Feature generation method, device and readable storage medium for multimedia resource recommendation

Abstract

The embodiments of the present application disclose a feature generation method, device, and readable storage medium for multimedia resource recommendation. The method includes: obtaining first object behavior data provided by a client data source, and generating at least two first object behavior statistical features associated with the first object in at least two different time periods based on the first object behavior data; obtaining second object behavior data provided by a training sample data source for multimedia resource recommendation, and generating at least two second object behavior statistical features associated with the second object in at least two different time periods based on the second object behavior data; generating object statistical features for each object based on the first object behavior statistical features associated with each first object and the second object behavior statistical features associated with each second object, and determining the object statistical features as input features of a multimedia resource recommendation model. By adopting the present application, the recommendation efficiency of multimedia resources can be improved and the applicability of multimedia resource recommendation can be enhanced.

Description

Technical Field

[0001] The present application relates to the field of Internet technology, and in particular to a feature generation method, device, and readable storage medium for multimedia resource recommendation. Background Art

[0002] As smartphones, tablets, and other devices become increasingly powerful, the functionality of applications (or apps) running on them is also becoming increasingly diverse. People can access information of interest through apps, and this diverse information provides a rich and varied user experience. Today, applications are becoming more user-friendly with technological advancements. To enhance the user experience, major applications are also offering more personalized multimedia resource recommendations.

[0003] Generally speaking, multimedia resource recommendations are based on user preferences for a particular type of multimedia resource, recommending similar multimedia resources to users, or recommending multimedia resources of potential interest to users based on the needs or preferences of user groups with similar interests or shared experiences. However, due to the large user base, the number of multimedia resources of interest to users in historical data can reach hundreds of billions, resulting in low accuracy and poor applicability of user-specific multimedia resource recommendations. Summary of the Invention

[0004] The embodiments of the present application provide a feature generation method, device, and computer-readable storage medium for multimedia resource recommendation, which can improve the recommendation efficiency of multimedia resources and enhance the applicability of multimedia resource recommendation.

[0005] In a first aspect, an embodiment of the present application provides a feature generation method for multimedia resource recommendation, the method comprising:

[0006] Obtaining first object behavior data provided by a client data source, and generating at least two first object behavior statistical features of at least two different time periods associated with the first object based on the first object behavior data, wherein the first objects are each object carried in the first object behavior data, and the first object behavior statistical features of any time period associated with any first object include at least two attribute features of an object attribute feature, a duration attribute feature, a resource content attribute feature, a resource type attribute feature, and an object resource behavior attribute feature;

[0007] Obtaining second object behavior data provided by a training sample data source for multimedia resource recommendation, and generating at least two second object behavior statistical features for at least two different time periods associated with the second object based on the second object behavior data, wherein the second objects are each object carried in the first object behavior data, and at least two object resource behavior attribute features are included in the second object behavior statistical features for any time period associated with any second object;

[0008] Based on the above-mentioned first object behavior statistical features associated with each of the above-mentioned first objects and the above-mentioned second object behavior statistical features associated with each of the above-mentioned second objects, object statistical features of each object are generated, and the above-mentioned object statistical features are determined as input features of a multimedia resource recommendation model, wherein the above-mentioned multimedia resource recommendation model is trained by the above-mentioned second object behavior data, and the above-mentioned multimedia resource model is used to output multimedia resource recommendation values ​​for each of the above-mentioned objects based on the object statistical features of each of the above-mentioned objects.

[0009] In a possible implementation, the generating of at least two first object behavior statistical features associated with the first object in at least two different time periods based on the first object behavior data includes:

[0010] generating at least two levels of object behavior features for at least two different time periods based on the generation time of each object behavior data carried in the first object behavior data, wherein the first level object behavior features are used to generate object behavior statistical features of each first object in a time period, and the generation time of the object behavior data included in different time periods do not overlap;

[0011] Based on the object identifiers of each first object carried in the above-mentioned at least two levels of object behavior features, at least two first object behavior statistical features indexed by the object identifiers of the above-mentioned each first object are generated, wherein any first object is associated with at least two first object behavior statistical features of the above-mentioned at least two different time periods indexed by the object identifier of any of the above-mentioned first objects.

[0012] In a possible implementation, generating at least two levels of object behavior features in at least two different time periods based on the generation time of each object behavior data carried in the first object behavior data includes:

[0013] Based on the generation time of each object behavior data carried in the first object behavior data, classifying the object behavior data of each first object belonging to the same time period into object behavior data of the same level to obtain at least two levels of object behavior data;

[0014] Based on the object behavior data of the above-mentioned first objects included in the object behavior data at all levels, at least two attribute characteristics of the object attribute characteristics, duration attribute characteristics, resource content attribute characteristics, resource type attribute characteristics and object resource behavior attribute characteristics of each first object are generated, and the above-mentioned at least two attribute characteristics associated with the above-mentioned first objects are spliced ​​to generate the object behavior characteristics of each level corresponding to the above-mentioned time periods, so as to obtain at least two levels of object behavior characteristics for at least two different time periods.

[0015] In one possible implementation, the object resource behavior attribute characteristics include one or more of the multimedia resource preference degree of the object or the object preference degree of the multimedia resource; and the generating of the object resource behavior attribute characteristics of each first object based on the object behavior data of each first object included in the object behavior data at each level includes:

[0016] Obtaining any object behavior statistic of any first object for any multimedia resource in the object behavior data of any first object included in the object behavior data of each level, and determining the ratio of the any object behavior statistic to the sum of the object behavior statistic of all objects for the any multimedia resource in the object behavior data of each level as the preference degree of the any first object for the any multimedia resource to obtain the multimedia resource preference degree of the any first object; or

[0017] Obtain any object behavior statistic of any first object in the object behavior data of any first object included in the object behavior data of each level for any multimedia resource, and determine the ratio of any object behavior statistic to the sum of the object behavior statistic of any first object in the object behavior data of each level for all multimedia resources as the preference degree of any first object for any multimedia resource to obtain the object preference degree of any multimedia resource.

[0018] In one possible implementation, the object resource behavior attribute characteristics include an expected number of clicks on a multimedia resource by an object; and generating the object resource behavior attribute characteristics of each first object based on the object behavior data of each first object included in the object behavior data at each level includes:

[0019] Obtaining a first object behavior statistic of any first object on a first multimedia resource in the object behavior data of any first object included in the object behavior data at each level, and obtaining a first object behavior operation ratio of all object users on the first multimedia resource in the object behavior data at each level;

[0020] Obtaining a second object behavior statistic of any first object on a second multimedia resource in the object behavior data of any first object at each level, and obtaining a second object behavior operation ratio of all user objects on the second multimedia resource in the object behavior data at each level;

[0021] Based on the first object behavior statistics, the first object behavior operation ratio, the second object behavior statistics and the second object behavior operation ratio, and the object behavior statistics of any first object for all multimedia resources, the expected number of clicks of any first object on the multimedia resources is generated to obtain the expected number of clicks of each first object on the multimedia resources.

[0022] In a possible implementation, generating at least two second object behavior statistical features of at least two different time periods associated with each object carried in the first object behavior data based on the second object behavior data includes:

[0023] Based on the generation time of each object behavior data carried in the second object behavior data, the second object behavior data is divided into a plurality of P1-level object behavior data, where the P1-level object behavior data is the object behavior data included in the minimum unit time length of the second object behavior data divided by the time interval;

[0024] generating at least two pieces of P2-level object behavior data for at least two different time periods based on accumulation of the plurality of P1-level object behavior data included in the plurality of minimum unit time periods, wherein one piece of P2-level object behavior data is accumulated from the plurality of P1-level object behavior data within a target time period, where the target time period is a positive integer multiple of the minimum unit time period;

[0025] Based on the object identifiers corresponding to each second object carried in each of the above-mentioned P2-level object behavior data, at least two second object behavior statistical features indexed by the object identifiers of the above-mentioned each second object are generated, wherein any second object is associated with at least two second object behavior statistical features of the above-mentioned at least two different time periods indexed by the object identifiers of any of the above-mentioned second objects.

[0026] In one possible implementation, the generating, based on the object identifiers corresponding to the second objects carried in the P2-level object behavior data, of at least two second object behavior statistical features indexed by the object identifiers of the second objects includes:

[0027] generating, based on the object identifiers corresponding to the respective second objects carried in the respective P2-level object behavior data, P2-level object behavior features associated with the respective second objects, wherein the P2-level object behavior features of a second object include at least two object resource behavior attribute features of the second object;

[0028] Using the object identifier of each second object as an index, concatenating at least two object resource behavior attribute features in the P2-level object behavior features associated with the object identifier of each second object to generate a second object behavior statistical feature associated with each object identifier;

[0029] The above-mentioned object resource behavior attribute characteristics include at least one object resource behavior-related characteristic of the object's click duration, number of clicks, click-through rate, number of exposures, viewing time, and interaction time for multimedia resources.

[0030] In a possible implementation, generating the object statistical feature of each object based on the first object behavior statistical feature associated with each first object and the second object behavior statistical feature associated with each second object includes:

[0031] Obtaining object identifiers corresponding to the first objects from the first object behavior statistical features, and obtaining object identifiers corresponding to the second objects from the second object behavior statistical features;

[0032] Using the object identifiers of the above-mentioned first objects and the object identifiers of the above-mentioned second objects as indexes, the above-mentioned first object behavior statistical features and the above-mentioned second object behavior statistical features are spliced ​​to generate object statistical features corresponding to each object identifier, wherein the above-mentioned first object behavior statistical features and the above-mentioned second object behavior statistical features with the same object identifier are spliced ​​into object statistical features of the same object.

[0033] In a second aspect, an embodiment of the present application provides a feature generation device for multimedia resource recommendation, the device comprising:

[0034] An acquisition module, configured to acquire first object behavior data provided by a client data source and acquire the first object behavior data and second object behavior data for the object behavior;

[0035] A first object behavior statistical feature generating module is configured to generate at least two first object behavior statistical features of at least two different time periods associated with the first object based on the first object behavior data acquired by the acquisition module, wherein the first object is each object carried in the first object behavior data, and the first object behavior statistical features of any time period associated with any first object include at least two attribute features among object attribute features, duration attribute features, resource content attribute features, resource type attribute features, and object resource behavior attribute features. The first object behavior data provided by a client data source is acquired, and based on the first object behavior data, at least two first object behavior statistical features of at least two different time periods associated with each object carried in the first object behavior data are generated;

[0036] The acquisition module is further configured to acquire second object behavior data provided by a training sample data source recommended by the multimedia resource;

[0037] The feature generation module for second object behavior statistics is further used to generate at least two second object behavior statistical features of at least two different time periods associated with the second object based on the second object behavior data acquired by the acquisition module, wherein the second object is each object carried in the first object behavior data, and at least two object resource behavior attribute features in the second object behavior statistical features of any time period associated with any second object are used to acquire the second object behavior data provided by the training sample data source for multimedia resource recommendation, and generate at least two second object behavior statistical features of at least two different data segments associated with each object carried in the first object behavior data based on the second object behavior data;

[0038] A feature aggregation module is used to generate object statistical features of each object based on the first object behavior statistical features associated with each first object and the second object behavior statistical features associated with each second object, and determine the object statistical features as input features of a multimedia resource recommendation model, wherein the multimedia resource recommendation model is trained by the second object behavior data, and the multimedia resource model is used to output multimedia resource recommendation values ​​for each object based on the object statistical features of each object. The object statistical features of each object are generated based on the first object behavior statistical features and the second object behavior statistical features associated with each object, and the object statistical features are determined as input features of the multimedia resource recommendation model.

[0039] In a third aspect, an embodiment of the present application provides a computer device, comprising: a processor, a memory, and a network interface;

[0040] The above-mentioned processor is connected to the above-mentioned memory and the above-mentioned network interface, wherein the above-mentioned network interface is used to provide data communication function, the above-mentioned memory is used to store program code, and the above-mentioned processor is used to call the above-mentioned program code to execute the method provided by the above-mentioned first aspect and any possible implementation method thereof.

[0041] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and the computer program includes program instructions. When the processor executes the program instructions, the method provided in the first aspect of the embodiment of the present application and any possible implementation thereof is executed.

[0042] In a fifth aspect, an embodiment of the present application provides a computer program product, which includes a computer program, and the computer program is stored in a computer-readable storage medium. The computer program is suitable for being read and executed by a processor, so that a computer device having the processor performs the method provided in the first aspect of the embodiment of the present application and any possible implementation thereof.

[0043] In an embodiment of the present application, a target client obtains first object behavior data provided by a client data source, and generates at least two levels of object behavior features for at least two different time periods based on the generation time of each object behavior data contained in the first object behavior data. The first objects are each object contained in the first object behavior data, and the first object behavior statistical features associated with any first object in any time period include at least two attribute features selected from the group consisting of object attribute features, duration attribute features, resource content attribute features, resource type attribute features, and object resource behavior attribute features. Next, the target client obtains second object behavior data provided by a training sample data source for multimedia resource recommendation, and generates at least two second object behavior statistical features associated with the second object in at least two different time periods based on the second object behavior data. The second objects are each object contained in the first object behavior data, and at least two object resource behavior attribute features are included in the second object behavior statistical features associated with any second object in any time period. Finally, object statistical features are generated for each object based on the first object behavior statistical features associated with each first object and the second object behavior statistical features associated with each second object, and the object statistical features are determined as input features for a multimedia resource recommendation model. Among them, the above-mentioned multimedia resource recommendation model is obtained by training the above-mentioned second object behavior data. The above-mentioned multimedia resource model is used to output the multimedia resource recommendation value of each object based on the object statistical characteristics of each object, which can provide a more comprehensive, clear and easily expandable statistical feature system, improve the utilization rate of computing resources, model training resources and storage resources for multimedia resource recommendation, improve the recommendation accuracy and efficiency of multimedia resources, and enhance applicability. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only 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.

[0045] Figure 1 This is a schematic diagram of the system architecture provided by the embodiment of the present application;

[0046] Figure 2 This is a flow chart of a feature generation method for media resource recommendation provided in an embodiment of the present application;

[0047] Figure 3 This is a schematic diagram of an application scenario of the feature generation method for multimedia resource recommendation provided in an embodiment of the present application;

[0048] Figure 4Schematic diagram of statistical characteristics of the feature generation method for multimedia resource recommendation provided in an embodiment of the present application;

[0049] Figure 5 This is a schematic diagram of another application scenario of the feature generation method for multimedia resource recommendation provided in an embodiment of the present application;

[0050] Figure 6 This is a schematic diagram of another application scenario of the feature generation method for multimedia resource recommendation provided in an embodiment of the present application;

[0051] Figure 7 This is a schematic diagram of another application scenario of the feature generation method for multimedia resource recommendation provided in an embodiment of the present application;

[0052] Figure 8 This is a schematic diagram of another application scenario of the feature generation method for multimedia resource recommendation provided in an embodiment of the present application;

[0053] Figure 9 This is a schematic diagram of another application scenario of the feature generation method for multimedia resource recommendation provided in an embodiment of the present application;

[0054] Figure 10 This is a schematic diagram of the structure of a feature generation device for multimedia resource recommendation provided in an embodiment of the present application;

[0055] Figure 11 It is a structural diagram of the computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0056] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0057] See Figure 1 , Figure 1 This is a schematic diagram of the system architecture provided by the embodiment of this application. Figure 1As shown, the system architecture may include a business server 100 and a terminal cluster, which may include: terminal devices 200a, 200b, 200c, ..., 200n, etc. Among them, the above-mentioned business server 100 can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers. It can also be a cloud server that provides basic cloud computing services such as cloud databases, cloud services, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The terminal devices (including terminal devices 200a, 200b, 200c, ..., 200n) can be smart terminals such as PDAs, smart phones, laptops, desktop computers, tablets, mobile internet devices (MIDs), wearable devices (such as smart watches, smart bracelets, etc.), smart computers, smart cars, etc. Among them, the business server 100 can establish a communication connection with each terminal device in the terminal cluster, and the terminal devices in the terminal cluster can also establish a communication connection with each other. In other words, the service server 100 can establish a communication connection with each of the terminal devices 200a, 200b, 200c, ..., 200n. For example, a communication connection can be established between the terminal device 200a and the service server 100. A communication connection can be established between the terminal devices 200a and 200b, and a communication connection can also be established between the terminal devices 200a and 200c. The above-mentioned communication connection is not limited to a connection method and can be directly or indirectly connected via a wired communication method or directly or indirectly connected via a wireless communication method. The specific connection method can be determined according to the actual application scenario and is not limited in this application.

