Feature extraction method and device, electronic equipment and storage medium

Abstract

The present disclosure relates to a feature extraction method and device, electronic equipment and storage medium, and belongs to the technical field of computers. The method comprises: obtaining association information, the association information comprising a plurality of objects and an association relationship between any two objects, the plurality of objects comprising an account registered on a server and a multimedia resource published through the server; determining a first object as an initial target object; selecting a next target object that satisfies an association condition from the association information based on a current target object until the number of determined target objects reaches a target number, thereby constructing a target object set; and performing feature extraction on the target object set to obtain an object feature of the first object. When the feature is extracted using the target object set, the influence of the association relationship between the objects on the first object can be fully considered, thereby improving the accuracy of the object feature.

Description

Technical Field

[0001] This disclosure relates to the field of computer technology, and in particular to a feature extraction method, apparatus, electronic device, and storage medium. Background Technology

[0002] With the development of computer technology and the increasing scale of the Internet, there are more and more types of objects on the Internet, such as users, videos, and authors. As the types of objects increase, there is also more and more information to consider when extracting the object characteristics of objects. How to determine the object characteristics of objects has become an urgent problem to be solved. Summary of the Invention

[0003] This disclosure provides a feature extraction method, apparatus, electronic device, and storage medium, which improves the accuracy of extracted object features.

[0004] According to one aspect of the embodiments of the present disclosure, a feature extraction method is provided, the feature extraction method comprising:

[0005] Obtain association information, which includes multiple objects and the association relationship between any two objects. The multiple objects include accounts registered on the server and multimedia resources published through the server.

[0006] The first object is determined as the initial target object, which is any account registered on the server or any multimedia resource published through the server;

[0007] From the association information, select the next target object that meets the association condition based on the current target object, until the number of determined target objects reaches the target number, and construct a target object set. The association condition refers to the association relationship between the next target object and the current target object being the currently determined target association relationship. The target object set includes the target number of target objects and the association relationship between any two target objects.

[0008] Feature extraction is performed on the target object set to obtain the object features of the first object.

[0009] The method provided in this disclosure, during the construction of a target object set, determines the next target object associated with the current target object based on the association relationships between different objects, so that the association relationship between the next target object and the current target object becomes the current target association relationship. It considers the impact of different association relationships on the target object, and determines multiple target objects through cross-relationships. Therefore, when extracting features using the target object set, it can fully consider the impact of the association relationships between each object on the first object, thereby improving the accuracy of object features.

[0010] In some embodiments, selecting the next target object that satisfies the association condition from the association information based on the current target object includes:

[0011] Based on the association information, various types of association relationships of the current target object are determined;

[0012] Select one type of association from the multiple types, and determine the selected association as the current target association;

[0013] From the associated information, select the object whose association relationship with the current target object is the current target association relationship, and determine the selected object as the next target object.

[0014] In this embodiment of the disclosure, one type of association relationship is selected from the various types of association relationships of the current target object, and the next target object is determined based on the selected association relationship. The influence of different association relationships on the two associated objects is taken into consideration. Compared with directly determining the next target object based on one type of association relationship, the influence of different association relationships on the target object has been considered in the process of selecting association relationships.

[0015] In some embodiments, selecting objects from the association information whose association with the current target object is the current target association, and determining the selected objects as the next target object, includes:

[0016] From the association information, select multiple candidate objects for the current target association relationship based on the association relationship between them;

[0017] Choose any one of the multiple candidate objects as the next target object.

[0018] In this embodiment of the disclosure, when the association relationship between the current target object and multiple candidate objects is a target association relationship, any one of the multiple candidate objects is selected to determine the next target object, thus ensuring that a target object that meets the association conditions can be selected.

[0019] In some embodiments, the feature extraction method further includes:

[0020] According to the various types of association relationships contained in the association information, the association information is divided into multiple association sub-information, each association sub-information contains multiple objects, and different objects in the same association sub-information have the same type of association relationship;

[0021] The step of extracting features from the target object set to obtain the object features of the first object includes:

[0022] From the plurality of associated sub-information, determine the target associated sub-information that contains the first object;

[0023] Feature extraction is performed on the target object set and the target associated sub-information to obtain the object features.

[0024] In this embodiment of the disclosure, the associated information is divided into multiple associated sub-information. Based on the target object set and the target associated sub-information, object features are extracted. While fully considering the influence of the association between each object on the first object, the influence of other objects corresponding to the same association on the first object is also considered, which further improves the accuracy of object features.

[0025] In some embodiments, the target association sub-information is multiple, and the step of extracting features from the target object set and the target association sub-information to obtain the object features includes:

[0026] For each target-related sub-information, feature extraction is performed on the target object set and the target-related sub-information to obtain the first feature of the first object;

[0027] Based on the first weights corresponding to multiple target association sub-information, the first features corresponding to the multiple target association sub-information are weighted to obtain the object features. The first weights corresponding to each target association sub-information characterize the degree of influence of the association relationship corresponding to the target association sub-information on the object features of the first object.

[0028] In this embodiment of the disclosure, when there are multiple target associated sub-information, considering that the degree of influence of the association relationship corresponding to different target associated sub-information on the first feature of the first object is different, the first features corresponding to each target associated sub-information are weighted and processed, and the features obtained by weighting are used as object features, which further improves the accuracy of object features.

[0029] In some embodiments, the target association sub-information is multiple, and the step of extracting features from the target object set and the target association sub-information to obtain the object features includes:

[0030] For each target-related sub-information, feature extraction is performed on the target object set and the target-related sub-information to obtain the first feature of the first object;

[0031] Based on the first weights corresponding to multiple target association sub-informations, the first features corresponding to the multiple target association sub-informations are weighted to obtain the second features of the first object. The first weights corresponding to each target association sub-information characterize the degree of influence of the association relationship corresponding to the target association sub-information on the object features of the first object.

[0032] Feature extraction is performed on the associated information to obtain the third feature of the first object;

[0033] The second feature is concatenated with the third feature to obtain the object feature.

[0034] In this embodiment of the disclosure, when extracting object features, feature extraction is performed on the associated information. This not only considers the impact of the association between various objects on the first object, and the impact of other objects corresponding to the same association on the first object, but also considers the global information corresponding to the first object, thereby further improving the accuracy of object features.

[0035] In some embodiments, the type to which the first object belongs is a target type. The step of extracting features from the target object set and the target associated sub-information for each target association sub-information to obtain a first feature of the first object includes:

[0036] For each target association sub-information, the target association sub-information is divided into multiple first sub-information. Each first sub-information contains the first object and other objects belonging to the reference type. The other objects and the first object constitute the association relationship corresponding to the target association sub-information. The reference types contained in different first sub-information are different, and the reference types are different from the target type.

[0037] The target object set is divided into multiple second sub-information, each second sub-information containing the first object and other objects, and the association relationship between the first object and the other objects includes at least two types, and the association relationship between the first object and the other objects in different second sub-information is not completely the same;

[0038] Feature extraction is performed on the plurality of first sub-information and the plurality of second sub-information respectively to obtain the fourth feature corresponding to the plurality of first sub-information and the fifth feature corresponding to the plurality of second sub-information;

[0039] Based on the second weights corresponding to the plurality of first sub-information and the second weights corresponding to the plurality of second sub-information, the fourth feature corresponding to the plurality of first sub-information and the fifth feature corresponding to the plurality of second sub-information are weighted to obtain the first feature of the first object. The second weights corresponding to each first sub-information represent the degree of influence of different objects under the same association on the object feature of the first object, and the second weights corresponding to each second sub-information represent the degree of influence of different objects under different associations on the object feature of the first object.

[0040] In this embodiment of the disclosure, the target association sub-information and the target object set are further divided to obtain multiple first sub-information corresponding to the target association sub-information and multiple second sub-information corresponding to the target object set. When extracting the fourth feature corresponding to each first sub-information, the influence of different types of objects associated with the first object on the first object can be fully considered. When extracting the fifth feature corresponding to each second sub-information, the influence of different association relationships on the first object can be fully considered. Compared with directly extracting features from the target association sub-information, the accuracy of the obtained first features is improved.

[0041] In some embodiments, dividing the target association sub-information into multiple first sub-information includes:

[0042] For each of the target-related sub-information, determine multiple reference types that are different from the target type;

[0043] Each time, one of the reference types is selected, the first object and other objects that meet the first selection condition are selected from the target association sub-information to form a first sub-information, until the multiple reference types are selected and multiple first sub-information are obtained. The selected object meets the first selection condition if the object belongs to the reference type and forms an association relationship with the first object corresponding to the target association sub-information.

[0044] In this embodiment of the disclosure, first sub-information is constructed according to reference type to ensure that each piece of first sub-information can reflect the degree of influence of the object belonging to the corresponding reference type on the first object.

[0045] In some embodiments, dividing the target object set into multiple second sub-information includes:

[0046] Determine the various types of associations included in the target object set;

[0047] Each time, at least two types of association relationships are selected from the multiple types. The first object and other objects that meet the second selection condition are selected from the target object set to form a second sub-information. This process continues until there are no more association relationships of at least two types that have not been selected at the same time, resulting in multiple second sub-information. The selected object meets the second selection condition if the association relationship between the selected object and the first object is one of the at least two types of association relationships.

[0048] In this embodiment of the disclosure, second sub-information is constructed according to different association relationships to ensure that each second sub-information can reflect the degree of influence of at least two types of association relationships on the first object.

[0049] In some embodiments, the step of extracting features from the target object set to obtain object features of the first object includes:

[0050] The feature extraction model is invoked to extract features from the target object set, thereby obtaining the object features of the first object.

[0051] In this embodiment of the disclosure, calling the feature extraction model can quickly and accurately extract the object features of the first object, thereby improving the efficiency of object feature extraction.

[0052] In some embodiments, the step of extracting features from the target object set and the target associated sub-information to obtain the object features includes:

[0053] The feature extraction model is invoked to extract features from the target object set and the target associated sub-information to obtain the object features.

[0054] In this embodiment of the disclosure, a feature extraction model is invoked to process the target object set and target associated sub-information, which improves the efficiency of object feature extraction and also improves the accuracy of object feature extraction.

[0055] In some embodiments, the feature extraction model includes multiple first feature extraction networks and second feature extraction networks, and the target association sub-information is multiple. The step of calling the feature extraction model to extract features from the target object set and the target association sub-information to obtain the object features includes:

[0056] Each first feature extraction network is invoked to extract features from the target object set and each target associated sub-information to obtain the first feature of the first object;

[0057] The second feature extraction network is invoked, and the first features corresponding to the multiple target association sub-information are weighted based on the first weights corresponding to the multiple target association sub-information to obtain the object features. The first weights corresponding to each target association sub-information characterize the degree of influence of the association relationship corresponding to the target association sub-information on the object features of the first object.

