Resource recommendation method and device, electronic equipment and storage medium

Abstract

This disclosure provides a resource recommendation method, apparatus, electronic device, and storage medium, relating to the field of data processing, and particularly to the field of information recommendation. The specific implementation scheme is as follows: Multiple object features of a target object and resource features of multiple resources to be recommended are obtained; based on the object information content of the multiple object features, an object embedding vector is determined for each object feature category, the vector dimension of which corresponds to the object information content; based on the resource information content of the resource features of the multiple resources to be recommended, a resource embedding vector is determined for each resource feature category, the vector dimension of which corresponds to the resource information content; based on the object embedding vector corresponding to each object feature category and the resource embedding vector corresponding to each resource feature category, the target resource is recommended to the target object.

Description

Technical Field

[0001] This disclosure relates to the field of data processing technology, and more particularly to the field of information recommendation, specifically to resource recommendation methods, apparatus, electronic devices, and storage media. Background Technology

[0002] With the development of internet technology, the internet can provide users with an increasing number of web servers, for example, to recommend video resources. Currently, recommendation systems are commonly used to recommend resources to users. Recommendation systems can use ranking models without an embedding layer or ranking models with an embedding layer.

[0003] However, ranking models without an embedding layer, such as single-objective recommendation models, are ineffective in large-scale business scenarios, especially in recommendation systems where identifier-based features are the main feature system. Without an embedding layer, the ranking model cannot remember all samples, thus failing to accurately recommend resources to users.

[0004] Furthermore, current ranking models with embedding layers, such as recommendation systems commonly used in industry, have strong memory and representation capabilities for identifier-type features, and can also perform high-order feature extraction and combination. However, these models use embedding vectors of the same dimension. For low-frequency features with little information, the dimension of the embedding vector is too high, resulting in a waste of storage and computing resources. On the other hand, for high-frequency features or features with a lot of information, the dimension of the embedding vector is too low, resulting in a weak ability to represent discrete features. Summary of the Invention

[0005] This disclosure provides a resource recommendation method, apparatus, electronic device, and storage medium.

[0006] According to one aspect of this disclosure, a resource recommendation method is provided, comprising: acquiring multiple object features of a target object and multiple resource features of resources to be recommended; determining an object embedding vector corresponding to each object feature category based on the object information content of the multiple object features, wherein the object embedding vector is an embedding vector obtained by constructing vectors of the object features contained in each object feature category through an embedding layer, and the vector dimension of the object embedding vector corresponds to the object information content; determining a resource embedding vector corresponding to each resource feature category based on the resource information content of the resource features of the multiple resources to be recommended, wherein the resource embedding vector is an embedding vector obtained by constructing vectors of the resource features contained in each resource feature category through an embedding layer, and the vector dimension of the resource embedding vector corresponds to the resource information content; and recommending target resources to the target object based on the object embedding vector corresponding to each object feature category and the resource embedding vector corresponding to each resource feature category.

[0007] According to another aspect of this disclosure, a resource recommendation apparatus is provided, comprising: a feature acquisition module, configured to acquire multiple object features of a target object and multiple resource features of resources to be recommended; an object vector determination module, configured to determine an object embedding vector corresponding to each object feature category based on the object information content of the multiple object features, wherein the object embedding vector is an embedding vector obtained by constructing vectors of the object features contained in each object feature category through an embedding layer, and the vector dimension of the object embedding vector corresponds to the object information content; a resource vector determination module, configured to determine a resource embedding vector corresponding to each resource feature category based on the resource information content of the resource features of the multiple resources to be recommended, wherein the resource embedding vector is an embedding vector obtained by constructing vectors of the resource features contained in each resource feature category through an embedding layer, and the vector dimension of the resource embedding vector corresponds to the resource information content; and a resource recommendation module, configured to recommend target resources to the target object based on the object embedding vector corresponding to each object feature category and the resource embedding vector corresponding to each resource feature category.

[0008] According to another aspect of this disclosure, an electronic device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the resource recommendation method described above.

[0009] According to another aspect of this disclosure, a non-transitory computer-readable storage medium is provided storing computer instructions for causing a computer to perform the resource recommendation method described above.

[0010] According to another aspect of this disclosure, a computer program product is provided, including a computer program that, when executed by a processor, implements the resource recommendation method described above.