[0058] It should be understood that Figure 1 Each terminal device in the terminal cluster shown can be installed with an application client. When the application client runs in each terminal device, it can be respectively connected to the above Figure 1The business servers 100 shown in the figure interact with each other so that the business server 100 can receive business data from each terminal device, or the business server 100 can push business data (such as multimedia resources) to each terminal device. Among them, the above-mentioned application client can be an application client with data information functions such as display text, images and videos, such as news applications, learning applications, social applications, instant messaging applications, live broadcast applications, short video applications, video applications, music applications, shopping applications, novel applications, payment applications, etc., which can be determined according to the actual application scenario requirements and are not limited here. Among them, the application client can be an independent client or an embedded sub-client integrated in a client (such as an instant messaging client, a social client, etc.), which can be determined according to the actual application scenario and are not limited here. For the convenience of description, taking the target client as an example, each operation object can view, click, collect and share multimedia resources in the target application through the terminal device during the process of using the target client through the terminal device. It can be understood that the above-mentioned multimedia resources can be any kind of multimedia data, which can specifically include but are not limited to audio, pictures or videos, etc., which can be determined according to the actual application scenario and are not limited here. The business server 100, as a server for multimedia resource recommendation applications, may be a collection of multiple servers including a backend server corresponding to the application client, a data processing server, and the like. The business server 100 may receive object behavior data from a terminal device (e.g., obtaining first object behavior data provided by a client data source and second object behavior data provided by a training sample data source for multimedia resource recommendation), generate object behavior statistical features associated with each operation object based on the object behavior data of each operation object, determine the object statistical features of each operation object as input features of a multimedia resource recommendation model, and thereby output objects with higher multimedia resource recommendation values ​​for each operation object, wherein the multimedia resource model is used to output multimedia resource recommendation values ​​for each object based on the object statistical features of each object.

[0059] It is understandable that in the specific implementation of this application, related data such as user information is involved. When the embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of relevant data must comply with relevant laws, regulations and standards of relevant countries and regions.

[0060] The feature generation method for multimedia resource recommendation provided in the embodiment of the present application (which may be referred to as the feature generation method or method for the convenience of description) is applicable to multimedia resource recommendation based on an application (such as the above-mentioned target client). It is understandable that the terminal devices to which the above-mentioned feature generation method is applicable include but are not limited to smartphones, computers, tablet computers, personal digital assistants (PDAs), mobile Internet devices (MIDs), and wearable devices. Optionally, the terminal device may also be a server corresponding to the above-mentioned smartphones, computers, tablet computers, PDAs, MIDs, and wearable devices, etc., which may be determined according to the actual application scenario and is not limited here. Correspondingly, the feature generation device for multimedia resource recommendation provided in the embodiment of the present application (or simply a multimedia resource recommendation device) includes but is not limited to smartphones, computers, tablet computers, PDAs, MIDs, and wearable devices. For the convenience of description, the multimedia resource recommendation device and / or terminal device provided in the embodiment of the present application will be described using a smartphone (or simply a mobile phone) as an example.

[0061] It is understandable that the feature generation method provided in the embodiment of the present application can be implemented as follows: Figure 1 The service server 100 shown in FIG. 1 may also be executed by a terminal device (such as Figure 1 The terminal device 200a, terminal device 200b, ..., terminal device 200n shown in the figure can be executed by any one of them), and can also be executed by the terminal device and the service server together. The specific implementation can be determined according to the actual application scenario and is not limited here. For the convenience of subsequent understanding and explanation, the embodiment of the present application can be Figure 1 A terminal device is selected from the terminal device cluster as a target terminal device, for example, the terminal device 200b is selected as the target terminal device.

[0062] It is understandable that statistical features based on object behavior are a very important feature in multimedia resource recommendation and ranking models. However, conventional multimedia resource recommendations are based on the past behavior data of the operating object, such as the click-through rate, viewing time, and number of views of a certain type of multimedia resource by the operating object, or based on the needs or preferences of a group of objects with similar interests or shared experiences to recommend multimedia resources that may be of interest to the operating object. However, due to the large base of operating objects, the number of multimedia resources that the operating objects are interested in in past data can reach hundreds of billions, resulting in low accuracy and poor applicability of multimedia resource recommendations for operating objects. The implementation of this application provides a comprehensive, clear and easily expandable statistical feature system used in a multimedia resource recommendation and ranking model. Based on object behavior, the statistical characteristics of the above object behavior are divided into five attribute characteristics, namely object attribute characteristics, duration attribute characteristics, resource content attribute characteristics, resource type attribute characteristics and object resource behavior attribute characteristics. In this way, multimedia resources can be recommended based on the statistical characteristics of each object behavior (which can be similar multimedia resource recommendations, author recommendations, preference recommendations, etc.) to enhance the recommendation effect of the object usage application for multimedia resources, improve the recommendation accuracy of multimedia resources, and enhance applicability.

[0063] The feature generation method provided in the embodiments of this application can be applied to applications that recommend various types of multimedia resources. These applications include, but are not limited to, multimedia applications, browser applications, game applications, shopping applications, tool applications, social applications, travel applications, and educational applications, all of which have multimedia resource processing capabilities. The same type of application may include multiple applications, without limitation. For example, multimedia applications may include video players, music players, photography applications, photo editing applications, and audio recording applications. Browser applications include, but are not limited to, QQ Browser. Game applications include, but are not limited to, Honor of Kings and QQ Speed. Shopping applications include, but are not limited to, movie ticket purchasing applications, restaurant reservation applications, and daily necessities purchasing applications. Tool applications include, but are not limited to, file editing, email, alarm clocks, calendars, photo albums, settings, and compasses. Social applications include, but are not limited to, WeChat and QQ. Travel applications include, but are not limited to, Railway 12306 and Ctrip. Educational applications include, but are not limited to, WeChat Reading and QQ Reading.

[0064] Further, see Figure 2 , Figure 2 This is a flow chart of the feature generation method for multimedia resource recommendation provided by the embodiment of the present application. For ease of understanding, the embodiment of the present application is described using a terminal device as an example. Figure 2The terminal device 200b in the example is described, and the service server can be the above Figure 1 The business server 100 of the corresponding embodiment. Figure 2 In the feature generation method for multimedia resource recommendation shown in FIG, each step of data processing can be represented by the above Figure 1 The business server 100 in the embodiment is used to execute the Figure 2 As shown, the feature generation method for multimedia resource recommendation may at least include the following steps S101 to S103.

[0065] S101, obtaining first object behavior data provided by a client data source, and generating at least two first object behavior statistical features of at least two different time periods associated with a first object based on the first object behavior data, wherein the first object is each object carried in the first object behavior data.

[0066] In some feasible implementations, the operation object (i.e., the terminal user) registers real-name information through the target client loaded on the terminal device 200b. After the real-name information registration is successful, the operation object can obtain the object identifier uniquely corresponding to the operation object through the above-mentioned target client, and obtain the first object behavior data provided by the client data source through the target client loaded on the terminal device 200b. The first object behavior data provided by the above-mentioned client data source is the original statistical data of the entire terminal generated based on the operation object behavior log. For the convenience of description, the embodiment of the present application takes the application client with the function of processing multimedia resources loaded on the terminal device 200b as an example for explanation, which is referred to as the target client for short. Please refer to Figure 3 , Figure 3 This is a schematic diagram of an application scenario of the feature generation method for multimedia resource recommendation provided by an embodiment of the present application. In an embodiment of the present application, after the operation object obtains the object identifier uniquely corresponding to the above-mentioned operation object through the above-mentioned target client, the above-mentioned object identifier can log in to the above-mentioned target client through the target client loaded on the terminal device 200b to implement the recommendation process for the target multimedia resource. This application refers to the application account for the target client registered in the target client by the operation object corresponding to the terminal device 200b as the object identifier. The object operation status on its operation interface (for example, interface 1) can be detected by the terminal device 200b. When the operation object clicks the icon of the target client on the operation interface of the terminal device 200b, the terminal device can be triggered to start the operation interface of the target client (for example, interface 2). At this time, the terminal device can detect the object operation instruction on its operation interface, and can determine that the application triggered by the object selection is the target client based on the click position of the object operation instruction. At this time, the terminal device can start the operation interface of the target client (for example, interface 2). As Figure 3As shown, the target client's operation interface includes multiple switchable interfaces such as the home page, my account, and extended functions. The operator can click different icons on the target client's operation interface to switch to different functional interfaces of the target client. For example, if the operator clicks the recommendation icon of the target client, it can trigger the terminal device to switch the target client's operation interface to the multimedia resource recommendation interface (such as interface 3). On interface 3, the popularity rankings of multiple multimedia resources can be displayed, and the corresponding multimedia resources can be recommended to the operator.

[0067] like Figure 3 As shown, the interface 3 can display a recommendation list that recommends multiple multimedia resources published on the target client according to a variety of sorting rules, including rankings, soaring lists, and premieres, etc. When the target client detects an object operation instruction on the display area of ​​any recommendation list, the target client switches to the recommendation page selected corresponding to the object operation instruction, and outputs the multimedia resource recommendation list on the recommendation list to the operating object. For example, when the operating object clicks on the operating area corresponding to "My", the target client can switch to the personal recommendation list corresponding to "My" for the multimedia resources recommended for the operation, that is, the process of the terminal device recommending multimedia resources to the operating object can be realized.

[0068] In some feasible implementations, after the operation object completes registration on the target client loaded on the terminal device 200b, it can log in to the target client through the object identifier assigned by the target client that uniquely corresponds to the operation object. After the operation object logs in to the target client, when the terminal device recommends the target multimedia resource to the operation object, it can first obtain the first object behavior data provided by the target client data source, and generate at least two first object behavior statistical features associated with the first object in at least two different time periods based on the first object behavior data. The first object behavior data carries the object behavior data of each operation object, and each operation object is an operation object that has successfully registered in the target client and obtained the object identifier (i.e., application account) for the target client. Please refer to Figure 4 , Figure 4It is a schematic diagram of the statistical characteristics of the feature generation method for multimedia resource recommendation provided in the embodiment of the present application. The feature generation method provided in the embodiment of the present application specifically divides the feature indicators of the above feature generation method into five attribute characteristics, and each attribute characteristic includes at least one dimension attribute. Among them, since the feature indicators of the above feature generation method can be specifically determined according to the actual application scenario, the attribute characteristics of the above feature generation method can also be determined according to the actual situation, and the dimension of each of the above attribute characteristics can also be determined according to the actual application scenario, and the present application does not limit this. In the embodiment of the present application, the first object behavior statistical characteristics generated based on the above first object behavior data include object attribute characteristics, duration attribute characteristics, resource content attribute characteristics, resource type attribute characteristics and object resource behavior attribute characteristics, and each of the above attribute characteristics includes at least one dimension. As Figure 4 The above-mentioned object attribute features include object features, content features, and object content features. Among them, the above-mentioned duration attribute features can be used to divide the above-mentioned first object behavior data into real-time object behavior features, short-term object behavior features, medium-term object behavior features, and long-term object behavior features based on the generation time of each object behavior data carried by the above-mentioned first object behavior data. The generation time of the object behavior data contained in the above-mentioned different time periods does not overlap, and the duration of different time periods can be determined according to the actual application scenario and is not limited here. Among them, the above-mentioned resource content attribute features include first-level resource content attribute features, second-level resource content attribute features, author attribute features, and tag attribute features, wherein the scope of first-level resource content is greater than that of second-level resource content. Among them, the above-mentioned resource type attribute features include graphic and text multimedia resource content attribute features and video multimedia resource content attribute features. Among them, the above-mentioned object resource behavior attribute features include the object's click duration, number of clicks, click-through rate, number of exposures, viewing time, interaction time, the object's preference for multimedia resources, the object's preference for multimedia resources, and the object's expected click volume on multimedia resources.

[0069] Please also see Figure 5 , Figure 5: This is another application scenario diagram of the feature generation method for multimedia resource recommendation provided by an embodiment of the present application. After each operation object successfully logs into the target client by being loaded on the terminal device 200b, the above-mentioned target client will obtain the first object behavior data provided by the target client data source, and the above-mentioned first object behavior data is composed of object behavior data associated with each first object. Wherein, the above-mentioned first object is each operation object carried in the above-mentioned first object behavior data, and the above-mentioned first object behavior data carries at least two or more object behavior data associated with the operation object. Wherein, each of the above-mentioned at least two or more operation objects completes real-name registration through the above-mentioned target client, and obtains an object identifier uniquely corresponding to each of the above-mentioned operation objects through the above-mentioned target client, and the above-mentioned object identifier can log in to the above-mentioned target client through the terminal device 200b. When the above-mentioned operation object successfully logs in to the above-mentioned target client, the above-mentioned target client can record the object behavior data of the above-mentioned operation object, such as the operation object's operations on viewing, collecting, sharing, etc. of multimedia resources. For the convenience of description, the above-mentioned first objects (i.e., each operation object) will be distinguished, and each first object will be named operation object 1, operation object 2, ..., operation object n, where n is a positive integer. It is understandable that the target multimedia resource recommended on the above-mentioned target client can be any kind of multimedia data, which may specifically include but is not limited to audio, pictures or videos, etc., which can be determined according to the actual application scenario and is not limited here. For the convenience of description, the embodiment of the present application will illustrate the type of target multimedia resource using video resources. When the terminal device receives the above-mentioned first object behavior data, based on the generation time of each object behavior data carried in the above-mentioned first object behavior data, the object behavior data of each operation object belonging to the same time period is divided into object behavior data of the same level to obtain at least two levels of object behavior data. The division of the above-mentioned first object behavior data based on the generation time of each object behavior data can be divided according to the actual application scenario, which is not limited here, that is, the above-mentioned first object behavior data can be divided into hourly object behavior data, daily object behavior data, weekly object behavior data, or the above-mentioned first object behavior data can be divided into monthly object behavior data, etc., which can be determined according to the actual application scenario and is not limited here. For the convenience of description here, the above-mentioned first object behavior data is divided into daily object behavior data and weekly object behavior data. As Figure 5As shown, based on the type of target multimedia resource (i.e., the above-mentioned video resource), the above-mentioned daily object behavior data can be divided into daily graphic object behavior data and daily video object behavior data, and the above-mentioned weekly object behavior data can be divided into weekly graphic object behavior data and weekly video object behavior data. The above-mentioned graphic is the graphic resource corresponding to the above-mentioned video resource. Based on the generation time of each object behavior data carried by the above-mentioned daily graphic object behavior data, the above-mentioned daily graphic object behavior data can be divided into 1-day graphic object behavior data, 2-day graphic object behavior data, 4-day graphic object behavior data, and 7-day graphic object behavior data. Based on the generation time of each object behavior data carried by the above-mentioned daily video object behavior data, the above-mentioned daily video object behavior data can be divided into 1-day video object behavior data, 2-day video object behavior data, 4-day video object behavior data, and 7-day video object behavior data. It is understandable that the generation time of the object behavior data contained in the above-mentioned different time periods does not overlap, and the division of the above-mentioned different time periods can be determined according to the actual application scenario, and is not limited here. That is, the above-mentioned daily object behavior data can also be divided into 3-day object behavior data, 5-day object behavior data, and 8-day object behavior data, etc. according to the actual application scenario. The specific division can be determined according to the actual application scenario, and is not limited here. Among them, for the convenience of description, the above-mentioned weekly graphic object behavior data can be divided into 12 weeks of graphic object behavior data based on the generation time of each object behavior data carried by the above-mentioned weekly graphic object behavior data. Among them, the above-mentioned weekly object behavior data is the object behavior data of the current operation object in the last 12 weeks, including 12 weeks of graphic object behavior data and 12 weeks of video object behavior data. For example, if the current time is xx / xx / xx, then the above-mentioned object behavior data within 12 weeks is the object behavior data of the operation object in the 12 weeks before xx / xx / xx. It can be understood that the above-mentioned division of weekly object behavior data based on the generation time of each object behavior data can also be determined according to the actual application scenario, and there is no restriction here, that is, the above-mentioned weekly object behavior data can also be divided into 2-week object behavior data and 4-week object behavior data according to the actual application scenario.