[0058] In some embodiments, the feature extraction model includes multiple first feature extraction networks, second feature extraction networks, and third feature extraction networks, and the target association sub-information is multiple. The step of calling the feature extraction model to extract features from the target object set and the target association sub-information to obtain the object features includes:

[0059] Each first feature extraction network is invoked to extract features from the target object set and each target associated sub-information to obtain the first feature of the first object;

[0060] The second feature extraction network is invoked, and the first features corresponding to the multiple target association sub-informations are weighted based on the first weights corresponding to the multiple target association sub-informations to obtain the second features of the first object. The first weights corresponding to each target association sub-information characterize the degree of influence of the association relationship corresponding to the target association sub-information on the object features of the first object.

[0061] The third feature extraction network is invoked to extract features from the associated information, thereby obtaining the third feature of the first object;

[0062] The second feature extraction network is invoked to concatenate the second feature with the third feature to obtain the object feature.

[0063] In some embodiments, before invoking the second feature extraction network to perform weighted processing on the first features corresponding to the multiple target association sub-information based on the first weights corresponding to the multiple target association sub-information to obtain the second features of the first object, the feature extraction method further includes:

[0064] The second feature extraction network is invoked to process the multiple target association sub-information to obtain the first weights corresponding to the multiple target association sub-information.

[0065] In some embodiments, the step of invoking each first feature extraction network to extract features from the target object set and each target-related sub-information to obtain the first feature of the first object includes:

[0066] For each target-related sub-information, the feature extraction layer in the first feature extraction network corresponding to the target-related sub-information is called to divide the target-related sub-information into multiple first sub-information. Each first sub-information contains the first object and other objects belonging to the reference type. The other objects and the first object constitute the association relationship corresponding to the target-related sub-information. The reference types contained in different first sub-information are different.

[0067] The feature extraction layer is invoked to divide the target object set into multiple second sub-information. Each second sub-information contains the first object and other objects, and the association relationship between the first object and the other objects includes at least two types. The association relationship between the first object and the other objects in different second sub-information is not completely the same.

[0068] The feature extraction layer is invoked to extract features from the plurality of first sub-information and the plurality of second sub-information respectively, to obtain the fourth feature corresponding to the plurality of first sub-information and the fifth feature corresponding to the plurality of second sub-information;

[0069] The concatenation layer in the first feature extraction network is invoked. Based on the second weights corresponding to the multiple first sub-information and the second weights corresponding to the multiple second sub-information, the fourth feature corresponding to the multiple first sub-information and the fifth feature corresponding to the multiple second sub-information are weighted to obtain the first feature of the first object. The second weights corresponding to each first sub-information represent the degree of influence of different objects under the same association on the object feature of the first object. The second weights corresponding to each second sub-information represent the degree of influence of different objects under different associations on the object feature of the first object.

[0070] In some embodiments, before invoking the concatenation layer in the first feature extraction network to perform weighted processing on the fourth feature corresponding to the plurality of first sub-information and the fifth feature corresponding to the plurality of second sub-information based on the second weights corresponding to the plurality of first sub-information and the second weights corresponding to the plurality of second sub-information to obtain the first feature of the first object, the feature extraction method further includes:

[0071] The attention layer in the first feature extraction network is invoked to process the plurality of first sub-information and the plurality of second sub-information respectively, so as to obtain the second weights corresponding to the plurality of first sub-information and the second weights corresponding to the plurality of second sub-information.

[0072] In some embodiments, the step of invoking the feature extraction layer in the first feature extraction network corresponding to the target associated sub-information to divide the target associated sub-information into multiple first sub-information includes:

[0073] The feature extraction layer is invoked to determine multiple reference types that are different from the target type;

[0074] The feature extraction layer is invoked, and one reference type is selected each time. The first object and other objects that meet the first selection condition are selected from the target association sub-information to form a first sub-information. This process continues until the selection of multiple reference types is completed, resulting in multiple first sub-information. The selected object meets the first selection condition if the object belongs to the reference type and forms an association relationship with the first object corresponding to the target association sub-information.

[0075] In some embodiments, invoking the feature extraction layer to divide multiple second sub-information from the target object set includes:

[0076] The feature extraction layer is invoked to determine the various types of relationships included in the target object set;

[0077] The feature extraction layer is invoked, and at least two types of association relationships are selected from the multiple types each time. The first object and other objects that meet the second selection condition are selected from the target object set to form a second sub-information. This process continues until there are no more association relationships of at least two types that have not been selected at the same time, resulting in multiple second sub-information. Here, the selected object meets the second selection condition if the association relationship between the selected object and the first object is one of the at least two types of association relationships.

[0078] In some embodiments, the training process of the feature extraction model is as follows:

[0079] Two sample objects are obtained from the sample association information, wherein the similarity between the two sample objects is greater than the reference similarity;

[0080] The feature extraction model is invoked to obtain the predicted features corresponding to the two sample objects respectively;

[0081] The predicted features corresponding to the two sample objects are processed using an objective function to obtain the loss value;

[0082] The feature extraction model is trained based on the loss value.

[0083] In some embodiments, after training the feature extraction model based on the loss value, the feature extraction method further includes:

[0084] If the loss value is less than the reference value, the training of the feature extraction model is terminated.

[0085] In some embodiments, the association information is a heterogeneous graph, which includes multiple object nodes and connecting lines between any two object nodes. An object node represents an object, and a connecting line represents the association relationship between the objects corresponding to the two object nodes connected by the connecting line.

[0086] The step of selecting the next target object that meets the association conditions from the association information based on the current target object, until the number of identified target objects reaches the target number, and constructing a target object set, includes:

[0087] From the heterogeneous graph, select the next target object node that meets the association condition based on the current target object node, until the number of determined target object nodes reaches the target number, and construct a target node network. The association condition refers to the connection line between the next target object node and the current target object node matching the currently determined target association relationship. The target node network includes the target number of target object nodes and the connection line between any two target object nodes.

[0088] The step of extracting features from the target object set to obtain the object features of the first object includes:

[0089] Feature extraction is performed on the target node network to obtain the object features.

[0090] In this embodiment of the disclosure, during the construction of the target node network, the next target object node associated with the current target object node is determined based on the association relationship of the connection lines between different object nodes. This ensures that the connection line between the next target object node and the current target object node matches the current target association relationship. By considering the impact of different association relationships on the target object nodes, the determination of multiple target object nodes across relationships is achieved. Thus, when extracting features using the target node network, the impact of the association relationship of the connection lines between various object nodes on the first object can be fully considered, thereby improving the accuracy of object features.

[0091] In some embodiments, the feature extraction method further includes:

[0092] Based on the various connection lines contained in the heterogeneous graph, the heterogeneous graph is divided into multiple association subgraphs. Each association subgraph contains multiple object nodes, and the connection lines between different object nodes in the same association subgraph match the same type of association.

[0093] The step of extracting features from the relationship network to obtain the object features of the first object includes:

[0094] From the plurality of association subgraphs, determine the target association subgraph that contains the first object node corresponding to the first object;

[0095] Feature extraction is performed on the target node network and the target association subgraph to obtain the object features.

[0096] In this embodiment of the disclosure, the heterogeneous graph is divided into multiple association subgraphs, and object features are extracted based on the target node set and the target association subgraph. While fully considering the influence of the association relationship of the connection lines between each object node on the first object node, the influence of other object nodes connected by the same type of connection line on the first object node is also considered, which further improves the accuracy of object features.

[0097] According to another aspect of the present disclosure, a feature extraction apparatus is provided, the feature extraction apparatus comprising:

[0098] The association information acquisition unit is configured to acquire association information, which includes multiple objects and the association relationship between any two objects. The multiple objects include accounts registered on the server and multimedia resources published through the server.

[0099] The initial object determination unit is configured to determine a first object as the initial target object, wherein the first object is any account registered on the server or any multimedia resource published through the server;

[0100] The object set construction unit is configured to perform the following: from the association information, select the next target object that meets the association condition based on the current target object, until the number of determined target objects reaches the target number, and construct a target object set. The association condition refers to the association relationship between the next target object and the current target object being the currently determined target association relationship. The target object set includes the target number of target objects and the association relationship between any two target objects.

[0101] The feature extraction unit is configured to perform feature extraction on the target object set to obtain the object features of the first object.

[0102] In some embodiments, the object collection construction unit includes:

[0103] The association determination subunit is configured to determine multiple types of associations of the current target object based on the association information;

[0104] The association selection subunit is configured to select one type of association from the multiple types and determine the selected association as the current target association.

[0105] The object determination subunit is configured to select an object from the association information that has an association relationship with the current target object as the current target association relationship, and determine the selected object as the next target object.

[0106] In some embodiments, the object determining subunit is configured to perform:

[0107] From the association information, select multiple candidate objects for the current target association relationship based on the association relationship between them;

[0108] Choose any one of the multiple candidate objects as the next target object.

[0109] In some embodiments, the feature extraction apparatus further includes:

[0110] The association information division unit is configured to divide the association information into multiple association sub-information according to the multiple types of association relationships contained in the association information. Each association sub-information contains multiple objects, and different objects in the same association sub-information have the same type of association relationship.

[0111] The feature extraction unit is configured to perform:

[0112] From the plurality of associated sub-information, determine the target associated sub-information that contains the first object;

[0113] Feature extraction is performed on the target object set and the target associated sub-information to obtain the object features.

[0114] In some embodiments, the target associated sub-information is multiple, and the feature extraction unit includes:

[0115] The first feature extraction subunit is configured to perform feature extraction on the target object set and the target association sub-information for each target association sub-information to obtain the first feature of the first object;

[0116] The object feature extraction subunit is configured to perform weighted processing on the first features corresponding to the multiple target association sub-information based on the first weights corresponding to the multiple target association sub-information to obtain the object features. The first weights corresponding to each target association sub-information characterize the degree of influence of the association relationship corresponding to the target association sub-information on the object features of the first object.

[0117] In some embodiments, the target associated sub-information is multiple, and the feature extraction unit includes:

[0118] The first feature extraction subunit is configured to perform feature extraction on the target object set and the target association sub-information for each target association sub-information to obtain the first feature of the first object;

[0119] The second feature determination subunit is configured to perform weighted processing on the first features corresponding to the multiple target association sub-informations based on the first weights corresponding to the multiple target association sub-informations to obtain the second features of the first object. The first weights corresponding to each target association sub-information characterize the degree of influence of the association relationship corresponding to the target association sub-information on the object features of the first object.