[0011] As can be seen from the above, in this disclosure, embedding vectors corresponding to object features are constructed based on the object information content of object features, and embedding vectors corresponding to resource features are constructed based on the resource information content of resource features. Moreover, the dimension of the embedding vector corresponding to the object feature is related to the object information content, and the dimension of the embedding vector corresponding to the resource feature is related to the resource information content. That is, in this disclosure, the dimension of the embedding vector corresponding to object features with different information content is different, and the resource content of the embedding vector corresponding to resource features with different information content is different. This allows features with high information content to be fully expressed, while features with low information content can also be expressed at a lower cost. This improves the ranking and recommendation effect while saving system resources, thereby achieving the goal of compatibility between feature representation ability and system resource overhead.

[0012] Therefore, the solution provided in this disclosure achieves the goal of accurately recommending resources to users, realizes the compatibility between feature representation capabilities and system resource overhead, and solves the problem in related technologies where feature representation capabilities and system resource overhead cannot be balanced.

[0013] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0014] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein:

[0015] Figure 1 This is a flowchart of a resource recommendation method according to an embodiment of the present disclosure;

[0016] Figure 2 This is a schematic diagram of a dual-tower coarse-ranking model in a resource recommendation system according to an embodiment of the present disclosure;

[0017] Figure 3 This is a schematic diagram of a resource recommendation device according to an embodiment of the present disclosure;

[0018] Figure 4 This is a schematic diagram of an electronic device according to an embodiment of the present disclosure. Detailed Implementation

[0019] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0020] It should be noted that the acquisition, storage, and application of user personal information involved in the technical solution disclosed herein comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0021] Example 1

[0022] According to one aspect of this disclosure, a resource recommendation method is provided, which can be executed on a server or a terminal device. In this embodiment, the terminal device is used as the execution subject for explanation and description. Figure 1 This is a flowchart of the resource recommendation method provided in this embodiment, such as... Figure 1 As shown, the method includes:

[0023] Step S102: Obtain multiple object features of the target object and multiple resource features of the resources to be recommended.

[0024] In step S102, the resource to be recommended can be video, text, image, audio, etc. The resource to be recommended can be stored in a cloud server. When the terminal device needs to recommend resources to the user (i.e., the target object), it can directly obtain the resource to be recommended from the cloud server. In this embodiment, video resources are used as an example for explanation.

[0025] Optionally, the FEED information stream is a continuously updated information stream that presents content to the user. For example, the FEED information stream will be continuously updated while the user refreshes a short video. The following explanation uses the process of a user refreshing a short video as an example. When a user (i.e., the target audience) refreshes a short video, the terminal device receives the user's resource refresh instruction and obtains the user's target characteristics, such as name, gender, occupation, and interests, through the user's registration information on the client playing the short video and / or the short videos published by the user through that client, and / or the user's registration information or published information on other clients. Simultaneously, the terminal device obtains the resource characteristics of multiple resources to be recommended from the resource server corresponding to that client, such as the resource identifier, resource content, resource author, and resource type (e.g., food video resources, funny video resources, story video resources, etc.) for each resource to be recommended.

[0026] Step S104: Based on the amount of object information possessed by multiple object features, determine the object embedding vector corresponding to each object feature category.

[0027] In step S104, the object embedding vector is an embedding vector obtained by constructing vectors from the object features contained in each object feature category through the embedding layer. The vector dimension of the object embedding vector corresponds to the object information content. Object features corresponding to the same object feature category have the same object information content within the same range. Object information content represents the frequency of use of the object feature or the amount of data contained in the object feature. In this embodiment, the example of representing the amount of data contained in an object feature by object information content is used for explanation. The scheme corresponding to the example of representing the amount of data contained in an object feature by object information content can also be applied to the scheme corresponding to the example of representing the frequency of use of an object feature by object information content.

[0028] Optionally, the vector dimension of the object embedding vector is positively correlated with the amount of information about the object. That is, when the amount of information about the object is large, the vector dimension of the object embedding vector corresponding to the object feature is large. For example, for features with less information such as gender, age, and occupation, a low-dimensional (e.g., 4-dimensional) embedding vector is constructed; for features with moderate information such as user interests and a list of installed applications, a medium-dimensional (e.g., 8-dimensional) embedding vector is constructed; and for features with more information such as user identifiers, a high-dimensional (e.g., 64-dimensional) embedding vector is constructed.

[0029] Step S106: Based on the amount of resource information possessed by the resource features of multiple resources to be recommended, determine the resource embedding vector corresponding to each resource feature category.