[0070] It is understandable that the above-mentioned operation objects are the operation objects that have successfully registered in the above-mentioned target client and obtained the object identification (i.e., application account) for the target client. It should be understood that since the time at which each operation object performs the operation on the multimedia resource after completing the real-name registration on the above-mentioned target client is different, the generation time of the object behavior data of the above-mentioned operation objects is also different, and the number of the above-mentioned operation objects is also different. The number of operation objects and the generation time of the object behavior data of each operation object can be determined according to the actual application scenario and are not limited here. For details, please refer to Figure 5, each of the above-mentioned operation objects carries at least one of the above-mentioned graphic object behavior data and the above-mentioned video object behavior data. For the convenience of description here, it is assumed that each operation object carries graphic object behavior data and video object behavior data. For example, it is assumed that the above-mentioned first object behavior data carries object behavior data of 4 different operation objects, and the above-mentioned 4 different operation objects are named operation object 1, operation object 2, operation object 3 and operation object 4 respectively. Assuming that the generation time of the object behavior data of the above-mentioned operation objects has been determined, for the convenience of description, the above-mentioned operation object 1 carries 14 days of object behavior data, the above-mentioned operation object 2 carries 2 days of object behavior data, the above-mentioned operation object 3 carries 4 days of object behavior data, and the above-mentioned operation object 4 carries 7 days of object behavior data. Based on the generation time of the graphic object behavior data and video object behavior data of the above-mentioned operation objects, the above-mentioned target client divides the graphic object behavior data and video object behavior data of each operation object belonging to the same time period into object behavior data of the same level. Specifically, based on the generation time of the graphic and text object behavior data carried by the operation object 1, the graphic and text object behavior data carried by the operation object 2, the graphic and text object behavior data carried by the operation object 3, and the graphic and text object behavior data carried by the operation object 4, the graphic and text object behavior data carried by the operation object 1, the operation object 2, the operation object 3, and the operation object 4 are divided into one day of graphic and text-level object behavior data to generate the one-day graphic and text-level object behavior data. Based on the generation time of the graphic and text object behavior data carried by the operation object 1, the graphic and text object behavior data carried by the operation object 3, and the graphic and text object behavior data carried by the operation object 4, the graphic and text object behavior data carried by the operation object 1, the operation object 3, and the operation object 4 are divided into two days of graphic and text-level object behavior data to generate the two-day graphic and text-level object behavior data. Based on the generation time of the graphic object behavior data carried by the aforementioned operation object 1 and the graphic object behavior data carried by the aforementioned operation object 4, the graphic object behavior data carried by the aforementioned operation object 1 and the aforementioned operation object 4 are divided into four days of graphic object behavior data to generate the aforementioned four days of graphic object behavior data. Based on the generation time of the graphic object behavior data carried by the aforementioned operation object 1, the graphic object behavior data carried by the aforementioned operation object 1 are divided into seven days of graphic object behavior data to generate the aforementioned seven days of graphic object behavior data. Similarly, based on the generation time of the video object behavior data carried by the aforementioned operation object 1, the video object behavior data carried by the aforementioned operation object 2, the video object behavior data carried by the aforementioned operation object 3, and the video object behavior data carried by the aforementioned operation object 4, the video object behavior data carried by the aforementioned operation objects 1, 2, 3, and 4 are divided into one day of video object behavior data to generate the aforementioned one day of video object behavior data.Based on the generation time of the video object behavior data carried by the operation object 1, the video object behavior data carried by the operation object 3, and the video object behavior data carried by the operation object 4, the video object behavior data carried by the operation object 1, the operation object 3, and the operation object 4 are divided into two days of video-level object behavior data to generate the two-day video-level object behavior data. Based on the generation time of the video object behavior data carried by the operation object 1 and the video object behavior data carried by the operation object 4, the video object behavior data carried by the operation object 1 and the operation object 4 are divided into four days of video-level object behavior data to generate the four-day video-level object behavior data. Based on the generation time of the video object behavior data carried by the operation object 1, the video object behavior data carried by the operation object 1 is divided into seven days of video-level object behavior data to generate the seven-day video-level object behavior data.

[0071] For ease of understanding, the following will use the above-mentioned daily graphic object behavior data to generate the daily graphic object behavior features as an example for illustration. The process of generating daily video object behavior features based on the above-mentioned daily video object behavior data is similar to the above-mentioned process of generating daily graphic object behavior features, and will not be elaborated on here. It is understandable that based on the daily graphic object behavior data of each of the above-mentioned operation objects included in the above-mentioned daily graphic object behavior data, at least two attribute features of the object attribute features, duration attribute features, resource content attribute features, resource type attribute features, and object resource behavior attribute features of each operation object can be generated. Among them, the number of dimensions of the object statistical features generated based on the object behavior data is the product of the number of dimensions of each of the above-mentioned five attribute features. It is understandable that the above-mentioned five attribute features may include multiple dimensions, and the total number of dimensions of the statistical features is the product of the number of dimensions of the five attribute features. Among them, the dimensions of the above-mentioned five attribute features included in the above-mentioned object statistical features are all one-dimensional, and ultimately an object statistical feature with a dimension of 1 (i.e., 1×1×1×1×1) can be generated. When the target client generates daily-level graphic and text object behavior data, the target client generates object attribute features, duration attribute features, resource content attribute features, resource type attribute features, and object resource behavior attribute features associated with each operation object based on the daily-level graphic and text object behavior data carried by the target client. Among them, the object attribute features may include object dimensions, the duration attribute features may include long-term object behavior features, the resource content attribute features may include secondary resource content attribute features, the resource type attribute features may include video multimedia resource attribute features, and the object resource behavior attribute features may include one of the following: the object's preference for multimedia resources, the object's preference for multimedia resources, and the object's expected number of clicks on multimedia resources. Among them, when generating the above-mentioned multimedia resource preference degree, obtain any day-level graphic object behavior statistical value of any of the above-mentioned operation objects included in the day-level graphic object behavior data for any multimedia resource, and determine the ratio of the above-mentioned any day-level graphic object behavior statistical value to the sum of the day-level graphic object behavior statistical values ​​of all operation objects in the above-mentioned day-level graphic object behavior data for any of the above-mentioned multimedia resources as the preference degree of any of the above-mentioned operation objects for any of the above-mentioned multimedia resources, so as to obtain the multimedia resource preference degree of any of the above-mentioned operation objects.When generating the object preference degree of any of the above-mentioned multimedia resources, obtain any object behavior statistic value of any of the above-mentioned operation objects included in the object behavior data of any of the above-mentioned operation objects for any multimedia resource, and determine the ratio of any of the above-mentioned object behavior statistic value to the sum of the object behavior statistic values ​​of any of the above-mentioned operation objects for all multimedia resources in the above-mentioned daily image and text object behavior data as the preference degree of any of the above-mentioned operation objects for any of the above-mentioned multimedia resources, so as to obtain the object preference degree of any of the above-mentioned multimedia resources. When generating the expected number of clicks of the operation object on the multimedia resource, obtain the first-level graphic object behavior statistics of any of the above-mentioned operation objects included in the daily-level graphic object behavior data on the first multimedia resource, and obtain the first-level graphic object behavior operation ratio of all operation objects on the first multimedia resource in the above-mentioned daily-level graphic object behavior data; obtain the second-level graphic object behavior statistics of any of the above-mentioned operation objects on the second multimedia resource in the daily-level graphic object behavior data, and obtain the second-level graphic object behavior operation ratio of all operation objects on the second multimedia resource in the above-mentioned daily-level graphic object behavior data; generate the expected number of clicks of any of the above-mentioned operation objects on the multimedia resource based on the above-mentioned first-level graphic object behavior statistics, the above-mentioned first-level graphic object behavior operation ratio, the above-mentioned second-level graphic object behavior statistics, the above-mentioned second-level graphic object behavior operation ratio, and the daily-level graphic object behavior statistics of any of the above-mentioned operation objects on all multimedia resources.

[0072] For example, suppose that the above-mentioned operation object 1 displays the above-mentioned multimedia resource 10 times in the first position and 20 times in the third position, and the click-through rate of the above-mentioned multimedia resource in the first position among all the above-mentioned operation objects (i.e., the above-mentioned operation object 1, the above-mentioned operation object 2, the above-mentioned operation object 3 and the above-mentioned operation object 4) is 0.2, and the click-through rate in the third position is 0.1. Assume that the multimedia resource is a set of nine-grid pictures, that is, the operation object 1 clicks on the first picture in the nine-grid 10 times and clicks on the third picture 20 times, and the click rates of all the operation objects on the first picture and the third picture are 0.2 and 0.1 respectively, then the expected number of clicks on the multimedia resource by the operation object 1 is 4 times, that is, ec_click = 10*0.2 + 20*0.1 = 4. If the operation object 1 clicks on the nine-grid pictures a total of 6 times, then the click-over expected click (COEC) is 1.5 (that is, COEC = 6 / 4 = 1.5).

[0073] For example, assuming that the one-dimensional statistical feature is the number of times each of the above-mentioned operation objects has been exposed to the secondary category content of the target video in the long term, that is, a one-dimensional statistical feature is generated based on the first object behavior data of each of the above-mentioned operation objects. In this case, the above-mentioned object attribute feature includes the object, the above-mentioned duration attribute feature includes the long term, the above-mentioned resource content attribute feature includes the secondary category, and the above-mentioned resource type attribute feature includes the video. When the above-mentioned operation object selects the number of exposures, the object's expected click volume on the multimedia resource in the above-mentioned object resource behavior data feature can be generated. The following will be described in detail using the above-mentioned one-day graphic-text object behavior data as an example. Based on the graphic-text object behavior data of operation object 1 carried in the above-mentioned one-day graphic-text object behavior data, the object features, long-term object behavior features, secondary category resource content attribute features, video multimedia resource attribute features, and the object's expected click volume on the multimedia resource associated with the above-mentioned operation object 1 are generated. The five attribute features associated with the above-mentioned operation object 1 are spliced ​​based on the object identifier corresponding to the above-mentioned operation object 1 to generate the one-day graphic-text object behavior features associated with the above-mentioned operation object 1. Specifically, when generating the object resource behavior data features of the above-mentioned operation object 1, obtain the first-day-level graphic and text object behavior statistics of the above-mentioned operation object 1 in the day-level graphic and text object behavior data of the above-mentioned operation object 1 included in the above-mentioned each day-level graphic and text object behavior data (that is, the above-mentioned 1-day-level graphic and text object behavior data, the above-mentioned 2-day-level graphic and text object behavior data, the above-mentioned 4-day-level graphic and text object behavior data, and the above-mentioned 7-day-level graphic and text object behavior data); obtain the first-day-level graphic and text object behavior operation ratio of all the above-mentioned operation objects (that is, the above-mentioned operation object 1, the above-mentioned operation object 2, the above-mentioned operation object 3, and the above-mentioned operation object 4) in the above-mentioned 1-day-level graphic and text object behavior data to the above-mentioned first multimedia resource; obtain the first-day-level graphic and text object behavior operation ratio of the above-mentioned operation object 1 in the above-mentioned each day-level graphic and text object behavior data (that is, the above-mentioned 1-day-level graphic and text object behavior data, the above-mentioned 2-day-level graphic and text object behavior data, the above-mentioned 4-day-level graphic and text object behavior data, and the above-mentioned 7-day-level graphic and text object behavior data). The second-level graphic object behavior statistics of the above-mentioned operation object 1 on the second multimedia resource in the daily graphic object behavior data are obtained, and the second-level graphic object behavior operation ratio of all operation objects in the above-mentioned daily graphic object behavior data (i.e., the above-mentioned operation object 1, the above-mentioned operation object 2, the above-mentioned operation object 3, and the above-mentioned operation object 4) on the above-mentioned second multimedia resource is obtained; based on the above-mentioned first-level graphic object behavior statistics of the above-mentioned operation object 1, the above-mentioned first-level graphic object behavior operation ratio, the above-mentioned second-level graphic object behavior statistics, the above-mentioned second-level graphic object behavior operation ratio, and the daily graphic object behavior statistics of the above-mentioned operation object 1 on all multimedia resources, the expected click volume of the above-mentioned operation object 1 on the daily graphic object side multimedia resource is generated. Similarly, based on the above-mentioned operations, the expected click volume of the daily graphic object side multimedia resource corresponding to the above-mentioned operation object 2, the above-mentioned operation object 3, and the above-mentioned operation object 4 is generated.After the target client generates five attribute features associated with operation object 1 based on the one-day-level graphic-text object behavior data of operation object 1, the five attribute features associated with operation object 1 are concatenated based on the object identifier of operation object 1 to generate a one-day graphic-text object behavior feature associated with operation object 1. Similarly, based on the one-day-level graphic-text object data of the other operation objects (i.e., operation object 2, operation object 3, and operation object 4) carried in the one-day-level graphic-text object behavior data, one-day graphic-text object behavior features associated with operation object 2, operation object 3, and operation object 4, respectively, are generated.

[0074] Similarly, based on the 2-day graphic-text object behavior data of operation object 1 carried in the 2-day graphic-text object behavior data, object characteristics, long-term object behavior characteristics, secondary resource content attribute characteristics, video multimedia resource attribute characteristics, and the object's expected number of clicks on multimedia resources associated with operation object 1 are generated. The five attribute characteristics associated with operation object 1 are then concatenated based on the object identifier corresponding to operation object 1 to generate a 2-day graphic-text object behavior characteristic associated with operation object 1. Similarly, based on the 2-day graphic-text object behavior data, 2-day graphic-text object behavior characteristics associated with operation objects 3 and 4 are respectively generated. Based on the graphic-text object behavior data of operation object 1 carried in the 4-day graphic-text object behavior data, object characteristics, long-term object behavior characteristics, secondary resource content attribute characteristics, video multimedia resource attribute characteristics, and the object's expected number of clicks on multimedia resources associated with operation object 1 are generated. The five attribute characteristics associated with operation object 1 are then concatenated based on the object identifier corresponding to operation object 1 to generate a 4-day graphic-text object behavior characteristic associated with operation object 1. Similarly, based on the aforementioned 4-day graphic-text-level object behavior data, a 4-day graphic-text-level object behavior feature associated with the aforementioned operation object 4 is generated. Based on the graphic-text object behavior data of operation object 1 carried in the aforementioned 7-day graphic-text-level object behavior data, object features, long-term object behavior features, secondary resource content attribute features, video multimedia resource attribute features, and the object's expected number of clicks on multimedia resources associated with the aforementioned operation object 1 are generated. The five attribute features associated with the aforementioned operation object 1 are concatenated based on the object identifier corresponding to the aforementioned operation object 1 to generate a 7-day graphic-text-level object behavior feature associated with the aforementioned operation object 1.

