under the second business objective;

[0023] respectively calculating sequence similarities between the second historical behavior sequence and the historical behavior sequence of each reference object under the second business goal;

[0024] From at least one reference object, a reference object corresponding to the greatest sequence similarity is selected; from the historical behavior sequence of the selected reference object under the first business goal, L resource identifiers are selected and dynamically added to the first historical behavior sequence.

[0025] In another specific embodiment, the object information of the target object includes: attribute information of the target object, and a target behavior sequence of the target object under each business goal; any target behavior sequence includes: a resource identifier of at least one resource;

[0026] The resource information of the target resource includes at least one of the following: attribute information of the target resource, attribute information of the resource entity used to push the target resource, and statistical information of the cross-relationship; wherein, the statistical information of the cross-relationship includes at least one of the following: the number of exposures generated by the resource entity for the target object within a first historical time period, and the number of times the target object performs corresponding resource feedback behavior on the resource entity within a second historical time period under at least one business goal.

[0027] In another specific embodiment, when the processing unit is used to call the multi-objective prediction model, perform multi-objective prediction processing on the target resource based on the object information of the target object and the resource information of the target resource, and obtain the prediction result of each prediction target, it can be specifically used to:

[0028] Using a common vector table, feature mapping is performed on the attribute information of the target object and the resource information of the target resource to obtain an object attribute vector and a resource feature vector;

[0029] Using the private vector table configured for each business goal, feature mapping is performed on the target behavior sequence of the target object under the corresponding business goal to obtain the behavior feature vector of the target object under each business goal;

[0030] Integrating the object attribute vector, the resource feature vector, and the behavior feature vector of the target object under each business goal to obtain a resource representation vector of the target resource;

[0031] The multi-objective prediction model is called to perform multi-objective prediction on the target resource according to the resource representation vector of the target resource to obtain a prediction result for each prediction target.

[0032] In another specific embodiment, when the processing unit is used to integrate the object attribute vector, the resource feature vector, and the behavior feature vector of the target object under each business goal to obtain the resource representation vector of the target resource, it can be specifically used to:

[0033] Invoking a behavior sequence representation extraction network to generate an interest representation vector for the target object under each business goal based on the similarity between the behavior feature vector of the target object under each business goal and the resource feature vector; each interest representation vector is used to represent: the interest of the target object in performing a resource feedback operation indicated by the corresponding business goal on the target resource;

[0034] The object attribute vector, the resource feature vector, and the interest representation vector of the target object under each business objective are concatenated to obtain a resource representation vector of the target resource.

[0035] In another specific implementation, before obtaining the multi-objective prediction model, the processing unit 602 may further be configured to:

[0036] Obtaining training samples and corresponding sample labels for a multi-objective prediction model, wherein the training samples include: object information of a sample object and resource information of a sample resource, wherein the sample resource is a resource that has been exposed to the sample object; the sample label is used to indicate whether the sample object has performed a resource feedback behavior indicated by each business objective on the sample resource;

[0037] Calling the multi-objective prediction model, performing multi-objective prediction processing on the sample resource according to various information in the training sample, and obtaining a multi-objective prediction result corresponding to the sample resource;

[0038] Based on the multi-objective prediction results corresponding to the sample labels and the sample resources, the multi-objective prediction model is updated.

[0039] In another specific embodiment, the multi-objective prediction model includes: a prediction network corresponding to each business objective, and the prediction objective of any prediction network is the prediction objective corresponding to the corresponding business objective; accordingly, when the processing unit is used to call the multi-objective prediction model and perform multi-objective prediction processing on the sample resource based on various information in the training sample to obtain the multi-objective prediction result corresponding to the sample resource, it can be specifically used to:

[0040] A reference business target is selected from the multiple business targets based on the sample labels; if the p-th business target has no dependent business targets, the p-th business target is determined to be the reference business target; or if the p-th business target has a dependent business target, and the sample object has performed a resource feedback action on the sample resource indicated by the business target on which the p-th business target depends, the p-th business target is determined to be the reference business target, where p is a positive integer;

[0041] The prediction network corresponding to each reference business objective in the multi-objective prediction model is called respectively, and the prediction processing of the corresponding prediction objective is performed on the sample resource according to each information in the training sample to obtain the multi-objective prediction result corresponding to the sample resource.

[0042] In another specific embodiment, the multi-objective prediction results corresponding to the sample resources include: prediction results of each prediction target corresponding to each reference business target; accordingly, when the processing unit 602 is used to update the multi-objective prediction model based on the sample labels and the multi-objective prediction results corresponding to the sample resources, it can be specifically used to:

[0043] Calling the loss function under each reference business objective respectively, performing loss value calculation based on the sample label and the prediction results of each prediction objective corresponding to the corresponding reference business objective, and obtaining the loss value under each reference business objective;

[0044] Integrating the loss values ​​under each reference business objective to obtain a model loss value generated by the multi-objective model through the training sample;

[0045] Update the network parameters of each network other than the target network in the target prediction model in a direction of reducing the model loss value; wherein the target network includes: the prediction network corresponding to the business objectives other than each reference business objective in the multiple business objectives.

[0046] In another aspect, an embodiment of the present application provides a computer device, the computer device including an input interface and an output interface, and the computer device further including:

[0047] processors and computer storage media;

[0048] The processor is suitable for implementing one or more instructions, the computer storage medium stores one or more instructions, and the one or more instructions are suitable for being loaded by the processor and executing the resource processing method mentioned above.

[0049] On the other hand, an embodiment of the present application provides a computer storage medium, which stores one or more instructions, and the one or more instructions are suitable for being loaded by a processor and executing the above-mentioned resource processing method.

[0050] On the other hand, an embodiment of the present application provides a computer program product, which includes one or more instructions; when the one or more instructions in the computer program product are executed by a processor, the above-mentioned resource processing method is implemented.

[0051] The embodiment of the present application determines multiple prediction targets of a multi-target prediction model based on the dependency relationship between multiple business targets, so that the prediction target corresponding to any dependent business target includes: the probability that the corresponding resource is subjected to the resource feedback behavior indicated by the business target on which the corresponding business target depends after the resource is subjected to the resource feedback behavior indicated by the business target on which the corresponding business target depends. It can be seen that such a prediction target takes into account the logical relationship that in the actual interaction process, the resource must be subjected to the resource feedback operation indicated by the business target on which any business target depends before the resource feedback operation indicated by the corresponding business target can be executed. It has high rationality, and reasonable prediction targets help to improve the accuracy of multi-target prediction. For the prediction target corresponding to any dependent business target, the multi-target prediction model can make predictions on the corresponding prediction target based on the object information of the target object and the resource information of the target resource. It can consider whether the target resource is executed by the resource feedback behavior indicated by the business target on which any business target depends, and the impact on the probability of the target resource being executed by the resource feedback behavior indicated by any business target, so that the predicted probability (i.e., the prediction result) is more accurate, thereby improving the accuracy of the multi-target prediction, so that the target resource can be recommended and decided based on the more accurate prediction results, thereby improving the accuracy of the resource recommendation decision. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0053] Figure 1a This is a schematic diagram of a resource details page provided in an embodiment of the present application;

[0054] Figure 1b This is a schematic diagram of the relationship between a subscription number message and a subscription number provided in an embodiment of the present application;

[0055] Figure 1c This is a logical diagram of a resource processing solution jointly executed by a terminal and a server provided in an embodiment of the present application;

[0056] Figure 1d This is another interactive logic diagram of a terminal and a server jointly executing a resource processing solution provided by an embodiment of the present application;

[0057] Figure 2 This is a flowchart of a resource processing method provided in an embodiment of the present application;

[0058] Figure 3This is a schematic diagram of a training process of a multi-objective prediction model provided in an embodiment of the present application;

[0059] Figure 4 This is a schematic diagram of the structure of a multi-objective prediction model provided in an embodiment of the present application;

[0060] Figure 5 This is a schematic diagram of a dynamic supplement of missing values ​​provided by an embodiment of the present application;

[0061] Figure 6 This is a schematic diagram of the structure of a resource processing device provided in an embodiment of the present application;

[0062] Figure 7 It is a structural diagram of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0063] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application.

[0064] In the embodiments of the present application, the resource recommendation service mentioned later refers to a service for recommending resources, which may be configured with multiple service objectives; the so-called service objectives refer to objectives set based on the service needs of the resource recommendation service, which may be, for example, click objectives, duration objectives, interaction objectives, purchase objectives, download objectives, and the like. A service objective indicates a resource feedback behavior, such as the resource feedback behavior indicated by a click objective includes click behavior, the resource feedback behavior indicated by a duration objective includes browsing behavior, the resource feedback behavior indicated by an interaction objective includes interaction behavior, the resource feedback behavior indicated by a purchase objective includes resource purchase behavior, the resource feedback behavior indicated by a download objective includes resource download behavior, and the like. Furthermore, interactive behaviors may include, but are not limited to: sharing behavior, forwarding behavior, liking behavior, commenting behavior, and collecting behavior; any interactive behavior can be accessed through the interactive entry in the resource details page (e.g. Figure 1a You can input through the sharing entrance, forwarding entrance, like entrance, comment entrance, collection entrance, etc. shown in the figure, or you can input through preset gestures or voice commands, without limitation.

[0065] It is understandable that when the resource feedback behavior indicated by any business goal can only be executed after the resource feedback behavior indicated by another business goal is executed, then any business goal can be considered to be dependent on another business goal. For example, usually after the click behavior is performed on the resource, the browsing behavior can be performed on the resource, then the duration goal for indicating the browsing behavior can be considered to be dependent on the click goal for indicating the click behavior; for another example, usually after the browsing behavior is performed on the resource, the interactive behavior can be performed on the resource, then the interactive behavior can be considered to be dependent on the duration goal for indicating the browsing behavior; for another example, usually after the browsing behavior is performed on the resource, the download behavior can be performed on the resource, then the download goal for indicating the download behavior can be considered to be dependent on the duration goal for indicating the browsing behavior, and so on.

[0066] The resources mentioned above may be any of the following: news information, items, multimedia data, and so on. Furthermore, the news information mentioned here may be subscription account messages (information disseminated through subscription accounts, which are accounts that disseminate information on social platforms), article information, advertising information, news information, etc.; the items mentioned here may be real items (such as clothes, shoes, and other commodities), or virtual items (such as game props, game characters, etc.); the multimedia data mentioned here may be audio data, video data, image data, etc. Any resource may be pushed directly to any object for exposure, or it may be pushed indirectly to any object for exposure through a resource entity; the so-called resource entity refers to the entrance for pushing resources, and the object can view the resources pushed by it by entering the resource entity. For example, if the resource is a subscription account message, the resource entity may be a subscription account, and the relationship between the subscription account and the subscription account message can be exemplified by referring to Figure 1b As shown; if the resource is an item, the resource entity can be a store; if the resource is multimedia data, the resource entity can be the account of the publisher of the multimedia data, and so on.

[0067] In order to improve the accuracy of resource recommendation decisions, the embodiment of the present application proposes a resource processing solution based on AI (Artificial Intelligence) technology. Among them, the so-called AI technology refers to the theory, method, technology and application system of using digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology in computer science; it mainly produces a new intelligent machine that can respond in a similar way to human intelligence by understanding the essence of intelligence, so that the intelligent machine has multiple functions such as perception, reasoning and decision-making. Correspondingly, AI technology is an interdisciplinary subject, which mainly includes several major directions such as computer vision technology (Computer Vision, CV), speech processing technology, natural language processing technology and machine learning (Machine Learning, ML) / deep learning.