[0030] In step S106, the resource embedding vector is an embedding vector obtained by constructing vectors from the resource features included in each resource feature category through the embedding layer. The vector dimension of the resource embedding vector corresponds to the amount of resource information. Resource features corresponding to the same resource feature category have the same amount of resource information, which represents the frequency of use of the resource feature or the amount of data contained in the resource feature. In this embodiment, the example of using the amount of resource information to represent the amount of data contained in a resource feature is used for illustration. The scheme corresponding to the example of using the amount of resource information to represent the amount of data contained in a resource feature can also be applied to the scheme corresponding to the example of using the amount of resource information to represent the frequency of use of a resource feature.

[0031] Optionally, the vector dimension of the resource embedding vector is positively correlated with the amount of resource information. That is, when there is more resource information, the vector dimension of the resource embedding vector corresponding to the resource feature is larger. For example, for features with less information, such as resource playback duration, a low-dimensional (e.g., 4-dimensional) embedding vector is constructed; for features with moderate information, such as resource title, a medium-dimensional (e.g., 8-dimensional) embedding vector is constructed; and for features with more information, such as resource identifier and resource author identifier, a high-dimensional (e.g., 64-dimensional) embedding vector is constructed.

[0032] Step S108: Based on the object embedding vector corresponding to each object feature category and the resource embedding vector corresponding to each resource feature category, recommend target resources to the target object.

[0033] It should be noted that, as can be seen from steps S104 to S106, the dimension of the embedding vector used in this disclosure is not fixed, but adaptively changes according to the amount of feature information. This allows features with high information content to be fully expressed, while features with low information content can also be expressed at a lower cost. This improves the ranking and recommendation effect while saving system resources, thereby achieving the goal of compatibility between feature representation ability and system resource overhead.

[0034] Based on the scheme defined in steps S102 to S108 above, it can be seen that this disclosure adopts an adaptive dimension embedding vector approach. After obtaining multiple object features of the target object and multiple resource features of the resources to be recommended, the object embedding vector corresponding to each object feature category is determined based on the object information content of the multiple object features, and the resource embedding vector corresponding to each resource feature category is determined based on the resource information content of the resource features of the multiple resources to be recommended.

[0035] Then, based on the object embedding vector corresponding to each object feature category and the resource embedding vector corresponding to each resource feature category, target resources are recommended to the target object. Here, the object embedding vector is...

[0036] The embedding vector is obtained by constructing vectors of object features contained in each object feature category through the embedding layer. The vector dimension of the object embedding vector corresponds to the amount of object information. The resource embedding vector is obtained by constructing vectors of resource features contained in each resource feature category through the embedding layer. The vector dimension of the resource embedding vector corresponds to the amount of resource information.

[0037] It is readily apparent that, in this disclosure, embedding vectors corresponding to object features are constructed based on the object information content of object features, and embedding vectors corresponding to resource features are constructed based on the resource information content of resource features. Furthermore, the dimension of the embedding vector corresponding to the object feature is related to the object information content, and the dimension of the embedding vector corresponding to the resource feature is related to the resource information content. That is, in this disclosure…

[0038] In this process, the dimensions of the embedding vectors corresponding to object features with different information content are different, and the resources of the embedding vectors corresponding to resource features with different information content are different. This allows features with high information content to be fully expressed, while features with low information content can also be expressed at a lower cost. This improves the ranking and recommendation effect while saving system resources, thus achieving the goal of compatibility between feature representation ability and system resource overhead.

[0039] Therefore, the solution provided in this disclosure achieves the goal of accurately recommending resources to users, realizes the compatibility between feature representation capabilities and system resource overhead, and thus solves the related technical problems.

[0040] During the procedure, there is a problem of not being able to balance feature representation capabilities with system resource overhead.

[0041] Example 2

[0042] According to one aspect of this disclosure, a method for generating a recognition model is also provided. In this embodiment...

[0043] In the example, combined Figure 1 The flowchart system for the generation method of the provided recognition model provides a detailed explanation of each step mentioned in Example 1, Section 5.

[0044] In one alternative embodiment, such as Figure 1 As shown, after obtaining multiple object features of the target object and multiple resource features of the resources to be recommended, the terminal device executes step S104, that is, based on the amount of object information possessed by the multiple object features, it determines the object embedding vector corresponding to each object feature category.

[0045] Specifically, the terminal device performs clustering processing on multiple object features based on the amount of object information they possess, resulting in multiple object feature categories, and determines the vector dimension corresponding to each object feature category, resulting in multiple object feature dimensions; then, based on the multiple object feature dimensions, it merges the object features contained in the corresponding object feature categories to obtain the object embedding vector corresponding to each object feature category.