[0120] The third feature extraction subunit is configured to perform feature extraction on the associated information to obtain the third feature of the first object;

[0121] The object feature extraction subunit is configured to concatenate the second feature with the third feature to obtain the object feature.

[0122] In some embodiments, the type to which the first object belongs is a target type, and the first feature extraction subunit is configured to perform:

[0123] For each target association sub-information, the target association sub-information is divided into multiple first sub-information. Each first sub-information contains the first object and other objects belonging to the reference type. The other objects and the first object constitute the association relationship corresponding to the target association sub-information. The reference types contained in different first sub-information are different, and the reference types are different from the target type.

[0124] The target object set is divided into multiple second sub-information, each second sub-information containing the first object and other objects, and the association relationship between the first object and the other objects includes at least two types, and the association relationship between the first object and the other objects in different second sub-information is not completely the same;

[0125] Feature extraction is performed on the plurality of first sub-information and the plurality of second sub-information respectively to obtain the fourth feature corresponding to the plurality of first sub-information and the fifth feature corresponding to the plurality of second sub-information;

[0126] Based on the second weights corresponding to the plurality of first sub-information and the second weights corresponding to the plurality of second sub-information, the fourth feature corresponding to the plurality of first sub-information and the fifth feature corresponding to the plurality of second sub-information are weighted to obtain the first feature of the first object. The second weights corresponding to each first sub-information represent the degree of influence of different objects under the same association on the object feature of the first object, and the second weights corresponding to each second sub-information represent the degree of influence of different objects under different associations on the object feature of the first object.

[0127] In some embodiments, the first feature extraction subunit is configured to perform:

[0128] For each of the target-related sub-information, determine multiple reference types that are different from the target type;

[0129] Each time, one of the reference types is selected, the first object and other objects that meet the first selection condition are selected from the target association sub-information to form a first sub-information, until the multiple reference types are selected and multiple first sub-information are obtained. The selected object meets the first selection condition if the object belongs to the reference type and forms an association relationship with the first object corresponding to the target association sub-information.

[0130] In some embodiments, the first feature extraction subunit is configured to perform:

[0131] Determine the various types of associations included in the target object set;

[0132] Each time, at least two types of association relationships are selected from the multiple types. The first object and other objects that meet the second selection condition are selected from the target object set to form a second sub-information. This process continues until there are no more association relationships of at least two types that have not been selected at the same time, resulting in multiple second sub-information. The selected object meets the second selection condition if the association relationship between the selected object and the first object is one of the at least two types of association relationships.

[0133] In some embodiments, the feature extraction unit is configured to execute a feature extraction model to extract features from the target object set to obtain the object features of the first object.

[0134] In some embodiments, the feature extraction unit is configured to execute a feature extraction model to extract features from the target object set and the target associated sub-information to obtain the object features.

[0135] In some embodiments, the feature extraction model includes multiple first feature extraction networks and second feature extraction networks, the target associated sub-information is multiple, and the feature extraction unit includes:

[0136] The first feature extraction subunit is configured to call each first feature extraction network to extract features from the target object set and each target associated sub-information respectively, so as to obtain the first feature of the first object;

[0137] The object feature extraction subunit is configured to execute the call to the second feature extraction network, and perform weighted processing on the first features corresponding to the multiple target association sub-information based on the first weights corresponding to the multiple target association sub-information to obtain the object features. The first weights corresponding to each target association sub-information characterize the degree of influence of the association relationship corresponding to the target association sub-information on the object features of the first object.

[0138] In some embodiments, the feature extraction model includes multiple first feature extraction networks, second feature extraction networks, and third feature extraction networks, the target associated sub-information is multiple, and the feature extraction unit includes:

[0139] The first feature extraction subunit is configured to call each first feature extraction network to extract features from the target object set and each target associated sub-information respectively, so as to obtain the first feature of the first object;

[0140] The second feature determination subunit is configured to execute the call to the second feature extraction network, and perform weighted processing on the first features corresponding to the multiple target association sub-informations based on the first weights corresponding to the multiple target association sub-informations to obtain the second feature of the first object. The first weights corresponding to each target association sub-information characterize the degree of influence of the association relationship corresponding to the target association sub-information on the object feature of the first object.

[0141] The third feature extraction subunit is configured to call the third feature extraction network to extract features from the associated information and obtain the third feature of the first object.

[0142] The object feature extraction subunit is configured to call the second feature extraction network to concatenate the second feature with the third feature to obtain the object feature.

[0143] In some embodiments, the feature extraction unit further includes:

[0144] The first weight determination subunit is configured to execute the call to the second feature extraction network to process the multiple target association sub-information and obtain the first weight corresponding to the multiple target association sub-information.

[0145] In some embodiments, the first feature extraction subunit is configured to perform:

[0146] For each target-related sub-information, the feature extraction layer in the first feature extraction network corresponding to the target-related sub-information is called to divide the target-related sub-information into multiple first sub-information. Each first sub-information contains the first object and other objects belonging to the reference type. The other objects and the first object constitute the association relationship corresponding to the target-related sub-information. The reference types contained in different first sub-information are different.

[0147] The feature extraction layer is invoked to divide the target object set into multiple second sub-information. Each second sub-information contains the first object and other objects, and the association relationship between the first object and the other objects includes at least two types. The association relationship between the first object and the other objects in different second sub-information is not completely the same.

[0148] The feature extraction layer is invoked to extract features from the plurality of first sub-information and the plurality of second sub-information respectively, to obtain the fourth feature corresponding to the plurality of first sub-information and the fifth feature corresponding to the plurality of second sub-information;

[0149] The concatenation layer in the first feature extraction network is invoked. Based on the second weights corresponding to the multiple first sub-information and the second weights corresponding to the multiple second sub-information, the fourth feature corresponding to the multiple first sub-information and the fifth feature corresponding to the multiple second sub-information are weighted to obtain the first feature of the first object. The second weights corresponding to each first sub-information represent the degree of influence of different objects under the same association on the object feature of the first object. The second weights corresponding to each second sub-information represent the degree of influence of different objects under different associations on the object feature of the first object.

[0150] In some embodiments, the feature extraction unit further includes:

[0151] The second weight determination subunit is configured to call the attention layer in the first feature extraction network to process the plurality of first sub-information and the plurality of second sub-information respectively, and obtain the second weights corresponding to the plurality of first sub-information and the second weights corresponding to the plurality of second sub-information.

[0152] In some embodiments, the first feature extraction subunit is configured to perform:

[0153] The feature extraction layer is invoked to determine multiple reference types that are different from the target type;

[0154] The feature extraction layer is invoked, and one reference type is selected each time. The first object and other objects that meet the first selection condition are selected from the target association sub-information to form a first sub-information. This process continues until the selection of multiple reference types is completed, resulting in multiple first sub-information. The selected object meets the first selection condition if the object belongs to the reference type and forms an association relationship with the first object corresponding to the target association sub-information.

[0155] In some embodiments, the first feature extraction subunit is configured to perform:

[0156] The feature extraction layer is invoked to determine the various types of relationships included in the target object set;

[0157] The feature extraction layer is invoked, and at least two types of association relationships are selected from the multiple types each time. A first object and other objects that meet the second selection condition are selected from the target object set to form a second sub-information. This process continues until there are no more association relationships of at least two types that have not been selected at the same time, resulting in multiple second sub-information. Here, the selected object meets the second selection condition if the association relationship between the selected object and the first object is one of the at least two types of association relationships.

[0158] In some embodiments, the feature extraction apparatus further includes:

[0159] The model training unit is configured to perform the acquisition of two sample objects from sample association information, wherein the similarity between the two sample objects is greater than the reference similarity.

[0160] The model training unit is also configured to call the feature extraction model to obtain the predicted features corresponding to the two sample objects respectively;

[0161] The model training unit is also configured to perform a target function to process the predicted features corresponding to the two sample objects to obtain a loss value;

[0162] The model training unit is also configured to train the feature extraction model based on the loss value.

[0163] In some embodiments, the model training unit is further configured to terminate the training of the feature extraction model in response to the loss value being less than a reference value.

[0164] In some embodiments, the association information is a heterogeneous graph, which includes multiple object nodes and connecting lines between any two object nodes. An object node represents an object, and a connecting line represents the association relationship between the objects corresponding to the two object nodes connected by the connecting line.

[0165] The object set construction unit is configured to perform the following operations: from the heterogeneous graph, select the next target object node that satisfies the association condition based on the current target object node, until the number of determined target object nodes reaches the target number, and construct a target node network. The association condition refers to the connection line between the next target object node and the current target object node matching the currently determined target association relationship. The target node network includes the target number of target object nodes and the connection line between any two target object nodes.

[0166] The feature extraction unit is configured to perform feature extraction on the target node network to obtain the object features.

[0167] In some embodiments, the feature extraction apparatus further includes:

[0168] The association information partitioning unit is configured to divide the heterogeneous graph into multiple association subgraphs according to the multiple connection lines contained in the heterogeneous graph. Each association subgraph contains multiple object nodes, and the connection lines between different object nodes in the same association subgraph match the same type of association relationship.

[0169] The feature extraction unit is configured to perform:

[0170] From the plurality of association subgraphs, determine the target association subgraph that contains the first object node corresponding to the first object;

[0171] Feature extraction is performed on the target node network and the target association subgraph to obtain the object features.

[0172] According to another aspect of the present disclosure, an electronic device is provided, the electronic device comprising:

[0173] One or more processors;

[0174] Memory for storing the one or more processor-executable instructions;

[0175] The one or more processors are configured to perform the feature extraction method described above.

[0176] According to another aspect of the present disclosure, a computer-readable storage medium is provided, wherein when instructions in the storage medium are executed by a processor of an electronic device, the electronic device is enabled to perform the feature extraction method described above.

[0177] According to another aspect of the present disclosure, a computer program product is provided, the computer program product including a computer program that is executed by a processor to implement the feature extraction method described above.

[0178] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0179] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure.

