Feature extraction method, device, computer device and storage medium

Abstract

The present application relates to a feature extraction method, apparatus, computer device, storage medium, and computer program product. The method includes: obtaining an object set including a plurality of objects; the object types of the plurality of objects being the same; the object type being one of the objects to be recommended or the recommended recipient objects in a recommendation scenario; for a target object in the object set, screening out neighbor objects similar to the target object from the object set, and determining the respective initial neighbor object features of the neighbor objects of the target object; respectively determining the aggregated neighbor object feature and the neighbor object environment feature of the neighbor objects of the target object based on the initial neighbor object features; determining the initial target object feature of the target object, and extracting a compatible feature from the initial target object feature that is compatible with the neighbor object environment feature; performing feature fusion based on the aggregated neighbor object feature and the compatible feature to obtain the enhanced feature of the target object. Using this method can enrich the information content contained in the features.

Description

Technical Field

[0001] The present application relates to the field of computer technologies, and in particular, to a feature extraction method, apparatus, computer device, and storage medium. Background Art

[0002] With the development of science and technology, a video playback platform can provide a function of video recommendation or a function of live broadcast room recommendation to users. When making a video recommendation or a live broadcast room recommendation to a user, an object to be recommended can be obtained, that is, a video or a live broadcast room to be recommended is obtained, and feature extraction is performed on the object to be recommended. Based on the extracted features, a target object is screened out from the objects to be recommended, and the screened target object is recommended to the user.

[0003] When feature extraction needs to be performed on an object to be recommended, it is often to digitize the content of the object to be recommended through a feature engineering technology in traditional machine learning technology and convert it into a vector form. However, the features proposed by the feature engineering technology have the disadvantage of containing less information. Summary of the Invention

[0004] Based on this, in view of the above technical problems, it is necessary to provide a feature extraction method, apparatus, computer device, computer-readable storage medium, and computer program product that can improve the information content contained in features.

[0005] In a first aspect, the present application provides a feature extraction method, and the method includes:

[0006] Obtain an object set including multiple objects; the multiple objects have the same object type; the object type is one of the objects to be recommended or recommended receiving objects in a recommendation scenario;

[0007] For a target object in the object set, screen out neighbor objects similar to the target object from the object set, and determine initial neighbor object features of the neighbor objects of the target object respectively;

[0008] Based on the initial neighbor object features, determine an aggregated neighbor object feature and a neighbor object environment feature of the neighbor objects of the target object respectively; the aggregated neighbor object feature represents the aggregation of the features contributed by the neighbor objects of the target object respectively; the neighbor object environment feature represents the overall environment feature constituted by the neighbor objects of the target object;

[0009] Determine an initial target object feature of the target object, and extract a compatible feature compatible with the neighbor object environment feature from the initial target object feature;

[0010] Perform feature fusion based on the aggregated neighbor object feature and the compatible feature to obtain an enhanced feature of the target object.

[0011] In one embodiment, screening out neighbor objects similar to the target object from the object set includes:

[0012] Obtaining the preference features and attribute information corresponding to each object in the object set;

[0013] For the target object and the remaining objects except the target object in the object set, determining the similarity between the preference features of the target object and the preference features of each of the remaining objects, and obtaining the preference similarity corresponding to each of the remaining objects;

[0014] Determining the similarity between the attribute information of the target object and the attribute information of each of the remaining objects, and obtaining the attribute similarity corresponding to each of the remaining objects;

[0015] According to each of the preference similarities and each of the attribute similarities, screening out neighbor objects similar to the target object from the remaining objects in the object set except the target object.

[0016] In one embodiment, according to each of the preference similarities and each of the attribute similarities, screening out neighbor objects similar to the target object from the remaining objects in the object set except the target object includes:

[0017] Performing similarity fusion processing on the preference similarity and the attribute similarity belonging to the same remaining object to obtain the target similarity corresponding to each of the remaining objects;

[0018] Taking the remaining objects in the object set whose target similarity meets the preset similarity condition as the neighbor objects of the target object.

[0019] In a second aspect, the present application further provides a feature extraction device, and the device includes:

[0020] A neighbor object determination module, configured to obtain an object set including a plurality of objects; the object types of the plurality of objects are the same; the object type is one of the objects to be recommended or the recommended receiving object in the recommendation scenario; for the target object in the object set, screening out neighbor objects similar to the target object from the object set, and determining the initial neighbor object features of each of the neighbor objects of the target object;

[0021] An aggregation gate module, configured to determine the aggregated neighbor object feature of the target object based on the initial neighbor object feature; the aggregated neighbor object feature represents the aggregation of the features contributed by each of the neighbor objects of the target object;

[0022] A filtering door module for determining the neighbor object environment feature of the target object based on the initial neighbor object feature; the neighbor object environment feature characterizes the overall environment feature of the object environment constituted by the neighbor objects of the target object.

[0023] A feature fusion module for determining the initial target object feature of the target object, extracting a compatible feature from the initial target object feature that is compatible with the neighbor object environment feature; performing feature fusion based on the neighbor object aggregation feature and the similar feature to obtain the enhanced feature of the target object.

[0024] In one embodiment, the neighbor object determination module is further configured to obtain the preference feature and attribute information corresponding to each object in the object set; for the target object and the remaining objects other than the target object in the object set, determine the similarity between the preference feature of the target object and the preference feature of each of the remaining objects to obtain the preference similarity corresponding to each of the remaining objects; determine the similarity between the attribute information of the target object and the attribute information of each of the remaining objects to obtain the attribute similarity corresponding to each of the remaining objects; and screen out the neighbor objects similar to the target object from the remaining objects other than the target object in the object set according to the preference similarities and the attribute similarities.

[0025] In one embodiment, the feature extraction device is further configured to obtain the preference information corresponding to each object in the object set and determine the preference vector corresponding to each preference information; obtain a preset dimensionality transformation matrix, and perform dimensionality reduction processing on each preference vector through the dimensionality transformation matrix to obtain the preference feature corresponding to each object.

[0026] In one embodiment, the attribute information includes a plurality of attribute sub-information; the feature extraction device is further configured to, for each of the remaining objects in the object set, determine the similarity between each attribute sub-information of the target object and the corresponding attribute sub-information of the current remaining object to obtain a plurality of sub-information similarities; perform an averaging process on the plurality of sub-information similarities to obtain the attribute similarity between the attribute information of the target object and the attribute information of the current remaining object.

[0027] In one embodiment, the attribute sub-information includes an attribute name and an attribute value; the feature extraction device is further configured to, for the plurality of attribute sub-information of the target object, determine the target attribute sub-information in the current remaining object that has the same attribute name as the current attribute sub-information of the target object; determine the similarity between the attribute value of the target attribute sub-information of the current remaining object and the attribute value of the current attribute sub-information of the target object to obtain the sub-information similarity.

[0028] In one embodiment, the neighbor object determination module is further configured to perform similarity fusion processing on the preference similarity and the attribute similarity that belong to the same remaining object, so as to obtain a target similarity corresponding to each of the remaining objects; and use the remaining objects in the object set whose target similarity meets a preset similarity condition as the neighbor objects of the target object.

[0029] In one embodiment, the neighbor object determination module is further configured to obtain the preference feature corresponding to each neighbor object of the target object, and obtain the attribute feature corresponding to each neighbor object of the target object; and perform feature fusion processing on the preference feature and the attribute feature corresponding to each neighbor object to obtain an initial neighbor object feature corresponding to each neighbor object of the target object.

[0030] In one embodiment, the feature extraction device is further configured to obtain the attribute information of the neighbor object; the attribute information includes a plurality of attribute sub-information; perform two-way interactive fusion processing and linear fusion processing on the plurality of attribute sub-information of the neighbor object to obtain corresponding two-way interactive fusion results and linear fusion results; and perform result fusion processing on the two-way interactive fusion result and the linear fusion result to obtain the attribute feature of the neighbor object.

[0031] In one embodiment, the feature extraction device is further configured to determine an attribute vector corresponding to each attribute sub-information of the neighbor object; perform vector fusion processing on every two of the plurality of attribute vectors to obtain a plurality of attribute fusion vectors, and perform superposition processing on the plurality of attribute fusion vectors to obtain the two-way interactive fusion result of the neighbor object.

[0032] In one embodiment, the aggregation gate module is further configured to determine a feature contribution weight corresponding to each neighbor object of the target object; and perform weighted average processing on the initial neighbor object features of the neighbor object according to the feature contribution weight to obtain the neighbor object aggregation feature of the target object.

[0033] In one embodiment, the aggregation gate module is further configured to determine the initial target object feature of the target object; splice the initial target object feature with the initial neighbor object features corresponding to each neighbor object of the target object respectively to obtain a plurality of object splicing features; multiply each of the object splicing features by a pre-trained aggregation matrix to obtain a plurality of aggregated splicing features, and respectively superimpose each of the aggregated splicing features with a pre-trained aggregation bias vector to obtain a plurality of aggregated superposition features; and perform activation processing on each of the aggregated superposition features through an activation function to obtain the feature contribution weight corresponding to each neighbor object of the target object.

[0034] In one embodiment, the filtering gate module is further configured to superimpose the respective initial neighbor object features of the neighbor objects of the target object to obtain a superimposed neighbor object feature; divide the initial superimposed feature by the number of neighbor objects of the target object to obtain the neighbor object environment feature of the target object.

[0035] In one embodiment, the filtering gate module is further configured to splice the initial target object feature and the neighbor object environment feature to obtain a spliced environment feature, multiply the pre-trained filtering matrix by the spliced environment feature to obtain a filtered spliced feature; superimpose the filtered spliced feature and the pre-trained filtering bias vector to obtain a filtered superimposed feature; perform an activation process on the filtered superimposed feature through an activation function to obtain the feature in the initial target object feature that is incompatible with the neighbor environment feature; filter the feature in the initial target object feature that is incompatible with the neighbor environment feature to obtain a compatible feature that is compatible with the neighbor environment feature.

[0036] In a third aspect, the present application further provides a computer device, which includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps in any one of the feature extraction methods provided in the embodiments of the present application are implemented.

[0037] In a fourth aspect, the present application further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps in any one of the feature extraction methods provided in the embodiments of the present application are implemented.

[0038] In a fifth aspect, the present application further provides a computer program product, which includes a computer program. When the computer program is executed by a processor, the steps in any one of the feature extraction methods provided in the embodiments of the present application are implemented.

[0039] The above-mentioned feature extraction method, device, computer device, storage medium, and computer program product can determine the target object in the object set and multiple neighbor objects similar to the target object by obtaining the object set. By determining the multiple neighbor objects, the initial neighbor object features corresponding to each neighbor object can be obtained, and thus the aggregated feature after aggregation can be obtained based on the respective initial neighbor object features, and the neighbor object environment feature reflecting the object environment constituted by the neighbor objects can be obtained. Since the neighbor object aggregated feature is obtained by fusing the features contributed by the neighbor objects of the target object respectively, the influence of the unimportant initial neighbor object aggregated feature on the neighbor object aggregated feature can be reduced, so that the obtained neighbor object aggregated feature is more accurate. By obtaining the neighbor object environment feature, the features compatible with the neighbor environment of the object environment can be extracted from the initial target object features of the target object, so as to obtain the compatible features having the same commonality with each neighbor object in the target object. When obtaining the compatible features having the same commonality with each neighbor object and the neighbor object aggregated feature reflecting the characteristics of the neighbor objects, the neighbor object aggregated feature and the compatible feature can be feature-fused to associate the target object with the neighbor objects, so as to obtain a reinforced feature including more information.

[0040] In addition, since in the process of feature extraction, not only the features of the target object are considered, but also the features of the neighbor objects similar to the target object are considered, the reinforced feature obtained by synthesizing the features of the target object and the neighbor objects can contain richer feature information. Since in the process of feature extraction, the object information of the target object and the neighbor objects will be transmitted, fused, filtered, etc., the utilization rate of the object information is improved, and the richness of the information included in the reinforced feature is further improved.

[0041] This application provides an object recommendation method, device, computer device, computer-readable storage medium, and computer program product that can improve the recommendation accuracy.

[0042] In a first aspect, this application provides an object recommendation method, and the method includes:

[0043] Obtain a recommended recipient object set and a recommended object set;

[0044] Extract features from each recommended recipient object in the recommended recipient object set and each recommended object in the recommended object set respectively to obtain corresponding recommended recipient object features and recommended object features;

[0045] According to the recommended recipient object features corresponding to each recommended recipient object and the recommended object features corresponding to each recommended object, determine the association relationship between each recommended recipient object and each recommended object;

[0046] Determine the recommended objects to be recommended corresponding to each recommended recipient according to the said association relationship;

[0047] When extracting features from at least one of the recommended recipients or the recommended objects, the feature extraction method described in any one of claims 1 to 13 is used for extraction.