[0046] Optionally, different object features have different amounts of object information. For example, the object information content corresponding to the gender feature is 2, i.e., male and female. Therefore, the terminal device can perform clustering processing based on the statistical object information content of each object feature to obtain multiple object feature categories. In this embodiment, the multiple object feature categories may include a first object feature category, a second object feature category, and a third object feature category. Among them, the first object feature category corresponds to less object information, such as gender, age, occupation, etc.; the second object feature category corresponds to a medium amount of object information, such as the user's interests and hobbies, etc.; and the third object feature category corresponds to more object information, such as the user's identifier, etc.

[0047] It should be noted that constructing object embedding vectors of corresponding dimensions for different object features not only saves the computational cost of the resource recommendation system, but also improves the system's memory and representation of important features, thereby improving the recommendation accuracy of the resource recommendation model.

[0048] In one alternative embodiment, such as Figure 1 As shown, after obtaining multiple object features of the target object and multiple resource features of the resources to be recommended, the terminal device executes step S106, that is, based on the amount of resource information possessed by the resource features of the multiple resources to be recommended, the resource embedding vector corresponding to each resource feature category is determined.

[0049] Specifically, the terminal device clusters multiple resource features based on the amount of resource information they contain, resulting in multiple resource feature categories, and determines the vector dimension corresponding to each resource feature category, resulting in multiple resource feature dimensions. Then, based on the multiple resource feature dimensions, the resource features contained in the corresponding resource feature categories are merged to obtain the resource embedding vector corresponding to each resource feature category.

[0050] Optionally, different resource features possess different amounts of resource information. Therefore, the terminal device can statistically analyze the amount of resource information for each resource feature and then perform clustering to obtain multiple resource feature categories. In this embodiment, the multiple resource feature categories may include a first resource feature category, a second resource feature category, and a third resource feature category. The first resource feature category corresponds to a smaller amount of resource information, such as resource playback duration; the second resource feature category corresponds to a medium amount of resource information, such as resource title; and the third resource feature category corresponds to a larger amount of resource information, such as resource identifier and resource author identifier.

[0051] It should be noted that constructing resource embedding vectors of corresponding dimensions for different resource features not only saves the computational cost of the resource recommendation system, but also improves the system's memory and representation of important features, thereby improving the recommendation accuracy of the resource recommendation model.

[0052] Furthermore, such as Figure 1 As shown, after obtaining the object embedding vector corresponding to each object feature category and the resource embedding vector corresponding to each resource feature category, the terminal device executes step S106 to recommend target resources to the target object based on the object embedding vector corresponding to each object feature category and the resource embedding vector corresponding to each resource feature category.

[0053] Specifically, the terminal device processes the object embedding vector corresponding to each object feature category through a first neural network to obtain object combination features; and processes the resource embedding vector corresponding to each resource feature category through a second neural network to obtain resource combination features; then, the object combination features and resource combination features are processed by inner product processing to obtain the completion rate corresponding to multiple resources to be recommended; finally, the target resources are recommended to the target objects based on the completion rate.

[0054] It should be noted that the first neural network and the second neural network described above have the same network structure. Optionally, Figure 2 A schematic diagram of a dual-tower coarse-ranking model in an alternative resource recommendation system is shown, by... Figure 2It can be seen that the dual-tower coarse-sorting model consists of a first embedding layer, a second embedding layer, a first neural network, and a second neural network. The first embedding layer is the embedding layer corresponding to the object embedding vector, and the second embedding layer is the embedding layer corresponding to the resource embedding vector.

[0055] Furthermore, the completion rate represents the probability that a recommended video resource will be played completely, reflecting the user's satisfaction with the recommended video resource. In practical applications, if the user's actual viewing time is greater than or equal to the product of the video resource's duration and a first preset coefficient a (e.g., 80%), the video resource is determined to have been played completely; if the user's actual viewing time is less than the product of the video resource's duration and a second preset coefficient b (e.g., 20%), the video resource is determined to have not been played. Where a + b = 1.

[0056] Optionally, the definition of a video resource completion playback can be expressed by the following formula:

[0057]

[0058] Here, play_dur represents the actual viewing time of the user, and video_dur represents the duration of the video resource.