[0180] Figure 1 This is a flowchart illustrating a feature extraction method according to an exemplary embodiment;

[0181] Figure 2 This is a flowchart illustrating another feature extraction method according to an exemplary embodiment;

[0182] Figure 3 This is a flowchart illustrating another feature extraction method according to an exemplary embodiment;

[0183] Figure 4 This is a flowchart illustrating another feature extraction method according to an exemplary embodiment;

[0184] Figure 5 This is a flowchart illustrating another feature extraction method according to an exemplary embodiment;

[0185] Figure 6 This is a schematic diagram illustrating the model structure of a feature extraction model according to an exemplary embodiment;

[0186] Figure 7 This is a schematic diagram illustrating the model structure of another feature extraction model according to an exemplary embodiment;

[0187] Figure 8 This is a flowchart illustrating another feature extraction method according to an exemplary embodiment;

[0188] Figure 9 This is a schematic diagram illustrating another feature extraction method according to an exemplary embodiment;

[0189] Figure 10 This is a block diagram illustrating a feature extraction apparatus according to an exemplary embodiment;

[0190] Figure 11 This is a block diagram illustrating another feature extraction apparatus according to an exemplary embodiment;

[0191] Figure 12 This is a structural block diagram of a terminal according to an exemplary embodiment;

[0192] Figure 13 This is a structural block diagram of a server according to an exemplary embodiment. Detailed Implementation

[0193] To enable those skilled in the art to better understand the technical solutions of this disclosure, the technical solutions in the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings.

[0194] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0195] It should be noted that the user information involved in this disclosure (including but not limited to user device information, user personal information, etc.) is all information authorized by the user or fully authorized by all parties.

[0196] To facilitate understanding of the feature extraction method provided in the embodiments of this disclosure, the keywords involved in the embodiments of this disclosure are explained as follows:

[0197] Network embedding technology refers to a technique that uses machine learning to transform complex networks into low-dimensional, dense feature vectors that preserve the original network topology. The feature vectors obtained through network embedding can be used for various operations, such as node classification, node prediction, and link prediction. Network embedding includes homogeneous graph-based network embedding and heterogeneous graph-based network embedding. A homogeneous graph is a network containing only a single type of node and a single type of edge. A heterogeneous graph is a graph where the sum of the number of node types and edge types is greater than two; for example, a heterogeneous graph may contain one type of node and multiple types of edges, or multiple types of nodes and one type of node, or multiple types of nodes and multiple types of edges.

[0198] The feature extraction method provided in this disclosure can be applied to various scenarios:

[0199] For example, this can be applied to scenarios where videos are recommended to users. Users register accounts on a server via their terminals. Terminals with accounts can then publish multimedia resources through the server, and these resources can be displayed on the terminals with accounts. Users interact with these multimedia resources through their terminals, creating associations between accounts or between different accounts. Alternatively, associations can arise between multimedia resources based on their type. Based on the accounts registered on the server, the multimedia resources published through the server, the associations between accounts, the associations between accounts and multimedia resources, and the associations between multimedia resources, association information encompassing accounts, multimedia resources, and various types of associations can be obtained.

[0200] For any given account, the association information is divided into multiple sub-associations based on the various types of associations included in the account's associated information. The target associated sub-association information containing the account is determined from these sub-associations. Based on this target associated sub-association information, the user characteristics of the user corresponding to the account are obtained. Then, the recommendation model or other recommendation methods are called to process the user characteristics and determine the videos recommended to that user.

[0201] Figure 1 This is a flowchart illustrating a feature extraction method according to an exemplary embodiment, see [link to flowchart]. Figure 1 This method, when applied to electronic devices, includes the following steps:

[0202] 101. Obtain related information.

[0203] The associated information includes various types of objects and the relationships between any two objects. These objects include accounts registered on the server and multimedia resources published through the server. Published multimedia resources include videos, audio, articles, or other forms of multimedia resources. The relationships include those between accounts, between accounts and multimedia resources, and between multimedia resources themselves. Examples include following relationships or other relationships between accounts, likes, clicks, comments, shares, publications, or other relationships between accounts and multimedia resources, and type relationships or other relationships between multimedia resources.

[0204] 102. Set the first object as the initial target object.

[0205] Here, the first object is any object in the associated information, that is, any account registered on the server or any multimedia resource published through the server. The initial target object refers to the multiple target objects related to the first object that are obtained sequentially.

[0206] 103. From the associated information, select the next target object that meets the association conditions based on the current target object, until the number of identified target objects reaches the target number, and construct a target object set.

[0207] Among them, the association condition refers to the association relationship between the next target object and the current target object being the currently determined target association relationship. The constructed target object set includes the target number of target objects and the association relationship between any two target objects.

[0208] The process of constructing the target object set is as follows: Taking the first object as the initial target object, determine the first association relationship of the first object from the various types of association relationships of the first object. From the multiple objects associated with the first object, select the second object whose association relationship with the first object is the first association relationship, and determine the selected second object as the next target object. Similarly, for the second object, determine the second association relationship of the second object from the various types of association relationships of the second object. From the multiple objects associated with the second object, select the third object whose association relationship with the second object is the second association relationship, and determine the selected third object as the next target object. Continue to determine the next target object based on the third object, until the number of determined target objects reaches the target number, thus obtaining the target object set.

[0209] 104. Extract features from the target object set to obtain the object features of the first object.

[0210] Here, object features are characteristics used to describe the first object, and object features can be in vector form, matrix form, or other forms. Since the constructed target object set starts with the first object and determines multiple target objects sequentially based on the relationships between objects, all determined target objects are objects that have a direct or indirect relationship with the first object. Therefore, by extracting features from the target object set, the object features of the first object can be obtained.

[0211] The method provided in this disclosure, during the construction of a target object set, determines the next target object associated with the current target object based on the association relationships between different objects, so that the association relationship between the next target object and the current target object becomes the current target association relationship. It considers the impact of different association relationships on the target object, and determines multiple target objects through cross-relationships. Therefore, when extracting features using the target object set, it can fully consider the impact of the association relationships between each object on the first object, thereby improving the accuracy of object features.

[0212] Figure 2 This is a flowchart illustrating another feature extraction method according to an exemplary embodiment, see [link to flowchart]. Figure 2 The method is executed by an electronic device and includes the following steps:

[0213] 201. Electronic devices acquire associated information.

[0214] The associated information includes multiple objects and the relationships between any two objects. The multiple objects include accounts registered on the server and multimedia resources published through the server. Published multimedia resources include videos, audio, articles, or other forms of multimedia resources. The relationships include relationships between accounts, between accounts and multimedia resources, and between multimedia resources themselves. Examples of relationships include following or other relationships between accounts, likes, clicks, comments, shares, publications, or other relationships between accounts and multimedia resources, and type relationships or other relationships between multimedia resources. Type relationships refer to different multimedia resources belonging to the same type; for example, two multimedia resources are both food videos.

[0215] Multimedia resources published through a server refer to resources that can be displayed on terminals logged into with an account. For example, if a terminal logged into account A publishes a video through the server, terminals logged into other accounts can also display the video and interact with it. Furthermore, interacting with the video establishes an association between the logged-in account and the video.

[0216] In some embodiments, the association between any two objects is determined based on an operation performed by one object on the other. For example, if any two objects are a user's account and a video, and the user likes the video through a terminal logged into the corresponding account, then a liking relationship is determined between the account and the video; or if any two objects are accounts belonging to two users, and one account follows the other account, then a following relationship is determined between the two accounts.

[0217] In some embodiments, the association information is sent to the electronic device by other electronic devices, or obtained by the electronic device based on an account registered on the server, multimedia resources published by the server, and the association relationship between objects. This disclosure does not limit the method of obtaining association information.

[0218] It should be noted that the embodiments of this disclosure do not limit the storage format of the associated information in the electronic device. For example, the associated information can be stored in a table format, a graph network format, or other formats. For example, when the associated information is stored in a graph network format, since the associated information includes multiple types of objects, the associated information is a heterogeneous graph. The heterogeneous graph includes multiple object nodes and connecting lines between any two object nodes. One object node represents one object, and one connecting line represents the association relationship between the objects corresponding to the two object nodes connected by the connecting line.

[0219] 202. The electronic device identifies the first object as the initial target object.

[0220] In this embodiment of the disclosure, the feature extraction process is described using the extraction of object features of a first object as an example.

[0221] Here, the first object is any account or any multimedia resource in the associated information. Taking the first object as the initial target object means that, starting from the first object, multiple target objects related to the first object are obtained sequentially according to the association relationships between the various objects. Subsequently, based on the obtained multiple target objects and the association relationships between them, the object features of the first object are extracted.

[0222] 203. The electronic device selects the next target object that meets the association conditions from the associated information based on the current target object, until the number of identified target objects reaches the target number, and constructs a target object set.

[0223] The association condition refers to the association relationship between the next target object and the current target object, which is the currently determined target association relationship. The target object set includes the target number of target objects and the association relationship between any two target objects. The target number can be any number, for example, 10, 20 or other numbers.

[0224] In some embodiments, the electronic device selects the next target object that satisfies the association conditions based on the current target object, including: the electronic device determining multiple types of association relationships of the current target object based on association information; selecting one type of association relationship from the determined multiple types, and determining the selected association relationship as the current target association relationship; then selecting objects from the association information whose association relationship with the current target object is the current target association relationship, and determining the selected objects as the next target object. Selecting one type of association relationship from the multiple types of the current target object and determining the next target object based on the selected association relationship considers the impact of different association relationships on the two associated objects. Compared with directly determining the next target object based on one type of association relationship, the process of selecting association relationships already considers the impact of different association relationships on the target object. Furthermore, in the process of selecting target objects multiple times, the target association relationships used are not completely the same; that is, multiple target objects are determined based on multiple target association relationships, realizing the determination of target objects across relationships. Moreover, this method of selecting the next target object ultimately selects multiple target objects that include not only objects directly associated with the first object, but also objects associated with the first object at intervals of one or more other target objects.

[0225] In some embodiments, after the electronic device determines the current target association relationship, it selects multiple candidate objects from the association information that have an association relationship with the current target object as the current target association relationship; and determines any one of the multiple candidate objects as the next target object, thereby realizing the selection of the next target object and ensuring that a target pair that meets the association conditions can be selected.

[0226] Taking a target quantity of m as an example, the process of constructing a target object set is explained as follows: For the first object, the electronic device acquires the first object, determines various types of association relationships for the first object, selects a first association relationship from these relationships, and then, based on the selected first association relationship, determines several candidate objects whose association relationship with the first object is that first association relationship. Any one of these candidate objects is selected as the second object. For the second object, the operation of determining the next target object from the first object is repeated to obtain a third object associated with the second object. This process continues until the m-th target object is obtained. Based on the acquired m target objects and the association relationships between them, the electronic device constructs a target object set. Here, m is a positive integer, and the first association relationship is any one of the various types of association relationships.