[0048] In a second aspect, the present application further provides an object recommendation device, and the device includes:

[0049] A feature extraction module, configured to obtain a set of recommended recipients and a set of recommended objects; extract features from each recommended recipient in the set of recommended recipients and each recommended object in the set of recommended objects respectively, to obtain corresponding recommended recipient features and recommended object features; when extracting features from at least one of the recommended recipients or the recommended objects, a feature extraction device is used for extraction;

[0050] An association relationship generation module, configured to determine the association relationship between each recommended recipient and each recommended object according to the recommended recipient features corresponding to each recommended recipient and the recommended object features corresponding to each recommended object;

[0051] A recommendation module, configured to determine the recommended objects to be recommended corresponding to each recommended recipient according to the said association relationship.

[0052] In one embodiment, the object recommendation device is further configured to obtain a first sample feature of a recommended recipient sample, a second sample feature of a recommended object sample, and an association relationship label; at least one of the first sample feature or the second sample feature is obtained by using a feature extraction device; input the first sample feature and the second sample feature into an association relationship generation model to be trained, to obtain a predicted association relationship between the recommended recipient sample and the recommended object sample; determine the difference between the predicted association relationship and the association relationship label, and construct an association relationship loss according to the difference; train the association relationship generation model through the association relationship loss until the training end condition is met and then end, to obtain a trained association relationship generation model.

[0053] In one embodiment, the association relationship generation module is further configured to, for each recommended recipient object in the recommended recipient object set, obtain the association relationship between the current recommended recipient object and each of the recommended objects according to the recommended recipient object characteristics of the current recommended recipient object and the recommended object characteristics corresponding to each of the recommended objects; the recommendation module is further configured to, for each recommended recipient object in the recommended recipient object set, screen out the recommended objects from the recommended object set whose association relationship with the current recommended recipient object meets the preset relationship condition; and use the screened-out recommended objects as the recommended objects to be recommended corresponding to the current recommended recipient object.

[0054] In one embodiment, the object recommendation device is further configured to output the recommended objects to be recommended corresponding to each user object; the recommended recipient objects in the recommended recipient object set are user objects; and the recommended objects in the recommended object set include at least one of videos and live broadcast rooms.

[0055] In a third aspect, the present application further provides a computer device, which includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps in any one of the object recommendation methods provided by the embodiments of the present application are implemented.

[0056] In a fourth aspect, the present application further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps in any one of the object recommendation methods provided by the embodiments of the present application are implemented.

[0057] In a fifth aspect, the present application further provides a computer program product, which includes a computer program. When the computer program is executed by a processor, the steps in any one of the object recommendation methods provided by the embodiments of the present application are implemented.

[0058] The above object recommendation method, device, computer device, storage medium, and computer program product can extract features from the recommended recipient objects in the recommended recipient object set and the recommended objects in the recommended object set to obtain the recommended recipient object features and the recommended object features. By obtaining the recommended recipient object features and the recommended object features, the association relationship between the recommended recipient object and the recommended object can be determined according to the recommended recipient object features and the recommended object features, so that the recommended object corresponding to the recommended recipient object can be determined according to the association relationship. Since when extracting features from at least one of the recommended recipient objects or the recommended objects, the above feature extraction method can be used for extraction, therefore, when this application performs feature extraction, it not only considers the features of a single object, but also considers the association features between objects, thereby enhancing the information content of the extracted features based on the association features between objects, and further enhancing the accuracy of the determined recommended objects based on the features with enhanced information content, effectively improving the personalized recommendation accuracy of the recommended objects.

[0059] In addition, since when using the feature extraction method for feature extraction, the preference information and attribute information of the object are considered simultaneously, it is possible to collaboratively process two major types of important information when making recommendations, thereby achieving a better recommendation effect. Since when using the feature extraction method for feature extraction, the initial neighbor object features will be aggregated and the features incompatible with the neighbor environment will be filtered out from the initial target object features, realizing the ipsilateral information association operation of the object, effectively improving the personalized recommendation accuracy of the recommended object. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 It is an application environment diagram of the feature extraction method in an embodiment;

[0061] Figure 2 It is a schematic flowchart of the feature extraction method in an embodiment;

[0062] Figure 3 It is a schematic diagram of the relationship between objects in an embodiment;

[0063] Figure 4 It is a schematic diagram of the feature extraction model in an embodiment;

[0064] Figure 5 It is a schematic flowchart of the object recommendation method in an embodiment;

[0065] Figure 6 It is a schematic diagram of video recommendation in an embodiment;

[0066] Figure 7 It is a schematic flowchart of the feature extraction method in a specific embodiment;

[0067] Figure 8 is a schematic flowchart of an object recommendation method in a specific embodiment;

[0068] Figure 9 is a structural block diagram of a feature extraction device in an embodiment;

[0069] Figure 10 is a structural block diagram of an object recommendation device in an embodiment;

[0070] Figure 11 is an internal structure diagram of a computer device in an embodiment. Detailed implementation manners

[0071] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0072] The feature extraction method provided by the embodiments of the present application can be applied to, for example, Figure 1 the application environment shown in the figure. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be integrated on the server 104, or placed in the cloud or on other servers. Both the terminal 102 and the server 104 can be used alone to execute the feature extraction method provided in the embodiments of the present application. The terminal 102 and the server 104 can also be used in cooperation to execute the feature extraction method provided in the embodiments of the present application. Taking the terminal 102 and the server 104 can be used in cooperation to execute the feature extraction method provided in the embodiments of the present application as an example for explanation, when the terminal 102 obtains an object set, it can send the object set to the server 104 so that the server 104 determines the target object in the object set and determines the neighbor objects of the target object. The server 104 determines the initial target object feature of the target object, and determines the initial neighbor object feature of the neighbor object, and obtains the enhanced feature of the target object according to the initial target object feature and the initial neighbor object feature. Among them, the terminal 102 can be, but is not limited to, various desktop computers, laptop computers, smart phones, tablet computers, Internet of Things devices and portable wearable devices. The Internet of Things devices can be smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, etc. The portable wearable devices can be smart watches, smart bracelets, head-mounted devices, etc. The server 104 can be implemented by an independent server or a server cluster composed of multiple servers.

[0073] In one embodiment, as Figure 2 shown, a feature extraction method is provided. Taking this method applied to a computer device as an example for explanation. The computer device can beFigure 1 The terminal or server in

[0074] Step 202: Obtain an object set including multiple objects; the object types of the multiple objects are the same; the object type is one of the objects to be recommended or the recommended recipient objects in the recommendation scenario.

[0075] Among them, the object set refers to a set including multiple objects, and the objects in the object set can belong to the same object type. The object type is one of the objects to be recommended or the recommended objects in the recommendation scenario. Among them, the object to be recommended refers to the object to be recommended. For example, in the video recommendation scenario of recommending videos to users, the object to be recommended can be the video recommended to the user, or in the live broadcast room recommendation scenario of recommending a live broadcast room to the user, the object to be recommended can be the live broadcast room recommended to the user. The recommended recipient object refers to the object that receives the recommendation. For example, in the video recommendation scenario, the recommended recipient object can be the user object that receives the recommended video.

[0076] Specifically, when feature extraction is required, the computer device can obtain an object set including multiple objects. For example, the computer device can use multiple videos in the video library as the object set, or the computer device can use multiple user objects who watch videos through the video playback platform as the object set.

[0077] Step 204: For the target object in the object set, screen out the neighbor objects similar to the target object from the object set, and determine the respective initial neighbor object features of the neighbor objects of the target object.

[0078] Among them, the neighbor object refers to the object in the object set that is similar to the target object. For example, the neighbor object can be an object with attributes similar to the target object; or, the neighbor object can be an object with preferences similar to the target object; or, the neighbor object can be an object with both attributes and preferences similar to the target object. Similar means that the similarity is greater than or equal to the similarity threshold. The initial neighbor object feature is the initial feature obtained after initial feature extraction of the neighbor object, and the enhanced feature can be obtained based on this initial feature.

[0079] Specifically, when the object set is obtained, the computer device can determine the target object in the object set. For example, the computer device can use any object in the object set as the target object, and screen out at least one neighbor object similar to the target object from the object set. When at least one neighbor object is obtained, the computer device can perform feature extraction processing on the neighbor object to obtain the respective initial neighbor object features corresponding to each neighbor object. By determining the neighbor object, when performing feature extraction, the neighbor object and the target object can be combined to obtain the enhanced feature of the target object, so that the information included in the extracted feature is richer.

[0080] In one embodiment, when an object set is obtained, the computer device can obtain the object information corresponding to each object. Among them, the object information includes attribute information and preference information. The attribute information can characterize the attributes of the object. For example, when the object is the recommended object in the recommendation scenario, the attribute information can include the label of the recommended object, the description information of the recommended object, etc. When the object is the recommendation recipient in the recommendation scenario, the attribute information can include information such as the age, gender, and region of the recommendation recipient. The preference information can characterize the preferences of the object. For example, in the recommendation scenario, the preference information of the recommended object reflects the degree to which the recommended object is favored by the recommendation recipient; the preference information of the recommendation recipient reflects the recommended objects preferred by the recommendation recipient. Further, the computer device can determine the neighbor objects similar to the target object in the object set according to the similarity between the object information corresponding to each object and the object information of the target object.

[0081] In one embodiment, when multiple target objects in the object set are determined, for each target object in the object set, the neighbor objects corresponding to each target object can be determined, so that the target object and the corresponding neighbor objects are connected by lines, thereby obtaining Figure 3 the relationship graph between objects as shown. Figure 3 Figure 1 shows a schematic diagram of an object graph in one embodiment. It is easy to understand that the computer device can determine the neighbor objects corresponding to each target object in parallel, thereby improving the determination efficiency of the neighbor objects.

[0082] In one embodiment, when determining the neighbor objects of the target object, the computer device can perform feature extraction processing on the object information of the neighbor objects through an initial feature extraction model to obtain the initial neighbor object features corresponding to each neighbor object. Among them, the initial neighbor object features can be in the form of vectors or matrices. The initial feature extraction model can digitize the object information and convert it into the form of a vector or convert it into the form of a matrix. The initial feature extraction model can be a general machine learning model with feature extraction capabilities that has been trained. When using the general initial feature extraction model for extraction in a specific scenario, the effect is not good. Therefore, it is necessary to further train and optimize the general initial feature extraction model with samples dedicated to the specific scenario.

[0083] Step 206, respectively determine the neighbor object aggregation feature and the neighbor object environment feature of the target object based on the initial neighbor object features; the neighbor object aggregation feature characterizes the aggregation of the features contributed by the neighbor objects of the target object; the neighbor object environment feature characterizes the overall environment feature of the object environment constituted by the neighbor objects of the target object.

[0084] Among them, the neighbor object aggregation feature represents the aggregation of the features contributed by each neighbor object of the target object. Since the relationships between the target object and different neighbor objects are different, the importance of neighbor objects to the target object is also different. In the case of different importance levels, the feature contribution weights corresponding to each neighbor object are also different, so that the features contributed by each neighbor object are different during the feature aggregation process. The neighbor object environmental feature represents the overall environmental feature of the object environment composed of the neighbor objects of the target object. When determining the neighbor objects of the target object, the object environment composed of the neighbor objects can be determined, and thus the neighbor object environmental feature is the overall environmental feature of this object environment. The object environment refers to the information set jointly formed by the object information of each neighbor object, and the overall environmental feature can be the feature obtained after feature extraction from the information set.

[0085] Specifically, when determining the initial neighbor object features corresponding to each neighbor object of the target object, the computer device can determine the feature contribution weight corresponding to each neighbor object according to the importance of the neighbor object to the target object, and perform feature fusion processing on the initial neighbor object features according to the feature contribution weight to obtain the neighbor object aggregation feature of the target object. Feature fusion refers to the process of merging multiple features into one feature. For example, weighted summation processing can be performed on the initial neighbor object features according to the feature contribution weight to obtain the neighbor object aggregation feature. Among them, for the current neighbor object among multiple neighbor objects, multiplying the initial neighbor object feature of the current neighbor object by the feature contribution weight of the current neighbor object can obtain the feature contributed by the current neighbor object in the process of generating the neighbor object aggregation feature.

[0086] Furthermore, when determining each neighbor object of the target object, the computer device can also determine the neighbor environment jointly formed by multiple neighbor objects, and determine the overall environmental feature of the neighbor environment according to the initial neighbor object features corresponding to each neighbor object.

[0087] Step 208: Determine the initial target object feature of the target object, and extract the compatible feature that is compatible with the neighbor object environmental feature from the initial target object feature.