[0059] Optional, such as Figure 2 As shown, the terminal device outputs the object embedding vector generated by the first embedding layer to the first neural network of the dual-tower coarse-ranking model, and the resource embedding vector generated by the second embedding layer to the second neural network of the dual-tower coarse-ranking model. After processing the object embedding vector and the resource embedding vector respectively, the two neural network models obtain vector combination features and resource embedding vectors.

[0060] The source combination features are used to calculate the inner product of the vector combination features and the resource combination features, thereby obtaining the completion rate of each resource to be recommended, and then recommending video resources to users based on the completion rate.

[0061] It should be noted that using a dual-tower coarse-ranking model to calculate the completion rate and then recommending video resources to users based on the completion rate can improve the efficiency and accuracy of matching the recommended resources with the target objects, thereby improving both the accuracy and efficiency of resource recommendation.

[0062] Furthermore, after determining the completion rate, the terminal device can then send data to the target audience based on the completion rate.

[0063] For example, recommending target resources. Specifically, the terminal device sorts multiple resources to be recommended based on their completion rate, obtains the ranking results, and recommends target resources to the target audience based on the ranking results.

[0064] It should be noted that, since the completion rate reflects the degree of user interest in a video resource,

[0065] Therefore, prioritizing recommended resources based on completion rate and then recommending five target resources to users based on the ranking results can prioritize recommending resources that users are most interested in, thereby improving user engagement with the product.

[0066] User experience on the client side.

[0067] In one alternative embodiment, the terminal device may respond to a resource update request from a target object and recommend multiple resources to be recommended to the target object in sequence according to the sorting results.

[0068] It should be noted that resources are recommended to users in order of sorting, prioritizing the resources that users are most interested in, thereby improving the user experience of the client.

[0069] Optionally, the aforementioned resource update request can be generated by user interaction, such as when a user switches video resources by swiping the client's user interface; alternatively, the aforementioned resource update request can also be generated based on the client playing the video.

[0070] The duration of the resource is generated based on the total duration of the video resources. For example, if the total duration of video resource 1 is 5.1 minutes, then when video resource 1 is played for 55 seconds, if the user does not perform any operation on video resource 1, the client generates a resource update request so that the client can automatically play the next recommended video resource.

[0071] The following explanation uses a user refreshing a short video feed via a client as an example. When a user refreshes the feed, the client on the terminal device first sends a request to the resource aggregation module in the resource recommendation system. The resource aggregation module then sends requests to various downstream queues to recall resources. Each queue recalls resources using different strategies based on their attributes. After scoring through stages such as coarse ranking, fine ranking, and re-ranking in the sorting phase, the resources and scores are returned to the aggregation module. The aggregation module performs deduplication, fusion, filtering, and sorting operations on the resources from each queue, generating the final distribution list and returning it to the client. The client then presents the final resource results that match the user's interests. Notably, the aggregation module and queues in the resource recommendation system are designed and optimized for a single product. Only the sorting phase utilizes machine learning and deep learning technologies with strong generalization capabilities, effectively combining the characteristics of different targets to improve sorting accuracy. Furthermore, a single sorting model models multiple targets within the recommendation scenario simultaneously, significantly saving machine resources.

[0072] In one alternative embodiment, such as Figure 2As shown, the first neural network includes at least a first factorization machine layer and a first fully connected layer. Specifically, in the process of processing the object embedding vectors corresponding to each object feature category through the first neural network to obtain object combination features, the terminal device performs inner product calculation on the object embedding vectors corresponding to each object feature category through the first factorization machine layer to obtain first cross features, and performs multiple gradient fitting processes on the object embedding vectors corresponding to each object feature category through the first fully connected layer to obtain first fitted features; then, the first cross features and the first fitted features are concatenated to obtain the object combination features.

[0073] Optionally, the first factorization machine layer can be an FM (Factorization Machines) layer, and the first fully connected layer consists of three layers, namely Layer-1, Layer-2 and Layer-3.

[0074] Specifically, after encoding, the training samples become offline sparse features, which are then input into the embedding layer. Based on the different information content of the features, they are transformed into dense feature vectors (i.e., embedding vectors) of different dimensions, and then input into the FM layer and the first fully connected layer, respectively. The feature vectors input to the FM layer are multiplied by different features, and the output inner product is the weight of the cross-features to extract second-order cross-feature information, thus obtaining the first cross-feature. The feature vectors input to the first fully connected layer are fitted with a function between the predicted target and the input vector through multi-layer gradient feedback to extract higher-order implicit information of the features, thus obtaining the first fitted feature. Finally, the first cross-feature and the first fitted feature are concatenated at the top layer of the layer to obtain the combined object feature.