[0227] In some embodiments, the (t+1)th target object associated with the t-th target object is determined using the following formula:

[0228]

[0229]

[0230] Where p(r) t+1 |r t v t p(v) represents the probability of selecting any one type of association from multiple types. t+1 |r t+1 v t ) represents the probability that the selected object from multiple candidate objects is the (t+1)th target object, r t This represents the first association relationship used to determine the t-th target object based on the (t-1)-th target object, r. t+1 This represents the second association relationship used to determine the (t+1)th target object based on the t-th target object, v t v represents the t-th target object. t+1 This represents the (t+1)th target object. N represents a set of associations that includes various types of relationships with the target object. r (v t () represents the first neighbor set of the t-th target object under the first association relationship. This first neighbor set includes multiple objects that have a first association relationship with the t-th target object. Let represent the second neighbor set of the t-th target object, which includes multiple objects with a definite second association relationship with the t-th target object.

[0231] In the process of constructing the target object set described above, when determining the (t+1)th target object associated with the t-th target object, the joint probability distribution satisfied by the (t+1)th target object is: p(r t+1 |r t v t )×p(v t+1 |r t+1 v t For example, if the t-th target object has k types of associations, and each type of association has l candidate objects associated with it, then the probability of selecting the (t+1)-th target object is:

[0232] 204. The electronic device divides the associated information into multiple associated sub-information according to the various types of association relationships contained in the associated information, and determines multiple target associated sub-information containing the first object from the multiple associated sub-information.

[0233] Each associated sub-information contains multiple objects, and different objects within the same associated sub-information have the same type of association relationship.

[0234] For example, if the associated information includes relationships such as likes, shares, and comments, then the associated information is divided into three sub-information. Each of the likes, shares, and comments corresponds to one sub-information. The relationship between each pair of objects in the sub-information corresponding to the likes relationship is called the likes relationship; the relationship between each pair of objects in the sub-information corresponding to the shares relationship is called the shares relationship; and the relationship between each pair of objects in the sub-information corresponding to the comments relationship is called the comments relationship.

[0235] The electronic device determines multiple target associated sub-information from multiple associated sub-information based on whether the associated sub-information contains the first object, and then extracts the features of the first object based on the multiple target associated sub-information.

[0236] In some embodiments, if only one associated sub-information contains the first object, a target associated sub-information is determined, and features of the first object are subsequently extracted based on the target associated sub-information.

[0237] 205. For each target-related sub-information, the electronic device performs feature extraction on the target object set and the target-related sub-information to obtain the first feature of the first object.

[0238] The first feature describes the characteristics of the first object under the influence of the corresponding association relationship. For example, if the association relationship corresponding to the target associated sub-information is a "like" relationship, then the first feature is determined based on other objects with a "like" relationship to the first object; or, if the association relationship corresponding to the target associated sub-information is a "comment" relationship, then the first feature is determined based on other objects with a "comment" relationship to the first object. These two different association relationships result in different first features.

[0239] In some embodiments, see Figure 3 The process by which an electronic device acquires the first feature is as follows:

[0240] 2051. For each target-related sub-information, the electronic device divides the target-related sub-information into multiple first sub-information.

[0241] Each of the first sub-information segments contains a first object and other objects belonging to the reference type. These other objects and the first object form an association relationship corresponding to the target association sub-information. Furthermore, the reference types contained in different first sub-information segments are different. The type to which the first object belongs is the target type, while the reference types are different from the target type.

[0242] In some embodiments, for each target associated sub-information, multiple reference types different from the target type are determined; each time, one reference type is selected, a first object and other objects that meet the first selection condition are selected from the target associated sub-information, and the first object and the selected other objects constitute a first sub-information, until multiple reference types are selected, and multiple first sub-information are obtained. The selected object meets the first selection condition if the object belongs to the reference type and forms an association relationship with the first object corresponding to the target associated sub-information.

[0243] For example, if the target associated sub-information includes three types of objects: account, audio, and video, and the association relationship corresponding to the target associated sub-information is a like relationship, and the first object is an account, then based on the audio type, the audio associated with the account in the target associated sub-information is selected, and the account and audio are combined into a first sub-information, in which the audio and account have a like relationship; based on the video type, the video associated with the account in the target associated sub-information is selected, and the account and video are combined into a first sub-information, in which the video and account have a like relationship.

[0244] It should be noted that each first sub-information may include only the first object belonging to the target type, or it may also include other objects belonging to the target type. For example, the first object may be the first account, and the first sub-information may also include a second account, with the first account and the second account respectively associated with the video.

[0245] 2052. Electronic devices divide the target object set into multiple second sub-information.

[0246] Each of the second sub-information contains a first object and other objects, and the association between the first object and other objects includes at least two types. The association between the first object and other objects in different second sub-information is not completely the same.

[0247] In some embodiments, the electronic device determines multiple types of association relationships included in the target object set; at least two types of association relationships are selected each time, and a first object and other objects that meet the second selection condition are selected from the target object set to form a second sub-information, until there are no more than two types of association relationships that have not been selected at the same time, and multiple second sub-information are obtained, wherein the selected object meets the second selection condition if the association relationship between the selected object and the first object is at least two types of association relationship.

[0248] For example, if the target object set includes like relationships, comment relationships, and repost relationships, and the first object is an account, then the first video that has a like relationship with the account and the second video that has a repost relationship with the account are selected to form a second sub-information. The first video and the second video can be the same or different.

[0249] 2053. The electronic device extracts features from multiple first sub-information and multiple second sub-information respectively, and obtains the fourth feature corresponding to the multiple first sub-information and the fifth feature corresponding to the multiple second sub-information.

[0250] The electronic device extracts features from each first sub-information to obtain a fourth feature corresponding to each first sub-information. The fourth feature corresponding to each first sub-information is used to describe the features of the first object under the influence of the reference type object included in the first sub-information. It also extracts features from each second sub-information to obtain a fifth feature corresponding to each second sub-information. The fifth feature corresponding to each second sub-information is used to describe the features of the first object under the influence of at least two types of association relationships included in the second sub-information.

[0251] 2054. The electronic device performs weighted processing on the fourth feature corresponding to the first sub-information and the fifth feature corresponding to the second sub-information based on the second weight corresponding to the first sub-information and the second weight corresponding to the second sub-information, to obtain the first feature of the first object.

[0252] Among them, the second weight corresponding to each first sub-information represents the degree of influence of different objects on the object characteristics of the first object under the same association relationship, and the second weight corresponding to each second sub-information represents the degree of influence of different objects on the object characteristics of the first object under different association relationships.

[0253] In the above Figure 3 In the process shown, the target association sub-information and the target object set are further divided to obtain multiple first sub-information corresponding to the target association sub-information and multiple second sub-information corresponding to the target object set. When extracting the fourth feature corresponding to each first sub-information, the influence of different types of objects associated with the first object on the first object can be fully considered. When extracting the fifth feature corresponding to each second sub-information, the influence of different association relationships on the first object can be fully considered. Compared with directly extracting features from the target association sub-information, the accuracy of the obtained first features is improved.

[0254] The above Figure 3 Taking only one target-related sub-information as an example, the above method can be applied to each target-related sub-information. Figure 3 The method shown determines the first feature of the first object corresponding to each target associated sub-information.

[0255] 206. The electronic device performs weighted processing on the first features corresponding to the multiple target association sub-information based on the first weights corresponding to the multiple target association sub-information to obtain the second features of the first object.

[0256] Among them, the first weight corresponding to each target association sub-information represents the degree of influence of the association relationship corresponding to the target association sub-information on the object features of the first object. For example, for a certain account, there is a like relationship and a forwarding relationship between the account and the video, and the account likes more videos but forwards fewer videos, which means that the like relationship has a greater impact on the features of the account, while the forwarding relationship has a smaller impact on the features of the account.

[0257] In some embodiments, the electronic device processes multiple target association sub-information based on an attention mechanism to obtain a first weight corresponding to each target association sub-information.

[0258] 207. The electronic device extracts features from the associated information to obtain the third feature of the first object.

[0259] Feature extraction is performed on the associated information, taking into account the global information corresponding to the first object. Therefore, the extracted third feature is used to characterize the global features of the first object in the entire association relationship.

[0260] 208. The electronic device combines the second feature with the third feature to obtain the object feature of the first object.

[0261] In some embodiments, the electronic device determines the sum of the second feature and the third feature as the object feature of the first object, so that the obtained first feature can more accurately describe the first object.

[0262] In some embodiments, the electronic device may skip steps 207 and 208 above and directly use the obtained second feature of the first object as the object feature of the first object. That is, the electronic device processes the target object set and multiple target association sub-information to obtain object features. This method of extracting object features fully considers the influence of the association relationship between each object on the first object, and also considers the influence of other objects corresponding to the same association relationship on the first object, further improving the accuracy of the extracted features.

[0263] Furthermore, in some embodiments, the obtained object features are processed based on the recommendation model to determine another object recommended to the first object. Since the accuracy of the object features obtained in this embodiment is more precise, the recommendation model can more accurately recommend another object of interest to the first object based on these features, thus improving the recommendation effect. The recommendation model can be any model used for object recommendation. Specifically, the recommendation model can be a content-based recommendation model, a collaborative filtering-based recommendation model, a neural network-based recommendation model, or other recommendation models.

[0264] The method provided in this disclosure, during the construction of a target object set, determines the next target object associated with the current target object based on the association relationships between different objects, so that the association relationship between the next target object and the current target object becomes the current target association relationship. It considers the impact of different association relationships on the target object, and determines multiple target objects through cross-relationships. Therefore, when extracting features using the target object set, it can fully consider the impact of the association relationships between each object on the first object, thereby improving the accuracy of object features.

[0265] Furthermore, when there are multiple target-related sub-information, considering that the influence of the association relationships corresponding to different target-related sub-information on the first feature of the first object is different, the first features corresponding to each target-related sub-information are weighted and the weighted features are used as object features, which further improves the accuracy of object features.

[0266] In some embodiments, taking the associated information as an example, such as a heterogeneous graph, the above... Figure 2 The feature extraction process shown is explained below:

[0267] Figure 4 This is a flowchart illustrating another feature extraction method according to an exemplary embodiment, see [link to flowchart]. Figure 3 The method is executed by an electronic device and includes the following steps:

[0268] 401. Obtaining heterogeneous graphs for electronic devices.

[0269] In a heterogeneous graph, there are multiple object nodes and connecting lines between any two object nodes. An object node represents an object, and a connecting line represents the association between the objects corresponding to the two object nodes connected by the connecting line.