[0088] Specifically, when determining the target object in the object set, the computer device can perform feature extraction processing on the object information of the target object through the initial feature extraction model to obtain the initial target object feature of the target object. Among them, the initial target object feature can be in the form of a vector or a matrix. Furthermore, when obtaining the initial target object feature, the computer device can also filter out the features in the initial target object feature that are incompatible with the neighbor object environmental feature to obtain the compatible feature that is compatible with the neighbor object environmental feature.

[0089] In one embodiment, "compatible" means the same or similar. The initial target object features may include sub-features with multiple dimensions. For example, they may include sub-features with an age dimension and sub-features with a gender dimension. Correspondingly, the neighbor object environmental features may also include sub-features with multiple dimensions. For example, they may also include sub-features with an age dimension and sub-features with a gender dimension. When the sub-features of the same dimension in the initial target object features are the same or similar to the sub-features of the corresponding dimension in the neighbor object environmental features, it can be considered that the sub-features of this dimension in the initial target object features are compatible with the sub-features of the corresponding dimension in the neighbor object environmental features. For example, when the age of the target object falls within 0 to 10 years old, and the average age of the neighbor object also falls within 0 to 10 years old, it can be considered that the sub-features of the age dimension in the initial target object features are compatible with the sub-features of the age dimension in the neighbor object environmental features. Therefore, the computer device can determine whether the sub-features of each dimension are compatible, retain the compatible sub-features in the initial target object features, and delete the incompatible sub-features.

[0090] In one embodiment, the computer device can extract compatible features that are compatible with the neighbor object environmental features from the initial target object features through a filtering gate module. By extracting compatible features that are compatible with the neighbor object environmental features from the initial target object features, the extracted compatible features can reflect the commonality between the target object and the object environment.

[0091] Step 210, perform feature fusion based on the neighbor object aggregated features and the compatible features to obtain the enhanced features of the target object.

[0092] Specifically, when the neighbor object aggregated features and the compatible features are obtained, the computer device can perform feature aggregation processing on the neighbor object aggregated features and the compatible features, so as to obtain the enhanced features of the target object. For example, the computer device can linearly superimpose the neighbor object aggregated features and the compatible features to obtain the enhanced features of the target object.

[0093] In one embodiment, the computer device determines the enhanced features of the target object through the following formula:

[0094]

[0095] Among them, LeakyReLU is the activation function; G a (u i ) is the neighbor object aggregated features of the target object; G f (u i ) is the compatible features of the target object; is the enhanced features of the target object.

[0096] In the above feature extraction method, by obtaining an object set, the target object in the object set and multiple neighbor objects similar to the target object can be determined. By determining multiple neighbor objects, the initial neighbor object features corresponding to each neighbor object can be obtained, so that the aggregated feature after aggregation can be obtained based on each initial neighbor object feature, and the neighbor object environment feature reflecting the object environment formed by the neighbor objects can be obtained. Since the neighbor object aggregation feature is obtained by fusing the features contributed by the neighbor objects of the target object respectively, the influence of unimportant initial neighbor object aggregation features on the neighbor object aggregation feature can be reduced, so that the obtained neighbor object aggregation feature is more accurate. By obtaining the neighbor object environment feature, the features compatible with the neighbor environment of the object environment can be extracted from the initial target object features of the target object, so as to obtain the compatible features in the target object that have the same commonality with each neighbor object. When obtaining the compatible features that have the same commonality with each neighbor object and the neighbor object aggregation feature reflecting the characteristics of the neighbor objects, the neighbor object aggregation feature and the compatible feature can be feature-fused to associate the target object with the neighbor objects, so as to obtain a strengthened feature including more information.

[0097] In addition, since in the process of feature extraction, not only the features of the target object are considered, but also the features of the neighbor objects similar to the target object are considered, the strengthened feature obtained by combining the features of the target object and the neighbor objects can contain richer feature information. Since during feature extraction, the object information of the target object and the neighbor objects will be transmitted, fused, filtered, etc., the utilization rate of the object information is improved, and further the richness of the information included in the strengthened feature is improved.

[0098] In one embodiment, screening out neighbor objects similar to the target object from the object set includes: obtaining the preference features and attribute information corresponding to each object in the object set; for the target object and the remaining objects other than the target object in the object set, determining the similarity between the preference features of the target object and the preference features of each remaining object respectively to obtain the preference similarity corresponding to each remaining object; determining the similarity between the attribute information of the target object and the attribute information of each remaining object respectively to obtain the attribute similarity corresponding to each remaining object; and screening out neighbor objects similar to the target object from the remaining objects other than the target object in the object set according to each preference similarity and each attribute similarity.

[0099] Specifically, when it is necessary to determine the neighbor objects of a target object, the computer device can obtain the preference features and attribute information corresponding to each object in the object set. Among them, the preference feature is the feature obtained by extracting features from the preference information. Further, when the preference features are obtained, for the target object and the remaining objects other than the target object in the object set, the computer device determines the similarity between the preference features of the target object and the preference features of each remaining object, and obtains the attribute similarity corresponding to each remaining object. For example, when the object set includes object A, object B, and object C, and object A is the target object, then object B and object C are the remaining objects. The computer device can determine the preference similarity between the preference features of object A and the preference features of object B and object C respectively, and obtain the preference similarity corresponding to object B and the preference similarity corresponding to object C. It is easy to understand that the preference similarity corresponding to object B refers to the similarity between the preference features of object B and the preference features of object A (also the target object). Correspondingly, the computer device can also determine the similarity between the attribute information of the target object and the attribute information of each remaining object, and obtain the attribute similarity corresponding to each remaining object. The attribute similarity corresponding to the remaining object is also the similarity between the attribute information of the remaining object and the attribute information of the target object.

[0100] Further, the computer device filters out the neighbor objects similar to the target object from the remaining objects other than the target object in the object set according to the preference similarity corresponding to each remaining object and the attribute similarity corresponding to each remaining object. For example, the computer device takes the remaining objects with a preference similarity greater than or equal to the preset preference similarity threshold and the remaining objects with an attribute similarity greater than or equal to the preset attribute similarity threshold as the neighbor objects of the target object.

[0101] In one embodiment, algorithm functions such as the cosine similarity algorithm, dot product algorithm, and Euclidean distance algorithm can be used to calculate the preference similarity between the preference features of the target object and the preference features of the remaining objects, and the above algorithm functions can also be used to calculate the attribute similarity between the attribute information of the target object and the attribute information of the remaining objects.

[0102] In the above embodiment, since the neighbor objects of the target object are determined by comprehensively considering the preference characteristics and attribute information, compared with determining the attribute similarity only through a single type of information, the neighbor objects determined in this embodiment can be more accurate.

[0103] In one embodiment, the step of determining the preference feature includes: obtaining the preference information corresponding to each object in the object set, and determining the preference vector corresponding to each preference information; obtaining a preset dimensionality transformation matrix, and respectively performing dimensionality reduction processing on each preference vector through the dimensionality transformation matrix to obtain the preference feature corresponding to each object.

[0104] Specifically, when it is necessary to determine the preference feature corresponding to each object, the computer device can obtain the preference information corresponding to each object and generate the preference vector corresponding to each preference information. The computer device can obtain a preset dimensionality transformation matrix, and respectively perform dimensionality reduction processing on each preference vector through the dimensionality transformation matrix to obtain the preference feature corresponding to each object.

[0105] In one embodiment, in a recommendation scenario, when the object is the recommended object, since the preference information of the recommended object reflects the degree to which the recommended object is favored by the recommended receiving object, therefore, the preference information of the recommended object can be determined from the first historical viewing record of the recommended object being viewed. For example, when the recommended object is a video, the preference information of the video can be determined from the click-play record of the video. Among them, the click-play record of the video can record the click-play results of different users on the video. For example, it can include that user A clicked to play video A, user B did not click to play video A, and user C clicked to play video A. Thus, the preference information obtained based on this video play record is (A: click; B: not click; C: click), and further, the preference vector generated according to this preference information is 101, where "1" represents click-play and "0" represents not click-play. Another example is that when the recommended object is a live broadcast room, the preference information of the live broadcast room can be determined from the click-play record of the live broadcast room.

[0106] When the object is a recommended recipient, since the preference information of the recommended recipient reflects the degree of preference of the recommended recipient for the recommended object, the preference information of the recommended recipient can be determined from the second historical viewing record of the recommended object viewed by the recommended recipient. For example, when the recommended recipient is a user object, the preference information of the user object can be determined from the click-play record of the user object clicking on the video. Among them, the click-play record of the video can record the click-play results of the same user on different videos. For example, it can include that user a clicked and played video A, user A did not click and play video B, and video C. Thus, the preference information obtained based on this video play record is (A: clicked; B: not clicked; C: not clicked). Furthermore, the preference vector of the user object generated according to this preference information is 100, where "1" represents click-play and "0" represents not click-play. Correspondingly, when the recommended recipient is a user object and the recommended object is a live room, the preference information of the user object can be determined from the click-play record of the user object clicking on the live room.

[0107] In one embodiment, the preference vector is mapped to a low-dimensional preference feature through a trainable dimensionality transformation matrix, so as to integrate the preference information of the object into the preference feature. Correspondingly, a matrix can be used to represent the preference features of multiple objects in the object set. For example, when the object is a recommended recipient, the preference features of multiple recommended recipients can be represented by the matrix When the object is a recommended object, the preference features of multiple recommended objects can be represented by the matrix where |U| is the number of recommended recipients, |V| is the number of recommended objects, and D is the dimension of the preference feature. The preference feature of recommended recipient i is the i-th row of P U denoted as The preference feature of recommended object j is the j-th row of P V denoted as

[0108] In the above embodiment, by performing dimensionality reduction processing on the preference vector, the amount of calculation for subsequent calculation of the preference vector can be reduced, thereby improving the efficiency of feature extraction based on the reduced amount of calculation.

[0109] In one embodiment, the attribute information includes multiple attribute sub-information; determining the similarity between the attribute information of the target object and the attribute information of each of the remaining objects respectively, to obtain the attribute similarity corresponding to each of the remaining objects, includes: for each of the remaining objects in the object set, determining the similarity between each attribute sub-information of the target object and the corresponding attribute sub-information of the current remaining object respectively, to obtain multiple sub-information similarities; performing an averaging process on the multiple sub-information similarities to obtain the attribute similarity between the attribute information of the target object and the attribute information of the current remaining object.

[0110] Specifically, the computer device can determine the attribute similarity between the attribute information of the target object and the attribute information of each of the remaining objects. The attribute information can include multiple attribute sub-information. For example, when the object is a recommended recipient, the attribute information can include the gender, region of belonging, age, etc. of the recommended recipient. Among them, gender, region of belonging, and age are all one attribute sub-information. Another example is that when the object is a recommended object, the attribute information can include tags, publishers, description information, etc. Among them, tags, publishers, and description information are all one attribute sub-information.

[0111] The computer device can simultaneously determine the attribute similarity between the attribute information of the target device and the attribute information of each of the remaining devices; it can also sequentially determine the attribute similarity between the attribute information of the target device and the attribute information of each of the remaining devices. For the sake of description convenience, the following takes the determination of the attribute similarity between the attribute information of the target device and the attribute information of the current remaining devices as an example for illustration. The computer device can determine the similarity between each attribute sub-information of the target object and the corresponding attribute sub-information of the current remaining objects respectively, and obtain the sub-information similarity corresponding to each attribute sub-information of the target object. The computer device performs an averaging process on the sub-information similarity corresponding to each attribute sub-information to obtain the attribute similarity between the attribute information of the target object and the attribute information of the current remaining objects.

[0112] In the above embodiment, since the attribute similarity is determined by comprehensively considering the similarity between various attribute sub-information, compared with determining the attribute similarity only through a single type of information, the accuracy of the determined attribute similarity can be improved in this embodiment.

[0113] In one of the embodiments, the attribute sub-information includes an attribute name and an attribute value; determining the similarity between each attribute sub-information of the target object and the corresponding attribute sub-information of the current remaining objects respectively, and obtaining a plurality of sub-information similarities, including: for the multiple attribute sub-information of the target object, determining the target attribute sub-information in the current remaining objects that has the same attribute name as the current attribute sub-information of the target object; determining the similarity between the attribute value of the target attribute sub-information of the current remaining objects and the attribute value of the current attribute sub-information of the target object to obtain the sub-information similarity.

[0114] Specifically, the computer device can simultaneously determine the similarity between multiple attribute sub-informations of the target device and multiple attribute sub-informations of the current remaining devices; it can also sequentially determine the similarity between multiple attribute sub-informations of the target device and the corresponding attribute sub-informations of the current remaining devices. For the convenience of description, the following takes the determination of the similarity between the current attribute sub-information of the target device and the corresponding attribute information of the current remaining devices as an example for illustration. The attribute sub-information may include an attribute name and an attribute value. For example, the attribute sub-information can be "Age: 20", where "Age" is the attribute name and "20" is the attribute value. When determining the current attribute sub-information of the target device, the computer device can determine the target attribute sub-information in the current remaining objects that has the same attribute name as the current attribute sub-information of the target object. For example, when the current attribute sub-information of the target object is "Age: 20", the computer device can use the attribute sub-information with the attribute name "Age" in the current remaining objects as the target attribute sub-information. Further, the computer device determines the similarity between the attribute value of the target attribute sub-information of the current remaining objects and the attribute value of the current attribute sub-information of the target object. For example, when the current attribute sub-information of the target device is "Age: 20" and the target attribute sub-information of the current remaining objects is "Age: 30", the computer device can determine the similarity between 20 and 30, thereby obtaining the sub-information similarity of the current attribute sub-information of the target object.