[0075] In some feasible implementations, when the target client generates daily-level graphic-text object behavior features associated with each of the above-mentioned operation objects, the daily-level graphic-text object behavior features of each operation object are spliced ​​based on the object identifiers corresponding to the above-mentioned operation objects to generate the first object graphic-text behavior statistical features associated with each of the above-mentioned operation objects. Specifically, when the target client generates the 1-day graphic-text object behavior features, the 2-day graphic-text object behavior features, the 4-day graphic-text object behavior features, and the 7-day graphic-text object behavior features associated with the above-mentioned operation object 1, the 1-day graphic-text object behavior features, the 2-day graphic-text object behavior features, the 4-day graphic-text object behavior features, and the 7-day graphic-text object behavior features are spliced ​​based on the object identifiers corresponding to the above-mentioned operation object 1 to generate the daily-level graphic-text object behavior features associated with the above-mentioned operation object 1. Similarly, daily-level graphic-text object behavior features associated with the above-mentioned operation object 2, the above-mentioned operation object 3, and the above-mentioned operation object 4 are generated respectively. Similarly, the implementation process of generating the daily video object behavior features respectively associated with the above-mentioned operation objects 1, the above-mentioned operation object 2, the above-mentioned operation object 3 and the above-mentioned operation object 4 based on the above-mentioned operation objects carried by the above-mentioned daily graphic and text object behavior data is similar to the above-mentioned daily graphic and text object behavior features, and will not be elaborated here.

[0076] Furthermore, based on the object identifiers of the operation objects carried in the daily graphic object behavior features associated with the above-mentioned operation objects and the daily video object behavior features associated with the above-mentioned operation objects, daily object behavior statistical features are generated with the object identifiers of the above-mentioned operation objects as indexes. Specifically, the daily graphic object behavior features generated by the above-mentioned target client include the daily graphic object behavior features associated with the above-mentioned operation object 1, the daily graphic object behavior features associated with the above-mentioned operation object 2, the daily graphic object behavior features associated with the above-mentioned operation object 3, and the daily graphic object behavior features associated with the above-mentioned operation object 4; the daily video object behavior features generated by the above-mentioned target client include the daily video object behavior features associated with the above-mentioned operation object 1, the daily video object behavior features associated with the above-mentioned operation object 2, the daily video object behavior features associated with the above-mentioned operation object 3, and the daily video object behavior features associated with the above-mentioned operation object 4; based on the daily graphic object behavior features and daily video object behavior features associated with the above-mentioned operation object 1, daily object behavior statistical features are generated with the object identifier of the above-mentioned operation object 1 as indexes. Similarly, based on the daily graphic object behavior characteristics and daily video object behavior characteristics associated with the above-mentioned operation object 2, the above-mentioned operation object 3, and the above-mentioned operation object 4, daily object behavior statistical characteristics are generated with the object identification of the above-mentioned operation objects (i.e., operation object 2, operation object 3, and operation object 4) as the index.

[0077] On the other hand, see Figure 5 , based on the generation time of the object behavior data of each operation object (i.e., the above-mentioned operation object 1, the above-mentioned operation object 2, the above-mentioned operation object 3, and the above-mentioned operation object 4) carried by the above-mentioned first object behavior data, the above-mentioned first object behavior data is divided into daily object behavior data and weekly object behavior data, and based on the type of the target multimedia resource (i.e., the above-mentioned video resource), the above-mentioned weekly object behavior data can be divided into weekly graphic object behavior data and weekly video object behavior data. Among them, the above-mentioned graphic is the graphic resource corresponding to the above-mentioned video resource. It can be understood that the division of the above-mentioned first object behavior data based on the generation time of each object behavior data can be divided according to the actual application scenario, and is not limited here. Among them, based on the generation time of each object behavior data carried by the above-mentioned weekly graphic object behavior data, the above-mentioned weekly graphic object behavior data can be divided into 12 weeks of graphic object behavior data. Similarly, based on the generation time of each object behavior data carried by the above-mentioned weekly video object behavior data, the above-mentioned weekly video object behavior data can be divided into 12 weeks of video object behavior data. It is understandable that the above-mentioned division of weekly object behavior data based on the generation time of each object behavior data can also be determined according to the actual application scenario, which is not limited here. Based on the weekly object behavior data of each operation object included in the above-mentioned weekly object behavior data, at least two attribute features of the object attribute characteristics, duration attribute characteristics, resource content attribute characteristics, resource type attribute characteristics and object resource behavior attribute characteristics of each operation object are generated. Among them, the total number of dimensions of the statistical features is the product of the number of dimensions of each of the above five attribute characteristics. When the above-mentioned target client generates the above-mentioned weekly graphic object behavior data and the above-mentioned weekly video object behavior data, 12 weeks of graphic object behavior characteristics and 12 weeks of video object behavior characteristics are generated based on the above-mentioned weekly graphic object behavior data and the above-mentioned weekly video object behavior data, wherein the above-mentioned 12 weeks of graphic object behavior characteristics include the above-mentioned object attribute characteristics, the above-mentioned duration attribute characteristics, the above-mentioned resource content attribute characteristics, the above-mentioned resource type attribute characteristics and the above-mentioned object resource behavior attribute characteristics. Among them, the above-mentioned object attribute characteristics include object dimensions, the above-mentioned duration attribute characteristics include long-term object behavior characteristics, the above-mentioned resource content attribute characteristics include secondary resource content attribute characteristics, and the above-mentioned resource type attribute characteristics include video multimedia resource attribute characteristics. When the above-mentioned operation object selects the number of exposures, the expected number of clicks on the multimedia resource by the object in the above-mentioned object resource behavior data characteristics can be generated. Below, the process of generating 12 weeks of graphic and text object behavior characteristics based on the above-mentioned 12 weeks of graphic and text object behavior data will be introduced in detail. The process of generating 12 weeks of video object behavior characteristics based on the above-mentioned 12 weeks of video object behavior data is similar to the above-mentioned process of generating 12 weeks of graphic and text object behavior characteristics, and will not be elaborated on here.

[0078] As mentioned above, the 12-week graphic and text object behavior data carries the 12-week graphic and text object behavior data for operation object 1, operation object 2, operation object 3, and operation object 4. Based on the 12-week graphic and text object behavior data for operation object 1 carried in the 12-week graphic and text object behavior data, object characteristics, long-term object behavior characteristics, secondary resource content attribute characteristics, video multimedia resource attribute characteristics, and the object's expected number of clicks on multimedia resources associated with operation object 1 are generated. The five attribute characteristics associated with operation object 1 are then concatenated based on the object identifier corresponding to operation object 1 to generate the 12-week graphic and text-level object behavior characteristics associated with operation object 1. Wherein, when generating the expected number of clicks of the object of the above-mentioned operation object 1 on the multimedia resource, the first 12 weeks of graphic and text object behavior statistics of the above-mentioned operation object 1 in the 12 weeks of graphic and text object behavior data on the first multimedia resource are obtained, and the first 12 weeks of graphic and text object behavior operation ratios of all the above-mentioned operation objects (i.e., the above-mentioned operation object 1, the above-mentioned operation object 2, the above-mentioned operation object 3 and the above-mentioned operation object 4) in the above-mentioned 12 weeks of graphic and text object behavior data on the above-mentioned first multimedia resource are obtained; the second 12 weeks of graphic and text object behavior statistics of the above-mentioned operation object 1 in the 12 weeks of graphic and text object behavior data on the second multimedia resource are obtained, and the second 12 weeks of graphic and text object behavior operation ratios of all the above-mentioned operation objects (i.e., the above-mentioned operation object 1, the above-mentioned operation object 2, the above-mentioned operation object 3 and the above-mentioned operation object 4) in the above-mentioned 12 weeks of graphic and text object behavior data on the above-mentioned second multimedia resource are obtained. Weekly graphic object behavior operation ratio; based on the above-mentioned first 12-week graphic object behavior statistics of the above-mentioned operation object 1, the above-mentioned first 12-week graphic object behavior operation ratio, the above-mentioned second 12-week graphic object behavior statistics, the above-mentioned second 12-week graphic object behavior operation ratio, and the 12-week graphic object behavior statistics of the above-mentioned operation object 1 for all multimedia resources, generate the expected click volume of the above-mentioned operation object 1 on the 12-week graphic-text side multimedia resources. Similarly, based on the above-mentioned operations, generate the expected click volume of the 12-week graphic-text side multimedia resources corresponding to the above-mentioned operation object 2, the above-mentioned operation object 3 and the above-mentioned operation object 4 respectively. After the above-mentioned target client generates five attribute features associated with the above-mentioned operation object 1 based on the 12-week graphic object behavior data of the above-mentioned operation object 1, based on the object identifier of the above-mentioned operation object 1, the five attribute features associated with the above-mentioned operation object 1 are spliced ​​to generate the 12-week graphic-text level object behavior features associated with the above-mentioned operation object 1.

[0079] Similarly, based on the 12 weeks of graphic object behavior data for operation object 2, operation object 3, and operation object 4 carried by the 12 weeks of graphic object behavior data, object characteristics, long-term object behavior characteristics, secondary resource content attribute characteristics, video multimedia resource attribute characteristics, and the object's expected number of clicks on multimedia resources associated with each of the above-mentioned operation objects are generated. The five attribute characteristics associated with each of the above-mentioned operation objects are then concatenated based on the object identifiers corresponding to each of the above-mentioned operation objects to generate the 12 weeks of graphic object behavior characteristics associated with operation object 2, the 12 weeks of graphic object behavior characteristics associated with operation object 3, and the 12 weeks of graphic object behavior characteristics associated with operation object 4. Similarly, the implementation process of generating the 12 weeks of video object behavior characteristics associated with operation object 1, operation object 2, operation object 3, and operation object 4, respectively, based on the above-mentioned operation objects carried by the 12 weeks of video object behavior data is similar to the 12 weeks of graphic object behavior characteristics, and will not be elaborated on here.

[0080] Furthermore, based on the object identification of each operation object carried in the 12-week graphic object behavior characteristics associated with each operation object and the 12-week video object behavior characteristics associated with each operation object, weekly object behavior statistical characteristics indexed by the object identification of each operation object are generated. Specifically, the 12-week graphic and text object behavior features generated by the above-mentioned target client include the 12-week graphic and text object behavior features associated with the above-mentioned operation object 1, the 12-week graphic and text object behavior features associated with the above-mentioned operation object 2, the 12-week graphic and text object behavior features associated with the above-mentioned operation object 3, and the 12-week graphic and text object behavior features associated with the above-mentioned operation object 4; the 12-week video object behavior features generated by the above-mentioned target client include the 12-week video object behavior features associated with the above-mentioned operation object 1, the 12-week video object behavior features associated with the above-mentioned operation object 2, the 12-week video object behavior features associated with the above-mentioned operation object 3, and the day-level video object behavior features associated with the above-mentioned operation object 4; based on the 12-week graphic and text object behavior features and the 12-week video object behavior features associated with the above-mentioned operation object 1, a 12-week object behavior statistical feature is generated with the object identifier of the above-mentioned operation object 1 as the index. Similarly, based on the 12-week graphic object behavior characteristics and 12-week video object behavior characteristics associated with the above-mentioned operation object 2, the above-mentioned operation object 3, and the above-mentioned operation object 4, a 12-week object behavior statistical feature is generated with the object identification of each of the above-mentioned operation objects (i.e., operation object 2, operation object 3, and operation object 4) as the index.

[0081] Furthermore, based on the object identification of each operation object carried by the above-mentioned daily object behavior statistical features and the above-mentioned 12-week object behavior statistical features, a first object behavior statistical feature is generated with the object identification of each operation object as the index. Specifically, based on the daily object behavior statistical features and the 12-week object behavior statistical features of the above-mentioned operation object 1, a first object behavior statistical feature associated with the above-mentioned operation object 1 is generated with the object identification of the above-mentioned operation object 1 as the index. Similarly, based on the daily object behavior statistical features and the 12-week object behavior statistical features of the above-mentioned operation objects (i.e., operation object 2, operation object 3, and operation object 4), a first object behavior statistical feature associated with the above-mentioned operation objects is generated with the object identification of each operation object as the index, that is, a first object behavior statistical feature associated with the above-mentioned operation object 2, a first object behavior statistical feature associated with the above-mentioned operation object 3, and a first object behavior statistical feature associated with the above-mentioned operation object 4 are generated respectively.

[0082] The feature generation method for multimedia resource recommendation provided in the embodiment of the present application can increase statistical features by adding attribute features or increasing the dimension of a certain attribute feature. The specific method can be determined according to the actual application scenario and is not limited here. It can provide a more comprehensive and expanded statistical feature usage method with a wide range of applications.

[0083] S102, obtaining second object behavior data provided by a training sample data source for multimedia resource recommendation, and generating at least two second object behavior statistical features of at least two different time periods associated with the second object based on the second object behavior data, wherein at least two object resource behavior attribute features are included in the second object behavior statistical features of any time period associated with any second object.