[0270] 402. The electronic device determines the first object node corresponding to the first object as the initial target object node.

[0271] In this embodiment of the disclosure, taking the extraction of object features of a first object as an example, the first object node corresponding to the first object is determined as the initial target object node.

[0272] 403. From the heterogeneous graph, the electronic device selects the next target object node that meets the association conditions based on the current target object node, until the number of identified target object nodes reaches the target number, thus constructing a target node network.

[0273] The association condition refers to the matching of the connection line between the next target object node and the current target object node with the currently determined target association relationship. The target node network includes the target number of target object nodes and the connection line between any two target object nodes.

[0274] In some embodiments, the electronic device selects the next target object node that satisfies the association conditions based on the current target object node, including: the electronic device determines multiple connection lines connecting the current target object node to other object nodes based on a heterogeneous graph; selects one connection line from the determined multiple connection lines and determines the selected connection line as the current target connection line; and then selects an object node from the heterogeneous graph whose connection line with the current target object node is the current target connection line and determines the selected object node as the next target object node.

[0275] In some embodiments, after the electronic device determines the current target connection line, it selects multiple candidate object nodes from the heterogeneous graph, which are the connection lines between the current target object node and the current target object node; and determines any one of the multiple candidate object nodes as the next target object node.

[0276] 404. The electronic device divides the heterogeneous graph into multiple relational subgraphs according to the various connection lines contained in the heterogeneous graph, and determines multiple target relational subgraphs containing the first object node from the multiple relational subgraphs.

[0277] Each association subgraph contains multiple object nodes, and the connecting lines between different object nodes in the same association subgraph match the same type of association.

[0278] The electronic device determines multiple target association subgraphs from multiple association subgraphs based on whether the association subgraph contains the first object node, and then extracts the features of the first object node based on these multiple target association subgraphs.

[0279] 405. For each target association subgraph, the electronic device performs feature extraction on the target node network and the target association subgraph to obtain the first feature corresponding to the first object node.

[0280] In some embodiments, see Figure 5 The process by which an electronic device acquires the first feature is as follows:

[0281] 4051. For each target association subgraph, the electronic device divides the target association subgraph into multiple first subgraphs.

[0282] Each first subgraph contains a first object node and other object nodes of the reference node type. The connecting line between any two object nodes in the same first subgraph matches the association relationship corresponding to the target association subgraph. Different first subgraphs contain reference nodes of different types. The first object node belongs to the target node type, while the reference node types differ from the target node types.

[0283] In some embodiments, for each target association subgraph, multiple reference node types different from the target type are determined; each time, one reference node type is selected, and a first object node and other object nodes that satisfy the third selection condition are selected from the target association subgraph. The first object node and other object nodes are combined to form a first subgraph, until multiple reference node types are selected, resulting in multiple first subgraphs. The selected object node satisfies the third selection condition if it belongs to a reference node type and has a connection line with the first object node.

[0284] 4052. Electronic devices divide the target node subgraph into multiple second subgraphs.

[0285] Each second subgraph contains a first object node and other object nodes, and the connection lines between the first object node and other object nodes match at least two types of association relationships. The association relationships between the first object node and other object nodes in different second subgraphs are not completely the same.

[0286] In some embodiments, the electronic device determines multiple types of association relationships included in the target node subgraph; at least two types of association relationships are selected each time, and other object nodes connected to the first object node are selected from the target node subgraph to form a second subgraph, wherein the connection lines between the other object nodes and the first object node are connection lines that match the at least two types of association relationships, until there are no more at least two types of association relationships that have not been selected simultaneously, thus obtaining multiple second subgraphs.

[0287] 4053. The electronic device extracts features from multiple first sub-images and multiple second sub-images respectively, and obtains the fourth feature corresponding to the multiple first sub-images and the fifth feature corresponding to the multiple second sub-images.

[0288] The electronic device performs feature extraction on each first subgraph to obtain a fourth feature corresponding to each first subgraph. The fourth feature corresponding to each first subgraph is used to describe the feature of the first object node under the influence of the reference node type object node included in the first subgraph. The electronic device also performs feature extraction on each second subgraph to obtain a fifth feature corresponding to each second subgraph. The fifth feature corresponding to each second subgraph is used to describe the feature of the first object node under the influence of at least two types of association relationships included in the second subgraph.

[0289] 4054. The electronic device performs weighted processing on the fourth features corresponding to the multiple first subgraphs and the fifth features corresponding to the multiple second subgraphs based on the second weights corresponding to the multiple first subgraphs and the second weights corresponding to the multiple second subgraphs, to obtain the first feature of the first object node.

[0290] Among them, the second weight corresponding to each first subgraph represents the degree of influence of different object nodes on the object features of the first object node under the same association relationship, and the second weight corresponding to each second subgraph represents the degree of influence of different object nodes on the object features of the first object node under different association relationships.

[0291] The above Figure 5 In the process shown, the fourth or fifth feature is extracted, realizing feature extraction at the meta-path level in the association subgraph. Based on the second weights corresponding to multiple first subgraphs and multiple second subgraphs, the multiple fourth and fifth features are weighted to obtain the features of the first object at the association level. The meta-path in the association subgraph indicates the subgraphs further subdivided from the association subgraph.

[0292] The above Figure 5 Taking only a single target association subgraph as an example, the above method can be applied to all target association subgraphs. Figure 5 The method shown determines the first feature of the first object node corresponding to each target association subgraph.

[0293] 406. The electronic device performs weighted processing on the first features corresponding to the multiple target association subgraphs based on the first weights corresponding to the multiple target association subgraphs, and obtains the second feature corresponding to the first object node.

[0294] Among them, the first weight corresponding to each target association subgraph represents the degree of influence of the association corresponding to the target association subgraph on the object features of the first object node.

[0295] 407. The electronic device extracts features from the heterogeneous graph to obtain the third feature corresponding to the first object node.

[0296] 408. The electronic device concatenates the second feature and the third feature to obtain the object feature corresponding to the first object node.

[0297] In some embodiments, the electronic device may not execute steps 407-408 above, but instead directly use the second feature corresponding to the first object node as the object feature corresponding to the first object node, that is, as the object feature of the first object.

[0298] In related technologies, heterogeneous graphs are split into different association subgraphs based on association relationships, and features are extracted only from the association subgraphs. However, this extraction method does not consider the mutual influence between different association relationships, resulting in low accuracy of the extracted features.

[0299] The method provided in this disclosure, during the construction of a target node network, determines the next target object node associated with the current target object node based on the association relationship of the connection lines between different object nodes. This ensures that the connection line between the next target object node and the current target object node matches the current target association relationship. By considering the impact of different association relationships on target object nodes, the method achieves the determination of multiple target object nodes across relationships. Thus, when extracting features using the target node network, the method can fully consider the impact of the association relationship of the connection lines between various object nodes on the first object, thereby improving the accuracy of object features.

[0300] Furthermore, the heterogeneous graph is divided into multiple relational subgraphs. Object features are extracted based on the target node set and the target relational subgraph. While fully considering the influence of the association relationship between the connecting lines of each object node on the first object node, the influence of other object nodes connected by the same type of connecting line on the first object node is also considered, which further improves the accuracy of object features.

[0301] Furthermore, in this embodiment of the disclosure, by constructing a target node network, semantic information in the heterogeneous graph is captured, and the heterogeneous graph is divided into different relational subgraphs. Differential expressions under different relational relationships are learned, that is, features corresponding to different relational relationships can be extracted, and global information in the heterogeneous graph can also be captured. When determining object features, the semantic information, global information and differential expressions under different relational relationships in the heterogeneous graph are comprehensively considered, which improves the accuracy of the extracted features.

[0302] In some embodiments, an electronic device invokes a feature extraction model to extract features of an object. This feature extraction model is trained and stored by the electronic device, or it is sent to the electronic device by another electronic device. The model structure of the feature extraction model is described below. Figure 6 The feature extraction model includes multiple first feature extraction networks 601 ( Figure 6(Taking two examples) the second feature extraction network 602 and the third feature extraction network 603.

[0303] In some embodiments, see Figure 7 Each of the first feature extraction networks 601 includes a feature extraction layer 611, an attention layer 621, and a splicing layer 631, and the second feature extraction network 602 includes an attention layer 612 and a splicing layer 622.

[0304] The process by which the electronic device invokes the feature extraction model to extract the object features of the first object is detailed below. Figure 6 The example shown:

[0305] Figure 8 This is a flowchart illustrating another feature extraction method according to an exemplary embodiment, see [link to flowchart]. Figure 8 The method is executed by an electronic device and includes the following steps:

[0306] 801. Electronic devices obtain associated information.

[0307] 802. The electronic device identifies the first object as the initial target object.

[0308] 803. The electronic device selects the next target object that meets the association conditions from the associated information based on the current target object, until the number of identified target objects reaches the target number, and constructs a target object set.

[0309] 804. The electronic device divides the associated information into multiple associated sub-information according to the various types of association relationships contained in the associated information, and determines multiple target associated sub-information containing the first object from the multiple associated sub-information.

[0310] The implementation methods of 801-804 in this disclosure are the same as those described above. Figure 2 The implementation methods of 201-204 in the illustrated embodiments are similar and will not be repeated here.

[0311] 805. The electronic device calls each first feature extraction network to extract features from the target object set and each target associated sub-information to obtain the first feature of the first object.

[0312] The first feature extraction network includes a feature extraction layer, an attention layer, and a splicing layer.

[0313] In some embodiments, for each target-related sub-information, the feature extraction layer in the first feature extraction network corresponding to the target-related sub-information is invoked to divide the target-related sub-information into multiple first sub-information; the feature extraction layer is invoked to divide the target object set into multiple second sub-information; the feature extraction layer is invoked to extract features from the multiple first sub-information and the multiple second sub-information respectively to obtain a fourth feature corresponding to the multiple first sub-information and a fifth feature corresponding to the multiple second sub-information; the attention layer in the first feature extraction network is invoked to process the multiple first sub-information and the multiple second sub-information respectively to obtain a second weight corresponding to the multiple first sub-information and a second weight corresponding to the multiple second sub-information; the concatenation layer is invoked to perform weighted processing on the fourth feature corresponding to the multiple first sub-information and the fifth feature corresponding to the multiple second sub-information based on the second weight corresponding to the multiple first sub-information and the second weight corresponding to the multiple second sub-information to obtain a first feature of the first object.