[0115] In one embodiment, if the attribute values of the target object and the current remaining objects are the same under the same attribute name, the sub-information similarity is 1; otherwise, it is 0.

[0116] In one embodiment, the computer device can determine the attribute similarity between the attribute information of the target object and the attribute information of the remaining objects through the following formula:

[0117]

[0118] Wherein, represents the nth attribute sub-information of the target object i; represents the nth attribute sub-information of the remaining object j; represents determining whether the attribute values belonging to the same attribute name of two objects are equal; N represents the number of attribute sub-informations.

[0119] In the above embodiment, since an object usually has multiple attribute sub-informations, and it is difficult to measure the similarity of different values under each attribute sub-information (for example, male and female in the gender attribute), therefore, in this embodiment, by using the coincidence degree of attribute values as the calculation method of attribute similarity, the determination of attribute similarity can be realized.

[0120] In one embodiment, screening out neighbor objects similar to a target object from the remaining objects in the object set according to each preference similarity and each attribute similarity includes: performing similarity fusion processing on the preference similarity and the attribute similarity belonging to the same remaining object to obtain a target similarity corresponding to each remaining object; and using the remaining objects in the object set whose target similarity meets a preset similarity condition as the neighbor objects of the target object.

[0121] Specifically, when obtaining the preference similarity and the attribute similarity corresponding to each remaining object, the computer device can comprehensively consider the preference similarity and the attribute similarity to obtain the target similarity corresponding to each remaining object. For example, the computer device can determine the preference similarity and the attribute similarity corresponding to the same remaining object, that is, determine the preference similarity and the attribute similarity belonging to the same remaining object, and perform similarity fusion on the preference similarity and the attribute similarity belonging to the same remaining object to obtain the target similarity corresponding to each remaining object. For example, the preference similarity and the attribute similarity belonging to the same remaining object can be added together to obtain the target similarity. Further, the computer device can screen out the objects in the remaining objects in the object set whose target similarity meets the preset similarity condition, and use the screened objects as the neighbor objects of the target object. Among them, the similarity condition can be freely set according to requirements. For example, the computer device uses the remaining objects whose target similarity is greater than or equal to a preset similarity threshold as the neighbor objects of the target object.

[0122] In one embodiment, when obtaining the preference similarity and the attribute similarity corresponding to each remaining object, the computer device can perform normalization processing on the preference similarity and the attribute similarity, and then add the preference similarity and the attribute similarity of the same remaining object element by element to obtain the target similarity.

[0123] In the above embodiment, since the neighbor objects similar to the target object are determined by comprehensively considering the preference similarity and the attribute similarity, compared with determining the similar neighbor objects only through single information, the neighbor objects determined in this embodiment can be more accurate.

[0124] In one embodiment, determining the initial neighbor object features of each neighbor object of the target object includes: obtaining the preference feature corresponding to each neighbor object of the target object, and obtaining the attribute feature corresponding to each neighbor object of the target object; and performing feature fusion processing on the preference feature and the attribute feature corresponding to each neighbor object to obtain the initial neighbor object feature of each neighbor object of the target object.

[0125] Specifically, before obtaining the neighbor aggregation feature and the neighbor environment feature, the computer device may also determine the initial neighbor object feature corresponding to each neighbor object. When it is necessary to determine the initial neighbor feature corresponding to each neighbor object, the computer device may obtain the preference feature corresponding to each neighbor object and the attribute feature corresponding to each target object. Among them, the preference feature can be implemented according to the above embodiments; the attribute feature can be obtained by extracting through a deep neural network. Further, the computer device performs feature fusion processing on the preference feature and the attribute feature corresponding to each neighbor object, so as to obtain the initial neighbor object feature corresponding to each neighbor object of the target object. For example, the computer device superimposes the preference feature and the attribute feature belonging to the same neighbor object to obtain the initial neighbor object feature of the neighbor object.

[0126] In one embodiment, when the initial neighbor object is the recommended recipient, the computer device obtains the initial neighbor object feature through the following formula:

[0127]

[0128] Wherein, represents the preference feature of the recommended recipient i; represents the attribute feature of the recommended recipient i; W U is a trainable parameter matrix; b U is a trainable bias vector; is the initial neighbor object feature of the recommended recipient.

[0129] In one embodiment, when the initial neighbor object is the recommended object, the computer device obtains the initial neighbor object feature through the following formula:

[0130]

[0131] Wherein, represents the preference feature of the recommended object i; represents the attribute feature of the recommended object i; W V is a trainable parameter matrix; b V is a trainable bias vector; is the initial neighbor object feature of the recommended object.

[0132] In the above embodiments, since the initial neighbor object feature is obtained by integrating the preference feature and the attribute feature, the determined initial neighbor object feature can contain richer information. Furthermore, based on the initial neighbor object feature including richer information, a reinforced feature including richer information is obtained.

[0133] In one embodiment, the step of determining the attribute characteristics of a neighbor object includes: obtaining the attribute information of the neighbor object; the attribute information includes a plurality of attribute sub-information; performing two-way interactive fusion processing and linear fusion processing on the plurality of attribute sub-information of the neighbor object to obtain corresponding two-way interactive fusion results and linear fusion results; and performing result fusion processing on the two-way interactive fusion results and the linear fusion results to obtain the attribute characteristics of the neighbor object.

[0134] Specifically, since one attribute information may include a plurality of attribute sub-information, when it is necessary to determine the attribute characteristics of a neighbor object, the computer device can perform two-way interactive fusion processing on the plurality of attribute sub-information to obtain two-way interactive fusion results, and perform linear fusion processing on the plurality of attribute sub-information to obtain linear fusion results. Among them, two-way interactive fusion processing refers to the process of fusing every two attribute sub-information and then further fusing the fused results. Linear fusion processing refers to the process of sequentially fusing attribute sub-information. When the two-way interactive fusion results and the linear fusion results are obtained, the computer device can fuse the two-way interactive fusion results and the linear fusion results, that is, perform result fusion processing, so as to obtain the attribute characteristics of the neighbor object. For example, the computer device can superimpose the two-way interactive fusion results and the linear fusion results to obtain the attribute characteristics of the neighbor object.

[0135] In one embodiment, the computer device can perform linear fusion processing through the following formula:

[0136]

[0137] where N represents the number of attribute sub-information; w k is a trainable parameter matrix; is the k-th attribute sub-information of neighbor object i. In one embodiment, can also be the attribute vector obtained after vectorizing the k-th attribute sub-information of neighbor object i.

[0138] In one embodiment, the computer device can perform result fusion processing through the following formula:

[0139]

[0140] where LeakyReLU is the activation function; and are both trainable parameter matrices; is the two-way interactive fusion result of neighbor object i; is the linear fusion result of object i.

[0141] In the above embodiments, regarding the attribute characteristics of neighbor objects, considering the correlation relationships among multiple attribute sub-information of neighbor objects, therefore, based on the two-way interactive fusion method of attribute sub-information intersection, new synthetic features formed after the intersection of attribute sub-information can be explicitly modeled, enriching the information included in the two-way interactive fusion result while reducing the requirements for the feature extraction ability of the model. In addition, by capturing the cross features in the attribute sub-information through two-way interactive fusion and then combining linear fusion to obtain the overall features of the attribute information, effective fusion of multiple attribute sub-information of neighbor objects can be achieved, thereby enhancing the richness of the information included in the attribute characteristics of neighbor objects.

[0142] In one embodiment, performing two-way interactive fusion processing on multiple attribute sub-information of neighbor objects includes: determining the attribute vectors respectively corresponding to each attribute sub-information of neighbor objects; performing vector fusion processing on every two of the multiple attribute vectors to obtain multiple attribute fusion vectors, and performing superposition processing on the multiple attribute fusion vectors to obtain the two-way interactive fusion result of neighbor objects.

[0143] Specifically, when multiple attribute sub-information of neighbor objects are obtained, the computer device can vectorize the attribute sub-information to obtain the attribute vectors respectively corresponding to each attribute sub-information. For example, when the attribute sub-information is gender: female, the computer device can determine the corresponding attribute vector as "10". Further, the computer device fuses every two of the multiple attribute vectors, that is, performs vector fusion processing, to obtain multiple attribute fusion vectors, and performs superposition processing on the multiple attribute fusion vectors to obtain the two-way interactive fusion result of neighbor objects.

[0144] In one embodiment, the computer device can perform two-way interactive fusion processing through the following formula:

[0145]

[0146] Where, represents the attribute vector corresponding to the k-th attribute sub-information of object i; represents the attribute vector corresponding to the l-th attribute sub-information of object i; N represents the number of attribute sub-information; w k and w l are both trainable parameter matrices; ⊙ represents element-wise multiplication.

[0147] In the above embodiments, through two-way interactive fusion processing, cross features in the attribute sub-information can be obtained, so that the attribute features obtained based on the two-way interactive fusion processing result can contain richer information.

[0148] In one embodiment, the step of determining the aggregated feature of the neighbor objects of the target object includes: determining the feature contribution weight corresponding to each neighbor object of the target object; and performing a weighted average process on the initial neighbor object features of the neighbor objects according to the feature contribution weights to obtain the aggregated feature of the neighbor objects of the target object.

[0149] Specifically, since the relationships between the target object and different neighbor objects are different, when aggregating the initial neighbor object features, it is necessary to consider not only the importance of the neighbor objects but also the importance of the more fine-grained information of the diversified information of each dimension of the initial neighbor object features. Therefore, when it is necessary to determine the aggregated feature of the neighbor objects of the target object, the computer device can determine the feature contribution weight corresponding to each neighbor object of the target object. Among them, the feature contribution weight refers to the proportion of the initial neighbor object feature in the aggregated feature of the neighbor objects. When determining the feature contribution weights corresponding to each neighbor object, the computer device can perform a weighted average process on the initial neighbor object features according to the feature contribution weights to obtain the aggregated feature of the neighbor objects of the target object.

[0150] In one embodiment, the computer device can determine the importance degree of the neighbor object according to the similarity degree between the target object and the neighbor object. For example, when the similarity degree is higher, the importance degree of the neighbor object is also higher, so that the feature contribution weight of the neighbor object is also higher.

[0151] In one embodiment, the computer device can determine the aggregated feature of the neighbor objects of the target object through the following formula:

[0152]

[0153] Where, represents the feature contribution weight of neighbor object K; represents the initial neighbor object feature of neighbor object K; represents the set of neighbor objects; ⊙ represents element-wise multiplication.

[0154] In one embodiment, the computer device can fuse the initial neighbor object features of each neighbor object through a pre-trained aggregation gate module to obtain the aggregated feature of the neighbor objects. Among them, the aggregation gate module is used to control which initial neighbor object features need to be fused.

[0155] In the above embodiment, by fusing the initial neighbor object features according to the feature contribution weights corresponding to each neighbor object, neighbor objects with high feature contribution weights can contribute more features, and neighbor objects with low feature contribution weights can contribute fewer features, so that the aggregated neighbor object aggregated feature is more accurate.

[0156] In one embodiment, determining the feature contribution weight corresponding to each neighbor object of the target object includes: determining the initial target object feature of the target object; splicing the initial target object feature with the initial neighbor object feature corresponding to each neighbor object of the target object respectively to obtain a plurality of object splicing features; multiplying the pre-trained aggregation matrix by each object splicing feature respectively to obtain a plurality of aggregated splicing features, and respectively superimposing each aggregated splicing feature with the pre-trained aggregation bias vector to obtain a plurality of aggregated superimposed features; and respectively performing activation processing on each aggregated superimposed feature through an activation function to obtain the feature contribution weight corresponding to each neighbor object of the target object.

[0157] Specifically, when determining the feature contribution matrix, the computer device can determine the initial target object feature of the target object; wherein, the determination method of the initial target object feature can be determined according to the determination method of the initial neighbor object feature. Further, the computer device can perform feature splicing processing on the target object feature and each neighbor object feature respectively to obtain a plurality of spliced object splicing features. The computer device obtains the pre-trained aggregation matrix, and multiplies the pre-trained aggregation matrix by each object splicing feature respectively to obtain a plurality of aggregated splicing features. The aggregation matrix is a trainable parameter matrix. Further, the computer device respectively adds each aggregated splicing feature with the pre-trained aggregation bias vector to obtain a plurality of aggregated superimposed features, and respectively performs activation processing on each aggregated superimposed feature through an activation function, so as to obtain the feature contribution weight corresponding to each neighbor object of the target object.