[0084] In some feasible implementations, the operation object registers real-name information through the target client installed on the terminal device 200b. After the operation object obtains the object identifier uniquely corresponding to the operation object through the target client, the target client installed on the terminal device 200b obtains the second object behavior data provided by the training sample data source for multimedia resource recommendation. The training sample data source for multimedia resource recommendation is the data samples of each terminal user who has completed registration on the target client, or the sample data of multimedia resources published by the target client. The specific data can be determined according to the actual application scenario and is not limited here. Please refer to Figure 6 , Figure 6: This is another application scenario diagram of the feature generation method for multimedia resource recommendation provided by an embodiment of the present application. When the target client obtains the second object behavior data provided by the training sample data source for multimedia resource recommendation, and based on the generation time of the second object behavior data carried in the second object behavior data, the second object behavior data is divided into multiple P1-level object behavior data. The P1-level object behavior data is the object behavior data included in the minimum unit time length of the second object behavior data divided by the time interval. Among them, the second object is the various objects carried in the first object behavior data. It is understandable that the second object can include the various objects carried in the first object behavior data, or it can be an object in the training sample. The second object can be determined according to the actual application scenario, and this application does not limit it here. For the convenience of description here, it is assumed that the second object includes the operation object carried in the first object behavior data, that is, the operation object 1, the operation object 2, the operation object 3 and the operation object 4 in the first object behavior data mentioned above, and the operation object 5. The division of the second object behavior data into time intervals can be determined based on the actual application scenario and is not limited in this application. That is, the P1-level object behavior data can be hourly or two-hourly. For ease of description, the P1-level object behavior data is determined as hourly. At least two P2-level object behavior data of at least two different time periods are generated based on the accumulation of multiple P1-level object behavior data included in multiple minimum unit durations. Each P2-level object behavior data is accumulated from multiple P1-level object behavior data within a target duration, where the target duration is a positive integer multiple of the minimum unit duration. It is understood that the target duration can be determined based on the actual application scenario and is not limited in this application. For example, the target duration can be 24 hours or one week. For ease of description, the target duration is set to 24 hours. It is understood that the duration of the P2-level object behavior data can be determined based on the actual application scenario and is not limited in this application. It can be one-week or two-week object behavior data. For the convenience of description, the at least two P2-level object behavior data are determined here as daily object behavior data and weekly object behavior data, such as 24-hour daily object behavior data and 1-week, 2-week, or 4-week weekly object behavior data. Specifically, the target client divides the second object behavior data into multiple hourly object behavior data based on the generation time of each object behavior data carried in the second object behavior data. The hourly object behavior data is the object behavior data included in the minimum unit time length of the second object behavior data divided by time interval.For example, assuming that the current time is xx / xx / xx, the above-mentioned hourly object behavior data is the object behavior data of each operation object before xx / xx / xx. When the target client generates the object behavior data of each operation object before xx / xx / xx based on the cumulative duration of the multiple hourly object behavior data included in the multiple minimum unit durations, the target client generates the day-level object behavior data including the time of xx / xx / xx (hereinafter referred to as the day for convenience of description). When the cumulative duration of the day-level object behavior data including the time of xx / xx / xx / xx reaches one week, the target client generates the object behavior data including the day (i.e., the object behavior data for a total of eight days). The target client generates the object behavior data based on the cumulative duration of the multiple hourly object behavior data reaching 24 hours (i.e., 24:00 every day), and generates the object behavior data of one week, two weeks, and four weeks respectively. Among them, the generation time of the object behavior data of one week including the day, the object behavior data of one week, the object behavior data of two weeks, and the object behavior data of four weeks do not overlap. Based on the object identifiers corresponding to the various operation objects (i.e., the above-mentioned operation object 1, the above-mentioned operation object 3, the above-mentioned operation object 4, and the above-mentioned operation object 5) carried in the above-mentioned 1-week object behavior data including the current day, the above-mentioned 1-week object behavior data, the above-mentioned 2-week object behavior data, and the above-mentioned 4-week object behavior data, the weekly object behavior statistical features are generated with the object identifiers of the above-mentioned various operation objects as indexes. It can be understood that at least two attribute features of the object attribute features, duration attribute features, resource content attribute features, resource type attribute features, and object resource behavior attribute features of each operation object are generated based on the weekly object behavior data of the above-mentioned various operation objects included in the above-mentioned weekly object behavior data. Among them, the total number of dimensions of the statistical features is the product of the number of dimensions of each of the above-mentioned five attribute features. It can be understood that the above-mentioned five attribute features may include multiple dimensions, and the total number of dimensions of the statistical features is the multiplication of the number of dimensions of the five attribute features. When the target client generates weekly object behavior data, the weekly object behavior data of each operation object carried by the target client generates object attribute characteristics, duration attribute characteristics, resource content attribute characteristics, resource type attribute characteristics and object resource behavior attribute characteristics related to each operation object. Among them, the object attribute characteristics include object dimensions, the duration attribute characteristics include long-term object behavior characteristics, the resource content attribute characteristics include secondary resource content attribute characteristics, the resource type attribute characteristics include video multimedia resource attribute characteristics, and the object resource behavior attribute characteristics include at least one object resource behavior-related characteristic of the click duration, number of clicks, click-through rate, number of exposures, viewing time and interaction time for the target multimedia resource. Among them, the number of clicks on the multimedia resource is the number of times the operation object clicks to view the multimedia resource within a certain period of time.The above-mentioned click-through rate for multimedia resources is the ratio of the number of clicks on the target multimedia resource by the operating object to the number of exposures of the target multimedia resource. The above-mentioned viewing time for multimedia resources is the total viewing time of the target multimedia resource by the operating object. The above-mentioned interaction time for multimedia resources is the interaction of the operating object with the target multimedia resource, such as likes, comments, sharing and other operations. Among them, when the above-mentioned click time for multimedia resources is greater than a certain threshold, a long click rate (LCR) for the target multimedia resource is generated. It can be understood that the above-mentioned threshold can be determined according to the actual application scenario, and this application does not impose any restrictions here. Based on the above-mentioned object resource behavior attribute characteristics including at least one object resource behavior-related feature of the click time, number of clicks, click rate, number of exposures, viewing time and interaction time for the target multimedia resource, a completion rate (PCR) for the above-mentioned target multimedia resource can be generated. Among them, the above-mentioned completion degree can be the ratio of the viewing time of the operation object for the above-mentioned target multimedia resource (such as a video resource) to the total viewing time of the above-mentioned target multimedia resource, or the proportion of the images that have been viewed in the above-mentioned target multimedia resource (such as a picture resource) by the operation object in the above-mentioned target multimedia resource. The actual meaning of the above-mentioned completion degree can be determined according to the actual application scenario, and this application does not limit it here. It can be understood that in an embodiment of the present application, the above-mentioned target client generates at least two object resource behavior attribute features in the object behavior features of each level associated with the object identifier of each operation object based on the above-mentioned object resource behavior attribute features including the click time, number of clicks, click-through rate, number of exposures, viewing time and interaction time for the target multimedia resource. Among them, the above-mentioned at least two object resource behavior attribute features are the long clicks and completions of the above-mentioned operation objects for the above-mentioned target multimedia resource. It can be understood that the above-mentioned operation objects can carry long clicks for the above-mentioned target multimedia resource, can carry the completion for the above-mentioned target multimedia resource, or can carry long clicks and completions for the above-mentioned target multimedia resource, and this application does not limit it here. For the convenience of description, it is assumed that each of the above-mentioned operation objects in the embodiment of the present application carries a long click and a completion degree for the above-mentioned target multimedia resource.

[0085] For example, assuming that when the target client obtains the object behavior data of each of the above-mentioned operation objects (i.e., the above-mentioned operation object 1, the above-mentioned operation object 3, the above-mentioned operation object 4, and the above-mentioned operation object 5) carried in the second object behavior data provided by the training sample data source recommended by the multimedia resource, based on the generation time of the object behavior data of each operation object carried in the second object behavior data, the object behavior data carried by each of the above-mentioned operation objects is divided into a plurality of hourly-level object behavior data. Among them, the plurality of hourly-level object behavior data are the object behavior data included in the minimum unit duration of the second object behavior data divided by the time interval. Specifically, based on the generation time of the object behavior data of the above-mentioned operation object 1 carried in the second object data, the object behavior data of the above-mentioned operation object 1 is divided into a plurality of hourly-level object behavior data. Specifically, assuming that the current time is xx / xx / xx, the hourly-level object behavior data are the object behavior data of each operation object before xx / xx / xx. When the target client generates the day-level object behavior data including the day of xx / xx / xx based on the accumulated duration of the hour-level object behavior data of the operation object 1, the target client generates the day-level object behavior data including the day of xx / xx / xx for the operation object 1 (hereinafter referred to as the day-level object behavior data including the day of xx / xx / xx). Wherein, the target duration is a positive integer multiple of the minimum unit duration. For the convenience of description, the target duration is determined to be 24 hours, that is, at 24 o'clock every day, the day-level object behavior data including the day of xx / xx / xx is generated for the operation object 1. When the target client generates the day-level object behavior data including the day of xx / xx / xx for the operation object 1, the day-level object behavior data including the day of xx / xx / xx is accumulated based on the day-level object behavior data including the day of xx / xx / xx for the operation object 1. When the accumulated duration of the day-level object behavior data including the day of xx / xx / xx reaches one week, the week-level object behavior data including the day of xx / xx / xx for the operation object 1 (that is, the object behavior data for a total of eight days) is generated. Based on the above-mentioned one-week object behavior data including the current day for the above-mentioned operation object 1, object characteristics, long-term object behavior characteristics, secondary resource content attribute characteristics, video multimedia resource attribute characteristics and click duration associated with the above-mentioned operation object 1 are generated, and the five attribute characteristics associated with the above-mentioned operation object 1 are spliced ​​based on the object identifier corresponding to the above-mentioned operation object 1 to generate the one-week object behavior characteristics including the current day associated with the above-mentioned operation object 1. Among them, the one-week object behavior characteristics including the current day generated in association with the above-mentioned operation object 1 include at least two object resource behavior attribute characteristics of the above-mentioned operation object 1. Specifically, based on the click duration of the above-mentioned operation object 1 on the above-mentioned target multimedia resource, the long click and completion degree of the above-mentioned operation object 1 on the target multimedia resource are respectively generated.

[0086] Similarly, based on the generation time of the object behavior data of the operation object 1 carried in the second object data, the object behavior data of the operation object 1 is divided into a plurality of hourly object behavior data. Specifically, assuming that the current time is xx / xx / xx, the hourly object behavior data is the object behavior data of each operation object one week before xx / xx / xx. When the cumulative duration of the hourly object behavior data of the operation object 1 by the target client reaches the target duration, daily object behavior data for the operation object 1 is generated. The target duration is a positive integer multiple of the minimum unit duration. For ease of description, the target duration is set to 24 hours, i.e., daily object behavior data for the operation object 1 is generated at 24:00 every day. When the target client generates daily object behavior data for the operation object 1, the daily object behavior data is accumulated based on the daily object behavior data for the operation object 1. When the cumulative duration of the daily object behavior data reaches one week, one week of object behavior data for the operation object 1 (i.e., a total of seven days of object behavior data) is generated. Based on the above-mentioned one-week object behavior data for the above-mentioned operation object 1, the object characteristics, long-term object behavior characteristics, secondary resource content attribute characteristics, video multimedia resource attribute characteristics and click duration associated with the above-mentioned operation object 1 are generated, and the five attribute characteristics associated with the above-mentioned operation object 1 are spliced ​​based on the object identifier corresponding to the above-mentioned operation object 1 to generate the one-week object behavior characteristics associated with the above-mentioned operation object 1. Among them, the one-week object behavior characteristics associated with the above-mentioned operation object 1 include at least two object resource behavior attribute characteristics of the above-mentioned operation object 1. Specifically, based on the click duration of the above-mentioned operation object 1 on the above-mentioned target multimedia resource, the long click and completion degree of the above-mentioned operation object 1 on the target multimedia resource are respectively generated.

[0087] Similarly, based on the object behavior data of the operation object 1 carried in the second object data, 2-week object behavior data for the operation object 1 (i.e., object behavior data for a total of seven days) is generated. Based on the 2-week object behavior data for the operation object 1, object features, long-term object behavior features, secondary resource content attribute features, video multimedia resource attribute features, and click duration associated with the operation object 1 are generated, and the five attribute features associated with the operation object 1 are spliced ​​based on the object identifier corresponding to the operation object 1 to generate the 2-week object behavior features associated with the operation object 1. Among them, the 2-week object behavior features associated with the operation object 1 include at least two object resource behavior attribute features of the operation object 1. Specifically, the long clicks and completion rate of the operation object 1 for the target multimedia resource are generated based on the click duration of the operation object 1 for the target multimedia resource. Based on the object behavior data of the operation object 1 carried in the second object data, 4-week object behavior data for the operation object 1 (i.e., object behavior data for a total of seven days) are generated. Based on the above-mentioned 4-week object behavior data for the above-mentioned operation object 1, object features, long-term object behavior features, secondary resource content attribute features, video multimedia resource attribute features and click duration associated with the above-mentioned operation object 1 are generated, and the five attribute features associated with the above-mentioned operation object 1 are spliced ​​based on the object identifier corresponding to the above-mentioned operation object 1 to generate the 4-week object behavior features associated with the above-mentioned operation object 1. Among them, the above-mentioned 4-week object behavior features associated with the above-mentioned operation object 1 include at least two object resource behavior attribute features of the above-mentioned operation object 1. Specifically, based on the click duration of the above-mentioned operation object 1 for the above-mentioned target multimedia resource, the long click and completion degree of the above-mentioned operation object 1 for the target multimedia resource are respectively generated. Among them, the generation time of the above-mentioned 1-week object behavior data including the current day, the above-mentioned 1-week object behavior data, the above-mentioned 2-week object behavior data and the above-mentioned 4-week object behavior data for the operation object 1 do not overlap.

[0088] Similarly, based on the object behavior data of the operation object 3 carried in the second object data, one-week object behavior data, one-week object behavior data, two-week object behavior data, and four-week object behavior data for the operation object 3, including the current day, are generated. The generation times of the one-week object behavior data, one-week object behavior data, two-week object behavior data, and four-week object behavior data for the operation object 3 do not overlap. Based on the one-week object behavior data, one-week object behavior data, two-week object behavior data, and four-week object behavior data for the operation object 3, object features, long-term object behavior features, secondary resource content attribute features, video multimedia resource attribute features, and click duration associated with the operation object 3 are generated. The five attribute features associated with the operation object 3 are then concatenated based on the object identifier corresponding to the operation object 3 to generate one-week object behavior features, one-week object behavior features, two-week object behavior features, and four-week object behavior features associated with the operation object 3, including the current day. The generated object behavior characteristics at each level associated with the operation object 3 include at least two object resource behavior attribute characteristics of the operation object 3. Specifically, based on the click duration of the target multimedia resource in the object behavior characteristics at each level of the operation object 3, the long click and completion degree of the target multimedia resource in the object behavior characteristics at each level are respectively generated.

[0089] Similarly, based on the object behavior data of the operation object 4 carried in the second object data, one-week-level object behavior data, one-week-level object behavior data, two-week-level object behavior data, and four-week-level object behavior data for the operation object 4, including the current day, are generated. The generation times of the one-week-level object behavior data, one-week-level object behavior data, two-week-level object behavior data, and four-week-level object behavior data for the operation object 4 do not overlap. Based on the one-week-level object behavior data, one-week-level object behavior data, two-week-level object behavior data, and four-week-level object behavior data for the operation object 4, object features, long-term object behavior features, secondary resource content attribute features, video multimedia resource attribute features, and click duration associated with the operation object 4 are generated. The five attribute features associated with the operation object 4 are then concatenated based on the object identifier corresponding to the operation object 4 to generate one-week-level object behavior features, one-week-level object behavior features, two-week-level object behavior features, and four-week-level object behavior features associated with the operation object 4, including the current day. The generated object behavior characteristics at each level associated with the operation object 4 include at least two object resource behavior attribute characteristics of the operation object 4. Specifically, based on the click duration of the target multimedia resource in the object behavior characteristics at each level of the operation object 4, the long click and completion degree of the target multimedia resource in the object behavior characteristics at each level are generated respectively.

[0090] Similarly, based on the object behavior data of the operation object 5 carried in the second object data, one-week-level object behavior data, one-week-level object behavior data, two-week-level object behavior data, and four-week-level object behavior data for the operation object 5, including the current day, are generated. The generation times of the one-week-level object behavior data, one-week-level object behavior data, two-week-level object behavior data, and four-week-level object behavior data for the operation object 5 do not overlap. Based on the one-week-level object behavior data, one-week-level object behavior data, two-week-level object behavior data, and four-week-level object behavior data for the operation object 5, object features, long-term object behavior features, secondary resource content attribute features, video multimedia resource attribute features, and click duration associated with the operation object 5 are generated. The five attribute features associated with the operation object 5 are then concatenated based on the object identifier corresponding to the operation object 5 to generate one-week-level object behavior features, one-week-level object behavior features, two-week-level object behavior features, and four-week-level object behavior features associated with the operation object 5, including the current day. The generated object behavior characteristics at each level associated with the operation object 5 include at least two object resource behavior attribute characteristics of the operation object 5. Specifically, based on the click duration of the target multimedia resource in the object behavior characteristics at each level of the operation object 5, the long click and completion degree of the target multimedia resource in the object behavior characteristics at each level are generated respectively.