[0068] Machine learning is the core of AI and the fundamental path to intelligent computing. Machine learning is a multidisciplinary field encompassing probability theory, statistics, approximation theory, convex analysis, and algorithmic complexity theory. It specifically studies how computing devices can simulate or implement human learning behaviors to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Deep learning, on the other hand, is a machine learning technique that utilizes deep neural network systems. Machine learning / deep learning typically encompasses various techniques, such as artificial neural networks, reinforcement learning (RL), and supervised learning. Supervised learning involves optimizing models using training samples with known categories (labeled samples). Sample labels can be understood as the annotated data used to train neural network models, such as the "0" / "1" sample labels indicating whether a training sample "belongs" to a particular category or not.

[0069] The resource processing scheme proposed in the embodiment of the present application mainly involves the machine learning / deep learning technology in the AI ​​technology mentioned above. Specifically, the general principle of the resource processing scheme proposed in the embodiment of the present application is as follows: First, the modeling of multiple business objectives of the resource recommendation business can be integrated into a multi-objective prediction model, and multiple prediction targets of the multi-objective prediction model can be designed based on the conditional dependency relationship between multiple business objectives (hereinafter referred to as dependency relationship). Specifically, a business objective indicates a resource feedback behavior, and multiple prediction targets may include a prediction target corresponding to each business objective. For any business objective with dependency, its corresponding prediction target can be designed as: after the resource is executed with the resource feedback behavior indicated by the business objective on which the corresponding business objective depends, the probability that the corresponding resource is executed with the resource feedback behavior indicated by the corresponding business objective; in this way, the corresponding prediction target takes into account the logical relationship that in the actual interaction process, the resource must be executed with the resource feedback operation indicated by the business objective on which any business objective depends before the resource feedback operation indicated by the corresponding business objective can be executed, thereby improving the rationality of the prediction target and thus improving the accuracy of the multi-objective prediction. Then, the multi-objective prediction model can be trained through machine learning / deep learning technology to update and optimize the model parameters of the multi-objective prediction model, thereby improving the prediction performance of the multi-objective prediction model, so that when recommending resources to the target object (any user), the object information of the target object and the resource information of the target resource to be recommended to the target object can be obtained, and the multi-objective prediction model can be called to perform multi-objective prediction processing on the target resource more accurately based on the object information of the target object and the resource information of the target resource, thereby improving the accuracy of the prediction results of each prediction target, and then performing recommendation decision processing on the target resource based on the more accurate prediction results, thereby improving the accuracy of the resource recommendation decision.

[0070] In a specific implementation, the resource processing scheme mentioned above can be executed by a computer device, which can be a terminal or a server. Alternatively, the resource processing scheme mentioned above can also be executed jointly by the terminal and the server. For example: the terminal used by the technician responds to the model setting operation input by the technician, integrates the modeling of multiple business objectives of the resource recommendation business into a multi-objective prediction model, and designs multiple prediction targets of the multi-objective prediction model based on the conditional dependency relationship between the multiple business objectives, and then sends the multi-objective prediction model with multiple prediction targets to the server; the server is responsible for training the multi-objective prediction model, and after obtaining the object information of the target object and the resource information of the target resource to be recommended to the target object, calls the multi-objective prediction model to perform multi-objective prediction processing on the target resource based on the obtained information, thereby performing recommendation decision processing on the target resource based on the prediction results of each prediction target, such as Figure 1cAs shown. Another example: the terminal used by the technician sends a multi-objective prediction model with multiple prediction targets to the server, and after the server trains the multi-objective prediction model, the server can also deploy the trained multi-objective prediction model to the device (terminal or server) that provides resource recommendation services, so that after the corresponding device obtains the object information of the target object and the resource information of the target resource to be recommended to the target object, it can call the multi-objective prediction model to perform multi-objective prediction processing on the target resource according to the obtained information, thereby performing recommendation decision processing on the target resource based on each prediction result, as shown in FIG. Figure 1d shown.

[0071] It is understood that the above is only an exemplary description of two implementation methods of the resource processing scheme proposed in the embodiment of the present application by the terminal and the server, and is not an exhaustive list. In addition, the terminals mentioned above can be smart phones, computers (such as tablets, laptops, desktop computers, etc.), smart wearable devices (such as smart watches, smart glasses), smart voice interaction devices, smart home appliances (such as smart TVs), vehicle terminals or aircraft, etc.; the server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN (Content Delivery Network), and basic cloud computing services such as big data and artificial intelligence platforms, etc. Furthermore, the terminal and server can be located within or outside the blockchain network, without limitation; further, the terminal and server can also upload any data stored internally to the blockchain network for storage to prevent the internally stored data from being tampered with and improve data security.

[0072] It is worth emphasizing that in the embodiments of the present application, when any method embodiment proposed in the embodiments of the present application is applied to a specific product or technology, user information and other related data (such as object information of the target object, any operation performed by any user, etc.) are involved, these related data are collected with the user's permission or consent, and the collection, use and processing of the relevant data comply with the relevant laws, regulations and standards of the relevant region.

[0073] Based on the relevant description of the above resource processing solution, the embodiment of the present application proposes a resource processing method. In the embodiment of the present application, the resource processing method can be executed by the computer device (terminal or server) mentioned above, or it can be executed by the terminal and the server together; for the sake of convenience, the following description will take the computer device executing the resource processing method as an example. Figure 2As shown, the resource processing method may include the following steps S201-S204:

[0074] S201: Acquire object information of a target object, and acquire resource information of a target resource to be recommended to the target object.

[0075] (1) The target object refers to any object (any user) to which resources need to be recommended. The object information of the target object may include at least one of the following: the attribute information of the target object, and the target behavior sequence of the target object under each business goal. Among them, ① the attribute information of the target object refers to the information used to characterize the object portrait of the target object, which may include but is not limited to: the age, gender, height, region of the target object, the number of exposure resources of the target object in a specified historical time period, etc.; the number of exposure resources mentioned here can be understood as the number of resources pushed to the target object, and the specified historical time period is a historical time period set according to experience or business needs, such as the past 1 day (1 day before the current time displayed by the system) or the past 7 days (i.e., the 7 days before the current time displayed by the system). ② The target behavior sequence of the target object under any business goal can be: the historical behavior sequence of the target object under the corresponding business goal; the so-called historical behavior sequence refers to: a sequence composed of the resource identifier of at least one resource for which the corresponding object has historically performed the resource feedback behavior indicated by the corresponding business goal. For example, the historical behavior sequence of the target object under the click target refers to: a sequence composed of the resource identifier of at least one resource on which the target object has historically performed click behavior; the historical behavior sequence of the target object under the interaction target refers to: a sequence composed of the resource identifier of at least one resource on which the target object has historically performed interactive behavior; the historical behavior sequence of the target object under the download target refers to: a sequence composed of the resource identifier of at least one resource on which the target object has historically performed download behavior, and so on.

[0076] It is understandable that, for any business goal, if the target object has not historically performed the resource feedback behavior indicated by the corresponding business goal, or the number of resources for which the target object has historically performed the resource feedback behavior indicated by the corresponding business goal is small, then the historical behavior sequence of the target object under any business goal (such as business goal a) may have missing values ​​(i.e., missing resource identifiers); in this case, the historical behavior sequence of the target object under other business goals can be used to supplement the corresponding missing values, or the historical behavior sequence of other objects similar to the target object under the same business goal (i.e., business goal a) can be used to supplement the corresponding missing values, so that the multi-objective prediction model can still perform modeling and reasoning under the corresponding business goal in the future, thereby improving the accuracy of multi-objective prediction.

[0077] Exemplarily, assuming that among multiple business objectives, a dependent business objective can be represented as the first business objective; then, the method of obtaining the target behavior sequence of the target object under the first business objective may include: s11, obtaining the first historical behavior sequence of the target object under the first business objective, the first historical behavior sequence refers to: a sequence composed of resource identifiers of at least one resource for which the target object has historically performed a first resource feedback behavior; the first resource feedback behavior refers to: the resource feedback behavior indicated by the first business objective. s12, detecting whether there are missing values ​​(i.e., missing resource identifiers) in the obtained first historical behavior sequence; if there are no missing values ​​in the obtained first historical behavior sequence, the first historical behavior sequence is used as the target behavior sequence of the target object under the first business objective. If there are L missing values ​​in the obtained first historical behavior sequence, the second historical behavior sequence of the target object under the second business objective is obtained, and based on the second historical behavior sequence, the first historical behavior sequence is dynamically supplemented with missing values; and the first historical behavior sequence after dynamic supplementation is used as the target behavior sequence of the target object under the first business objective.

[0078] Wherein, ① the aforementioned L is a positive integer. The value of L can specifically be equal to the difference between the length of the first historical behavior sequence (used to indicate the total number of resource identifiers required to be included in the first historical behavior sequence) and the number of resource identifiers already included in the first historical behavior sequence. It is understandable that when the first historical behavior sequence is empty, since the number of resource identifiers already included in the first historical behavior sequence is 0, L in this case is equal to the length of the first historical behavior sequence. ② The aforementioned second business goal refers to the business goal on which the first business goal depends. For example, if the first business goal is an interaction goal, the second business goal can be a click goal or a duration goal; for another example, if the first business goal is a browsing goal, the second business goal can be a click goal, and so on. ③ The aforementioned second historical behavior sequence refers to a sequence consisting of the resource identifiers of at least one resource for which the target object historically performed the second resource feedback behavior. The second resource feedback behavior refers to the resource feedback behavior indicated by the second business goal. For example, if the second business goal is a click goal, then the second historical behavior sequence refers to a sequence consisting of the resource identifiers of at least one resource for which the target object historically performed the click behavior.

[0079] Furthermore, based on the second historical behavior sequence, a specific implementation method for dynamically supplementing the missing values ​​of the first historical behavior sequence may be: using the resources corresponding to each resource identifier in the second historical behavior sequence as candidate resources; and respectively counting the total number of times each candidate resource has been executed with the first resource feedback behavior in history. Then, the top L candidate resources can be screened out from the candidate resources in descending order of the total number of times; and the resource identifiers of the top L candidate resources can be dynamically supplemented to the first historical behavior sequence. Since the first resource feedback behavior is executed after the second resource feedback behavior, each resource that has been executed with the second resource feedback behavior by the target object has a greater probability of being continued to be executed with the first resource feedback behavior by the target object. Therefore, the rationality of the first historical behavior sequence can be improved by selecting L candidate resources from the resources that have been executed with the second resource feedback behavior by the target object to fill in the first historical behavior sequence.

[0080] Alternatively, another specific implementation of dynamically supplementing missing values ​​in the first historical behavior sequence based on the second historical behavior sequence can include: determining at least one reference object, where the reference object satisfies the following conditions: neither the historical behavior sequence under the first business objective nor the historical behavior sequence under the second business objective contains missing values. Sequence similarity is calculated between the second historical behavior sequence and the historical behavior sequence under the second business objective for each reference object, and the reference object corresponding to the greatest sequence similarity is selected from the at least one reference object. Then, L resource identifiers can be selected from the historical behavior sequence under the first business objective of the selected reference object and dynamically supplemented into the first historical behavior sequence. Since similarity between the historical behavior sequences of two objects under the second business objective indicates that the resources affected by the second resource feedback behaviors previously executed by the two objects are similar, it can also indicate that the two objects have similar interests. Furthermore, it can also indicate that the resources affected by the first resource feedback behaviors subsequently executed by the two objects after executing the second resource feedback behaviors are generally similar. Therefore, by supplementing the first historical behavior sequence through sequence similarity, the rationality of the first historical behavior sequence can be improved.