[0158] In one embodiment, the computer device can determine the feature contribution weight through the following formula:

[0159]

[0160] where, represents the initial target object feature of target object i; represents the initial neighbor object feature of neighbor object k; W a represents the aggregation matrix; b a represents the aggregation bias vector; σ represents the activation function; represents the feature contribution weight of neighbor object k.

[0161] In the above embodiment, since the feature contribution weight is obtained by comprehensively considering the initial target object feature and the initial neighbor object, the determined feature contribution weight can characterize the importance degree of the initial neighbor object to the target object, so that the neighbor object aggregation feature can be obtained based on the importance degree subsequently.

[0162] In one embodiment, the step of determining the neighbor object environmental feature of the target object includes: superimposing the respective initial neighbor object features of the neighbor objects of the target object to obtain a neighbor object superimposed feature; dividing the initial superimposed feature by the number of neighbor objects of the target object to obtain the neighbor object environmental feature of the target object.

[0163] Specifically, since the neighbor object environmental feature reflects the overall environmental feature of the neighbor environment constituted by the neighbor objects, the computer device can use the average feature of each neighbor object feature as the neighbor object environmental feature of the target object. For example, the computer device can superimpose the respective initial neighbor object features of the neighbor objects of the target object to obtain a neighbor object superimposed feature, and divide the initial superimposed feature by the number of neighbor objects of the target object to obtain the neighbor object environmental feature of the target object.

[0164] In one embodiment, the computer device can determine the neighbor object environmental feature of the target object through the following formula.

[0165]

[0166] Wherein, represents the number of neighbor objects; represents the initial neighbor object feature of the kth neighbor object; S represents the neighbor object environmental feature of the target object.

[0167] In the above embodiment, by performing an averaging process on each neighbor object, the commonality of the object environment can be reflected based on the average feature, so as to obtain the neighbor object environmental feature that reflects the overall environmental feature of the object environment based on the commonality of the object environment.

[0168] In one embodiment, extracting a compatible feature from the initial target object feature that is compatible with the neighbor object environmental feature includes: splicing the initial target object feature and the neighbor object environmental feature to obtain an environmental splicing feature, and multiplying the pre-trained filtering matrix by the environmental splicing feature to obtain a filtered splicing feature; superimposing the filtered splicing feature and the pre-trained filtering bias vector to obtain a filtered superimposed feature; performing an activation process on the filtered superimposed feature through an activation function to obtain the feature in the initial target object feature that is incompatible with the neighbor environmental feature; filtering the feature in the initial target object feature that is incompatible with the neighbor environmental feature to obtain a compatible feature that is compatible with the neighbor environmental feature.

[0169] Specifically, when it is necessary to fuse the neighbor object aggregation features and the compatible features, the computer device can splice the initial target object features and the neighbor object environment features to obtain the spliced environment features after splicing, and obtain a filtering matrix. Multiply the filtering matrix by the spliced environment features to obtain the filtered spliced features. The filtering matrix is a trainable parameter matrix. Further, the computer device performs an activation process on the filtered superimposed features through an activation function to obtain the features in the initial target object features that are incompatible with the neighbor environment features, that is, the features in the initial target object features that are neither the same nor similar to the neighbor environment features. The computer device filters out the features in the initial target object features that are incompatible with the neighbor environment features, and can obtain the compatible features that are compatible with the neighbor environment features.

[0170] In one embodiment, the computer device can obtain the compatible features through the following formula:

[0171]

[0172]

[0173] Where represents the features in the initial target object features of target object i that are incompatible with the neighbor environment features; represents the neighbor environment features; represents the initial target object features; W f represents the filtering matrix; b f represents the filtering bias vector; G f (u i ) represents the compatible features of target object i; ⊙ represents element-wise multiplication.

[0174] In the above embodiment, by subtracting the features in the initial target object features that are incompatible with the neighbor environment features, the compatible features can be quickly obtained, thereby improving the determination efficiency of the compatible features.

[0175] In one embodiment, referring to Figure 4 , the feature extraction method can be performed through a feature extraction model. The feature extraction model can include an initial feature extraction module, an aggregation gate module, a filtering gate module, and a feature fusion module. The initial target object features and the initial neighbor object features can be obtained through the initial feature extraction module, the neighbor object aggregation features can be obtained through the aggregation gate module, the compatible features can be obtained through the filtering gate module, and the enhanced features can be obtained through the feature fusion module. The computer device can train the feature extraction model with training samples to adjust the parameter matrix and bias vector in the feature extraction model, so as to obtain a trained feature extraction model. Figure 4 Shows a schematic diagram of the feature extraction model in one embodiment.

[0176] In one embodiment, as Figure 5 shown, an object recommendation method is provided. Taking the application of this method to a computer device as an example for illustration. The computer device can be Figure 1 a terminal or a server in

[0177] Step 502, obtain a set of recommended recipient objects and a set of objects to be recommended.

[0178] Specifically, when it is necessary to recommend corresponding objects to be recommended to the recommended recipient objects, the computer device can obtain a set of recommended recipient objects and a set of objects to be recommended. Among them, the set of recommended recipient objects includes multiple recommended recipient objects, and the set of objects to be recommended may also include multiple objects to be recommended.

[0179] Step 504, extract features for each recommended recipient object in the set of recommended recipient objects and each object to be recommended in the set of objects to be recommended respectively, to obtain corresponding recommended recipient object features and object to be recommended features; among them, when extracting features for at least one of the recommended recipient objects or the objects to be recommended, the feature extraction method according to any one of claims 1 to 13 is used for extraction.

[0180] Specifically, when the set of recommended recipient objects and the set of objects to be recommended are obtained, the computer device can perform first feature extraction on each recommended recipient object in the set of recommended recipient objects respectively, to obtain recommended recipient object features corresponding to each recommended recipient object. And the computer device can perform second feature extraction on each object to be recommended in the set of objects to be recommended respectively, to obtain object to be recommended features corresponding to each object to be recommended. Among them, at least one of the first feature extraction or the second feature extraction can be performed using the above-mentioned feature extraction method. That is, both the first feature extraction and the second feature extraction can be performed using the above-mentioned feature extraction method, or either the first feature extraction or the second feature extraction process can be performed using the above-mentioned feature extraction method.

[0181] In one of the embodiments, when the above-mentioned feature extraction method is used for feature extraction, the extracted feature is an enhanced feature. For example, when the above-mentioned feature extraction method is used for feature extraction of the recommended recipient object, the enhanced feature of the recommended recipient object can be obtained, and this enhanced feature is also the recommended recipient object feature. When the above-mentioned feature extraction method is not used for feature extraction processing, the extracted feature is an initial object feature. For example, the initial object feature can be obtained according to the feature extraction method of the initial neighbor object feature or the initial target object feature.

[0182] Step 506: Determine the association relationship between each recommended recipient and each recommended object according to the characteristics of each recommended recipient and the characteristics of each recommended object respectively.

[0183] Step 508: Determine the recommended objects to be recommended for each recommended recipient according to the association relationship.

[0184] Specifically, when the characteristics of each recommended recipient and the characteristics of each recommended object are obtained, the computer device can input the characteristics of each recommended recipient and the characteristics of each recommended object into the association relationship generation model to obtain the association relationship between each recommended recipient and each recommended object. Among them, the association relationship generation model can be a pre-trained machine learning model.

[0185] Furthermore, when the association relationship between each recommended recipient and each recommended object is determined, the computer device can obtain the recommended objects corresponding to each recommended recipient according to the association relationship, so as to recommend the recommended objects to the corresponding recommended recipients.

[0186] In one embodiment, the computer device can calculate the association relationship between the two through association relationship calculation algorithms such as cosine similarity calculation function, dot product function, Euclidean distance calculation function, etc.

[0187] In one embodiment, determining the association relationship between each recommended recipient and each recommended object according to the characteristics of each recommended recipient and the characteristics of each recommended object respectively includes: for each recommended recipient in the recommended recipient set, obtaining the association relationship between the current recommended recipient and each recommended object according to the characteristics of the current recommended recipient and the characteristics of each recommended object respectively; determining the recommended objects to be recommended for each recommended recipient according to the association relationship includes: for each recommended recipient in the recommended recipient set, screening out the recommended objects from the recommended object set whose association relationship with the current recommended recipient meets the preset relationship conditions; using the screened recommended objects as the recommended objects to be recommended for the current recommended recipient.

[0188] Specifically, the computer device can determine the association relationship between the recommended recipient and each recommended object simultaneously; or it can determine the association relationship between a single recommended recipient and each recommended object in sequence. For the sake of description, the following takes the determination of the association relationship between the current recommended recipient and each recommended object as an example for illustration. The computer device can input the recommended recipient characteristics of the current recommended recipient and the recommended object characteristics corresponding to each recommended object into a pre-trained association relationship generation model, so as to obtain the association relationship between the current recommended recipient and the recommended object.

[0189] Furthermore, when the association relationship between the current recommended recipient and the recommended object is obtained, the computer device can screen out the recommended objects from the set of recommended objects whose association relationship with the current recommended recipient meets the preset relationship conditions, and use the screened-out recommended objects as the recommended objects corresponding to the current recommended object, that is, recommend the screened-out recommended objects to the current recommended recipient. Among them, the preset relationship conditions can be freely set according to requirements. For example, the recommended objects whose association relationship is greater than or equal to the preset relationship threshold can be used as the recommended objects to be recommended to the current recommended recipient. Another example is that the computer device can sort the recommended objects according to the association relationship from large to small, and use the top preset number of recommended objects as the recommended objects to be recommended corresponding to the current recommended recipient.

[0190] In one embodiment, the computer device can use the association relationship between the recommended recipient characteristics of the current recommended recipient and the recommended recipient of the current recommended object as the association relationship between the current recommended recipient and the current recommended recipient.

[0191] In this embodiment, by obtaining the set of recommended recipients and the set of recommended objects, the characteristics of the recommended recipients in the set of recommended recipients can be extracted and the characteristics of the recommended objects in the set of recommended objects can be extracted to obtain the recommended recipient characteristics and the recommended object characteristics. By obtaining the recommended recipient characteristics and the recommended object characteristics, the association relationship between the recommended recipient and the recommended object can be determined according to the recommended recipient characteristics and the recommended object characteristics, so that the recommended object corresponding to the recommended recipient can be determined according to this association relationship. Since when extracting the characteristics of at least one of the recommended recipient or the recommended object, the above-mentioned feature extraction method can be used for extraction, therefore, when this application performs feature extraction, it not only considers the characteristics of a single object, but also considers the association characteristics between objects, thereby enhancing the information content of the extracted characteristics based on the association characteristics between objects, and then enhancing the accuracy of the determined recommended object based on the characteristics with enhanced information content, effectively improving the personalized recommendation accuracy of the recommended object.

[0192] In addition, when using the feature extraction method for feature extraction, both the preference information and attribute information of the object are considered. Therefore, when making recommendations, two major types of important information can be processed collaboratively, thus achieving a better recommendation effect. When using the feature extraction method for feature extraction, the initial neighbor object features are aggregated and the features incompatible with the neighbor environment are filtered out from the initial target object features, realizing the ipsilateral information association operation of the object and effectively improving the personalized recommendation accuracy of the recommended object.

[0193] In one embodiment, the association relationship between the recommended receiving object and the recommended object is determined by an association relationship generation model; the training steps of the association relationship generation model include: obtaining the first sample feature of the recommended receiving object sample, the second sample feature of the recommended object sample, and the association relationship label; at least one of the first sample feature or the second sample feature is extracted by using the feature extraction method as described in any one of claims 1 to 13; inputting the first sample feature and the second sample feature into the association relationship generation model to be trained to obtain the predicted association relationship between the recommended receiving object sample and the recommended object sample; determining the difference between the predicted association relationship and the association relationship label, and constructing an association relationship loss according to the difference; training the association relationship generation model through the association relationship loss until the training end condition is met and then ending to obtain the trained association relationship generation model.

[0194] Specifically, before determining the association relationship between the recommended receiving object and the recommended object through the association relationship generation model, the computer device can also train the association relationship generation model. The computer device can obtain the recommended receiving object sample and the recommended object sample, and perform third feature extraction on the recommended receiving object sample to obtain the first sample feature, and perform fourth feature extraction on the recommended object sample to obtain the second sample feature. Among them, at least one of the third feature extraction or the fourth feature extraction uses the above-mentioned feature extraction method, that is, both the third feature extraction and the fourth feature extraction use the above-mentioned feature extraction method, or the third feature extraction or the fourth feature extraction uses the above-mentioned feature extraction method. Further, the computer device can input the first sample feature and the second sample feature into the association relationship generation model to be trained, and output the predicted association relationship between the recommended receiving object sample and the recommended object sample through the association relationship generation model to be trained. The computer device obtains the association relationship label, that is, obtains the standard association relationship, determines the difference between the predicted association relationship and the association relationship label, constructs an association relationship loss according to the difference, and trains the association relationship generation model through the association relationship loss until the training end condition is met and then ending to obtain the trained association relationship generation model. Among them, the training end condition can be freely set according to requirements.