[0091] When the target client generates object behavior features at various levels associated with the above-mentioned operation objects based on the object behavior data of the above-mentioned operation objects carried by the second object behavior data, at least two object resource behavior attribute features in the object behavior features at various levels associated with the object identifiers of the above-mentioned operation objects are spliced ​​to generate second object behavior statistical features associated with the above-mentioned operation object identifiers. Specifically, with the object identifier of the above-mentioned operation object 1 as an index, the long clicks and completions for the above-mentioned target multimedia resources included in the 1-week object behavior features associated with the above-mentioned operation object 1 are spliced. Similarly, with the object identifier of the above-mentioned operation object 1 as an index, the long clicks and completions for the above-mentioned target multimedia resources included in the 1-week object behavior features associated with the above-mentioned operation object 1 are spliced. Similarly, with the object identifier of the above-mentioned operation object 1 as an index, the long clicks and completions for the above-mentioned target multimedia resources included in the 2-week object behavior features associated with the above-mentioned operation object 1 are spliced. Similarly, using the object identifier of the operation object 1 as an index, the long clicks and completions for the target multimedia resource included in the 4-week object behavior characteristics associated with the operation object 1 are spliced ​​together. Using the object identifier of the operation object 3 as an index, the long clicks and completions for the target multimedia resource included in the 1-week object behavior characteristics including the current day, the 1-week object behavior characteristics, the 2-week object behavior characteristics, and the 4-week object behavior characteristics associated with the operation object 3 are spliced ​​together respectively. Using the object identifier of the operation object 4 as an index, the long clicks and completions for the target multimedia resource included in the 1-week object behavior characteristics including the current day, the 1-week object behavior characteristics, the 2-week object behavior characteristics, and the 4-week object behavior characteristics associated with the operation object 4 are spliced ​​together respectively. Using the object identifier of the above-mentioned operation object 5 as an index, the long clicks and completion rates for the above-mentioned target multimedia resources included in the above-mentioned 1-week object behavior characteristics of the day, the above-mentioned 1-week object behavior characteristics, the 2-week object behavior characteristics, and the above-mentioned 4-week object behavior characteristics associated with the above-mentioned operation object 5 are spliced ​​respectively.

[0092] Please see Figure 6Based on the object identifiers of the aforementioned operation objects, the object behavior features at all levels associated with the aforementioned operation objects are concatenated using the object identifiers of the aforementioned operation objects to generate second object behavior statistical features associated with the object identifiers of the aforementioned operations. Specifically, based on the aforementioned 1-week-level object behavior features, the aforementioned 1-week-level object behavior features, the aforementioned 2-week-level object behavior features, and the aforementioned 4-week-level object behavior features associated with the aforementioned operation object 1, using the object identifier of the aforementioned operation object 1 as an index, the aforementioned 1-week-level object behavior features, the aforementioned 1-week-level object behavior features, the aforementioned 2-week-level object behavior features, and the aforementioned 4-week-level object behavior features are concatenated to generate a second object behavior statistical feature associated with the object identifier of the aforementioned operation object 1. Similarly, based on the one-week-level object behavior features, one-week-level object behavior features, two-week-level object behavior features, and four-week-level object behavior features associated with operation object 3, the one-week-level object behavior features, one-week-level object behavior features, two-week-level object behavior features, and four-week-level object behavior features are concatenated using the object identifier of operation object 3 as an index to generate a second object behavior statistical feature associated with the object identifier of operation object 3. Similarly, based on the one-week-level object behavior features, one-week-level object behavior features, two-week-level object behavior features, and four-week-level object behavior features associated with operation object 4, the one-week-level object behavior features, one-week-level object behavior features, two-week-level object behavior features, and four-week-level object behavior features are concatenated using the object identifier of operation object 4 as an index to generate a second object behavior statistical feature associated with the object identifier of operation object 4. Similarly, based on the above-mentioned 1-week level object behavior characteristics including the current day, the above-mentioned 1-week level object behavior characteristics, the above-mentioned 2-week level object behavior characteristics and the above-mentioned 4-week level object behavior characteristics associated with the above-mentioned operation object 5, with the object identification of the above-mentioned operation object 5 as the index, the above-mentioned 1-week level object behavior characteristics including the current day, the above-mentioned 1-week level object behavior characteristics, the above-mentioned 2-week level object behavior characteristics and the above-mentioned 4-week level object behavior characteristics are spliced ​​to generate a second object behavior statistical feature associated with the object identification of the above-mentioned operation object 5.

[0093] S103, generating object statistical features of each object based on the first object behavior statistical features associated with each of the above-mentioned first objects and the second object behavior statistical features associated with each of the above-mentioned second objects, and determining the input features of the multimedia resource recommendation model with the above-mentioned object statistical features, wherein the above-mentioned multimedia resource recommendation model is trained by the above-mentioned second object behavior data, and the above-mentioned multimedia resource model is used to output multimedia resource recommendation values ​​for each of the above-mentioned objects based on the object statistical features of each of the above-mentioned objects.

[0094] See Figure 7 , Figure 7: This is another application scenario diagram of the feature generation method for multimedia resource recommendation provided by an embodiment of the present application. In some feasible implementations, after obtaining the first object behavior data provided by the target client data source, the target client generates first object behavior statistical features associated with each of the first objects based on the object behavior data of the first object carried by the first object behavior data. After obtaining the second object behavior data provided by the training sample data source for multimedia resource recommendation, the target client generates second object behavior statistical features associated with the second object based on the object behavior data of the second object carried by the second object behavior data. The target client obtains the object identifiers corresponding to each of the first objects from the first object behavior statistical features, and obtains the object identifiers corresponding to each of the second objects from the second object behavior statistical features. Specifically, the target client obtains the object identifiers corresponding to the operation object 1, the operation object 2, the operation object 3, and the operation object 4 from the first object behavior statistical features. The target client obtains the object identifiers corresponding to the operation object 1, the operation object 3, the operation object 4, and the operation object 5 from the second object behavior statistical features. The target client uses the object identifiers of the first objects and the object identifiers of the second objects as indexes to splice the first object behavior statistical features and the second object behavior statistical features to generate object statistical features corresponding to the object identifiers. The first object behavior statistical features and the second object behavior statistical features with the same object identifier are spliced ​​into object statistical features of the same object. It can be understood that each of the operation objects may have only the first object behavior statistical features associated with the operation object, may have only the second object behavior statistical features associated with the operation object, or may have the first object behavior statistical features and the second object behavior statistical features associated with the operation object. Specifically, the target client uses the object identifier corresponding to the operation object 1 as an index to splice the first object behavior statistical features and the second object behavior statistical features associated with the operation object 1 to generate object statistical features corresponding to the object identifier of the operation object 1. The target client uses the object identifier corresponding to the operation object 2 as an index to generate object statistical features corresponding to the object identifier of the operation object 2 based on the first object behavior statistical features associated with the operation object 2. The target client uses the object identifier corresponding to the operation object 3 as an index to concatenate the first object behavior statistical feature and the second object behavior statistical feature associated with the operation object 3 to generate an object statistical feature corresponding to the object identifier of the operation object 3.The target client uses the object identifier corresponding to the operation object 4 as an index to concatenate the first object behavior statistical feature and the second object behavior statistical feature associated with the operation object 4 to generate an object statistical feature corresponding to the object identifier of the operation object 4. The target client uses the object identifier corresponding to the operation object 5 as an index to generate an object statistical feature corresponding to the object identifier of the operation object 5 based on the second object behavior statistical feature associated with the operation object 5.

[0095] After the target client generates object statistical features associated with each of the operation objects based on the first object behavior statistical features and the second object behavior statistical features of each of the operation objects, the target client uses the object identifier of each of the operation objects as an index and uses the object behavior features associated with the object identifier of each of the operation objects as input features for determining a multimedia resource recommendation model. Furthermore, the target client uses the object identifier of each of the operation objects as an index and writes the object behavior features associated with each of the operation objects into a statistical feature storage unit. The feature extraction and online ranking service can read the object statistical features in the statistical feature storage unit and then score them using the multimedia resource recommendation ranking model to output multimedia resource recommendation values ​​for each of the operation objects. The target multimedia resource is recommended to the user based on the multimedia resource recommendation values. The training sample data for the scene itself can then be updated. The updated training sample data can then be used for model training, thereby improving the model's expressiveness and enhancing data resource utilization. Optionally, multimedia resource recommendation values ​​for each of the above-mentioned operation objects are output based on the object statistical characteristics of each of the above-mentioned operation objects. After the above-mentioned target client obtains the multimedia resource recommendation values ​​for each of the above-mentioned operation objects, each of the multimedia resources can be sorted based on the multimedia resource recommendation values ​​for each of the above-mentioned operation objects. Furthermore, the above-mentioned target client can sort the multimedia resources for each of the above-mentioned operation objects in a descending order of recommendation values ​​based on the multimedia resource recommendation values ​​of each of the above-mentioned operation objects to obtain a multimedia resource recommendation list for each of the above-mentioned operation objects, and based on the multimedia resource recommendation list, recommend the top n multimedia resources to each of the above-mentioned operation objects in a descending order of recommendation values, where n is a positive integer.

[0096] Optionally, after the target client obtains the multimedia resource recommendation value of each operation object, the target client can sort the multimedia resources based on the multimedia resource recommendation value of each operation object. The target client can sort the multimedia resources of each operation object in descending order based on the multimedia resource recommendation value of each operation object to obtain a multimedia resource recommendation list for each operation object, and extract the top n multimedia resources based on the multimedia resource recommendation list. The top n multimedia resources are recommended to each operation object in a random order, where n is a positive integer, so as to improve the diversity of the multimedia resource recommendation list.

[0097] In some feasible implementations, based on the implementation provided by steps S101 to S103 above, 335-dimensional object statistical features can be generated after screening, including 260-dimensional video-side client statistical features of the object and 75-dimensional in-scene long-click completion statistical features of the training sample. For ease of description, the terms appearing in the following examples are explained:

[0098] pv: number of exposures; clk: number of clicks; longClk: number of long clicks; crSum: sum of completions; time: playback duration; ec: expected number of clicks; allCtr: total click-through rate of the current user in a certain time period; allLongCtr: total long click-through rate of the current user in a certain time period; allCrAvg: total average completion rate of the current user in a certain time period; allTimePerPv: average playback duration of the current user in a certain time period; sumCtr: total click-through rate of all users in a certain time period; sumLongCtr: total long click-through rate of all users in a certain time period; sumCrAvg: total average completion rate of all users in a certain time period; sumTimePerPv: average playback duration of all users in a certain time period; pvFea: min(ln(pv+1),11.0) / 11.0, indicating the number of exposures and its normalization; ctrFea: min(clk / (pv+0.01), 1.0) represents the click-through rate and its normalization; longCtrFea: min(longClk / (pv+0.01), 1.0) represents the long click-through rate and its normalization; crAvgFea: min(crSum / (pv+0.01), 5.0) / 5.0 represents the average completion rate of each exposure and its normalization; timePerPvFea: min(time / (pv+0.01), 2000) / 2000 represents the average playback time of each exposure and its normalization; coecFea: min(clk / (ec+0.01), 10.0) / 10.0 represents the COEO features of the operation object and its normalization.

[0099] Inner preference level:

[0100] allCoec: min(clk / (pv*allCtr+0.01),10.0) / 10.0, which indicates the click preference of the current operation object for the current content and its normalization;

[0101] allLoel: min(longClk / (pv*allLongCtr+0.01),10.0) / 10.0, which indicates the normalized long click preference of the current operation object for the current content.

[0102] allRoer: min(crSum / (pv*allCrAvg+0.01),10.0) / 10.0, which indicates the completion preference of the current operation object for the current content and its normalization;

[0103] allToet: min(time / (pv*allTimePerPv+0.01),10.0) / 10.0, which indicates the preference of the current operation object for the playback duration of the current content and its normalization for the operation object.

[0104] Inter preference level:

[0105] sumCoec: min(clk / (pv*sumCtr+0.01),10.0) / 10.0, which represents the click preference of the current operation object for the current content among all operation objects and its normalization;

[0106] sumLoel: min(longClk / (pv*sumLongCtr+0.01),10.0) / 10.0, which represents the long click preference of the current operation object for the current content and its normalization among all operation objects;

[0107] sumRoer: min(crSum / (pv*sumCrAvg+0.01),10.0) / 10.0, which indicates the completion preference of the current operation object for the current content among all operation objects and its normalization;

[0108] sumToet: min(time / (pv*sumTimePerPv+0.01),10.0) / 10.0, which indicates the preference of the current operation object for the playback duration of the current content and its normalization among all operation objects.

[0109] For example, assume that the target client is used on July 15, and the video statistical features are 150 dimensions, as described below:

[0110] Dimensions 0-5: pvFea, ctrFea, timePerPvFea, coecFea, sumCoec, sumToet of the current operation object on July 14th;

[0111] Dimensions 6-11: pvFea, ctrFea, timePerPvFea, coecFea, sumCoec, sumToet of the current operation object in the 2-day period from July 12 to July 14;

[0112] Dimensions 12-17: pvFea, ctrFea, timePerPvFea, coecFea, sumCoec, and sumToet of the current operation object in the 4-day period from July 8 to July 11;

[0113] Dimensions 18-24: pvFea, ctrFea, timePerPvFea, coecFea, sumCoec, sumToet of the current operation object in the 7-day period from July 1 to July 7;

[0114] Dimensions 24-29: pvFea, ctrFea, timePerPvFea, coecFea, sumCoec, sumToet of the current operation object in the last 12 weeks;

[0115] Dimensions 40-47: pvFea, ctrFea, timePerPvFea, coecFea, allCoec, allToet, sumCoec, sumToet of the first-level classification of the current video within 1 day on July 14th;

[0116] Dimensions 48-45: pvFea, ctrFea, timePerPvFea, coecFea, allCoec, allToet, sumCoec, and sumToet of the first-level classification of the current video within the time period of July 12th to July 14th.

[0117] Dimensions 46-54: pvFea, ctrFea, timePerPvFea, coecFea, allCoec, allToet, sumCoec, and sumToet of the first-level classification of the current video within the time period of July 8th to July 11th.

[0118] Dimensions 54-61: pvFea, ctrFea, timePerPvFea, coecFea, allCoec, allToet, sumCoec, and sumToet of the first-level classification of the current video within the time period of July 1st to July 7th.

[0119] Dimensions 62-69: pvFea, ctrFea, timePerPvFea, coecFea, allCoec, allToet, sumCoec, sumToet of the secondary classification of the current video within 1 day on July 14th;

[0120] Dimensions 70-77: pvFea, ctrFea, timePerPvFea, coecFea, allCoec, allToet, sumCoec, and sumToet of the secondary classification of the current video within the time period of July 12th to July 14th.

[0121] Dimensions 78-85: pvFea, ctrFea, timePerPvFea, coecFea, allCoec, allToet, sumCoec, and sumToet of the secondary classification of the current video within the 4-day time period from July 8 to July 11.

[0122] Dimensions 86-94: pvFea, ctrFea, timePerPvFea, coecFea, allCoec, allToet, sumCoec, and sumToet of the secondary classification of the current video within the time period of July 1st to July 7th.

[0123] Dimensions 94-101: pvFea, ctrFea, timePerPvFea, coecFea, allCoec, allToet, sumCoec, sumToet of the current operation object on the current video author within 1 day on July 14;

[0124] Dimensions 102-109: pvFea, ctrFea, timePerPvFea, coecFea, allCoec, allToet, sumCoec, sumToet of the current operation object for the current video author within the 2-day period from July 12 to July 14;

[0125] Dimensions 110-117: pvFea, ctrFea, timePerPvFea, coecFea, allCoec, allToet, sumCoec, sumToet of the current operation object for the current video author within the 4-day period from July 8th to July 11th;

[0126] Dimensions 118-125: pvFea, ctrFea, timePerPvFea, coecFea, allCoec, allToet, sumCoec, sumToet of the current operation object for the current video author within the 7-day period from July 1 to July 7;

[0127] Dimensions 126-144: pvFea, ctrFea, timePerPvFea, coecFea, allCoec, allToet, sumCoec, sumToet of the first-level classification of the current video to which the current operation object belongs in the last 12 weeks;

[0128] Dimensions 144-141: pvFea, ctrFea, timePerPvFea, coecFea, allCoec, allToet, sumCoec, sumToet of the secondary classification of the current video to which the current operation object belongs in the last 12 weeks;

[0129] Dimensions 142-159: pvFea, ctrFea, timePerPvFea, coecFea, allCoec, allToet, sumCoec, sumToet of the current operation object on the current video author in the last 12 weeks.