[0075] It should be noted that the solution provided in this embodiment constructs a neural network dual-tower model for predicting the completion target through an embedding layer, a fully connected layer, and an FM layer. By constructing embedding layers of different dimensions to adaptively learn various information features, the prediction accuracy is improved while saving a lot of system resources.

[0076] In another alternative embodiment, such as Figure 2 As shown, the second neural network includes at least a second factorization layer and a second fully connected layer. Specifically, in the process of processing the resource embedding vectors corresponding to each resource feature category through the second neural network to obtain resource combination features, the terminal device performs inner product calculation on the resource embedding vectors corresponding to each resource feature category through the second factorization layer to obtain the second cross feature; and performs multiple gradient fitting processes on the resource embedding vectors corresponding to each resource feature category through the second fully connected layer to obtain the second fitted feature; finally, the second cross feature and the second fitted feature are concatenated to obtain the resource combination feature.

[0077] It should be noted that the process by which the terminal device processes the resource embedding vector through the second neural network is similar to the process by which it processes the object embedding vector through the first neural network; the only difference is the type of embedding vector being processed. The process by which the object embedding vector is processed through the first neural network has been described in detail above and will not be repeated here.

[0078] As can be seen from the above, this disclosure provides a resource recommendation method that can solve the problem that the embedding layer cannot fully represent features with different information content in the coarse ranking stage of the recommendation system. It proposes an adaptive dimension embedding layer, which can fully express features with high information content while also representing features with low information content at a lower cost, thereby improving the coarse ranking recommendation effect and saving system resources.

[0079] Example 3

[0080] According to one aspect of this disclosure, a resource recommendation device is also provided, wherein, Figure 3 This is a schematic diagram of the resource recommendation device, such as... Figure 3 As shown, the device includes: a feature acquisition module 301, an object vector determination module 303, a resource vector determination module 305, and a resource recommendation module 307.

[0081] The feature acquisition module 301 is used to acquire multiple object features of the target object and multiple resource features of the resources to be recommended; the object vector determination module 303 is used to determine the object embedding vector corresponding to each object feature category based on the object information content of the multiple object features, wherein the object embedding vector is an embedding vector obtained by constructing vectors of the object features contained in each object feature category through the embedding layer, and the vector dimension of the object embedding vector corresponds to the object information content; the resource vector determination module 305 is used to determine the resource embedding vector corresponding to each resource feature category based on the resource information content of the resource features of the multiple resources to be recommended, wherein the resource embedding vector is an embedding vector obtained by constructing vectors of the resource features contained in each resource feature category through the embedding layer, and the vector dimension of the resource embedding vector corresponds to the resource information content; the resource recommendation module 307 is used to recommend target resources to the target object based on the object embedding vector corresponding to each object feature category and the resource embedding vector corresponding to each resource feature category.

[0082] It should be noted that the feature acquisition module 301, object vector determination module 303, resource vector determination module 305, and resource recommendation module 307 mentioned above correspond to steps S102 to S108 of the above embodiments. The four modules and the corresponding steps implement the same instances and application scenarios, but are not limited to the content disclosed in the above embodiments.

[0083] Optionally, the object vector determination module includes: a first clustering module, a first dimension determination module, and a first merging module. The first clustering module is used to cluster multiple object features based on the amount of object information they possess, obtaining multiple object feature categories; the first dimension determination module is used to determine the vector dimension corresponding to each object feature category, obtaining multiple object feature dimensions; the first merging module is used to merge the object features contained in the corresponding object feature categories based on the multiple object feature dimensions, obtaining the object embedding vector corresponding to each object feature category.

[0084] Optionally, the resource vector determination module includes: a second clustering module, a second dimension determination module, and a second merging module. The second clustering module is used to cluster multiple resource features based on the amount of resource information they possess, resulting in multiple resource feature categories. The second dimension determination module is used to determine the vector dimension corresponding to each resource feature category, resulting in multiple resource feature dimensions. The second merging module is used to merge the resource features contained in the corresponding resource feature categories based on the multiple resource feature dimensions, resulting in a resource embedding vector corresponding to each resource feature category.