[0195] In one embodiment, the computer device can determine the predicted association relationship between the recommended object sample and the object to be recommended sample through the following formula:

[0196]

[0197] Wherein, is the first sample feature; is the second sample feature; MLP is a multi-layer perceptron model; b u and b v and μ respectively represent the trainable bias value of the recommended recipient object sample, the bias value of the object to be recommended sample, and the global bias value; represents the predicted association relationship between the recommended recipient object sample i and the object to be recommended sample j. It is easy to understand that when the training is completed, the computer device can also determine the association relationship between the recommended recipient object and the object to be recommended through the above formula.

[0198] In one embodiment, the loss function of the association relationship generation model can be:

[0199]

[0200] Wherein, is the predicted association relationship; R i,j is the association relationship label; U represents the set of recommended recipient object samples; V represents the set of objects to be recommended.

[0201] In the above embodiment, by training the association relationship generation model, a trained association relationship generation model can be obtained, so that subsequently, a more accurate association relationship between the recommended recipient object and the object to be recommended can be obtained based on the trained association relationship generation model.

[0202] In one embodiment, the recommended recipient objects in the recommended recipient object set are user objects; the objects to be recommended in the object to be recommended set include at least one of videos and live rooms; the object recommendation method is executed by an object recommendation model; the objects to be recommended corresponding to each user object output by the object recommendation model include at least one of videos and live rooms.

[0203] Specifically, the recommended recipient can specifically be a user object. For example, the recommended recipient can be a logged-in user who logs in to the video playback platform through a logged-in account, etc. The object to be recommended can specifically be a video. For example, the object to be recommended can be a video uploaded to the video playback platform. The computer device can execute the above object recommendation method through the object recommendation model, so as to obtain the videos to be recommended corresponding to each user object, and then realize the personalized recommendation of videos. For example, the object to be recommended can be a live broadcast room on the live broadcast platform. The computer device can execute the above object recommendation method through the object recommendation model, so as to obtain the live broadcast rooms to be recommended corresponding to each user object, and then realize the personalized recommendation of the live broadcast rooms. Also, for example, the object to be recommended can include at least one of a video and a live broadcast room. The computer device can execute the above object recommendation method through the object recommendation model, so as to obtain at least one of a live broadcast room and a video corresponding to each user object. For example, a video to be recommended to user A can be obtained, a live broadcast room to be recommended to user B can be obtained, and a video and a live broadcast room to be recommended to user C can be obtained.

[0204] Among them, the object recommendation model includes a feature extraction module, a correlation generation module, and a recommendation module. The feature extraction module is also the above-mentioned feature extraction model. The feature extraction model can execute the above-mentioned feature extraction method to obtain enhanced features. The correlation generation module is also the above-mentioned correlation generation model. Through the correlation generation module, the correlation between the recommended recipient and the object to be recommended can be obtained. For example, the correlation between the user object and the video can be obtained, or the correlation between the user object and the live broadcast room can be obtained. Thus, the recommendation model can obtain at least one of a video and a live broadcast room to be recommended to the user object based on the correlation. For example, referring to Figure 6 , the object recommendation model can be carried on the video playback platform, so as to recommend videos to the user object through the object recommendation model. Figure 6 Fig. shows a schematic diagram of video recommendation in an embodiment.

[0205] In one embodiment, the user object can be a virtual object, that is, an object simulated by a computer device according to a natural person.

[0206] In this embodiment, when the recommended recipient is a user object and the object to be recommended is a video, the videos to be recommended to the user object can be obtained, thereby improving the accuracy of personalized video recommendation.

[0207] In one embodiment, the recommended object may be a user object logged in to a reading platform, and the object to be recommended may be a book. By executing the above object recommendation method, the books to be recommended to the user object can be obtained. In one embodiment, the recommended object may be a user object logged in to a music playing platform, and the object to be recommended may be music. By executing the above object recommendation method, the music to be recommended to the user object can be obtained. It is easy to understand that the object to be recommended may also be a picture, a game, etc.; the recommended receiving object may also be an animal, a plant, etc. For example, music can be recommended to the animals in the breeding farm through the above object recommendation method, so as to promote the growth of the animals through the recommended music.

[0208] The above application scenarios are only illustrative. It can be understood that the application of the object recommendation method provided by each embodiment of the present application is not limited to the above scenarios.

[0209] In a specific embodiment, referring to Figure 7 , a feature extraction method is provided, including:

[0210] S702, obtaining an object set including multiple objects; the object types of the multiple objects are the same; the object type is one of the objects to be recommended or the recommended receiving objects in the recommendation scenario.

[0211] S704, obtaining the preference information corresponding to each object in the object set, and determining the preference vector corresponding to each preference information; obtaining a preset dimensionality transformation matrix, and respectively performing dimensionality reduction processing on each preference vector through the dimensionality transformation matrix to obtain the preference features corresponding to each object.

[0212] S706, obtaining the attribute information of the neighbor object; the attribute information includes multiple attribute sub-information; determining the attribute vector corresponding to each attribute sub-information of the neighbor object.

[0213] S708, performing vector fusion processing on every two of the multiple attribute vectors to obtain multiple attribute fusion vectors, and performing superposition processing on the multiple attribute fusion vectors to obtain the two-way interaction fusion result of the neighbor object.

[0214] S710, performing linear fusion processing on the multiple attribute sub-information of the neighbor object to obtain a linear fusion result; performing result fusion processing on the two-way interaction fusion result and the linear fusion result to obtain the attribute feature of the neighbor object.

[0215] S712, obtaining the preference features and attribute information corresponding to each object in the object set; for the target object and the remaining objects other than the target object in the object set, determining the similarity between the preference feature of the target object and the preference feature of each remaining object to obtain the preference similarity corresponding to each remaining object.

[0216] S714. Determine the similarity between the attribute information of the target object and the attribute information of each of the remaining objects, and obtain the attribute similarity corresponding to each of the remaining objects.

[0217] S716. Perform similarity fusion processing on the preference similarity and the attribute similarity belonging to the same remaining object to obtain the target similarity corresponding to each of the remaining objects; use the remaining objects in the object set whose target similarity meets the preset similarity condition as the neighbor objects of the target object.

[0218] S718. Determine the feature contribution weight corresponding to each neighbor object of the target object; perform weighted average processing on the initial neighbor object features of the neighbor objects according to the feature contribution weight to obtain the aggregated neighbor object features of the target object.

[0219] S720. Stack the initial neighbor object features of the neighbor objects of the target object to obtain the stacked neighbor object features; divide the initial stacked features by the number of neighbor objects of the target object to obtain the neighbor object environment features of the target object.

[0220] S722. Concatenate the initial target object features and the neighbor object environment features to obtain the concatenated environment features, and multiply the pre-trained filtering matrix by the concatenated environment features to obtain the filtered concatenated features.

[0221] S724. Stack the filtered concatenated features and the pre-trained filtered bias vector to obtain the filtered stacked features; perform activation processing on the filtered stacked features through an activation function to obtain the features in the initial target object features that are incompatible with the neighbor environment features.

[0222] S726. Filter out the features in the initial target object features that are incompatible with the neighbor environment features to obtain the compatible features that are compatible with the neighbor environment features; perform feature fusion based on the aggregated neighbor object features and the compatible features to obtain the enhanced features of the target object.

[0223] In the above embodiments, since during the feature extraction process, not only the features of the target object are considered, but also the features of the neighbor objects similar to the target object are considered, the enhanced features obtained by combining the features of the target object and the neighbor objects can contain richer feature information.

[0224] In a specific embodiment, refer to Figure 8 , a method for object recommendation is provided, including:

[0225] S802. Obtain a recommended recipient object set and a recommended object set; the recommended recipient objects in the recommended recipient object set are user objects; the recommended objects in the recommended object set are videos or live broadcast rooms.

[0226] S804. Extract features for each recommended recipient in the recommended recipient set and each recommended object in the recommended object set respectively to obtain corresponding recommended recipient features and recommended object features. When extracting features for at least one of the recommended recipients or recommended objects, use the above-mentioned feature extraction method for extraction.

[0227] S806. For each recommended recipient in the recommended recipient set, based on the recommended recipient features of the current recommended recipient and the recommended object features corresponding to each recommended object, obtain the association relationship between the current recommended recipient and each recommended object.

[0228] S808. For each recommended recipient in the recommended recipient set, filter out the recommended objects from the recommended object set whose association relationship with the current recommended recipient meets the preset relationship conditions. Use the filtered recommended objects as the recommended objects to be recommended corresponding to the current recommended recipient.

[0229] S810. Determine the recommended objects to be recommended corresponding to each recommended recipient according to the association relationship.

[0230] In this embodiment, since when extracting features, not only the features of a single object are considered, but also the association features between objects are considered, the amount of information contained in the extracted features is improved based on the association features between objects. Furthermore, the accuracy of the determined recommended objects is improved based on the features with increased information amount, effectively improving the personalized recommendation accuracy of the recommended objects.

[0231] It should be understood that although the steps in the flowcharts involved in the above-described embodiments are shown in sequence according to the arrows, these steps do not necessarily need to be executed in the order indicated by the arrows. Unless there is a clear indication in this article, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, at least some of the steps in the flowcharts involved in the above-described embodiments may include multiple steps or multiple stages. These steps or stages do not necessarily need to be executed at the same time, but can be executed at different times. The execution order of these steps or stages does not necessarily need to be sequential, but can be executed alternately or in turn with at least some of the steps or stages in other steps or other steps.

[0232] Based on the same inventive concept, an embodiment of this application further provides a feature extraction device for implementing the feature extraction method involved above. The solution provided by this device to solve the problem is similar to the solution described in the above method. Therefore, the specific limitations in one or more of the following embodiments of the feature extraction device can refer to the limitations on the feature extraction method in the above text and will not be elaborated here.

[0233] In one embodiment, as Figure 9 shown, a feature extraction device 900 is provided, including: a neighbor object determination module 902, an aggregation gate module 904, a filtering gate module 906, and a feature fusion module 909:

[0234] The neighbor object determination module 902 is configured to obtain an object set including multiple objects; the object types of the multiple objects are the same; the object type is one of the objects to be recommended or the recommended recipient objects in the recommendation scenario; for the target object in the object set, similar neighbor objects to the target object are screened out from the object set, and the respective initial neighbor object features of the neighbor objects of the target object are determined;

[0235] The aggregation gate module 904 is configured to determine an aggregated neighbor object feature of the neighbor objects of the target object based on the initial neighbor object features; the aggregated neighbor object feature represents the aggregation of the features contributed by the neighbor objects of the target object respectively;

[0236] The filtering gate module 906 is configured to determine an environmental feature of the neighbor objects of the target object based on the initial neighbor object features; the environmental feature of the neighbor objects represents the overall environmental feature of the object environment constituted by the neighbor objects of the target object;

[0237] The feature fusion module 908 is configured to determine an initial target object feature of the target object, extract compatible features from the initial target object features that are compatible with the environmental feature of the neighbor objects; based on the aggregated neighbor object feature and the similar features, perform feature fusion to obtain an enhanced feature of the target object.

[0238] In one of the embodiments, the neighbor object determination module 902 obtains the respective preference features and attribute information corresponding to each object in the object set; for the target object and the remaining objects except the target object in the object set, determines the similarity between the preference feature of the target object and the preference feature of each remaining object respectively, to obtain the respective preference similarity corresponding to each remaining object; determines the similarity between the attribute information of the target object and the attribute information of each remaining object respectively, to obtain the respective attribute similarity corresponding to each remaining object; and screens out neighbor objects similar to the target object from the remaining objects except the target object in the object set according to the respective preference similarities and the respective attribute similarities.

[0239] In one embodiment, the feature extraction device 900 obtains the preference information corresponding to each object in the object set and determines the preference vector corresponding to each preference information; obtains a preset dimensionality transformation matrix, and respectively performs dimensionality reduction processing on each preference vector through the dimensionality transformation matrix to obtain the preference features corresponding to each object.

[0240] In one embodiment, the attribute information includes multiple attribute sub-information; for each remaining object in the object set, the feature extraction device 900 determines the similarity between each attribute sub-information of the target object and the corresponding attribute sub-information of the current remaining object, obtaining multiple sub-information similarities; performs an averaging process on the multiple sub-information similarities to obtain the attribute similarity between the attribute information of the target object and the attribute information of the current remaining object.