[0130] Assume that the target client is used on July 15, and the statistical features of the image and text corresponding to the video are 110 dimensions, as described below:

[0131] Dimensions 0-5: pvFea, ctrFea, timePerPvFea, coecFea, sumCoec, sumToet of the current operation object on July 14th;

[0132] Dimensions 6-11: pvFea, ctrFea, timePerPvFea, coecFea, sumCoec, sumToet of the current operation object in the 2-day time period from July 12 to July 14;

[0133] Dimensions 12-17: pvFea, ctrFea, timePerPvFea, coecFea, sumCoec, sumToet of the current operation object in the 4-day time period from July 8th to July 11th;

[0134] Dimensions 18-24: pvFea, ctrFea, timePerPvFea, coecFea, sumCoec, sumToet of the current operation object in the 7-day period from July 1 to July 7;

[0135] Dimensions 24-29: pvFea, ctrFea, timePerPvFea, coecFea, sumCoec, sumToet of the current operation object in the last 12 weeks;

[0136] Dimensions 40-47: pvFea, ctrFea, timePerPvFea, coecFea, allCoec, allToet, sumCoec, sumToet of the first-level classification of the image and text corresponding to the current video within 1 day on July 14th;

[0137] Dimensions 48-45: pvFea, ctrFea, timePerPvFea, coecFea, allCoec, allToet, sumCoec, and sumToet of the first-level classification of the image and text corresponding to the current video within the time period of July 12th to July 14th.

[0138] Dimensions 46-54: pvFea, ctrFea, timePerPvFea, coecFea, allCoec, allToet, sumCoec, and sumToet of the first-level classification of the image and text corresponding to the current video within the time period of July 8th to July 11th.

[0139] Dimensions 54-61: pvFea, ctrFea, timePerPvFea, coecFea, allCoec, allToet, sumCoec, sumToet of the first-level classification of the image and text corresponding to the current video within the time period of July 1st to July 7th;

[0140] Dimensions 62-69: pvFea, ctrFea, timePerPvFea, coecFea, allCoec, allToet, sumCoec, sumToet of the secondary classification of the image and text corresponding to the current video within 1 day on July 14th;

[0141] Dimensions 70-77: pvFea, ctrFea, timePerPvFea, coecFea, allCoec, allToet, sumCoec, sumToet of the secondary classification of the image and text corresponding to the current video within the time period of July 12th to July 14th;

[0142] Dimensions 78-85: pvFea, ctrFea, timePerPvFea, coecFea, allCoec, allToet, sumCoec, and sumToet of the secondary classification of the image and text corresponding to the current video within the 4-day period from July 8 to July 11 of the current operation object;

[0143] Dimensions 86-94: pvFea, ctrFea, timePerPvFea, coecFea, allCoec, allToet, sumCoec, and sumToet of the secondary classification of the image and text corresponding to the current video within the time period of July 1st to July 7th.

[0144] Dimensions 94-101: pvFea, ctrFea, timePerPvFea, coecFea, allCoec, allToet, sumCoec, sumToet of the first-level classification of the image and text corresponding to the current video by the current operation object in the last 12 weeks;

[0145] Dimensions 102-109: pvFea, ctrFea, timePerPvFea, coecFea, allCoec, allToet, sumCoec, sumToet of the secondary classification of the image and text corresponding to the current video by the current operation object in the last 12 weeks.

[0146] Assume that the target client is used by the operation object on July 15, and the statistical features of long click, completion, and duration in the scene are 75 dimensions, as described below:

[0147] Dimensions 0-4: pvFea, longCtrFea, timePerPvFea, and crAvgFea of ​​the current operation object within 8 days from July 8 to July 15;

[0148] Dimensions 4-7: pvFea, longCtrFea, timePerPvFea, and crAvgFea of ​​the current operation object within 7 days from July 1 to July 7;

[0149] Dimensions 8-14: pvFea, longCtrFea, timePerPvFea, crAvgFea, allLoel, allToet, and allRoer of the first-level classification of the current video for the current operation object within 8 days from July 8 to July 15;

[0150] Dimensions 15-21: pvFea, longCtrFea, timePerPvFea, crAvgFea, allLoel, allToet, and allRoer of the secondary classification of the current video for the current operation object within 8 days from July 8 to July 15;

[0151] Dimensions 22-28: pvFea, longCtrFea, timePerPvFea, crAvgFea, allLoel, allToet, and allRoer of the current operation object for the current video author within 8 days from July 8 to July 15;

[0152] Dimensions 29-45: pvFea, longCtrFea, timePerPvFea, crAvgFea, allLoel, allToet, and allRoer of the first-level category to which the current video belongs within 7 days from July 1 to July 7;

[0153] Dimensions 46-42: pvFea, longCtrFea, timePerPvFea, crAvgFea, allLoel, allToet, and allRoer of the secondary classification of the current video within 7 days from July 1 to July 7;

[0154] Dimensions 44-49: pvFea, longCtrFea, timePerPvFea, crAvgFea, allLoel, allToet, and allRoer of the current operation object on the current video author within 7 days from July 1 to July 7;

[0155] Dimensions 50-54: pvFea, longCtrFea, timePerPvFea, and crAvgFea of ​​the current operation object within 14 days from June 17 to June 40;

[0156] Dimensions 54-60: pvFea, longCtrFea, timePerPvFea, crAvgFea, allLoel, allToet, and allRoer of the first-level category to which the current video belongs within 14 days from June 17 to June 40;

[0157] Dimensions 61-67: pvFea, longCtrFea, timePerPvFea, crAvgFea, allLoel, allToet, and allRoer of the secondary classification of the current video within 14 days from June 17 to June 40;

[0158] Dimensions 68-74: pvFea, longCtrFea, timePerPvFea, crAvgFea, allLoel, allToet, and allRoer of the current user for the current video author within 14 days from June 17 to June 40.

[0159] In some feasible implementations, after generating object statistical features based on the implementation provided in the above embodiment, the object statistical features can also be passed through a statistical feature convolution layer and then concatenated with the output features of the last network layer of the multimedia resource recommendation model as estimated features for multimedia resource recommendation. For details, see Figure 8 , Figure 8 It is another application scenario diagram of the feature generation method for multimedia resource recommendation provided by the embodiment of the present application. The above-mentioned multimedia resource recommendation model can splice the embedded features with the cross-features of the object side and the content side, the embedded features of the object side features and the embedded features of the content side features, and generate the network layer output features of the multimedia resources through the network layer of the above-mentioned multimedia resource recommendation model. After generating the object statistical features associated with the above-mentioned operation objects (i.e., 335-dimensional object behavior features), the above-mentioned target client passes the object statistical features associated with the above-mentioned operation objects through a statistical feature convolution layer, and splices them with the network layer output features of the above-mentioned multimedia resource recommendation model as the recommendation value estimation features of the multimedia resources, which are used to obtain the recommendation value of the multimedia resources. The feature generation method for multimedia resource recommendation provided by the embodiment of the present application can not only optimize the multimedia resource recommendation model, but also improve the accuracy of the object statistical features associated with the above-mentioned operation objects.

[0160] In some feasible implementations, after generating object statistical features based on the implementation method provided in the above embodiment, the above object statistical features can also be interacted with other important features. It can be understood that the above other important features can be determined according to the actual scenario application and are not limited here. Specifically, after the above target client generates the object statistical features associated with the above operation objects, the object statistical features associated with the above operation objects are interacted with the multimedia resource identification features to generate interactive features, and the above interactive features, multimedia resource identification features, and the embedded features of the above object-side content-side cross features, the embedded features of the object-side features, and the embedded features of the content-side features are spliced ​​as the network layer input features of the multimedia resource recommendation model. Please refer to Figure 9 , Figure 9This is another application scenario diagram of the feature generation method for multimedia resource recommendation provided in an embodiment of the present application. Optionally, the network layer of the multimedia resource recommendation model may include a CONCAT layer, an EXPERT layer, an EXPERT output weighted sum layer, an LCR DNN, and a PCR DNN. The CONCAT layer represents connecting the various features of the previous layer to form a 1000+ dimensional vector. The EXPERT layer represents the expert network layer, and each expert network is a multi-layer perceptron. It can be understood that the number of the expert network layers can be determined according to the actual scenario application and is not limited here. For the convenience of description, it is assumed that the expert network has 3 expert networks. The EXPERT output weighted sum layer performs weighted summation on the outputs of each expert network in the previous layer. The LCR DNN is a multi-layer perceptron for the long click task layer, and the PCR DNN is a multi-layer perceptron for the completion task layer. Specifically, the multimedia resource recommendation model can splice the embedded features with object-side and content-side cross-features, the embedded features of object-side features, and the embedded features of content-side features as the embedded features of the network layer of the multimedia resource recommendation model. The network layer of the multimedia resource recommendation model can splice the embedded features with object-side and content-side cross-features, the embedded features of object-side features, and the embedded features of content-side features through the CONCAT layer, and call the EXPERT layer to output the feature values ​​of each expert network for the long click and the feature values ​​of each expert network for completion based on the spliced ​​features of the previous layer. The multimedia resource recommendation model calls the EXPERT output weighted summation layer to perform weighted summation on the feature values ​​of the expert networks for long clicks and the feature values ​​of the expert networks for completion based on the feature values ​​of the expert networks for long clicks and the feature values ​​of the expert networks for completion, so as to generate the feature summation values ​​of the expert networks for long clicks and the feature summation values ​​of the expert networks for completion, and inputs the feature summation values ​​of the expert networks for long clicks and the feature summation values ​​of the expert networks for completion into LCR DNN and PCRDNN respectively to generate the network layer output features of the multimedia resource recommendation model. After generating the object statistical features associated with the above-mentioned operation objects (i.e., 335-dimensional object behavior features), the target client passes the object statistical features associated with the above-mentioned operation objects through a statistical feature convolution layer, and splices them with the network layer output features of the multimedia resource recommendation model as the recommendation value estimation features of the multimedia resources.

[0161] Optional, see Figure 9After the target client generates the object statistical features associated with each of the operation objects, the object statistical features associated with each of the operation objects are interacted with the multimedia resource identification features to generate interactive features. The interactive features, the multimedia resource identification features, the embedded features of the object-side content-side cross features, the embedded features of the object-side features, and the embedded features of the content-side features are concatenated to serve as the network-layer embedded features of the multimedia resource recommendation model. The multimedia resource recommendation model concatenates the embedded features through the CONCAT layer and calls the EXPERT layer to output the feature values ​​of each expert network for the long click and the feature values ​​of each expert network for the completion degree based on the concatenated features of the previous layer. The multimedia resource recommendation model calls the EXPERT output weighted summation layer to perform weighted summation on the feature values ​​of the expert networks for long clicks and the feature values ​​of the expert networks for completion based on the feature values ​​of the expert networks for long clicks and the feature values ​​of the expert networks for completion, and generates feature summation values ​​of the expert networks for long clicks and completion respectively, and inputs the feature summation values ​​of the expert networks for long clicks and completion respectively into the LCR DNN and PCR DNN to generate the network layer output features of the multimedia resource recommendation model. After generating the object statistical features associated with the above-mentioned operation objects (i.e., 335-dimensional object behavior features), the target client passes the object statistical features associated with the above-mentioned operation objects through a statistical feature convolution layer, and splices them with the network layer output features of the multimedia resource recommendation model as the recommendation value estimation features of the multimedia resource. Through practice, it can be known that after the relevant policies are launched, the feature generation method for multimedia resource recommendation proposed in the embodiment of this application can bring multiple beneficial effects in the multimedia resource recommendation model. Taking Tencent News as an example, the launch of multiple strategies related to statistical features has brought about a significant increase of 19.33% in the average deep consumption VV per capita across all terminals, a significant increase of 1.79% in the average stay time per capita across all terminals, and a significant increase of 0.16% in the next-day retention rate. By adopting this application, the expression ability of the model can be greatly improved, so that the target client can recommend multimedia resources that the object is more interested in, thereby greatly improving business indicators. At the same time, the embodiment of this application systematically proposes a clear, comprehensive, and easily expandable statistical feature system that can be used in the multimedia resource recommendation model, which greatly improves the efficiency of computing resources and storage resources and reduces the cost of resource use.

[0162] The feature generation method for multimedia resource recommendation provided in the embodiments of this application has the advantages of systematicity, comprehensiveness, and scalability, improving the model's expressiveness and providing more personalized services for the user. Furthermore, the feature generation method for multimedia resource recommendation provided in the embodiments of this application has efficient computing and storage solutions, significantly improving the efficiency of computing resources, model training resources, and storage resources.

[0163] Based on the description of the above-mentioned feature generation method embodiment of multimedia resource recommendation, the present application embodiment also discloses a feature generation device for multimedia resource recommendation. The feature generation device for multimedia resource recommendation can be applied to Figures 1 to 9 In the feature generation method of the embodiment shown, the steps in the feature generation method for performing multimedia resource recommendation are used. Here, the feature generation device for multimedia resource recommendation can be the above-mentioned Figures 1 to 9 The service server or terminal device in the embodiment shown, that is, the feature generation device for multimedia resource recommendation can be the above-mentioned Figures 1 to 9 The execution body of the feature generation method for multimedia resource recommendation in the embodiment shown. Figure 10 , Figure 10 Schematic diagram of the structure of the feature generation device for multimedia resource recommendation provided in the embodiment of the present application. In the embodiment of the present application, the device can run the following modules:

[0164] An acquisition module 31 is configured to acquire first object behavior data provided by a client data source;

[0165] The acquisition module 31 is further configured to acquire the second object behavior data provided by the training sample data source recommended by the multimedia resource;

[0166] a feature generation module 32 for generating, based on the first object behavior data acquired by the acquisition module 31, at least two first object behavior statistical features for at least two different time periods associated with the first object, wherein the first objects are the objects carried in the first object behavior data, and the first object behavior statistical features for any time period associated with any first object include at least two attribute features of an object attribute feature, a duration attribute feature, a resource content attribute feature, a resource type attribute feature, and an object resource behavior attribute feature;

[0167] The feature generation module 32 is further configured to generate at least two second object behavior statistical features for at least two different time periods associated with the second object based on the second object behavior data acquired by the acquisition module, wherein the second object is each object carried in the first object behavior data, and at least two object resource behavior attribute features are included in the second object behavior statistical features for any time period associated with any second object;

[0168] The feature aggregation module 33 is used to generate object statistical features of each object based on the first object behavior statistical features associated with each of the above-mentioned first objects and the second object behavior statistical features associated with each of the above-mentioned second objects, and determine the above-mentioned object statistical features as input features of the multimedia resource recommendation model, wherein the above-mentioned multimedia resource recommendation model is trained by the above-mentioned second object behavior data, and the above-mentioned multimedia resource model is used to output the multimedia resource recommendation value of each of the above-mentioned objects based on the object statistical features of each of the above-mentioned objects.

[0169] According to the above Figure 2 The corresponding embodiment, Figure 2 The implementation method described in steps S101 to S103 of the feature generation method for multimedia resource recommendation shown in FIG. Figure 10 Each module of the device shown in FIG. Figure 2 The implementation method described in step S101 of the feature generation method for multimedia resource recommendation shown in FIG. Figure 10 The device shown in FIG. 1 is executed by the acquisition module 31 and the feature generation module 32. The implementation method described in step S102 can be executed by the acquisition module 31 and the feature generation module 32. The implementation method described in step S103 can be executed by the feature aggregation module 33. The implementation methods executed by the acquisition module 31, the feature generation module 32 and the feature aggregation module 33 can be referred to in the above. Figure 2 The implementation methods provided in each step of the corresponding embodiment will not be repeated here.