[0085] Optionally, the resource recommendation module includes: a first processing module, a second processing module, a first calculation module, and a first recommendation module. The first processing module processes the object embedding vector corresponding to each object feature category using a first neural network to obtain object combination features; the second processing module processes the resource embedding vector corresponding to each resource feature category using a second neural network to obtain resource combination features, wherein the first neural network and the second neural network have the same network structure; the first calculation module performs inner product processing on the object combination features and the resource combination features to obtain the completion rates corresponding to multiple resources to be recommended, wherein the completion rate represents the probability that the video resource to be recommended will be completed; the first recommendation module recommends target resources to target objects based on the completion rates.

[0086] Optionally, the first recommendation module includes a sorting module and a second recommendation module. The sorting module is used to sort multiple resources to be recommended based on their completion rate, obtaining a sorting result; the second recommendation module is used to recommend target resources to the target audience based on the sorting result.

[0087] Optionally, the first neural network includes at least a first factorization machine layer and a first fully connected layer, and the first processing module includes a second calculation module, a first fitting module, and a first concatenation module. Specifically, the second calculation module is used to calculate the inner product of the object embedding vectors corresponding to each object feature category through the first factorization machine layer to obtain the first cross feature; the first fitting module is used to perform multiple gradient fitting processes on the object embedding vectors corresponding to each object feature category through the first fully connected layer to obtain the first fitted feature; and the first concatenation module is used to concatenate the first cross feature and the first fitted feature to obtain the object combination feature.

[0088] Optionally, the second neural network includes at least a second factorization layer and a second fully connected layer, and the second processing module includes a third calculation module, a second fitting module, and a second concatenation module. Specifically, the third calculation module is used to calculate the inner product of the resource embedding vectors corresponding to each resource feature category through the second factorization layer to obtain the second cross feature; the second fitting module is used to perform multiple gradient fitting processes on the resource embedding vectors corresponding to each resource feature category through the second fully connected layer to obtain the second fitted feature; and the second concatenation module is used to concatenate the second cross feature and the second fitted feature to obtain the resource combination feature.

[0089] Optionally, the second recommendation module includes a third recommendation module, which is used to respond to the resource update request of the target object and recommend multiple resources to be recommended to the target object in sequence according to the ranking result.

[0090] Example 4

[0091] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0092] Figure 4 A schematic block diagram of an example electronic device 400 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0093] like Figure 4As shown, device 400 includes a computing unit 401, which can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) 402 or a computer program loaded from storage unit 408 into random access memory (RAM) 403. RAM 403 may also store various programs and data required for the operation of device 400. The computing unit 401, ROM 402, and RAM 403 are interconnected via bus 404. Input / output (I / O) interface 405 is also connected to bus 404.

[0094] Multiple components in device 400 are connected to I / O interface 405, including: input unit 406, such as keyboard, mouse, etc.; output unit 407, such as various types of monitors, speakers, etc.; storage unit 408, such as disk, optical disk, etc.; and communication unit 409, such as network card, modem, wireless transceiver, etc. Communication unit 409 allows device 400 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0095] The computing unit 401 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 401 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 401 performs the various methods and processes described above, such as the resource recommendation method. For example, in some embodiments, the resource recommendation method may be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 408. In some embodiments, part or all of the computer program may be loaded and / or installed on device 400 via ROM 402 and / or communication unit 409. When the computer program is loaded into RAM 403 and executed by the computing unit 401, one or more steps of the resource recommendation method described above may be performed. Alternatively, in other embodiments, the computing unit 401 may be configured to perform the resource recommendation method by any other suitable means (e.g., by means of firmware).

[0096] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0097] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0098] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0099] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0100] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with embodiments of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0101] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0102] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0103] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A resource recommendation method, comprising: The system obtains multiple object features of the target object and multiple resource features of the resources to be recommended. The object features are obtained through the target object's registration information on the client playing short videos, the short videos published on the client, and the publication information on other clients. Based on the object information content of the multiple object features, the multiple object features are clustered to obtain multiple object feature categories, and the object information content represents the usage frequency of the object features. Based on multiple object feature dimensions, the object features contained in the corresponding object feature categories are merged to obtain the object embedding vector corresponding to each object feature category. The object feature dimension is determined according to the vector dimension corresponding to the object feature category, and the vector dimension of the object embedding vector corresponds to the object information content. The object embedding vector corresponding to each object feature category is processed by the first neural network to obtain the object combination feature. The object combination feature is obtained by concatenating the first cross feature and the first fitting feature. The first cross feature is obtained by calculating the inner product of the object embedding vectors, and the first fitting feature is obtained by performing multiple gradient fitting processes on the object embedding vectors corresponding to each object feature category. Based on the resource information content of the resource features, a resource embedding vector corresponding to each resource feature category is determined. The resource embedding vector is obtained by constructing vectors of the resource features contained in each resource feature category through an embedding layer. The vector dimension of the resource embedding vector corresponds to the resource information content. The completion rates of multiple resources to be recommended are obtained by performing inner product processing on the object combination features and resource combination features. Generate a resource update request based on the duration of video playback on the client and the total duration of the video resources; In response to the resource update request, target resources are recommended to the target object based on the completion rate.