[0081] (2) Target resource refers to any resource to be recommended to the target object in the resource recommendation system. The resource information of the target resource may include the attribute information of the target resource, and the attribute information of the target resource refers to information that can be used to indicate the historical situation of resource feedback behavior being executed on the target resource; for example, it may include but is not limited to: the retention time of the target resource (i.e., the interval between the time when the target resource is generated and the current display time of the system, such as the number of hours since the generation), the number of exposures of the target resource, the number of times the target resource has been subjected to any resource feedback behavior (such as the number of click behaviors on the target resource, the number of interactive behaviors on the target resource, the number of download behaviors on the target resource, etc.), the browsing time of the target resource, etc.

[0082] Optionally, if the target resource is indirectly pushed to any object for exposure via a resource entity, the resource information of the target resource may further include at least one of the following: attribute information of the resource entity used to push the target resource, and cross-relationship statistical information. ① The attribute information of the resource entity used to push the target resource may include, but is not limited to: the entity identifier of the corresponding resource entity, the number of followers of the corresponding resource entity (i.e., the number of objects that follow the corresponding resource entity), the number of resources pushed by the corresponding resource entity within a preset period (e.g., 7 days, 30 days, etc.), the number of resource feedback actions (e.g., clicks, browsing, interactions, etc.) performed on the corresponding resource entity within the preset period, etc. ② The cross-relationship statistical information may include at least one of the following: the number of exposures generated by the resource entity for the target object within a first historical time period (i.e., the number of times the target object was exposed), and the number of resource feedback actions (e.g., clicks, interactions, etc.) performed by the target object on the resource entity within a second historical time period under at least one business objective. The first historical time period and the second historical time period herein both refer to the time period before the current system display moment; these two historical time periods may be the same or different, without limitation.

[0083] S202, obtaining a multi-objective prediction model.

[0084] Among them, the multiple prediction targets of the multi-target prediction model are determined based on the dependency relationship between multiple business targets. The business target that any business target depends on refers to: the business target that needs to be achieved before the corresponding business target is achieved; a business target indicates a resource feedback behavior, so achieving the corresponding business target can be understood as: executing the resource feedback behavior indicated by the corresponding business target. Further, the multiple prediction targets determined based on the dependency relationship between multiple business targets may include: the prediction targets corresponding to each business target. Among them, ① the prediction target corresponding to a business target with no dependency (that is, a business target that does not depend on any business target) may include: the probability that the resource is executed with the resource feedback behavior indicated by the corresponding business target. ② The prediction target corresponding to any business target with dependency (that is, a business target that depends on other business targets) includes: the probability that the corresponding resource is executed with the resource feedback behavior indicated by the corresponding business target after the resource is executed with the resource feedback behavior indicated by the business target that the corresponding business target depends on.

[0085] For example, if multiple business goals include click goals and duration goals, and the duration goal depends on the click goal, then in this case, the click goal is a business goal without any dependency, while the duration goal is a business goal with a dependency. Then, since the resource feedback behavior indicated by the click behavior includes click behavior, and the resource feedback behavior indicated by the duration goal includes browsing behavior, the prediction target corresponding to the click goal may include: the probability that the resource is clicked; the prediction target corresponding to the duration goal may include: the probability that the corresponding resource is browsed after the resource is clicked. Similarly, if multiple business goals include duration goals and interaction goals, and the interaction goal depends on the duration goal, then in this case, the prediction target corresponding to the duration goal may include: the probability that the resource is browsed; the prediction target corresponding to the interaction goal may include: the probability that the corresponding resource is interactive after the resource is browsed. If multiple business goals include click goals and download goals (or purchase goals), and the download goal (or purchase goal) depends on the click goal, then in this case, the prediction target corresponding to the click target may include: the probability of the resource being clicked; the prediction target corresponding to the download target (or purchase target) may include: the probability of the corresponding resource being downloaded (or purchased) after the resource is clicked.

[0086] For another example, if multiple business goals include a click goal, a duration goal, and an interaction goal, and the duration goal depends on the click goal, and the interaction goal depends on the duration goal, then in this case, the click goal is a business goal without any dependencies, while the duration goal and the interaction goal are both dependent businesses. Then, the prediction target corresponding to the click goal may include: the probability that the resource is clicked; the prediction target corresponding to the duration goal may include: the probability that the corresponding resource is browsed after the resource is clicked; the prediction target corresponding to the interaction goal may include: the probability that the corresponding resource is interactive after the resource is browsed. Further, based on this example, if multiple business goals also include: a purchase goal or a download goal that depends on the interaction goal, then the prediction target corresponding to the purchase goal or the download goal may include: the probability that the corresponding resource is purchased or downloaded after the resource is interactive.

[0087] It should be noted that the above is only an example of several specific implementations of multiple business goals, and is not exhaustive. In actual applications, the above multiple business goals may vary according to different application scenarios. When multiple business goals include business goals other than the above mentioned goals, their specific implementation methods are similar to the specific implementation methods of the above mentioned goals, and will not be repeated here. In addition, in the examples listed above, for the duration goal, the related prediction goal focuses on whether the resource is browsed; based on this, in an optional implementation, N duration thresholds can be configured for the duration goal, where N is a positive integer, so that the prediction goals related to the duration goal are optimized based on the N duration thresholds, so that the corresponding prediction goals not only focus on whether the resource is browsed, but also on the duration of the resource being browsed.

[0088] Specifically, if a duration goal is configured with N duration thresholds, browsing behavior may include browsing operations corresponding to each duration threshold. The browsing operation corresponding to the i-th duration threshold is defined as an operation with a browsing duration greater than the i-th duration threshold, where i∈[1,N]. In this case, the number of prediction targets corresponding to the duration goal is N. If, among multiple business goals, no business goals are dependent on the duration goal, the i-th prediction target corresponding to the duration goal includes the probability that the browsing operation corresponding to the i-th duration threshold will be executed on the resource. If, among multiple business goals, a business goal is dependent on the duration goal, the i-th prediction target corresponding to the duration goal includes the probability that the browsing operation corresponding to the i-th duration threshold will be executed on the resource after the resource feedback behavior indicated by the business goal on which the duration goal depends is executed. For example, if the business goal on which the duration goal depends is a click goal, the i-th prediction target corresponding to the duration goal includes the probability that the browsing operation corresponding to the i-th duration threshold will be executed on the resource after the click behavior is executed on the resource.

[0089] Furthermore, if the browsing behavior includes browsing operations corresponding to N duration thresholds, and a business goal among multiple business goals is dependent on a duration target, then the number of prediction targets corresponding to the corresponding business goals that are dependent on the duration target may also be N, and the i-th prediction target may include: after the browsing operation corresponding to the i-th duration threshold is executed on the resource, the probability of the corresponding resource performing the resource feedback behavior indicated by the corresponding business goal that is dependent on the duration target is executed. For example, taking the business goal that is dependent on the duration target as an interaction target, the i-th prediction target corresponding to the interaction target includes: after the browsing operation corresponding to the i-th duration threshold is executed on the resource, the probability of the corresponding resource performing the interaction behavior is executed.

[0090] Based on the above description, let's take an example: suppose the duration goal is configured with three duration thresholds: 0 seconds, 7 seconds, 26 seconds, and 85 seconds; and suppose the duration goal is dependent on the click goal, and the interaction goal is dependent on the duration goal. In this case, the click goal can correspond to one prediction goal, and both the duration goal and the interaction goal can correspond to four prediction goals. For details, see Table 1 below:

[0091] Table 1

[0092]

[0093] S203 , calling a multi-objective prediction model, performing multi-objective prediction processing on the target resource according to the object information of the target object and the resource information of the target resource, and obtaining a prediction result for each prediction target.

[0094] In a specific implementation, the computer device can generate a resource representation vector of the target resource based on the object information of the target object and the resource information of the target resource. In one embodiment, the resource representation vector of the target resource can be generated by feature extraction; specifically, the computer device can call a feature extraction network to perform feature extraction on the object information of the target object and the resource information of the target resource, respectively, to obtain a first feature vector corresponding to the target object and a second feature vector corresponding to the target resource, and then concatenate the first feature vector and the second feature vector to obtain a resource representation vector of the target resource. In another embodiment, the resource representation vector of the target resource can be generated by feature mapping based on a vector table (Embedding table). The so-called vector table refers to a list consisting of multiple information and a feature vector (Embedding) corresponding to each information. Embedding is a numerical vector composed of multiple floating-point numbers, which can be used to map the original features of any information. After end-to-end learning of the neural network, it describes the various attributes and properties of the corresponding information in high-dimensional space. Specifically, a public vector table (also called a basic vector table, specifically a vector table that can be used to perform feature mapping on any information) can be obtained, and the public vector table can be used to perform feature mapping on the object information of the target object and the resource information of the target resource respectively (that is, to search the corresponding vector table for the feature vector corresponding to the information to be mapped), and the object feature vector and the resource feature vector are obtained. Then, the object feature vector and the resource feature vector are spliced ​​to obtain the resource representation vector of the target resource.

[0095] Furthermore, considering that the target object's object information may include: the target object's attribute information and the target object's target behavior sequence for each business objective, any target behavior sequence includes: the resource identifier of at least one resource; and resource identifiers often overlap with feature vectors in a common vector table. Using a common vector table to perform feature mapping on the target behavior sequence may result in low feature mapping accuracy. Therefore, to avoid using the same feature vector for feature values ​​with significantly different meanings, thereby improving the accuracy of feature mapping results, and to fine-grainedly distinguish between resources that have been subjected to different resource feedback behaviors, additional vector tables can be used for feature mapping of target behavior sequences under different business objectives. In this case, the specific implementation method for generating the resource representation vector of the target resource based on the object information of the target object and the resource information of the target resource can include the following steps: s21, using a public vector table to perform feature mapping on the attribute information of the target object and the resource information of the target resource, respectively, to obtain an object attribute vector and a resource feature vector; s22, using a private vector table configured for each business target (i.e., a vector table used only for feature mapping the behavior sequence under the corresponding business target), respectively, to perform feature mapping on the target behavior sequence of the target object under the corresponding business target, to obtain the behavior feature vector of the target object under each business target. s23, integrating the object attribute vector, resource feature vector, and the behavior feature vector of the target object under each business target to obtain the resource representation vector of the target resource.

[0096] Among them, the specific implementation method of step s23 can be: directly splicing the object attribute vector, resource feature vector and the behavior feature vector of the target object under each business goal to obtain the resource representation vector of the target resource. Alternatively, the behavior sequence representation extraction network can be called to generate the interest representation vector of the target object under each business goal based on the similarity between the behavior feature vector and the resource feature vector of the target object under each business goal; any interest representation vector is used to represent: the interest of the target object in performing the resource feedback operation indicated by the corresponding business goal on the target resource. Then, the object attribute vector, resource feature vector and the interest representation vector of the target object under each business goal are spliced ​​to obtain the resource representation vector of the target resource. In this way, the resource representation vector of the target resource can include the interest representation vector of the target object under each business goal, which can effectively enrich the amount of information of the resource representation vector of the target resource, so that the subsequent multi-target prediction model can more accurately perform multi-target prediction based on the resource representation vector of the target resource, thereby improving the accuracy of the multi-target prediction.