[0314] In some embodiments, a feature extraction layer is invoked to determine multiple reference types that are different from the target type; the feature extraction layer is invoked again, and one of the reference types is selected each time. A first object and other objects that meet the first selection condition are selected from the target association sub-information. The first object and the selected other objects are combined to form a first sub-information. This process is repeated until multiple reference types are selected, resulting in multiple first sub-information. The selected object meets the first selection condition if it belongs to a reference type and forms an association relationship with the first object corresponding to the target association sub-information.

[0315] In some embodiments, a feature extraction layer is invoked to determine the various types of associations included in the target object set; the feature extraction layer is invoked again to select at least two types of associations from the various types each time, and a first object and other objects that meet the second selection condition are selected from the target object set to form a second sub-information, until there are no more associations of at least two types that have not been selected at the same time, and multiple second sub-information are obtained, wherein the selected object meets the second selection condition if the association between the selected object and the first object is at least two types of association.

[0316] 806. The electronic device calls the second feature extraction network, and performs weighted processing on the first features corresponding to the multiple target association sub-information based on the first weights corresponding to the multiple target association sub-information, to obtain the second feature of the first object.

[0317] The second feature extraction network includes an attention layer and a splicing layer.

[0318] In some embodiments, the electronic device invokes the attention layer in the second feature extraction network to process multiple target association sub-information to obtain the first weights corresponding to the multiple target association sub-information; it then invokes the splicing layer to perform weighted processing on the first features corresponding to the multiple target association sub-information based on the first weights, to obtain the second feature of the first object.

[0319] 807. The electronic device calls the third feature extraction network to extract features from the associated information and obtain the third feature of the first object.

[0320] 808. The electronic device calls the second feature extraction network to concatenate the second feature with the third feature to obtain the object feature of the first object.

[0321] In some embodiments, the electronic device invokes the splicing layer in the second feature extraction network to splice the second feature with the third feature to obtain the object feature of the first object.

[0322] In some embodiments, step 807 is executed first, followed by step 806, whereby the electronic device calls the second feature extraction network to perform weighted processing on the first features corresponding to the multiple target association sub-information based on the first weights corresponding to the multiple target association sub-information, to obtain the second feature of the first object, and then concatenates the second feature with the third feature to obtain the object feature of the first object.

[0323] In addition, in some embodiments, when the association information is a heterogeneous graph, the target node network and multiple target association subgraphs are determined based on the heterogeneous graph. The feature extraction model is then invoked to process the target node network and multiple target association subgraphs to obtain the object features corresponding to the first object node.

[0324] For example, see Figure 9Accounts are categorized into two types: user accounts and author accounts. The heterogeneous graph includes user nodes corresponding to user accounts, author nodes corresponding to author accounts, and video nodes corresponding to videos. Based on this heterogeneous graph, a target node network, a comment relationship subgraph, and a follow relationship subgraph are determined. The target node network and the comment relationship subgraph are input into a first feature extraction network. Through the feature extraction layer in the first feature extraction network, features are extracted from the target node network and the comment relationship subgraph, yielding the fourth feature corresponding to the two first subgraphs obtained from the comment relationship subgraph and the fifth feature corresponding to the one second subgraph obtained from the target node network. Through the attention layer in the first feature extraction network, the second weights corresponding to the two first subgraphs and the second subgraph are obtained. Finally, through the concatenation layer in the first feature extraction layer, the first feature corresponding to the comment relationship subgraph is obtained. Similarly, for the second target relationship network, the target node network and the second target relationship network are input into a second first feature extraction network. Through the second first feature extraction network, the first feature corresponding to the follow relationship subgraph is obtained.

[0325] Then, the first features corresponding to the obtained comment relationship subgraph and the first features corresponding to the attention relationship subgraph are input into the second feature extraction network. After passing through the attention layer in the second feature extraction network, the first weights corresponding to the two target relationship subgraphs are obtained. After passing through the concatenation layer in the second feature extraction network, the second feature corresponding to the target user node is obtained. Simultaneously, the heterogeneous graph is input into the third feature extraction network, which outputs the third feature corresponding to the target user node. The third feature is then input into the second feature extraction network, and after passing through the concatenation layer in the second feature extraction network, the second and third features are concatenated to obtain the user feature corresponding to the target user node.

[0326] The method provided in this disclosure, during the construction of a target object set, determines the next target object associated with the current target object based on the association relationships between different objects, so that the association relationship between the next target object and the current target object becomes the current target association relationship. It considers the impact of different association relationships on the target object, and determines multiple target objects through cross-relationships. Therefore, when extracting features using the target object set, it can fully consider the impact of the association relationships between each object on the first object, thereby improving the accuracy of object features.

[0327] In this embodiment of the disclosure, before calling the feature extraction model to extract object features, the feature extraction model needs to be trained first. The training process of the feature extraction model is as follows: The electronic device acquires two sample objects from the sample association information, and the similarity between the two sample objects is greater than the reference similarity; the feature extraction model is called to obtain the predicted features corresponding to the two sample objects respectively; the predicted features corresponding to the two sample objects are processed by the objective function to obtain the loss value; based on the loss value, the feature extraction model is trained, that is, the parameters in the feature extraction model are adjusted based on the loss value, and the adjusted feature extraction model is used to continue training based on the sample objects until the obtained loss value is less than the reference value, and the training of the feature extraction model ends.

[0328] In some embodiments, when the sample object is a sample object node in a sample heterogeneous graph, the electronic device pre-sets the pattern S of the metapath with the association relationship r: φ(v1)→φ(v2)→φ(v3)…φ(v l ), where l is the length of the meta-path, meaning that l object nodes need to be selected from the sample heterogeneous graph based on this meta-path. The electronic device uses the following formula to determine two similar sample object nodes based on this meta-path:

[0329]

[0330] in, This indicates that two identical sample object nodes are defined as follows: and This indicates that when the association relationship is r, the sample object node v i The set of neighboring nodes, O(V) t+1 ) indicates and A collection of object nodes that belong to the same node type. express The node type it belongs to, E indicates that the sample object node is and The sample association relationships that are satisfied between them. That is, the sample object nodes. and The samples need to satisfy the correlation relationship, and It needs to belong to the sample node type.

[0331] In some embodiments, the Skip-Gram model is invoked to determine the relationships between other object nodes associated with any object node. This involves sampling the path excluding object node v. i Other object nodes besides v i In the context of object node v i The context C is represented as C = {v j |v i∈S, |ji|≤δ, j≠i}, where δ represents the threshold of the moving window size radius.

[0332] Furthermore, the loss value is determined using the following objective function:

[0333]

[0334] Where θ represents the parameters in the model, v j and v i Let C represent two similar sample object nodes, and let C represent object node v in the sampling path. i A collection of other object nodes besides the one mentioned above.

[0335]

[0336] Where H represents the loss value, c j It is the object node v j The corresponding object characteristics, c k It is the object node v k The corresponding object feature, σ(·), is the Sigmoid function (activation function), expressed as: It is the probability of negative sampling, v k Representative from P neg The object nodes obtained by sampling Represents object node v i In the relationship R l The following are the object features. A smaller H indicates a more accurate feature extraction model.

[0337] It should be noted that other methods can also be used to train the feature extraction model. For example, the sample features of the sample object can be obtained, and the parameters of the feature extraction model can be adjusted according to the difference between the obtained predicted features and the sample features, so that the difference is continuously reduced, thereby training the feature extraction model.

[0338] Figure 10 This is a block diagram illustrating a feature extraction apparatus according to an exemplary embodiment. See also... Figure 10 The device includes:

[0339] The association information acquisition unit 1001 is configured to acquire association information, which includes multiple objects and the association relationship between any two objects. The multiple objects include accounts registered on the server and multimedia resources published through the server.

[0340] The initial object determination unit 1002 is configured to determine a first object as the initial target object, wherein the first object is any account registered on the server or any multimedia resource published through the server.

[0341] The object set construction unit 1003 is configured to select the next target object that meets the association condition from the association information based on the current target object, until the number of determined target objects reaches the target number, and construct the target object set. The association condition refers to the association relationship between the next target object and the current target object being the currently determined target association relationship. The target object set includes the target number of target objects and the association relationship between any two target objects.

[0342] The feature extraction unit 1004 is configured to perform feature extraction on the target object set to obtain the object features of the first object.

[0343] The apparatus provided in this disclosure, during the process of constructing a set of target objects, determines the next target object associated with the current target object based on the association relationship between different objects, so that the association relationship between the next target object and the current target object becomes the current target association relationship. It considers the impact of different association relationships on the target objects and determines multiple target objects through cross-relationships. Therefore, when extracting features using the set of target objects, it can fully consider the impact of the association relationship between each object on the first object and improve the accuracy of object features.

[0344] In some embodiments, see Figure 11 The object collection construction unit 1003 includes:

[0345] The association determination subunit 1013 is configured to perform the determination of multiple types of associations of the current target object based on association information;

[0346] The association selection subunit 1023 is configured to select one type of association from multiple types and determine the selected association as the current target association.

[0347] The object determination subunit 1033 is configured to select an object from the association information that has an association relationship with the current target object as the current target association relationship, and determine the selected object as the next target object.

[0348] In some embodiments, see Figure 11 The object is identified as subunit 1033, which is configured to execute:

[0349] From the associated information, select multiple candidate objects whose associations with the current target object are used as the current target associations;

[0350] Choose any one of the multiple candidate objects as the next target object.

[0351] In some embodiments, see Figure 11The feature extraction device also includes:

[0352] The association information division unit 1005 is configured to divide the association information into multiple association sub-information according to the multiple types of association relationships contained in the association information. Each association sub-information contains multiple objects, and different objects in the same association sub-information have the same type of association relationship.

[0353] Feature extraction unit 1004 is configured to perform:

[0354] From multiple related sub-information, determine the target related sub-information that contains the first object;

[0355] Feature extraction is performed on the target object set and its associated sub-information to obtain object features.

[0356] In some embodiments, there are multiple target associated sub-information items, see [link to relevant documentation]. Figure 11 Feature extraction unit 1004 includes:

[0357] The first feature extraction subunit 1014 is configured to perform feature extraction on the target object set and the target associated sub-information for each target associated sub-information, and obtain the first feature of the first object;

[0358] The object feature extraction subunit 1024 is configured to perform weighted processing on the first features corresponding to multiple target association sub-information based on the first weights corresponding to multiple target association sub-information to obtain object features. The first weights corresponding to each target association sub-information characterize the degree of influence of the association relationship corresponding to the target association sub-information on the object features of the first object.