[0241] In one embodiment, the attribute sub-information includes an attribute name and an attribute value; for the multiple attribute sub-information of the target object, the feature extraction device 900 determines the target attribute sub-information in the current remaining object that has the same attribute name as the current attribute sub-information of the target object; determines the similarity between the attribute value of the target attribute sub-information of the current remaining object and the attribute value of the current attribute sub-information of the target object to obtain the sub-information similarity.

[0242] In one embodiment, the neighbor object determination module 902 performs a similarity fusion process on the preference similarity and the attribute similarity belonging to the same remaining object to obtain the target similarity corresponding to each remaining object; uses the remaining objects in the object set whose target similarity meets the preset similarity condition as the neighbor objects of the target object.

[0243] In one embodiment, the neighbor object determination module 902 obtains the preference features corresponding to each neighbor object of the target object, and obtains the attribute features corresponding to each neighbor object of the target object; performs a feature fusion process on the preference features and the attribute features corresponding to each neighbor object to obtain the initial neighbor object features corresponding to each neighbor object of the target object.

[0244] In one embodiment, the feature extraction device 900 obtains the attribute information of the neighbor object; the attribute information includes multiple attribute sub-information; performs a two-way interactive fusion process and a linear fusion process on the multiple attribute sub-information of the neighbor object to obtain the corresponding two-way interactive fusion result and linear fusion result; performs a result fusion process on the two-way interactive fusion result and the linear fusion result to obtain the attribute features of the neighbor object.

[0245] In one embodiment, the feature extraction device 900 determines the attribute vectors corresponding to each piece of attribute sub-information of the neighbor objects; performs vector fusion processing on every two of the multiple attribute vectors to obtain multiple attribute fusion vectors, and performs superposition processing on the multiple attribute fusion vectors to obtain the two-way interaction fusion result of the neighbor objects.

[0246] In one embodiment, the aggregation gate module 904 determines the feature contribution weights corresponding to each neighbor object of the target object; according to the feature contribution weights, performs weighted average processing on the initial neighbor object features of the neighbor objects to obtain the neighbor object aggregation feature of the target object.

[0247] In one embodiment, the aggregation gate module 904 determines the initial target object feature of the target object; splices the initial target object feature with the initial neighbor object features corresponding to each neighbor object of the target object respectively to obtain multiple object splicing features; multiplies each object splicing feature by the pre-trained aggregation matrix respectively to obtain multiple aggregated splicing features, and respectively superimposes each aggregated splicing feature with the pre-trained aggregation bias vector to obtain multiple aggregated superposition features; performs activation processing on each aggregated superposition feature through an activation function respectively to obtain the feature contribution weights corresponding to each neighbor object of the target object.

[0248] In one embodiment, the filtering gate module 906 superimposes the initial neighbor object features of the neighbor objects of the target object to obtain a neighbor object superposition feature; divides the initial superposition feature by the number of neighbor objects of the target object to obtain the neighbor object environment feature of the target object.

[0249] In one embodiment, the filtering gate module 906 splices the initial target object feature with the neighbor object environment feature to obtain an environment splicing feature, and multiplies the environment splicing feature by the pre-trained filtering matrix to obtain a filtered splicing feature; superimposes the filtered splicing feature with the pre-trained filtering bias vector to obtain a filtered superposition feature; performs activation processing on the filtered superposition feature through an activation function to obtain the features in the initial target object feature that are incompatible with the neighbor environment feature; filters the features in the initial target object feature that are incompatible with the neighbor environment feature to obtain the compatible features that are compatible with the neighbor environment feature.

[0250] Based on the same inventive concept, an embodiment of the present application further provides an object recommendation device for implementing the object recommendation method involved above. The implementation solution provided by this device for solving problems is similar to the implementation solution described in the above method. Therefore, the specific limitations in one or more of the following feature extraction device embodiments can refer to the limitations on the feature extraction method in the above text, and will not be repeated here.

[0251] In one embodiment, asFigure 10 As shown, an object recommendation device 1000 is provided, including: a feature extraction module 1002, a correlation relationship generation module 1004, and a recommendation module 1006:

[0252] The feature extraction module 1002 is configured to obtain a set of recommended recipient objects and a set of objects to be recommended; extract features for each recommended recipient object in the set of recommended recipient objects and each object to be recommended in the set of objects to be recommended respectively, to obtain corresponding recommended recipient object features and object to be recommended features; wherein, when extracting features for at least one of the recommended recipient objects or the objects to be recommended, a feature extraction device is used for extraction;

[0253] The correlation relationship generation module 1004 is configured to determine the correlation relationship between each recommended recipient object and each object to be recommended respectively according to the recommended recipient object features corresponding to each recommended recipient object and the object to be recommended features corresponding to each object to be recommended;

[0254] The recommendation module 1006 is configured to determine the objects to be recommended corresponding to each recommended recipient object according to the correlation relationship.

[0255] In one embodiment, the object recommendation device 1000 is further configured to obtain the first sample features of the recommended recipient object samples, the second sample features of the objects to be recommended samples, and the correlation relationship labels; at least one of the first sample features or the second sample features is obtained by using a feature extraction device; input the first sample features and the second sample features into a correlation relationship generation model to be trained, to obtain the predicted correlation relationship between the recommended recipient object samples and the objects to be recommended samples; determine the difference between the predicted correlation relationship and the correlation relationship labels, and construct a correlation relationship loss according to the difference; train the correlation relationship generation model through the correlation relationship loss until the training end condition is satisfied and then end, to obtain a trained correlation relationship generation model.

[0256] In one embodiment, the correlation relationship generation module 1004 is further configured to, for each recommended recipient object in the set of recommended recipient objects, obtain the correlation relationship between the current recommended recipient object and each object to be recommended respectively according to the recommended recipient object features of the current recommended recipient object and the object to be recommended features corresponding to each object to be recommended; the recommendation module is further configured to, for each recommended recipient object in the set of recommended recipient objects, screen out the objects to be recommended whose correlation relationship with the current recommended recipient object satisfies a preset relationship condition from the set of objects to be recommended; use the screened-out objects to be recommended as the objects to be recommended corresponding to the current recommended recipient object.

[0257] In one embodiment, the object recommendation device 1000 is further configured to output the recommended objects to be recommended corresponding to each user object; the recommended receiving objects in the recommended receiving object set are user objects; the recommended objects in the recommended object set include at least one of videos and live rooms.

[0258] Each module in the above-mentioned feature extraction device and object recommendation device can be implemented in whole or in part by software, hardware, and their combination. Each of the above modules can be embedded in the processor of the computer device in hardware form or be independent of it, or can be stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to each of the above modules.

[0259] In one embodiment, a computer device is provided. The computer device can be a server, and its internal structure diagram can be as Figure 11 shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O), and a communication interface. Among them, the processor, the memory, and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store feature extraction data and object recommendation data. The input / output interface of the computer device is used for the processor to exchange information with external devices. The communication interface of the computer device is used to communicate with external terminals through a network connection. When the computer program is executed by the processor, it implements a feature extraction method and an object recommendation method.

[0260] Those skilled in the art can understand that Figure 11 the structure shown in

[0261] is only a block diagram of some structures related to the solution of this application, and does not constitute a limitation on the computer device to which the solution of this application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements.

[0262] In one embodiment, a computer device is further provided, including a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, the steps in the above-mentioned method embodiments are implemented.

[0263] In one embodiment, a computer program product or a computer program is provided. The computer program product or the computer program includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, causing the computer device to perform the steps in the foregoing method embodiments.

[0264] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use, and processing of relevant data need to comply with the relevant laws, regulations, and standards of relevant countries and regions.

[0265] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, database, or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memories. Non-volatile memories can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memories can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The databases involved in the embodiments provided in the present application can include at least one of relational databases and non-relational databases. Non-relational databases can include distributed databases based on blockchain, etc., without limitation. The processors involved in the embodiments provided in the present application can be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, data processing logics based on quantum computing, etc., without limitation.

[0266] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0267] The above-described embodiments merely represent several implementation manners of the present application. The description is relatively specific and detailed, but it should not be construed as a limitation on the patent scope of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A feature extraction method, characterized in that, the method includes: obtaining an object set including a plurality of objects; the object types of the plurality of objects are the same; the object type is one of the objects to be recommended or the recommended recipient in the recommendation scenario; for a target object in the object set, screening out neighbor objects similar to the target object from the object set, and determining the respective initial neighbor object features of the neighbor objects of the target object; the neighbor objects are objects with attributes and preferences similar to those of the target object; respectively determining an aggregated neighbor object feature and an environmental neighbor object feature of the neighbor objects of the target object based on the initial neighbor object features; the aggregated neighbor object feature represents the aggregation of the features contributed by the neighbor objects of the target object; the environmental neighbor object feature represents the overall environmental feature constituted by the neighbor objects of the target object; determining an initial target object feature of the target object, and filtering out the features incompatible with the environmental neighbor object feature in the initial target object feature to obtain compatible features compatible with the environmental neighbor object feature; the compatibility means the same or similar; performing feature fusion based on the aggregated neighbor object feature and the compatible features to obtain an enhanced feature of the target object.

2. The method according to claim 1, characterized in that, the screening out neighbor objects similar to the target object from the object set includes: obtaining the respective preference features and attribute information corresponding to each object in the object set; for the target object and the remaining objects other than the target object in the object set, determining the similarity between the preference feature of the target object and the preference feature of each of the remaining objects to obtain the respective preference similarities corresponding to each of the remaining objects; determining the similarity between the attribute information of the target object and the attribute information of each of the remaining objects to obtain the respective attribute similarities corresponding to each of the remaining objects; screening out neighbor objects similar to the target object from the remaining objects other than the target object in the object set according to the respective preference similarities and the respective attribute similarities.

3. The method according to claim 2, characterized in that, the determining step of the preference feature includes: obtaining the respective preference information corresponding to each object in the object set, and determining the respective preference vectors corresponding to each preference information; obtaining a preset dimensionality transformation matrix, and respectively performing dimensionality reduction processing on each preference vector through the dimensionality transformation matrix to obtain the respective preference features corresponding to each object.

4. The method according to claim 2, characterized in that, the attribute information includes a plurality of attribute sub-information; the determining the similarity between the attribute information of the target object and the attribute information of each of the remaining objects to obtain the respective attribute similarities corresponding to each of the remaining objects includes: for each of the remaining objects in the object set, determining the similarity between each attribute sub-information of the target object and the corresponding attribute sub-information of the current remaining object to obtain a plurality of sub-information similarities; Performing an averaging process on the similarities of the multiple sub-information to obtain the attribute similarity between the attribute information of the target object and the attribute information of the current remaining objects.

5. The method according to claim 4, wherein, the attribute sub-information includes an attribute name and an attribute value; the determining the similarity between each attribute sub-information of the target object and the corresponding attribute sub-information of the current remaining objects respectively to obtain multiple sub-information similarities includes: For multiple attribute sub-information of the target object, determining the target attribute sub-information in the current remaining objects that has the same attribute name as the current attribute sub-information of the target object; Determining the similarity between the attribute value of the target attribute sub-information of the current remaining objects and the attribute value of the current attribute sub-information of the target object to obtain the sub-information similarity.

6. The method according to claim 1, wherein, the determining the respective initial neighbor object features of the neighbor objects of the target object includes: Obtaining the respective preference features corresponding to each neighbor object of the target object, and obtaining the respective attribute features corresponding to each neighbor object of the target object; Performing a feature fusion process on the respective preference features and attribute features corresponding to each neighbor object to obtain the respective initial neighbor object features of each neighbor object of the target object.

7. The method according to claim 6, wherein, the determining step of the attribute features of the neighbor object includes: Obtaining the attribute information of the neighbor object; the attribute information includes multiple attribute sub-information; Performing a two-way interaction fusion process and a linear fusion process on the multiple attribute sub-information of the neighbor object to obtain the corresponding two-way interaction fusion result and linear fusion result; Performing a result fusion process on the two-way interaction fusion result and the linear fusion result to obtain the attribute features of the neighbor object.

8. The method according to claim 7, wherein, the performing a two-way interaction fusion process on the multiple attribute sub-information of the neighbor object includes: Determining the respective attribute vectors corresponding to each attribute sub-information of the neighbor object; Performing a vector fusion process on every two of the multiple attribute vectors to obtain multiple attribute fusion vectors, and performing a superposition process on the multiple attribute fusion vectors to obtain the two-way interaction fusion result of the neighbor object.

9. The method according to claim 1, wherein, the determining step of the aggregated feature of the neighbor objects of the target object includes: Determining the respective feature contribution weights corresponding to each neighbor object of the target object; Performing a weighted averaging process on the initial neighbor object features of the neighbor object according to the feature contribution weights to obtain the aggregated feature of the neighbor objects of the target object.