[0097] Optionally, for the pth (p is a positive integer) business goal, the behavior feature vector of the target object under the pth business goal may include: the identification feature vectors of each resource identifier in the target behavior sequence of the target object under the pth business goal; accordingly, the similarity between the behavior feature sequence vector and the resource feature vector may include: the similarity between each identification feature vector and the resource feature vector in the corresponding feature sequence vector. Based on this, the method of generating the interest representation vector of the target object under the pth business goal may be: according to the principle that similarity is positively correlated with weight, according to the similarity between each identification feature vector and the resource feature vector in the behavior feature vector of the target object under the pth business goal, weights are assigned to the corresponding identification feature vectors; based on the weights of each identification feature vector, the corresponding identification feature vectors are weightedly aggregated to obtain the interest representation vector of the target object under the pth business goal. Generating the interest representation vector in this way can solve the problem of interest drift in the target behavior sequence. The so-called interest drift refers to the change of the interest of the object over time.

[0098] After generating the resource representation vector of the target resource in the above manner, the computer device can call the multi-objective prediction model to perform multi-objective prediction on the target resource according to the resource representation vector of the target resource, and obtain the prediction result of each prediction target. Specifically, the multi-objective prediction model may include a prediction network corresponding to each business target, and the prediction target of any prediction network is the prediction target corresponding to the corresponding business target. The computer device can input the resource representation vector of the target resource into the prediction network corresponding to each business target respectively, so as to call the prediction network corresponding to each business target respectively, and perform prediction processing of the corresponding prediction target on the target object according to the resource representation vector of the target resource, and obtain the prediction result of each prediction target. Alternatively, the multi-objective prediction model can be called to perform feature extraction on the resource representation vector of the target resource to obtain the high-order feature vector of the target resource, and then the high-order feature vector of the target resource can be input into the prediction network corresponding to each business target respectively, so as to call the prediction network corresponding to each business target respectively, and perform prediction processing of the corresponding prediction target on the target object according to the high-order feature vector of the target resource, and obtain the prediction result of each prediction target.

[0099] It can be understood that: from the above, any prediction target includes probability, so the prediction processing for any prediction target refers to: predicting the value of the probability in the corresponding prediction target (that is, the probability value); based on this, it can be seen that the prediction result of any prediction target includes: the predicted probability value.

[0100] S204: Based on the prediction results of each prediction target, a recommendation decision process is performed on the target resource.

[0101] In one specific implementation of step S204, the computer device may detect whether the probability values ​​in the prediction results for each prediction target are greater than a probability threshold; if so, a decision may be made to recommend the target resource to the target object; if not, a decision may be made not to recommend the target resource to the target object. Alternatively, based on the relationship between the probability values ​​in the prediction results for each prediction target and the probability threshold, the number of target prediction results may be counted, where the target prediction results are those with probability values ​​greater than the probability threshold. If the number of target prediction results is greater than the threshold, a decision may be made to recommend the target resource to the target object; otherwise, a decision may be made not to recommend the target resource to the target object.

[0102] In another specific implementation of step S204, the probability values ​​in the prediction results of each prediction target can be integrated to obtain a recommendation value for the target resource; if the recommendation value is greater than a preset value, a decision can be made to recommend the target resource to the target object; otherwise, a decision can be made not to recommend the target resource to the target object. Alternatively, the recommendation values ​​of other resources to be recommended to the target object can be obtained, and the recommendation values ​​of other resources can be calculated in the same manner as the recommendation value of the target resource; in descending order of recommendation value, based on the recommendation values ​​of other resources and the recommendation value of the target resource, the other resources and the target resource are sorted to obtain a sorted sequence, and the target number of resources are sequentially selected from the sorted sequence, and the selected resources are recommended to the target object.

[0103] It should be noted that the embodiment of the present application does not limit the specific implementation method of the computer device integrating the probability values ​​in the prediction results of each prediction target to obtain the recommended value of the target resource. For example, the probability values ​​in the prediction results of each prediction target can be directly averaged or weighted summed to obtain the recommended value of the target resource. For example, for any business target, the probability values ​​in the prediction results of each prediction target corresponding to the corresponding business target can be multiplied to obtain the score of the target resource under the corresponding business target; then, the score of the target resource under each business target can be integrated (such as summation or multiplication) to obtain the recommended value of the target resource.

[0104] The embodiment of the present application determines multiple prediction targets of a multi-target prediction model based on the dependency relationship between multiple business targets, so that the prediction target corresponding to any dependent business target includes: the probability that the corresponding resource is subjected to the resource feedback behavior indicated by the business target on which the corresponding business target depends after the resource is subjected to the resource feedback behavior indicated by the business target on which the corresponding business target depends. It can be seen that such a prediction target takes into account the logical relationship that in the actual interaction process, the resource must be subjected to the resource feedback operation indicated by the business target on which any business target depends before the resource feedback operation indicated by the corresponding business target can be executed. It has high rationality, and reasonable prediction targets help to improve the accuracy of multi-target prediction. For the prediction target corresponding to any dependent business target, the multi-target prediction model can make predictions on the corresponding prediction target based on the object information of the target object and the resource information of the target resource. It can consider whether the target resource is executed by the resource feedback behavior indicated by the business target on which any business target depends, and the impact on the probability of the target resource being executed by the resource feedback behavior indicated by any business target, so that the predicted probability (i.e., the prediction result) is more accurate, thereby improving the accuracy of the multi-target prediction, so that the target resource can be recommended and decided based on the more accurate prediction results, thereby improving the accuracy of the resource recommendation decision.

[0105] In practical applications, the computer equipment performs the above Figure 2 Before the method embodiment shown, the multi-objective prediction model can be trained to improve the model performance of the multi-objective prediction model, so that Figure 2 The method embodiment shown can improve the accuracy of the prediction results obtained by multi-objective prediction model, thereby improving the accuracy of resource recommendation decision. Figure 3 The flowchart shown in the figure illustrates the training process of the multi-objective prediction model. Figure 3 As shown, the training process of the multi-objective prediction model may include the following steps S301-S303:

[0106] S301, obtaining training samples and corresponding sample labels of a multi-objective prediction model.

[0107] The multiple prediction targets of the multi-target prediction model are determined based on the dependencies between multiple business targets, with each business target indicating a resource feedback behavior. Specifically, the multiple prediction targets include a prediction target corresponding to each business target, and the prediction target corresponding to any dependent business target includes: the probability that the corresponding resource will perform the resource feedback behavior indicated by the business target after the resource performs the resource feedback behavior indicated by the business target on which the corresponding business target depends.

[0108] The training samples of the multi-objective prediction model may include: object information of the sample object and resource information of the sample resource. Similar to the object information of the target object mentioned above, the object information of the sample object may include at least one of the following: attribute information of the sample object, and the sample behavior sequence of the sample object under each business target; the sample behavior sequence of the sample object under any business target may be: the historical behavior sequence of the sample object under the corresponding business target. It is understandable that the method for obtaining the historical behavior sequence of the sample object under any business target can refer to the relevant description of the historical behavior sequence of the target object under any business target mentioned above, which will not be repeated here. In addition, the sample resource is a resource that has been exposed to the sample object, and similar to the resource information of the sample resource mentioned above, the resource information of the sample resource may include at least one of the following: attribute information of the sample resource, attribute information of the resource entity used to push the sample resource, and statistical information of the cross-relationship, etc.

[0109] The sample label of a training sample indicates whether the sample subject has performed the resource feedback behavior indicated by each business objective on the sample resource. Specifically, the sample label may include the label value of the training sample under each business objective; the label value of the training sample under any business objective indicates whether the sample subject has performed the resource feedback behavior indicated by the corresponding business objective on the sample resource. For example, if the business objective is a click objective, if the label value of the training sample under the click objective is a valid value (e.g., 1), it indicates that the sample subject has performed a click behavior on the sample resource; if the label value of the training sample under the click objective is an invalid value (e.g., 0), it indicates that the sample subject has not performed a click behavior on the sample resource. It is worth noting that if multiple business objectives include a duration objective, and the duration objective is configured with N duration thresholds, the browsing behavior indicated by the duration objective specifically includes the browsing operations corresponding to each duration threshold. In this case, the label value of the training sample under the duration objective can be used not only to indicate whether the training subject has performed a browsing behavior on the sample resource, but also, if the training subject has performed a browsing behavior on the sample resource, to indicate the specific browsing operation performed by the training subject.

[0110] S302 , calling a multi-objective prediction model, performing multi-objective prediction processing on sample resources according to various information in the training samples, and obtaining a multi-objective prediction result corresponding to the sample resources.

[0111] In a specific implementation, when executing step S302, the computer device may generate a resource representation vector of the sample resource based on various information in the training sample. The specific implementation method may refer to the specific implementation method of generating the resource representation vector of the target resource mentioned above, which will not be described in detail here. Then, the computer device may call the multi-objective prediction model to perform multi-objective prediction on the sample resource based on the resource representation vector of the sample resource, and obtain a multi-objective prediction result corresponding to the sample resource. Specifically, the computer device may input the resource representation vector of the sample resource into the prediction network corresponding to each business goal respectively, and call the prediction network corresponding to each business goal respectively to perform prediction processing on the sample object according to the resource representation vector of the sample resource, and obtain a multi-objective prediction result. In this specific implementation, the multi-objective prediction result may include the prediction result of each prediction target.

[0112] In another specific implementation, it is considered that the multi-objective prediction model includes: a prediction network corresponding to each business objective, and the prediction target of any prediction network is the prediction target corresponding to the corresponding business objective; the prediction target corresponding to any dependent business objective includes: after the resource is executed with the resource feedback behavior indicated by the business objective on which the corresponding business objective depends, the probability that the corresponding resource is executed with the resource feedback behavior indicated by the corresponding business objective. For any dependent business objective, if the sample object does not execute the resource feedback behavior indicated by the business objective on which the corresponding business objective depends on the sample resource, then the sample object will certainly not execute the resource feedback behavior indicated by the corresponding business objective on the sample resource. In this case, if the sample resource is still used to train the prediction network corresponding to the corresponding business objective, it is easy to cause the corresponding prediction network to fit the wrong knowledge, thereby affecting the model training effect.

[0113] Based on this, to improve model training effectiveness, when executing step S302, the computer device can filter reference business targets from multiple business targets based on sample labels. If the p-th business target has no dependent business targets, then the p-th business target is determined to be the reference business target. If the p-th business target has dependent business targets, and the sample object has performed a resource feedback behavior on the sample resource indicated by the business target on which the p-th business target depends, then the p-th business target is determined to be the reference business target. For example, if the multiple business targets include a click target and an interaction target, and the interaction target depends on the click target, then the click target can be determined as the reference business target. Furthermore, since the interaction target depends on the click target, and the resource feedback behavior indicated by the click target is a click behavior, if the sample label indicates that the sample object has performed a click behavior on the sample resource, then the interaction target can also be determined as the reference business target. After determining the reference business targets, the prediction network corresponding to each reference business target in the multi-target prediction model can be separately called. Based on the information in the training sample, the sample resource is predicted for the corresponding prediction target to obtain the multi-target prediction result corresponding to the sample resource. Specifically, a resource representation vector for the sample resource can be generated based on various information in the training sample. This resource representation vector is then input into the prediction network corresponding to each reference business objective. The prediction network corresponding to each reference business objective is then called to perform prediction processing on the sample object for the corresponding prediction objective based on the resource representation vector of the sample resource, thereby obtaining a multi-objective prediction result. In this specific implementation, the multi-objective prediction result may include the prediction results for each prediction objective corresponding to each participating business objective.

[0114] S303: Update the multi-objective prediction model based on the multi-objective prediction results corresponding to the sample labels and sample resources.