2. The method according to claim 1, wherein, Based on the amount of object information possessed by the multiple object features, determining the object embedding vector corresponding to each object feature category further includes: The vector dimension corresponding to each object feature category is determined to obtain the multiple object feature dimensions.

3. The method according to claim 1, wherein, Based on the amount of resource information possessed by the resource features of the multiple resources to be recommended, a resource embedding vector corresponding to each resource feature category is determined, including: Based on the amount of resource information possessed by multiple resource features, clustering is performed on the multiple resource features to obtain multiple resource feature categories; Determine the vector dimension corresponding to each resource feature category to obtain multiple resource feature dimensions; Based on the multiple resource feature dimensions, the resource features contained in the corresponding resource feature categories are merged to obtain the resource embedding vector corresponding to each resource feature category.

4. The method according to claim 1, wherein, Recommending the target resource to the target object based on the completion rate includes: The multiple resources to be recommended are sorted according to the completion rate to obtain the sorting result; The target resource is recommended to the target object based on the ranking result.

5. The method according to claim 1, wherein, The first neural network includes at least a first factorization machine layer and a first fully connected layer. It processes the object embedding vector corresponding to each object feature category to obtain combined object features, including: The first cross feature is obtained by performing an inner product calculation on the object embedding vector corresponding to each object feature category through the first factorization machine layer; The first fully connected layer performs multiple gradient fitting processes on the object embedding vector corresponding to each object feature category to obtain the first fitted feature.

6. The method according to claim 1, wherein, The second neural network includes at least a second factorization machine layer and a second fully connected layer. It processes the resource embedding vectors corresponding to each resource feature category to obtain resource combination features, including: The second cross feature is obtained by performing an inner product calculation on the resource embedding vector corresponding to each resource feature category through the second factorization machine layer; The second fully connected layer performs multiple gradient fitting processes on the resource embedding vector corresponding to each resource feature category to obtain the second fitted feature.

7. The method according to claim 4, wherein, Recommend target resources to the target object based on the ranking results, including: In response to the resource update request of the target object, the plurality of resources to be recommended are recommended to the target object in sequence according to the sorting result.

8. A resource recommendation device, comprising: The feature acquisition module is used to acquire multiple object features of the target object and multiple resource features of the resources to be recommended. The object features are acquired through the target object's registration information on the client playing short videos, the short videos published on the client, and the publication information on other clients. The object vector determination module is used to cluster multiple object features based on the object information content of the multiple object features to obtain multiple object feature categories, wherein the object information content represents the usage frequency of the object features; and to merge the object features contained in the corresponding object feature categories based on multiple object feature dimensions to obtain an object embedding vector corresponding to each object feature category, wherein the object feature dimensions are determined according to the vector dimensions corresponding to the object feature categories, and the vector dimensions of the object embedding vector correspond to the object information content. The first processing module is used to process the object embedding vector corresponding to each object feature category through a first neural network to obtain object combination features. The object combination features are obtained by concatenating a first cross feature and a first fitting feature. The first cross feature is obtained by calculating the inner product of the object embedding vectors, and the first fitting feature is obtained by performing multiple gradient fitting processes on the object embedding vectors corresponding to each object feature category. The resource vector determination module is used to determine the resource embedding vector corresponding to each resource feature category based on the resource information content of the resource features. The resource embedding vector is an embedding vector obtained by constructing vectors of the resource features contained in each resource feature category through an embedding layer. The vector dimension of the resource embedding vector corresponds to the resource information content. The first calculation module is used to perform inner product processing on the object combination features and resource combination features to obtain the completion rate corresponding to the multiple resources to be recommended. The first recommendation module is used to generate a resource update request based on the duration of video resource playback on the client and the total duration of video resources; in response to the resource update request, it recommends target resources to the target object based on the completion rate.

9. An electronic device, comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the resource recommendation method according to any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to execute the resource recommendation method according to any one of claims 1 to 7.

11. A computer program product comprising a computer program that, when executed by a processor, implements the resource recommendation method according to any one of claims 1 to 7.

Images