[0170] In an embodiment of the present application, a feature generation device for multimedia resource recommendation can receive first object behavior data provided by the target client data source and second object behavior data provided by a training sample data source for multimedia resource recommendation. When the acquisition module receives the first object behavior data, it triggers the first object behavior statistical feature generation module to acquire the first object behavior data provided by the client data source and, based on the first object behavior data, generate at least two first object behavior statistical features for at least two different time periods associated with each object carried in the first object behavior data. When the acquisition module receives the second object behavior data, it triggers the second object behavior statistical feature generation module to acquire second object behavior data provided by the training sample data source for multimedia resource recommendation and, based on the second object behavior data, generate at least two second object behavior statistical features for at least two different data segments associated with each object carried in the first object behavior data. When the target client detects the generation of the first and second object behavior statistical features, it triggers the aggregation module to generate object statistical features for each object based on the first and second object behavior statistical features associated with each object, and uses the object statistical features as input features for the multimedia resource recommendation model. The feature generation method for multimedia resource recommendation provided by the embodiment of the present application proposes a comprehensive, clear and easily expandable statistical feature system, while greatly improving the efficiency of computing resources, model training resources and storage resources, and improving the accuracy of recommendations and enhancing applicability.

[0171] In an embodiment of the present application, each module in the device shown in the above figure can be separately or all merged into one or several other modules to constitute, or a certain module (or modules) thereof can also be further split into multiple smaller modules in function to constitute, which can achieve the same operation without affecting the realization of the technical effect of the embodiment of the present application. The above modules are divided based on logical functions. In practical applications, the function of a module can also be realized by multiple modules, or the function of multiple modules can be realized by one module. In other feasible implementations of the present application, the above device can also include other modules. In practical applications, these functions can also be implemented with the assistance of other modules, and can be implemented by the collaboration of multiple modules, which is not limited here.

[0172] See Figure 11 , Figure 11 This is a schematic diagram of the structure of the computer device provided in the embodiment of the present application. As shown in FIG11 , the computer device 1000 can be the above-mentioned Figure 2-Figure 9The terminal device in the corresponding embodiment. The computer device 1000 may include: a processor 1001, a network interface 1004 and a memory 1005. In addition, the computer device 1000 may also include: a user interface 1003, and at least one communication bus 1002. The communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 may include a display screen (Display), a keyboard (Keyboard), and the user interface 1003 may optionally include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface). The memory 1005 may be a high-speed RAM memory, or a non-volatile memory (non-volatile memory), such as at least one disk memory. The memory 1005 may optionally also be at least one storage device located away from the aforementioned processor 1001. As Figure 11 As shown, the memory 1005 as a computer-readable storage medium may include an operating system, a network communication module, a user interface module, and a device control application.

[0173] The network interface 1004 in the computer device 1000 can also be connected to the above Figure 1 The terminal 200b in the corresponding embodiment is connected to the network, and the optional user interface 1003 may also include a display screen (Display) and a keyboard (Keyboard). Figure 11 In the computer device 1000 shown, the network interface 1004 can provide network communication functions; the user interface 1003 is mainly used to provide an input interface for developers; and the processor 1001 can be used to call the device control application stored in the memory 1005 to implement the above Figure 2 The feature generation method for multimedia resource recommendation in the corresponding embodiment.

[0174] It should be understood that the computer device 1000 described in the embodiment of the present application can execute the above Figure 2 The description of the feature generation method for multimedia resource recommendation in the corresponding embodiment will not be repeated here. In addition, the description of the beneficial effects of adopting the same method will not be repeated here either.

[0175] In addition, it should be pointed out that the embodiment of the present application further provides a computer-readable storage medium, and the computer-readable storage medium stores a computer program executed by the device for generating content recommendation based on the characteristics of multimedia resource recommendation mentioned above, and the computer program includes program instructions. When the processor executes the program instructions, the above-mentioned Figure 2The description of the feature generation method for multimedia resource recommendation in the corresponding embodiment will not be repeated here. In addition, the description of the beneficial effects of using the same method will not be repeated here. For technical details not disclosed in the computer-readable storage medium embodiment involved in this application, please refer to the description of the method embodiment of this application.

[0176] In addition, it should be noted that: the embodiment of the present application also provides a computer program product, which may include a computer program, and the computer program may be stored in a computer-readable storage medium. The processor of the computer device reads the computer program from the computer-readable storage medium, and the processor may execute the computer program, so that the computer device performs the above Figures 2 to 9 The description of the feature generation method for multimedia resource recommendation in the corresponding embodiment will not be repeated here. In addition, the description of the beneficial effects of using the same method will not be repeated here. For technical details not disclosed in the computer program product embodiments involved in this application, please refer to the description of the method embodiments of this application.

[0177] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing related hardware through a computer program. The above-described program can be stored in a computer-readable storage medium. When executed, the program can include the processes in the above-described method embodiments. The above-described storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0178] The above disclosure is only a preferred embodiment of the present application, and certainly cannot be used to limit the scope of rights of the present application. Therefore, equivalent changes made according to the claims of the present application are still within the scope covered by the present application.

Claims

1. A feature generation method for multimedia resource recommendation, characterized in that: The method comprises: Obtaining first object behavior data provided by a client data source, and generating at least two first object behavior statistical features of at least two different time periods associated with the first object based on the first object behavior data, wherein the first objects are each object carried in the first object behavior data, and the first object behavior statistical features of any time period associated with any first object include at least two attribute features of an object attribute feature, a duration attribute feature, a resource content attribute feature, a resource type attribute feature, and an object resource behavior attribute feature; Obtaining second object behavior data provided by a training sample data source for multimedia resource recommendation, and generating at least two second object behavior statistical features for at least two different time periods associated with a second object based on the second object behavior data, wherein the second objects are each object carried in the first object behavior data, and at least two object resource behavior attribute features are included in the second object behavior statistical features for any time period associated with any second object; generating object statistical features for each object based on the first object behavior statistical features associated with each first object and the second object behavior statistical features associated with each second object, and determining the object statistical features as input features of a multimedia resource recommendation model, wherein the multimedia resource recommendation model is trained using the second object behavior data; The multimedia resource recommendation model is used to splice the embedded features with cross-features of the object side and the content side, the embedded features of the object side features and the embedded features of the content side features through the CONCAT layer to obtain spliced ​​features; the multimedia resource recommendation model is also used to output feature values ​​for the spliced ​​features through multiple expert networks for long clicks, output feature values ​​for the spliced ​​features through multiple expert networks for completion, perform weighted summation on the feature values ​​output by multiple expert networks for long clicks through the EXPERT output weighted summation layer, obtain the feature sum value of multiple expert networks for long clicks, and perform weighted summation on the feature values ​​output by multiple expert networks for completion. The feature values ​​output by multiple expert networks are weightedly summed to obtain the feature sum values ​​of multiple expert networks for completion; the multimedia resource recommendation model is also used to generate network layer output features based on the feature sum values ​​of multiple expert networks for long clicks and the feature sum values ​​of multiple expert networks for completion through the multi-layer perceptron of the long click task layer and the multi-layer perceptron of the completion task layer; the multimedia resource recommendation model is also used to splice the features of the object statistical features processed by the statistical feature convolution layer with the network layer output features to obtain recommendation value estimation features, and the recommendation value estimation features are used to obtain multimedia resource recommendation values ​​for each object.

2. The method according to claim 1, characterized in that The generating, based on the first object behavior data, at least two first object behavior statistical features in at least two different time periods associated with the first object includes: generating at least two levels of object behavior features for at least two different time periods based on the generation time of each object behavior data carried in the first object behavior data, wherein the first level object behavior features are used to generate object behavior statistical features of each first object in a time period, and the generation time of the object behavior data included in different time periods do not overlap; Based on the object identifiers of each first object carried in the at least two levels of object behavior features, at least two first object behavior statistical features indexed by the object identifiers of each first object are generated, wherein any first object is associated with at least two first object behavior statistical features of the at least two different time periods indexed by the object identifier of any first object.

3. The method according to claim 2, characterized in that The generating of at least two levels of object behavior features in at least two different time periods based on the generation time of each object behavior data carried in the first object behavior data includes: Based on the generation time of each object behavior data carried in the first object behavior data, classifying the object behavior data of each first object belonging to the same time period into object behavior data of the same level to obtain at least two levels of object behavior data; Based on the object behavior data of each first object included in the object behavior data at each level, at least two attribute characteristics of the object attribute characteristics, duration attribute characteristics, resource content attribute characteristics, resource type attribute characteristics and object resource behavior attribute characteristics of each first object are generated, and the at least two attribute characteristics associated with each first object are spliced ​​to generate object behavior characteristics of each level corresponding to each time period, so as to obtain at least two levels of object behavior characteristics for at least two different time periods.

4. The method according to claim 3, characterized in that The object resource behavior attribute characteristics include one or more of the multimedia resource preference degree of the object or the object preference degree of the multimedia resource; Generating the object resource behavior attribute characteristics of each first object based on the object behavior data of each first object included in the object behavior data at each level includes: Obtaining any object behavior statistic of any first object for any multimedia resource in the object behavior data of any first object included in the object behavior data of each level, and determining a ratio of the any object behavior statistic to the sum of the object behavior statistic of all objects for the any multimedia resource in the object behavior data of each level as the preference degree of the any first object for the any multimedia resource to obtain the multimedia resource preference degree of the any first object; or Obtain any object behavior statistic of any first object in the object behavior data of any first object included in the object behavior data of each level for any multimedia resource, and determine the ratio of any object behavior statistic to the sum of the object behavior statistic of any first object in the object behavior data of each level for all multimedia resources as the preference degree of any first object for any multimedia resource to obtain the object preference degree of any multimedia resource.

5. The method according to claim 3, characterized in that The object resource behavior attribute characteristics include the expected number of clicks of the object on the multimedia resource; the object resource behavior attribute characteristics of each first object generated based on the object behavior data of each first object included in the object behavior data at each level include: Obtaining a first object behavior statistic of any first object on a first multimedia resource in the object behavior data of any first object included in the object behavior data of each level, and obtaining a first object behavior operation ratio of all objects on the first multimedia resource in the object behavior data of each level; Obtaining a statistical value of a second object behavior of any first object on a second multimedia resource in the object behavior data of any first object at each level, and obtaining a proportion of second object behavior operations of all objects on the second multimedia resource in the object behavior data of each level; Based on the first object behavior statistics, the first object behavior operation ratio, the second object behavior statistics and the second object behavior operation ratio, and the object behavior statistics of any object for all multimedia resources, the expected number of clicks of any first object on the multimedia resources is generated to obtain the expected number of clicks of each first object on the multimedia resources.

6. The method according to any one of claims 1 to 5, characterized in that The step of generating at least two second object behavior statistical features of at least two different time periods associated with each object carried in the first object behavior data based on the second object behavior data includes: Based on the generation time of each object behavior data carried in the second object behavior data, the second object behavior data is divided into a plurality of P1-level object behavior data, where the P1-level object behavior data is the object behavior data included in the minimum unit time length of the second object behavior data divided by the time interval; generating at least two pieces of P2-level object behavior data of at least two different time periods based on accumulation of a plurality of the P1-level object behavior data included in a plurality of minimum unit durations, wherein one piece of P2-level object behavior data is accumulated from a plurality of the P1-level object behavior data within a target duration, where the target duration is a positive integer multiple of the minimum unit duration; Based on the object identifier corresponding to each second object carried in each of the P2-level object behavior data, at least two second object behavior statistical features indexed by the object identifier of each second object are generated, wherein any second object is associated with at least two second object behavior statistical features of the at least two different time periods indexed by the object identifier of any second object.

7. The method according to claim 6, characterized in that The generating, based on the object identifiers corresponding to the respective second objects carried in the respective P2-level object behavior data, at least two second object behavior statistical features indexed by the object identifiers of the respective second objects includes: generating, based on the object identifiers corresponding to the respective second objects carried in the respective P2-level object behavior data, P2-level object behavior features associated with the respective second objects, wherein the P2-level object behavior features of a second object include at least two object resource behavior attribute features of the second object; Using the object identifier of each second object as an index, concatenating at least two object resource behavior attribute features in the P2-level object behavior features associated with the object identifier of each second object to generate a second object behavior statistical feature associated with each object identifier; The object resource behavior attribute characteristics include at least one object resource behavior-related characteristic of the object's click duration, number of clicks, click rate, number of exposures, viewing time, and interaction time for multimedia resources.

8. The method according to claim 7, characterized in that Generating the object statistical feature of each object based on the first object behavior statistical feature associated with each first object and the second object behavior statistical feature associated with each second object includes: Obtaining an object identifier corresponding to each first object from the first object behavior statistical feature, and obtaining an object identifier corresponding to each second object from the second object behavior statistical feature; Using the object identifiers of the first objects and the object identifiers of the second objects as indexes, the first object behavior statistical features and the second object behavior statistical features are spliced ​​to generate object statistical features corresponding to the respective object identifiers, wherein the first object behavior statistical features and the second object behavior statistical features with the same object identifier are spliced ​​into object statistical features of the same object.

9. A feature generation device for multimedia resource recommendation, characterized in that: include: An acquisition module, configured to acquire first object behavior data provided by a client data source; a feature generation module, configured to generate at least two first object behavior statistical features for at least two different time periods associated with the first object based on the first object behavior data acquired by the acquisition module, wherein the first objects are each object carried in the first object behavior data, and the first object behavior statistical features for any time period associated with any first object include at least two attribute features of an object attribute feature, a duration attribute feature, a resource content attribute feature, a resource type attribute feature, and an object resource behavior attribute feature; The acquisition module is further configured to acquire second object behavior data provided by a training sample data source recommended by the multimedia resource; The feature generation module is further configured to generate at least two second object behavior statistical features of at least two different time periods associated with the second object based on the second object behavior data acquired by the acquisition module, wherein the second objects are each object carried in the first object behavior data, and at least two object resource behavior attribute features are included in the second object behavior statistical features of any time period associated with any second object; a feature aggregation module, configured to generate object statistical features for each object based on the first object behavior statistical features associated with each first object and the second object behavior statistical features associated with each second object, and determine the object statistical features as input features of a multimedia resource recommendation model, wherein the multimedia resource recommendation model is trained using the second object behavior data; The multimedia resource recommendation model is used to splice the embedded features with cross-features of the object side and the content side, the embedded features of the object side features and the embedded features of the content side features through the CONCAT layer to obtain spliced ​​features; the multimedia resource recommendation model is also used to output feature values ​​for the spliced ​​features through multiple expert networks for long clicks, output feature values ​​for the spliced ​​features through multiple expert networks for completion, perform weighted summation on the feature values ​​output by multiple expert networks for long clicks through the EXPERT output weighted summation layer, obtain the feature sum value of multiple expert networks for long clicks, and perform weighted summation on the feature values ​​output by multiple expert networks for completion. The feature values ​​output by multiple expert networks are weightedly summed to obtain the feature sum values ​​of multiple expert networks for completion; the multimedia resource recommendation model is also used to generate network layer output features based on the feature sum values ​​of multiple expert networks for long clicks and the feature sum values ​​of multiple expert networks for completion through the multi-layer perceptron of the long click task layer and the multi-layer perceptron of the completion task layer; the multimedia resource recommendation model is also used to splice the features of the object statistical features processed by the statistical feature convolution layer with the network layer output features to obtain recommendation value estimation features, and the recommendation value estimation features are used to obtain multimedia resource recommendation values ​​for each object.

10. A computer device, characterized in that: include: processor, memory, and network interface; The processor is connected to the memory and the network interface, wherein the network interface is used to provide a data communication function, the memory is used to store program code, and the processor is used to call the program code to execute the method according to any one of claims 1 to 8.

11. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and the computer program is suitable for being loaded by a processor and executing the method according to any one of claims 1 to 8.

12. A computer program product, characterized in that The computer program product comprises a computer program stored in a computer-readable storage medium. The computer program is suitable for being read and executed by a processor, so that a computer device having the processor executes the method according to any one of claims 1 to 8.

Images