[0115] In a specific implementation, if the multi-objective prediction results corresponding to the sample resources include: the prediction results of each prediction target; then the specific implementation method of step S303 can be: calling the loss function under each business target respectively, performing loss value calculations based on the sample labels and the prediction results of each prediction target corresponding to the corresponding business target, and obtaining the loss value under each business target; integrating the loss values ​​under each business target to obtain the model loss value; and updating each parameter in the target prediction model in the direction of reducing the model loss value.

[0116] In another specific implementation, if the multi-objective prediction results corresponding to the sample resources include: the prediction results of each prediction target corresponding to each reference business target; then the specific implementation method of step S303 can be: calling the loss function under each reference business target respectively, performing loss value calculation according to the sample label and the prediction results of each prediction target corresponding to the corresponding reference business target, and obtaining the loss value under each reference business target; integrating the loss values ​​under each reference business target to obtain the model loss value generated by the multi-objective model through the training sample; updating the network parameters of each network other than the target network in the target prediction model in the direction of reducing the model loss value; wherein the target network includes: the prediction network corresponding to the business targets other than each reference business target in multiple business targets.

[0117] It should be noted that the embodiments of the present application do not limit the loss function mentioned above; for example, it can specifically be a binary cross entropy loss function. In addition, the above-mentioned "in the direction of reducing the model loss value" refers to: the model optimization direction with the goal of minimizing the model loss value; through this direction, the model optimization is performed so that the model loss value generated again by the training target after each optimization is less than the model loss value generated by the training target before optimization. For example, the model loss value calculated this time is 0.85, then after optimizing the model parameters of the training target in the direction of reducing the model loss value, the model loss value generated by the optimized training target (that is, the intermediate model corresponding to the pth round) should be less than 0.85.

[0118] The embodiment of the present application integrates the modeling of multiple business objectives of the resource recommendation business into a multi-objective prediction model, and designs multiple prediction objectives of the multi-objective prediction model by utilizing the dependency relationship between multiple business objectives, which can improve the rationality of multiple prediction objectives and help improve the model performance of the multi-objective prediction model. In addition, during the training process, by screening the reference business objectives and calling the prediction network corresponding to the reference business objectives for prediction processing, the label sparsity problem of the downstream business objectives can be alleviated, thereby improving the effect of model training and further improving the model performance of the multi-objective prediction model. Furthermore, during the model training process, by introducing the historical behavior sequence of sample resources under each business objective as the sample behavior sequence, the multi-objective prediction model can learn more information during the prediction process, thereby enhancing the prediction accuracy of the multi-objective prediction model; for historical behavior sequences with missing values, a dynamic supplement mechanism for missing values ​​is introduced to alleviate the problem of low feature coverage of some business objectives. The so-called feature coverage refers to the proportion of samples whose feature values ​​are not the default values ​​(or missing values) in the entire sample set; the higher the coverage of the data features, the more complete the information described.

[0119] Based on the above Figure 2 and Figure 3 The relevant description of the method embodiment shown in the figure, the embodiment of the present application proposes a multi-objective prediction model based on click targets, duration targets and interaction targets; this multi-objective prediction model can be applied to the recommendation ranking of information flow content products, such as the recommendation ranking of subscription account message boxes, the recommendation ranking of message cards (including those subscribed by users themselves and those actively recommended by the platform), etc. The optimization goal of this multi-objective prediction model is: when users use information flow products, the comprehensive optimization of the click-through rate, browsing (reading) stay time and interaction rate of the content; the optimization of the first two goals focuses on improving content consumption, and the optimization of the interaction rate goal focuses on improving the user experience.

[0120] The following uses subscription account messages as an example to illustrate the multi-objective prediction model proposed in this embodiment from multiple dimensions, including model input, model structure, prediction target, loss function, and model output.

[0121] (1) Model input.

[0122] During the model training process, each training sample input to the model represents the basic features of a sample object for a subscription account message exposed to it (i.e., the sample resource mentioned above). During the model inference process, each piece of data input to the model represents the basic features of the target object and the subscription account message to be recommended to it. Specifically, the basic features can be abstractly represented as: <attribute information of the object, attribute information of the subscription account to which the subscription account message belongs, attribute information of the subscription account message, statistical information of the cross-relationship, historical click behavior sequence, historical interaction behavior sequence>. For example:

[0123] The object's attribute information includes: object ID (denoted as uin), object age, object gender, object location, number of messages exposed to the object in the past day, number of messages exposed to the object in the past 7 days, etc.

[0124] The attribute information of the subscription account to which the subscription account message belongs may include: the subscription account identifier (denoted as bizuin), the number of subscription account followers, the number of messages created by the subscription account in the past 7 days, the number of clicks and reads of the subscription account in the past 7 days, etc.

[0125] Subscription account message attribute information may include: the number of hours since the subscription account message was sent (i.e., retention time), the number of impressions of the message in the past hour, the number of clicks on the message in the past hour, etc.

[0126] Cross-relationship statistics may include: the number of exposures of the subject to the subscription account in the past 28 days, the number of clicks on the subscription account by the subject in the past 28 days, etc.

[0127] The click history behavior sequence may include: a resource identifier of at least one resource on which the object has performed click behavior within the past period of time (such as 30 days);

[0128] The interactive history behavior sequence may include: the resource identifier of at least one resource on which the object has performed interactive behavior in the past period of time (such as 30 days); it can be subdivided into like sequence, collection sequence, etc. according to the specific operations of the interactive behavior.

[0129] (2)Model structure.

[0130] See for example Figure 4 As shown, the model structure of the multi-objective prediction model proposed in the embodiment of the present application is roughly as follows:

[0131] The input is the Input Features layer. Each piece of data input to this layer contains all the information mentioned in the model input section above.

[0132] Next comes the Embedding Lookup (feature mapping) layer: Basic features such as object attribute information, resource attribute information (such as the attributes of subscription account messages and the attributes of the subscription accounts to which subscription account messages belong), and cross-relationship statistics are mapped using the same Embedding table (i.e., the public vector table) to obtain object attribute vectors and resource feature vectors. However, since the resource identifiers of each resource in the historical behavior sequence often overlap with the feature values ​​in the public vector table, to avoid assigning the same feature vector to feature values ​​with significantly different meanings and to provide a fine-grained distinction between common click items and interactive items, two additional Embedding tables (i.e., Click ID Emb (a private vector table configured for click targets) and Interactive ID Emb (a private vector table configured for interactive targets)) are used for feature mapping for click history sequences and interactive history sequences, respectively, to obtain the corresponding behavior feature vectors. Embedding Lookup is a technique used to process discrete features (such as categorical features and identifiers). Its primary function is to map discrete features into a continuous vector space, enabling neural networks to better process and learn these features. This mapping process is typically achieved by searching a predefined embedding list. Specifically, the Embedding Lookup operation involves the following steps: 1) Initialization: First, an embedding list of size (V, D) is defined, where V represents the number of possible feature values ​​(such as the number of identifiers) and D represents the dimension of the embedding vector. 2) Discrete Feature Encoding: Discrete features are converted to integer encodings, typically by mapping feature values ​​to a continuous integer range (such as 0 to V-1). For example, a movie ID (identifier) ​​might be assigned an integer code, such as 0 for movie 1, 1 for movie 2, and so on. 3) Embedding Vector Search: For each discrete feature value, the corresponding row in the embedding category is searched, resulting in a D-dimensional vector, which is the embedding vector (i.e., feature vector) for that feature value.

[0133] After the Embedding Look Up layer, in the Sequence Modeling (interest representation) layer, the behavior feature vectors corresponding to each historical behavior sequence are processed using the behavior sequence representation extraction network to obtain the interest representation vector of the object preference click (that is, the interest representation vector of the object under the click target) and the interest representation vector of the object preference interaction (that is, the interest representation vector of the object under the interaction target). It is understandable that if the behavior sequences such as likes and favorites are subdivided, then multiple interest representation vectors can be obtained under the interaction target. It is worth noting that since interactive behaviors account for a very small proportion of the entire data, the interactive historical behavior sequences introduced in the embodiment of the present application often have a low coverage rate (<10%). In order to enable the multi-target prediction model to have the ability to model objects that have never generated interactive behaviors, the embodiment of the present application further introduces a dynamic weight module for missing value supplementation, such as Figure 5 As shown. Specifically, the principle of the dynamic weight module for missing value supplementation is as follows: when the object has historical interaction resources (that is, the interaction history behavior sequence is not empty), the behavior feature vector of the historical interaction behavior sequence is passed into the behavior sequence representation extraction network for processing; when the object has no historical interaction resources, the object's historical click resources can be used in combination with popular interaction information for replacement. Specifically, the number of interactions (global interaction number) of each resource indicated by the click history interaction behavior sequence from all objects is stored and dynamically updated in the feature system; then, the resource identifiers of each resource ranked in the top Q in the global interaction number are filtered out from the click history behavior sequence of the corresponding object (such as the sample object or the target object) (when the length of the interaction history behavior sequence is Q), as a replacement sequence for the missing interaction sequence of the corresponding object, so that the behavior feature vector of the replacement sequence is also passed into the behavior sequence representation extraction network for processing.

[0134] Next is the Embedding (vector) splicing layer, which splices the object attribute vector, resource feature vector, and processed interest representation vectors into a tensor to obtain the resource representation vector of the corresponding resource.

[0135] This is followed by the Feature Extraction layer, which performs feature extraction on the resource representation vector of the corresponding resource to obtain the high-order feature vector of the corresponding resource.

[0136] Finally, the output layers utilize multiple parallel prediction networks to predict the corresponding prediction targets based on the high-order feature vectors of the corresponding resources. Each prediction network corresponds to a business objective. The prediction network mentioned here can be an MLP (Multi-Layer Perceptron) network. An MLP is a basic feedforward neural network consisting of at least three layers of neurons: an input layer, a hidden layer, and an output layer. Each layer combines linear and nonlinear transformations. It is also commonly referred to as a Fully Connected Layer (FCN).

[0137] (3) Prediction target.

[0138] Since in information flow products, the ratio of clicks to exposures is usually on the order of one to ten, while the ratio of interactions to clicks is usually on the order of one to one hundred; therefore, if exposure data is directly used as the training set for all targets, the problem of sparse positive sample labels may be faced, which is not conducive to accurate prediction of targets. Based on this, the embodiment of the present application utilizes the dependency relationship between the three business targets of click targets, duration targets, and interaction targets to design the prediction targets of the multi-target prediction model to alleviate the problem of sparse labels for downstream targets. Specifically, let the feature be X, the exposure data set be D_{X}, and the click data set be D_{Y=1, X}. For feature X, the positive click sample is represented by Y=1 and the negative sample is Y=0. For click samples, three time thresholds are divided according to their browsing time (reading time), such as 7 seconds, 26 seconds, and 85 seconds. Whether the i-th time threshold is exceeded is represented by Z=i (i∈{0, 1, 2, 3}), where Z=0 indicates that the browsing time is less than 7 seconds, Z=1 indicates that the browsing time is greater than 7 seconds, Z=2 indicates that the browsing time is greater than 26 seconds, and Z=3 indicates that the browsing time is greater than 85 seconds. For click samples (i.e., training samples belonging to sample resources that have historically been clicked), interactive positive samples are represented by I=1, and negative samples are represented by I=0.

[0139] For the three business goals of click targets, duration targets, and engagement targets, we use exposure data, click data, and click data with different viewing durations as training sample spaces. Based on this, we can derive eight estimated targets. The specific meanings of these eight estimated targets can be found in Table 2 below:

[0140] Table 2

[0141]

[0142] (4) Loss function.