10. The method according to claim 9, wherein, the determining the respective feature contribution weights corresponding to each neighbor object of the target object includes: Determining the initial target object feature of the target object; Concatenating the initial target object feature with the respective initial neighbor object features corresponding to each neighbor object of the target object to obtain multiple object concatenated features; Multiply the pre-trained aggregation matrix by each of the object concatenated features to obtain a plurality of aggregated concatenated features, and respectively superimpose each of the aggregated concatenated features with the pre-trained aggregation bias vector to obtain a plurality of aggregated superimposed features; Perform activation processing on each of the aggregated superimposed features through an activation function to obtain the feature contribution weights corresponding to each of the neighbor objects of the target object.

11. The method according to claim 1, wherein, the step of determining the neighbor object environmental features of the target object includes: Superimpose the respective initial neighbor object features of the neighbor objects of the target object to obtain a neighbor object superimposed feature; Divide the neighbor object superimposed feature by the number of neighbor objects of the target object to obtain the neighbor object environmental features of the target object.

12. The method according to claim 1, wherein, the filtering of the features in the initial target object feature that are incompatible with the neighbor object environmental features to obtain compatible features that are compatible with the neighbor object environmental features includes: Concatenate the initial target object feature with the neighbor object environmental feature to obtain an environmental concatenated feature, and multiply the pre-trained filtering matrix by the environmental concatenated feature to obtain a filtered concatenated feature; Superimpose the filtered concatenated feature and the pre-trained filtering bias vector to obtain a filtered superimposed feature; Perform activation processing on the filtered superimposed feature through an activation function to obtain the features in the initial target object feature that are incompatible with the neighbor object environmental features; Filter out the features in the initial target object feature that are incompatible with the neighbor object environmental features to obtain compatible features that are compatible with the neighbor object environmental features.

13. An object recommendation method, wherein, the method includes: Obtain a set of recommended receiving objects and a set of objects to be recommended; Extract features for each recommended receiving object in the set of recommended receiving objects and each object to be recommended in the set of objects to be recommended respectively to obtain corresponding recommended receiving object features and object to be recommended features; Determine the association relationship between each recommended receiving object and each object to be recommended according to the recommended receiving object features corresponding to each recommended receiving object and the object to be recommended features corresponding to each object to be recommended; Determine the object to be recommended corresponding to each recommended receiving object according to the association relationship; wherein, when extracting features for at least one of the recommended receiving objects or the objects to be recommended, the feature extraction method described in any one of claims 1 to 12 is used for extraction.

14. The method according to claim 13, wherein, the association relationship between the recommended receiving object and the object to be recommended is determined by an association relationship generation model; the training steps of the association relationship generation model include: Obtain the first sample features of the recommended recipient samples, the second sample features of the recommended object samples, and the association relationship labels; at least one of the first sample features or the second sample features is obtained by using the feature extraction method described in any one of claims 1 to 12; Input the first sample features and the second sample features into the association relationship generation model to be trained, and obtain the predicted association relationship between the recommended recipient samples and the recommended object samples; Determine the difference between the predicted association relationship and the association relationship label, and construct an association relationship loss according to the difference; Train the association relationship generation model through the association relationship loss until the training end condition is met and then end, to obtain a trained association relationship generation model.

15. The method according to claim 13, wherein, determining the association relationship between each recommended recipient and each recommended object according to the recommended recipient features corresponding to each recommended recipient and the recommended object features corresponding to each recommended object respectively, includes: For each recommended recipient in the recommended recipient set, obtain the association relationship between the current recommended recipient and each recommended object according to the recommended recipient features of the current recommended recipient and the recommended object features corresponding to each recommended object respectively; The determining the recommended objects to be recommended corresponding to each recommended recipient according to the association relationship includes: For each recommended recipient in the recommended recipient set, screen out the recommended objects from the recommended object set whose association relationship with the current recommended recipient meets the preset relationship conditions; Take the screened recommended objects as the recommended objects to be recommended corresponding to the current recommended recipient.

16. The method according to any one of claims 13 to 15, wherein, The recommended recipients in the recommended recipient set are user objects; the recommended objects in the recommended object set are at least one of videos and live rooms; the object recommendation method is executed by an object recommendation model; the recommended objects corresponding to each user object output by the object recommendation model include at least one of videos and live rooms.

17. A feature extraction device, wherein, The device includes: A neighbor object determination module, configured to obtain an object set including a plurality of objects; the object types of the plurality of objects are the same; the object type is one of the recommended objects or recommended recipients in the recommendation scenario; for the target object in the object set, screen out the neighbor objects similar to the target object from the object set, and determine the initial neighbor object features of the neighbor objects of the target object respectively; the neighbor objects are objects with attributes and preferences similar to the target object; An aggregation gate module, configured to determine the aggregated neighbor object features of the target object based on the initial neighbor object features; the aggregated neighbor object features represent the aggregation of the features contributed by the neighbor objects of the target object respectively; A filtering door module, configured to determine the neighbor object environment feature of the target object based on the initial neighbor object feature; the neighbor object environment feature represents the overall environment feature of the object environment constituted by the neighbor objects of the target object. A feature fusion module, configured to determine the initial target object feature of the target object, filter the features in the initial target object feature that are incompatible with the neighbor object environment feature, and obtain the compatible features that are compatible with the neighbor object environment feature; the compatibility means being the same or similar; perform feature fusion based on the neighbor object aggregation feature and the compatible features to obtain the enhanced feature of the target object.

18. The apparatus according to claim 17, wherein, the neighbor object determination module is further configured to obtain the preference feature and the attribute information respectively corresponding to each object in the object set; for the target object and the remaining objects except the target object in the object set, determine the similarity between the preference feature of the target object and the preference feature of each of the remaining objects, and obtain the preference similarity respectively corresponding to each of the remaining objects; determine the similarity between the attribute information of the target object and the attribute information of each of the remaining objects, and obtain the attribute similarity respectively corresponding to each of the remaining objects; and screen out the neighbor objects similar to the target object from the remaining objects except the target object in the object set according to the respective preference similarities and the respective attribute similarities.

19. The apparatus according to claim 18, wherein, the feature extraction device is further configured to obtain the preference information respectively corresponding to each object in the object set, and determine the preference vector respectively corresponding to each preference information; obtain a preset dimensionality transformation matrix, and perform dimensionality reduction processing on each preference vector through the dimensionality transformation matrix to obtain the preference feature respectively corresponding to each object.

20. The apparatus according to claim 18, wherein, the attribute information includes a plurality of attribute sub-information; the feature extraction device is further configured to, for each of the remaining objects in the object set, determine the similarity between each attribute sub-information of the target object and the corresponding attribute sub-information of the current remaining object, and obtain a plurality of sub-information similarities; perform an averaging process on the plurality of sub-information similarities to obtain the attribute similarity between the attribute information of the target object and the attribute information of the current remaining object.

21. The apparatus according to claim 20, wherein, the attribute sub-information includes an attribute name and an attribute value; the feature extraction device is further configured to, for the plurality of attribute sub-information of the target object, determine the target attribute sub-information in the current remaining object that has the same attribute name as the current attribute sub-information of the target object; determine the similarity between the attribute value of the target attribute sub-information of the current remaining object and the attribute value of the current attribute sub-information of the target object to obtain the sub-information similarity.

22. The apparatus according to claim 17, wherein, The neighbor object determination module is further configured to obtain the preference features corresponding to each neighbor object of the target object, and obtain the attribute features corresponding to each neighbor object of the target object; perform feature fusion processing on the preference features and attribute features corresponding to each neighbor object to obtain the initial neighbor object features corresponding to each neighbor object of the target object.

23. The apparatus according to claim 22, wherein, the feature extraction apparatus is further configured to obtain the attribute information of the neighbor object; the attribute information includes a plurality of attribute sub-information; perform two-way interactive fusion processing and linear fusion processing on the plurality of attribute sub-information of the neighbor object to obtain corresponding two-way interactive fusion results and linear fusion results; perform result fusion processing on the two-way interactive fusion result and the linear fusion result to obtain the attribute feature of the neighbor object.

24. The apparatus according to claim 23, wherein, the feature extraction apparatus is further configured to determine the attribute vector corresponding to each attribute sub-information of the neighbor object; perform vector fusion processing on every two of the plurality of attribute vectors to obtain a plurality of attribute fusion vectors, and perform superposition processing on the plurality of attribute fusion vectors to obtain the two-way interactive fusion result of the neighbor object.

25. The apparatus according to claim 17, wherein, the aggregation gate module is further configured to determine the feature contribution weight corresponding to each neighbor object of the target object; perform weighted average processing on the initial neighbor object features of the neighbor object according to the feature contribution weight to obtain the neighbor object aggregation feature of the target object.

26. The apparatus according to claim 25, wherein, the aggregation gate module is further configured to determine the initial target object feature of the target object; splice the initial target object feature with the initial neighbor object features corresponding to each neighbor object of the target object respectively to obtain a plurality of object splicing features; multiply the pre-trained aggregation matrix by each of the object splicing features respectively to obtain a plurality of aggregation splicing features, and perform superposition processing on each of the aggregation splicing features and the pre-trained aggregation bias vector respectively to obtain a plurality of aggregation superposition features; perform activation processing on each aggregation superposition feature through an activation function to obtain the feature contribution weight corresponding to each neighbor object of the target object.

27. The apparatus according to claim 17, wherein, the filtering gate module is further configured to superimpose the initial neighbor object features of the neighbor objects of the target object to obtain a neighbor object superposition feature; divide the neighbor object superposition feature by the number of neighbor objects of the target object to obtain the neighbor object environment feature of the target object.

28. The apparatus according to claim 17, wherein, the filtering gate module is further configured to splice the initial target object feature with the neighbor object environment feature to obtain an environment splicing feature, and multiply the pre-trained filtering matrix by the environment splicing feature to obtain a filtering splicing feature; Superimpose the filtered splicing feature and the pre-trained filtered bias vector to obtain a filtered superimposed feature; perform an activation process on the filtered superimposed feature through an activation function to obtain the feature in the initial target object feature that is incompatible with the neighbor object environment feature; Filter out the features in the initial target object feature that are incompatible with the neighbor object environment feature to obtain compatible features that are compatible with the neighbor object environment feature.

29. An object recommendation device, characterized in that, the device includes: A feature extraction module, configured to obtain a set of recommended receiving objects and a set of recommended objects; extract features for each recommended receiving object in the set of recommended receiving objects and each recommended object in the set of recommended objects respectively, to obtain corresponding recommended receiving object features and recommended object features; wherein, when extracting features for at least one of the recommended receiving object or the recommended object, the feature extraction device as described in claim 17 is used for extraction; An association relationship generation module, configured to determine the association relationship between each recommended receiving object and each recommended object according to the recommended receiving object feature corresponding to each recommended receiving object and the recommended object feature corresponding to each recommended object; A recommendation module, configured to determine the recommended object to be recommended corresponding to each recommended receiving object according to the association relationship.

30. The device according to claim 29, characterized in that, the association relationship between the recommended receiving object and the recommended object is determined by an association relationship generation model; the object recommendation device is further configured to obtain a first sample feature of a recommended receiving object sample, a second sample feature of a recommended object sample, and an association relationship label; at least one of the first sample feature or the second sample feature is obtained by using the feature extraction method as described in any one of claims 1 to 12; input the first sample feature and the second sample feature into the association relationship generation model to be trained to obtain the predicted association relationship between the recommended receiving object sample and the recommended object sample; determine the difference between the predicted association relationship and the association relationship label, and construct an association relationship loss according to the difference; train the association relationship generation model through the association relationship loss until the training end condition is met and then end to obtain a trained association relationship generation model.

31. The device according to claim 29, characterized in that, the association relationship generation module is further configured to, for each recommended receiving object in the set of recommended receiving objects, obtain the association relationship between the current recommended receiving object and each recommended object according to the recommended receiving object feature of the current recommended receiving object and the recommended object features corresponding to each recommended object; the recommendation module is further configured to, for each recommended receiving object in the set of recommended receiving objects, screen out the recommended objects from the set of recommended objects whose association relationship with the current recommended receiving object meets the preset relationship condition; The selected recommended object is used as the recommended object to be recommended corresponding to the current recommended recipient object.

32. The apparatus according to claim 29, wherein, the recommended recipient objects in the recommended recipient object set are user objects; the recommended objects in the recommended object set are at least one of videos and live broadcast rooms.

33. A computer device, comprising a memory and a processor, the memory storing a computer program, wherein, when the processor executes the computer program, the steps of the method according to any one of claims 1 to 16 are implemented.

34. A computer-readable storage medium, having a computer program stored thereon, wherein, when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 16 are implemented.

35. A computer program product, comprising a computer program, wherein, when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 16 are implemented.

Images