[0143] The multi-objective prediction model uses a model loss function to calculate the model loss value corresponding to the training sample, and updates the model parameters through back-propagation training. Specifically, the model loss function (Loss) can be derived by combining the three hyperparameters α, β, and γ and the binary cross entropy loss function under the three business objectives. It can be specifically shown in the following formula:

[0144]

[0145] Among them, 1(Z=i) indicates whether Z is equal to i, and if so, the value is 1, otherwise it is 0. The optimal value combination of α, β, and γ is selected through experiments.

[0146] (5) Model output.

[0147] Based on the above, it can be seen that the model output may include the prediction results of at least one prediction target. During the model training phase, each prediction result output by the model can be applied to the model loss function to calculate the model loss value, thereby optimizing and updating the model. During the model inference phase, each prediction result output by the model needs to be fused through a formula to obtain the recommended value (F) of the corresponding resource before it can be applied to online sorting. Exemplarily, the fusion formula can be as follows:

[0148]

[0149] Where Π represents the quadrature operation, Ti and Hi are fractional transformation functions containing hyperparameters:

[0150]

[0151]

[0152] The optimal values ​​of a, b, c, and d are obtained through experimental search.

[0153] Based on the above description, the embodiment of the present application can alleviate the label sparsity problem of downstream targets by integrating the modeling of multiple targets such as clicks, browsing time, and interactions (such as likes, comments, collections, etc.) into one model, and using the conditional dependencies between targets to design estimated targets. In addition, the accuracy of the estimated interactive targets can be enhanced by introducing the behavioral sequence features of each interactive behavior. For users who lack historical interactive behaviors, a dynamic weight module for missing value supplementation is introduced to alleviate the problem of low coverage of interactive features.

[0154] Based on the description of the above resource processing method embodiment, the present application embodiment also discloses a resource processing device; the resource processing device can be a computer program (including one or more instructions) running on a computer device, and the resource processing device can execute each step in the above method flow. Figure 6, the resource processing device can run the following units:

[0155] An acquiring unit 601 is configured to acquire object information of a target object and resource information of a target resource to be recommended to the target object;

[0156] The acquisition unit 601 is further configured to acquire a multi-objective prediction model, wherein multiple prediction targets of the multi-objective prediction model are determined based on dependency relationships between multiple business objectives, where each business objective indicates a resource feedback behavior; the multiple prediction targets include a prediction target corresponding to each business objective, and the prediction target corresponding to any dependent business objective includes: a probability that the corresponding resource will execute the resource feedback behavior indicated by the business objective on which the corresponding business objective depends, after the resource executes the resource feedback behavior indicated by the business objective on which the corresponding business objective depends;

[0157] The processing unit 602 is configured to call the multi-objective prediction model, perform multi-objective prediction processing on the target resource according to the object information of the target object and the resource information of the target resource, and obtain a prediction result for each prediction target;

[0158] The processing unit 602 is further configured to perform a recommendation decision process on the target resource based on the prediction result of each prediction target.

[0159] In a specific embodiment, the object information of the target object includes at least: a target behavior sequence of the target object under each business objective; any target behavior sequence includes: a resource identifier of at least one resource; wherein, among the multiple business objectives, a dependent business objective is represented as a first business objective; when the acquisition unit 601 is used to acquire the target behavior sequence of the target object under the first business objective, it can be specifically used to:

[0160] Obtaining a first historical behavior sequence of the target object under the first business objective, where the first historical behavior sequence is a sequence consisting of resource identifiers of at least one resource for which the target object has historically executed a first resource feedback behavior; the first resource feedback behavior is the resource feedback behavior indicated by the first business objective;

[0161] If there are L missing values ​​in the obtained first historical behavior sequence, then obtain a second historical behavior sequence of the target object under the second business goal, where the second business goal refers to the business goal on which the first business goal depends; the second historical behavior sequence refers to a sequence consisting of resource identifiers of at least one resource for which the target object has historically executed a second resource feedback behavior; the second resource feedback behavior refers to the resource feedback behavior indicated by the second business goal; and L is a positive integer.

[0162] Based on the second historical behavior sequence, the first historical behavior sequence is dynamically supplemented with missing values; and the dynamically supplemented first historical behavior sequence is used as the target behavior sequence of the target object under the first business goal.

[0163] In another specific implementation, when the acquisition unit 601 is used to dynamically supplement missing values ​​of the first historical behavior sequence based on the second historical behavior sequence, it can be specifically used to:

[0164] Taking the resources corresponding to the resource identifiers in the second historical behavior sequence as candidate resources; and counting the total number of times the first resource feedback behavior has been executed on each candidate resource in the past;

[0165] Filter out the first L candidate resources from the candidate resources in descending order of the total number of times; and dynamically add the resource identifiers of the first L candidate resources to the first historical behavior sequence.

[0166] In another specific implementation, when the acquisition unit 601 is used to dynamically supplement missing values ​​of the first historical behavior sequence based on the second historical behavior sequence, it can be specifically used to:

[0167] Determine at least one reference object, where the reference object refers to an object that meets the following conditions: no missing values ​​exist in the historical behavior sequence under the first business objective and the historical behavior sequence under the second business objective;

[0168] respectively calculating sequence similarities between the second historical behavior sequence and the historical behavior sequence of each reference object under the second business goal;

[0169] From at least one reference object, a reference object corresponding to the greatest sequence similarity is selected; from the historical behavior sequence of the selected reference object under the first business goal, L resource identifiers are selected and dynamically added to the first historical behavior sequence.

[0170] In another specific embodiment, when the processing unit 602 is used to call the multi-objective prediction model, perform multi-objective prediction processing on the target resource based on the object information of the target object and the resource information of the target resource, and obtain the prediction result of each prediction target, it can be specifically used to:

[0171] Using a common vector table, feature mapping is performed on the attribute information of the target object and the resource information of the target resource to obtain an object attribute vector and a resource feature vector;

[0172] Using the private vector table configured for each business goal, feature mapping is performed on the target behavior sequence of the target object under the corresponding business goal to obtain the behavior feature vector of the target object under each business goal;

[0173] Integrating the object attribute vector, the resource feature vector, and the behavior feature vector of the target object under each business goal to obtain a resource representation vector of the target resource;

[0174] The multi-objective prediction model is called to perform multi-objective prediction on the target resource according to the resource representation vector of the target resource to obtain a prediction result for each prediction target.

[0175] In another specific embodiment, when the processing unit 602 is used to integrate the object attribute vector, the resource feature vector, and the behavior feature vector of the target object under each business goal to obtain the resource representation vector of the target resource, it can be specifically used to:

[0176] Invoking a behavior sequence representation extraction network to generate an interest representation vector for the target object under each business goal based on the similarity between the behavior feature vector of the target object under each business goal and the resource feature vector; each interest representation vector is used to represent: the interest of the target object in performing a resource feedback operation indicated by the corresponding business goal on the target resource;

[0177] The object attribute vector, the resource feature vector, and the interest representation vector of the target object under each business objective are concatenated to obtain a resource representation vector of the target resource.

[0178] In another specific implementation, before obtaining the multi-objective prediction model, the processing unit 602 may further be configured to:

[0179] Obtaining training samples and corresponding sample labels for a multi-objective prediction model, wherein the training samples include: object information of a sample object and resource information of a sample resource, wherein the sample resource is a resource that has been exposed to the sample object; the sample label is used to indicate whether the sample object has performed a resource feedback behavior indicated by each business objective on the sample resource;

[0180] Calling the multi-objective prediction model, performing multi-objective prediction processing on the sample resource according to various information in the training sample, and obtaining a multi-objective prediction result corresponding to the sample resource;

[0181] Based on the multi-objective prediction results corresponding to the sample labels and the sample resources, the multi-objective prediction model is updated.

[0182] In another specific embodiment, the multi-objective prediction model includes: a prediction network corresponding to each business objective, and the prediction objective of any prediction network is the prediction objective corresponding to the corresponding business objective; accordingly, when the processing unit 602 is used to call the multi-objective prediction model and perform multi-objective prediction processing on the sample resource based on various information in the training sample to obtain the multi-objective prediction result corresponding to the sample resource, it can be specifically used to:

[0183] A reference business target is selected from the multiple business targets based on the sample labels; if the p-th business target has no dependent business targets, the p-th business target is determined to be the reference business target; or if the p-th business target has a dependent business target, and the sample object has performed a resource feedback action on the sample resource indicated by the business target on which the p-th business target depends, the p-th business target is determined to be the reference business target, where p is a positive integer;

[0184] The prediction network corresponding to each reference business objective in the multi-objective prediction model is called respectively, and the prediction processing of the corresponding prediction objective is performed on the sample resource according to each information in the training sample to obtain the multi-objective prediction result corresponding to the sample resource.

[0185] In another specific embodiment, the multi-objective prediction results corresponding to the sample resources include: prediction results of each prediction target corresponding to each reference business target; accordingly, when the processing unit 602 is used to update the multi-objective prediction model based on the sample labels and the multi-objective prediction results corresponding to the sample resources, it can be specifically used to:

[0186] Calling the loss function under each reference business objective respectively, performing loss value calculation based on the sample label and the prediction results of each prediction objective corresponding to the corresponding reference business objective, and obtaining the loss value under each reference business objective;

[0187] Integrating the loss values ​​under each reference business objective to obtain a model loss value generated by the multi-objective model through the training sample;

[0188] Update the network parameters of each network other than the target network in the target prediction model in a direction of reducing the model loss value; wherein the target network includes: the prediction network corresponding to the business objectives other than each reference business objective in the multiple business objectives.

[0189] According to another embodiment of the present application, Figure 6The various units in the resource processing device shown can be individually or all combined into one or several other units to form, or one (or some) of the units can be further divided into multiple smaller functional units to form, which can achieve the same operation without affecting the realization of the technical effects of the embodiments of the present application. The above-mentioned units are divided based on logical functions. In actual applications, the functions of one unit can also be implemented by multiple units, or the functions of multiple units can be implemented by one unit. In other embodiments of the present application, other units can also be included based on the resource processing device. In actual applications, these functions can also be implemented with the assistance of other units, and can be implemented by the collaboration of multiple units.

[0190] According to another embodiment of the present application, a computer program (including one or more instructions) capable of executing the steps involved in the above method embodiment can be constructed by running the computer program (including one or more instructions) on a general computing device such as a computer including processing elements and storage elements such as a central processing unit (CPU), a random access memory (RAM), and a read-only memory (ROM). Figure 6 The resource processing device shown in the embodiment of the present application is used to implement the resource processing method of the embodiment of the present application. The computer program can be recorded on a computer-readable storage medium, for example, and loaded into the above-mentioned computing device through the computer-readable storage medium and run therein.

[0191] It is worth noting that, in the embodiments of the present application, the term "module" or "unit" refers to a computer program or a part of a computer program with a predetermined function, and works together with other related parts to achieve a predetermined goal, and can be implemented in whole or in part by using software, hardware (such as processing circuits or memories) or a combination thereof. Similarly, a processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can include a part of the overall module or unit of the module or unit function.

[0192] The embodiment of the present application determines multiple prediction targets of a multi-target prediction model based on the dependency relationship between multiple business targets, so that the prediction target corresponding to any dependent business target includes: the probability that the corresponding resource is subjected to the resource feedback behavior indicated by the business target on which the corresponding business target depends after the resource is subjected to the resource feedback behavior indicated by the business target on which the corresponding business target depends. It can be seen that such a prediction target takes into account the logical relationship that in the actual interaction process, the resource must be subjected to the resource feedback operation indicated by the business target on which any business target depends before the resource feedback operation indicated by the corresponding business target can be executed. It has high rationality, and reasonable prediction targets help to improve the accuracy of multi-target prediction. For the prediction target corresponding to any dependent business target, the multi-target prediction model can make predictions on the corresponding prediction target based on the object information of the target object and the resource information of the target resource. It can consider whether the target resource is executed by the resource feedback behavior indicated by the business target on which any business target depends, and the impact on the probability of the target resource being executed by the resource feedback behavior indicated by any business target, so that the predicted probability (i.e., the prediction result) is more accurate, thereby improving the accuracy of the multi-target prediction, so that the target resource can be recommended and decided based on the more accurate prediction results, thereby improving the accuracy of the resource recommendation decision.

[0193] Based on the description of the above method embodiment and apparatus embodiment, the present application embodiment also provides a computer device. Figure 7 , the computer device at least includes a processor 701, an input interface 702, an output interface 703 and a computer storage medium 704. Among them, the processor 701, input interface 702, output interface 703 and computer storage medium 704 in the computer device can be connected via a bus or other means. The computer storage medium 704 can be stored in the memory of the computer device, and the computer storage medium 704 is used to store a computer program, and the computer program includes one or more instructions. The processor 701 is used to execute one or more instructions in the computer program stored in the computer storage medium 704. The processor 701 (or CPU (Central Processing Unit)) is the computing core and control core of the computer device, which is suitable for implementing one or more instructions, and is specifically suitable for loading and executing one or more instructions to realize the corresponding method flow or corresponding function.

[0194] In one embodiment, the processor 701 described in the embodiment of the present application can be used to perform a series of resource processing on the target object, specifically including: obtaining object information of the target object, and obtaining resource information of the target resource to be recommended to the target object; obtaining a multi-objective prediction model, wherein the multiple prediction targets of the multi-objective prediction model are determined based on the dependency relationship between multiple business targets, and one business target indicates a resource feedback behavior; the multiple prediction targets include a prediction target corresponding to each business target, and the prediction target corresponding to any dependent business target includes: after the resource is executed with the resource feedback behavior indicated by the business target on which the corresponding business target depends, the probability that the corresponding resource is executed with the resource feedback behavior indicated by the corresponding business target; calling the multi-objective prediction model, and performing multi-objective prediction processing on the target resource according to the object information of the target object and the resource information of the target resource to obtain a prediction result for each prediction target; based on the prediction result of each prediction target, performing recommendation decision processing on the target resource, and so on.

[0195] The embodiment of the present application also provides a computer storage medium (Memory), which is a memory device in a computer device for storing computer programs and data. It is understandable that the computer storage medium here can include both built-in storage media in the computer device and, of course, extended storage media supported by the computer device. The computer storage medium provides a storage space, which stores the operating system of the computer device. In addition, a computer program is also stored in the storage space, which includes one or more instructions suitable for being loaded and executed by the processor 701, and these instructions can be one or more program codes. It should be noted that the computer storage medium here can be a high-speed RAM memory, or a non-volatile memory (non-volatile memory), such as at least one disk storage; optionally, it can also be at least one computer storage medium located away from the aforementioned processor.

[0196] In one embodiment, a processor may load and execute one or more instructions stored in a computer storage medium to implement the corresponding steps in the above method embodiment. In a specific implementation, the processor may load and execute the following steps:

[0197] Acquiring object information of a target object and acquiring resource information of a target resource to be recommended to the target object;

[0198] Obtaining a multi-objective prediction model, wherein multiple prediction objectives of the multi-objective prediction model are determined based on dependency relationships between multiple business objectives, where each business objective indicates a resource feedback behavior; the multiple prediction objectives include a prediction objective corresponding to each business objective, and the prediction objective corresponding to any dependent business objective includes: a probability that the corresponding resource will perform the resource feedback behavior indicated by the business objective on which the corresponding business objective depends, after the resource performs the resource feedback behavior indicated by the business objective on which the corresponding business objective depends;

[0199] Calling the multi-objective prediction model, performing multi-objective prediction processing on the target resource according to the object information of the target object and the resource information of the target resource, and obtaining a prediction result for each prediction target;

[0200] Based on the prediction result of each prediction target, a recommendation decision process is performed on the target resource.

[0201] In a specific embodiment, the object information of the target object includes at least: a target behavior sequence of the target object under each business goal; any target behavior sequence includes: a resource identifier of at least one resource; wherein, among the multiple business goals, a dependent business goal is represented as a first business goal; when obtaining the target behavior sequence of the target object under the first business goal, the one or more instructions can be loaded and specifically executed by the processor:

[0202] Obtaining a first historical behavior sequence of the target object under the first business objective, where the first historical behavior sequence is a sequence consisting of resource identifiers of at least one resource for which the target object has historically executed a first resource feedback behavior; the first resource feedback behavior is the resource feedback behavior indicated by the first business objective;

[0203] If there are L missing values ​​in the obtained first historical behavior sequence, then obtain a second historical behavior sequence of the target object under the second business goal, where the second business goal refers to the business goal on which the first business goal depends; the second historical behavior sequence refers to a sequence consisting of resource identifiers of at least one resource for which the target object has historically executed a second resource feedback behavior; the second resource feedback behavior refers to the resource feedback behavior indicated by the second business goal; and L is a positive integer.

[0204] Based on the second historical behavior sequence, the first historical behavior sequence is dynamically supplemented with missing values; and the dynamically supplemented first historical behavior sequence is used as the target behavior sequence of the target object under the first business goal.

[0205] In another specific implementation, when dynamically supplementing missing values ​​of the first historical behavior sequence based on the second historical behavior sequence, the one or more instructions may be loaded and specifically executed by the processor:

[0206] Taking the resources corresponding to the resource identifiers in the second historical behavior sequence as candidate resources; and counting the total number of times the first resource feedback behavior has been executed on each candidate resource in the past;

[0207] Filter out the first L candidate resources from the candidate resources in descending order of the total number of times; and dynamically add the resource identifiers of the first L candidate resources to the first historical behavior sequence.

[0208] In another specific implementation, when dynamically supplementing missing values ​​of the first historical behavior sequence based on the second historical behavior sequence, the one or more instructions may be loaded and specifically executed by the processor:

[0209] Determine at least one reference object, where the reference object refers to an object that meets the following conditions: no missing values ​​exist in the historical behavior sequence under the first business objective and the historical behavior sequence under the second business objective;

[0210] respectively calculating sequence similarities between the second historical behavior sequence and the historical behavior sequence of each reference object under the second business goal;

[0211] From at least one reference object, a reference object corresponding to the greatest sequence similarity is selected; from the historical behavior sequence of the selected reference object under the first business goal, L resource identifiers are selected and dynamically added to the first historical behavior sequence.

[0212] In another specific embodiment, when calling the multi-objective prediction model, performing multi-objective prediction processing on the target resource based on the object information of the target object and the resource information of the target resource, and obtaining a prediction result for each prediction target, the one or more instructions may be loaded and specifically executed by the processor:

[0213] Using a common vector table, feature mapping is performed on the attribute information of the target object and the resource information of the target resource to obtain an object attribute vector and a resource feature vector;

[0214] Using the private vector table configured for each business goal, feature mapping is performed on the target behavior sequence of the target object under the corresponding business goal to obtain the behavior feature vector of the target object under each business goal;

[0215] Integrating the object attribute vector, the resource feature vector, and the behavior feature vector of the target object under each business goal to obtain a resource representation vector of the target resource;

[0216] The multi-objective prediction model is called to perform multi-objective prediction on the target resource according to the resource representation vector of the target resource to obtain a prediction result for each prediction target.

[0217] In another specific embodiment, when the object attribute vector, the resource feature vector, and the behavior feature vector of the target object under each business objective are integrated to obtain the resource representation vector of the target resource, the one or more instructions may be loaded and specifically executed by the processor:

[0218] Invoking a behavior sequence representation extraction network to generate an interest representation vector for the target object under each business goal based on the similarity between the behavior feature vector of the target object under each business goal and the resource feature vector; each interest representation vector is used to represent: the interest of the target object in performing a resource feedback operation indicated by the corresponding business goal on the target resource;

[0219] The object attribute vector, the resource feature vector, and the interest representation vector of the target object under each business objective are concatenated to obtain a resource representation vector of the target resource.

[0220] In another specific implementation, before obtaining the multi-objective prediction model, the one or more instructions may be loaded and specifically executed by the processor:

[0221] Obtaining training samples and corresponding sample labels for a multi-objective prediction model, wherein the training samples include: object information of a sample object and resource information of a sample resource, wherein the sample resource is a resource that has been exposed to the sample object; the sample label is used to indicate whether the sample object has performed a resource feedback behavior indicated by each business objective on the sample resource;

[0222] Calling the multi-objective prediction model, performing multi-objective prediction processing on the sample resource according to various information in the training sample, and obtaining a multi-objective prediction result corresponding to the sample resource;

[0223] Based on the multi-objective prediction results corresponding to the sample labels and the sample resources, the multi-objective prediction model is updated.

[0224] In another specific embodiment, the multi-objective prediction model includes: a prediction network corresponding to each business objective, and the prediction objective of any prediction network is the prediction objective corresponding to the corresponding business objective; accordingly, when calling the multi-objective prediction model and performing multi-objective prediction processing on the sample resource based on various information in the training sample to obtain the multi-objective prediction result corresponding to the sample resource, the one or more instructions can be loaded and specifically executed by the processor:

[0225] A reference business target is selected from the multiple business targets based on the sample labels; if the p-th business target has no dependent business targets, the p-th business target is determined to be the reference business target; or if the p-th business target has a dependent business target, and the sample object has performed a resource feedback action on the sample resource indicated by the business target on which the p-th business target depends, the p-th business target is determined to be the reference business target, where p is a positive integer;

[0226] The prediction network corresponding to each reference business objective in the multi-objective prediction model is called respectively, and the prediction processing of the corresponding prediction objective is performed on the sample resource according to each information in the training sample to obtain the multi-objective prediction result corresponding to the sample resource.

[0227] In another specific embodiment, the multi-objective prediction results corresponding to the sample resources include: prediction results of each prediction target corresponding to each reference business target; accordingly, when the multi-objective prediction model is updated based on the sample labels and the multi-objective prediction results corresponding to the sample resources, the one or more instructions can be loaded and specifically executed by the processor:

[0228] Calling the loss function under each reference business objective respectively, performing loss value calculation based on the sample label and the prediction results of each prediction objective corresponding to the corresponding reference business objective, and obtaining the loss value under each reference business objective;

[0229] Integrating the loss values ​​under each reference business objective to obtain a model loss value generated by the multi-objective model through the training sample;

[0230] Update the network parameters of each network other than the target network in the target prediction model in a direction of reducing the model loss value; wherein the target network includes: the prediction network corresponding to the business objectives other than each reference business objective in the multiple business objectives.

[0231] The embodiment of the present application determines multiple prediction targets of a multi-target prediction model based on the dependency relationship between multiple business targets, so that the prediction target corresponding to any dependent business target includes: the probability that the corresponding resource is subjected to the resource feedback behavior indicated by the business target on which the corresponding business target depends after the resource is subjected to the resource feedback behavior indicated by the business target on which the corresponding business target depends. It can be seen that such a prediction target takes into account the logical relationship that in the actual interaction process, the resource must be subjected to the resource feedback operation indicated by the business target on which any business target depends before the resource feedback operation indicated by the corresponding business target can be executed. It has high rationality, and reasonable prediction targets help to improve the accuracy of multi-target prediction. For the prediction target corresponding to any dependent business target, the multi-target prediction model can make predictions on the corresponding prediction target based on the object information of the target object and the resource information of the target resource. It can consider whether the target resource is executed by the resource feedback behavior indicated by the business target on which any business target depends, and the impact on the probability of the target resource being executed by the resource feedback behavior indicated by any business target, so that the predicted probability (i.e., the prediction result) is more accurate, thereby improving the accuracy of the multi-target prediction, so that the target resource can be recommended and decided based on the more accurate prediction results, thereby improving the accuracy of the resource recommendation decision.

[0232] It should be noted that, according to one aspect of the present application, a computer program product or computer program is also provided, which includes one or more instructions, and the one or more instructions are stored in a computer storage medium. The processor of the computer device reads the one or more instructions from the computer storage medium, and the processor executes the one or more instructions, so that the computer device performs the methods provided in various optional ways in the above-mentioned method embodiments. It should be understood that what is disclosed above is only a preferred embodiment of the present application, and it is certainly not used to limit the scope of the rights of the present application. Therefore, equivalent changes made in accordance with the claims of the present application are still within the scope covered by the present application.

Claims

1. A resource processing method, characterized in that: include: Acquiring object information of a target object and acquiring resource information of a target resource to be recommended to the target object; Acquire a multi-objective prediction model, wherein multiple prediction objectives of the multi-objective prediction model are determined based on dependency relationships between multiple business objectives, and each business objective indicates a resource feedback behavior; The multiple prediction targets include a prediction target corresponding to each business target, and the prediction target corresponding to any dependent business target includes: after the resource performs the resource feedback behavior indicated by the business target on which the corresponding business target depends, the probability that the corresponding resource will perform the resource feedback behavior indicated by the corresponding business target; Calling the multi-objective prediction model, performing multi-objective prediction processing on the target resource according to the object information of the target object and the resource information of the target resource, and obtaining a prediction result for each prediction target; Based on the prediction result of each prediction target, a recommendation decision process is performed on the target resource.

2. The method according to claim 1, wherein The multiple business goals include: a click goal and a duration goal, and the duration goal is dependent on the click goal; the resource feedback behavior indicated by the click goal includes a click behavior, and the resource feedback behavior indicated by the duration goal includes a browsing behavior; The duration target is configured with N duration thresholds, where N is a positive integer; the browsing behavior includes: browsing operations corresponding to each duration threshold, and the browsing operation corresponding to the i-th duration threshold is: an operation with a browsing duration greater than the i-th duration threshold, i∈[1,N]; The number of prediction targets corresponding to the duration target is N, and the i-th prediction target corresponding to the duration target includes: after the resource is clicked, the probability of the corresponding resource being browsed corresponding to the i-th duration threshold.

3. The method according to claim 2, wherein The multiple business goals also include: an interactive goal that depends on the duration goal, and the resource feedback behavior indicated by the interactive goal includes an interactive behavior; The number of prediction targets corresponding to the interaction target is N, and the i-th prediction target corresponding to the interaction target includes: the probability of executing the interaction behavior on the corresponding resource after the browsing operation corresponding to the i-th duration threshold is executed on the resource.

4. The method according to any one of claims 1 to 3, wherein The object information of the target object includes at least: a target behavior sequence of the target object under each business goal; any target behavior sequence includes: a resource identifier of at least one resource; Among the multiple business objectives, a dependent business objective is represented as a first business objective; and a method of obtaining a target behavior sequence of the target object under the first business objective includes: Obtaining a first historical behavior sequence of the target object under the first business objective, where the first historical behavior sequence is a sequence consisting of resource identifiers of at least one resource for which the target object has historically executed a first resource feedback behavior; the first resource feedback behavior is the resource feedback behavior indicated by the first business objective; If there are L missing values ​​in the obtained first historical behavior sequence, then obtain a second historical behavior sequence of the target object under the second business goal, where the second business goal refers to the business goal on which the first business goal depends; the second historical behavior sequence refers to a sequence consisting of resource identifiers of at least one resource for which the target object has historically executed a second resource feedback behavior; the second resource feedback behavior refers to the resource feedback behavior indicated by the second business goal; and L is a positive integer. Based on the second historical behavior sequence, the first historical behavior sequence is dynamically supplemented with missing values; and the dynamically supplemented first historical behavior sequence is used as the target behavior sequence of the target object under the first business goal.

5. The method according to claim 4, wherein The dynamically supplementing missing values ​​of the first historical behavior sequence based on the second historical behavior sequence includes: Taking the resources corresponding to the resource identifiers in the second historical behavior sequence as candidate resources; and counting the total number of times the first resource feedback behavior has been executed on each candidate resource in the past; Filter out the first L candidate resources from the candidate resources in descending order of the total number of times; and dynamically add the resource identifiers of the first L candidate resources to the first historical behavior sequence.

6. The method according to claim 4, wherein The dynamically supplementing missing values ​​of the first historical behavior sequence based on the second historical behavior sequence includes: Determine at least one reference object, where the reference object refers to an object that meets the following conditions: no missing values ​​exist in the historical behavior sequence under the first business objective and the historical behavior sequence under the second business objective; respectively calculating sequence similarities between the second historical behavior sequence and the historical behavior sequence of each reference object under the second business goal; From at least one reference object, a reference object corresponding to the greatest sequence similarity is selected; from the historical behavior sequence of the selected reference object under the first business goal, L resource identifiers are selected and dynamically added to the first historical behavior sequence.

7. The method according to any one of claims 1 to 3, wherein: The object information of the target object includes: attribute information of the target object, and a target behavior sequence of the target object under each business goal; any target behavior sequence includes: a resource identifier of at least one resource; The resource information of the target resource includes at least one of the following: attribute information of the target resource, attribute information of the resource entity used to push the target resource, and statistical information of the cross-relationship; wherein, the statistical information of the cross-relationship includes at least one of the following: the number of exposures generated by the resource entity for the target object within a first historical time period, and the number of times the target object performs corresponding resource feedback behavior on the resource entity within a second historical time period under at least one business goal.

8. The method according to claim 7, wherein The calling of the multi-objective prediction model, performing multi-objective prediction processing on the target resource according to the object information of the target object and the resource information of the target resource, and obtaining a prediction result for each prediction target includes: Using a common vector table, feature mapping is performed on the attribute information of the target object and the resource information of the target resource to obtain an object attribute vector and a resource feature vector; Using the private vector table configured for each business goal, feature mapping is performed on the target behavior sequence of the target object under the corresponding business goal to obtain the behavior feature vector of the target object under each business goal; Integrating the object attribute vector, the resource feature vector, and the behavior feature vector of the target object under each business goal to obtain a resource representation vector of the target resource; The multi-objective prediction model is called to perform multi-objective prediction on the target resource according to the resource representation vector of the target resource to obtain a prediction result for each prediction target.

9. The method according to claim 8, wherein The integrating the object attribute vector, the resource feature vector, and the behavior feature vector of the target object under each business goal to obtain the resource representation vector of the target resource includes: Invoking a behavior sequence representation extraction network to generate an interest representation vector for the target object under each business goal based on the similarity between the behavior feature vector of the target object under each business goal and the resource feature vector; each interest representation vector is used to represent: the interest of the target object in performing a resource feedback operation indicated by the corresponding business goal on the target resource; The object attribute vector, the resource feature vector, and the interest representation vector of the target object under each business objective are concatenated to obtain a resource representation vector of the target resource.

10. The method according to any one of claims 1 to 3, wherein Before obtaining the multi-objective prediction model, the method further includes: Obtaining training samples and corresponding sample labels for a multi-objective prediction model, wherein the training samples include: object information of a sample object and resource information of a sample resource, wherein the sample resource is a resource that has been exposed to the sample object; the sample label is used to indicate whether the sample object has performed a resource feedback behavior indicated by each business objective on the sample resource; Calling the multi-objective prediction model, performing multi-objective prediction processing on the sample resource according to various information in the training sample, and obtaining a multi-objective prediction result corresponding to the sample resource; Based on the multi-objective prediction results corresponding to the sample labels and the sample resources, the multi-objective prediction model is updated.

11. The method according to claim 10, wherein The multi-objective prediction model includes: a prediction network corresponding to each business objective, and the prediction objective of any prediction network is the prediction objective corresponding to the corresponding business objective; calling the multi-objective prediction model, performing multi-objective prediction processing on the sample resource based on various information in the training sample, and obtaining a multi-objective prediction result corresponding to the sample resource, including: A reference business target is selected from the multiple business targets based on the sample labels; if the p-th business target has no dependent business targets, the p-th business target is determined to be the reference business target; if the p-th business target has a dependent business target, and the sample object has performed a resource feedback action on the sample resource indicated by the business target on which the p-th business target depends, the p-th business target is determined to be the reference business target, where p is a positive integer; The prediction network corresponding to each reference business objective in the multi-objective prediction model is called respectively, and the prediction processing of the corresponding prediction objective is performed on the sample resource according to each information in the training sample to obtain the multi-objective prediction result corresponding to the sample resource.

12. The method according to claim 11, wherein The multi-objective prediction results corresponding to the sample resources include: prediction results of each prediction target corresponding to each reference business target; The updating of the multi-objective prediction model based on the multi-objective prediction results corresponding to the sample labels and the sample resources includes: Calling the loss function under each reference business objective respectively, performing loss value calculation based on the sample label and the prediction results of each prediction objective corresponding to the corresponding reference business objective, and obtaining the loss value under each reference business objective; Integrating the loss values ​​under each reference business objective to obtain a model loss value generated by the multi-objective model through the training sample; Update the network parameters of each network other than the target network in the target prediction model in a direction of reducing the model loss value; wherein the target network includes: the prediction network corresponding to the business objectives other than each reference business objective in the multiple business objectives.

13. A resource processing device, characterized in that: include: an acquiring unit, configured to acquire object information of a target object and resource information of a target resource to be recommended to the target object; The acquisition unit is further configured to acquire a multi-objective prediction model, wherein the multiple prediction objectives of the multi-objective prediction model are determined based on the dependency relationship between the multiple business objectives, and each business objective indicates a resource feedback behavior; The multiple prediction targets include a prediction target corresponding to each business target, and the prediction target corresponding to any dependent business target includes: after the resource performs the resource feedback behavior indicated by the business target on which the corresponding business target depends, the probability that the corresponding resource will perform the resource feedback behavior indicated by the corresponding business target; a processing unit, configured to call the multi-objective prediction model, perform multi-objective prediction processing on the target resource according to the object information of the target object and the resource information of the target resource, and obtain a prediction result for each prediction target; The processing unit is further configured to make a recommendation decision on the target resource based on the prediction result of each prediction target.

14. A computer device comprising an input interface and an output interface, characterized in that: Also includes: processors and computer storage media; The processor is adapted to implement one or more instructions, the computer storage medium stores one or more instructions, and the one or more instructions are adapted to be loaded by the processor and executed by the resource processing method according to any one of claims 1 to 12.

15. A computer storage medium, characterized in that The computer storage medium stores one or more instructions, and the one or more instructions are suitable for being loaded by a processor and executed by the resource processing method according to any one of claims 1 to 12.

16. A computer program product, characterized in that The computer program product includes one or more instructions; when the one or more instructions in the computer program are executed by a processor, the resource processing method according to any one of claims 1 to 12 is implemented.

Images