[0359] In some embodiments, there are multiple target associated sub-information items, see [link to relevant documentation]. Figure 11 Feature extraction unit 1004 includes:

[0360] The first feature extraction subunit 1014 is configured to perform feature extraction on the target object set and the target associated sub-information for each target associated sub-information, and obtain the first feature of the first object;

[0361] The second feature determination subunit 1034 is configured to perform weighted processing on the first features corresponding to multiple target association sub-informations based on the first weights corresponding to multiple target association sub-informations to obtain the second feature of the first object. The first weights corresponding to each target association sub-information characterize the degree of influence of the association relationship corresponding to the target association sub-information on the object feature of the first object.

[0362] The third feature extraction subunit 1044 is configured to perform feature extraction on the associated information to obtain the third feature of the first object;

[0363] The object feature extraction subunit 1024 is configured to concatenate the second feature with the third feature to obtain the object feature.

[0364] In some embodiments, the type of the first object is the target type, see [link to documentation]. Figure 11 The first feature extraction subunit 1014 is configured to perform:

[0365] For each target-related sub-information, the target-related sub-information is divided into multiple first sub-information. Each first sub-information contains a first object and other objects belonging to the reference type. The other objects and the first object constitute the association relationship corresponding to the target-related sub-information. The reference types contained in different first sub-information are different, and the reference types are different from the target types.

[0366] The target object set is divided into multiple second sub-information. Each second sub-information contains a first object and other objects. The association between the first object and other objects includes at least two types. The association between the first object and other objects in different second sub-information is not completely the same.

[0367] Feature extraction is performed on multiple first sub-informations and multiple second sub-informations respectively to obtain the fourth feature corresponding to the multiple first sub-informations and the fifth feature corresponding to the multiple second sub-informations;

[0368] Based on the second weights corresponding to multiple first sub-information and the second weights corresponding to multiple second sub-information, the fourth feature corresponding to multiple first sub-information and the fifth feature corresponding to multiple second sub-information are weighted to obtain the first feature of the first object. The second weights corresponding to each first sub-information represent the degree of influence of different objects under the same association on the object feature of the first object, and the second weights corresponding to each second sub-information represent the degree of influence of different objects under different associations on the object feature of the first object.

[0369] In some embodiments, see Figure 11 The first feature extraction subunit 1014 is configured to perform:

[0370] For each target-related sub-information, determine multiple reference types that are different from the target type;

[0371] Each time, one of the reference types is selected, the first object and other objects that meet the first selection condition are selected from the target associated sub-information to form a first sub-information. This process continues until multiple reference types are selected, resulting in multiple first sub-information. The selected object meets the first selection condition if it belongs to the reference type and forms an association relationship with the first object corresponding to the target associated sub-information.

[0372] In some embodiments, see Figure 11 The first feature extraction subunit 1014 is configured to perform:

[0373] Identify the various types of relationships included in the target object set;

[0374] Each time, at least two types of association relationships are selected from multiple types. A first object and other objects that meet the second selection condition are selected from the target object set to form a second sub-information. This process continues until there are no more association relationships of at least two types that have not been selected at the same time, resulting in multiple second sub-information. Here, the selected object meets the second selection condition if the association relationship between the selected object and the first object is at least two types of association relationship.

[0375] In some embodiments, the feature extraction unit 1004 is configured to execute a feature extraction model to extract features from a set of target objects to obtain object features of a first object.

[0376] In some embodiments, the feature extraction unit 1004 is configured to execute a feature extraction model to extract features from the target object set and target associated sub-information to obtain object features.

[0377] In some embodiments, the feature extraction model includes multiple first feature extraction networks and second feature extraction networks, and the target associated sub-information is multiple, see [link to relevant documentation]. Figure 11 Feature extraction unit 1004 includes:

[0378] The first feature extraction subunit 1014 is configured to call each first feature extraction network to extract features from the target object set and each target associated sub-information to obtain the first feature of the first object;

[0379] The object feature extraction subunit 1024 is configured to execute the call to the second feature extraction network, and perform weighted processing on the first features corresponding to the multiple target association sub-information based on the first weights corresponding to the multiple target association sub-information to obtain object features. The first weights corresponding to each target association sub-information represent the degree of influence of the association relationship corresponding to the target association sub-information on the object features of the first object.

[0380] In some embodiments, the feature extraction model includes multiple first feature extraction networks, second feature extraction networks, and third feature extraction networks, and the target associated sub-information is multiple, see [link to documentation]. Figure 11 Feature extraction unit 1004 includes:

[0381] The first feature extraction subunit 1014 is configured to call each first feature extraction network to extract features from the target object set and each target associated sub-information to obtain the first feature of the first object;

[0382] The second feature determination subunit 1034 is configured to execute the second feature extraction network, and based on the first weights corresponding to the multiple target association sub-informations, perform weighted processing on the first features corresponding to the multiple target association sub-informations to obtain the second feature of the first object. The first weights corresponding to each target association sub-information characterize the degree of influence of the association relationship corresponding to the target association sub-information on the object feature of the first object.

[0383] The third feature extraction subunit 1044 is configured to call the third feature extraction network to extract features from the associated information and obtain the third feature of the first object.

[0384] The object feature extraction subunit 1024 is configured to call the second feature extraction network to concatenate the second feature with the third feature to obtain the object feature.

[0385] In some embodiments, see Figure 11 The feature extraction unit 1004 also includes:

[0386] The first weight determination subunit 1054 is configured to execute the call to the second feature extraction network to process multiple target association sub-information and obtain the first weights corresponding to the multiple target association sub-information.

[0387] In some embodiments, see Figure 11 The first feature extraction subunit 1014 is configured to perform:

[0388] For each target-related sub-information, the feature extraction layer in the first feature extraction network corresponding to the target-related sub-information is called to divide the target-related sub-information into multiple first sub-information. Each first sub-information contains a first object and other objects belonging to the reference type. The other objects and the first object constitute the association relationship corresponding to the target-related sub-information. The reference types contained in different first sub-information are different.

[0389] The feature extraction layer is invoked to divide the target object set into multiple second sub-information. Each second sub-information contains the first object and other objects, and the association between the first object and other objects includes at least two types. The association between the first object and other objects in different second sub-information is not completely the same.

[0390] The feature extraction layer is invoked to extract features from multiple first sub-informations and multiple second sub-informations respectively, resulting in the fourth feature corresponding to the multiple first sub-informations and the fifth feature corresponding to the multiple second sub-informations.

[0391] The concatenation layer in the first feature extraction network is invoked. Based on the second weights corresponding to multiple first sub-information and the second weights corresponding to multiple second sub-information, the fourth feature corresponding to multiple first sub-information and the fifth feature corresponding to multiple second sub-information are weighted to obtain the first feature of the first object. The second weights corresponding to each first sub-information represent the degree of influence of different objects under the same association on the object feature of the first object. The second weights corresponding to each second sub-information represent the degree of influence of different objects under different associations on the object feature of the first object.

[0392] In some embodiments, see Figure 11 The feature extraction unit 1004 also includes:

[0393] The second weight determination subunit 1064 is configured to call the attention layer in the first feature extraction network to process multiple first sub-information and multiple second sub-information respectively, and obtain the second weights corresponding to the multiple first sub-information and the second weights corresponding to the multiple second sub-information.

[0394] In some embodiments, see Figure 11 The first feature extraction subunit 1014 is configured to perform:

[0395] The feature extraction layer is invoked to identify multiple reference types that are different from the target type;

[0396] The feature extraction layer is invoked, and one reference type is selected each time. The first object and other objects that meet the first selection condition are selected from the target association sub-information to form a first sub-information. This process continues until multiple reference types are selected, resulting in multiple first sub-information. The selected object meets the first selection condition if it belongs to the reference type and forms an association relationship with the first object corresponding to the target association sub-information.

[0397] In some embodiments, see Figure 11 The first feature extraction subunit 1014 is configured to perform:

[0398] The feature extraction layer is invoked to determine the various types of relationships included in the target object set;

[0399] The feature extraction layer is invoked, and at least two types of association relationships are selected from multiple types each time. The first object and other objects that meet the second selection condition are selected from the target object set to form a second sub-information. This process continues until there are no more association relationships of at least two types that have not been selected at the same time, resulting in multiple second sub-information. Here, the selected object meets the second selection condition if the association relationship between the selected object and the first object is at least two types of association relationship.

[0400] In some embodiments, see Figure 11 The feature extraction device also includes:

[0401] Model training unit 1006 is configured to perform the task of obtaining information about two sample objects in the sample association, where the similarity between the two sample objects is greater than the reference similarity.

[0402] Model training unit 1006 is also configured to execute the feature extraction model to obtain the predicted features corresponding to the two sample objects respectively;

[0403] The model training unit 1006 is also configured to process the predicted features corresponding to the two sample objects using an objective function to obtain the loss value;

[0404] Model training unit 1006 is also configured to perform loss-based training of the feature extraction model.

[0405] In some embodiments, the model training unit 1006 is further configured to terminate the training of the feature extraction model in response to a loss value being less than a reference value.

[0406] In some embodiments, the association information is a heterogeneous graph, which includes multiple object nodes and connecting lines between any two object nodes. An object node represents an object, and a connecting line represents the association relationship between the objects corresponding to the two object nodes connected by the connecting line.

[0407] The object set construction unit 1003 is configured to perform the following operations: from the heterogeneous graph, select the next target object node that satisfies the association condition based on the current target object node, until the number of determined target object nodes reaches the target number, and construct a target node network. The association condition refers to the connection line between the next target object node and the current target object node matching the currently determined target association relationship. The target node network includes the target number of target object nodes and the connection line between any two target object nodes.

[0408] The feature extraction unit 1004 is configured to perform feature extraction on the target node network to obtain object features.

[0409] In some embodiments, see Figure 11 The feature extraction device also includes:

[0410] The association information partitioning unit 1005 is configured to divide the heterogeneous graph into multiple association subgraphs according to the various connection lines contained in the heterogeneous graph. Each association subgraph contains multiple object nodes, and the connection lines between different object nodes in the same association subgraph match the same type of association relationship.

[0411] Feature extraction unit 1004 is configured to perform:

[0412] From multiple association subgraphs, determine the target association subgraph that contains the first object node corresponding to the first object;

[0413] Feature extraction is performed on the target node network and the target association subgraph to obtain object features.

[0414] Regarding the apparatus in the above embodiments, the specific manner in which each unit performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0415] In an exemplary embodiment, an electronic device is provided, the electronic device including one or more processors and a memory for storing processor-executable instructions; wherein the one or more processors are configured to perform the feature extraction method described in the above embodiments.

[0416] In some embodiments, the electronic device is provided as a terminal. Figure 12 This is a structural block diagram of a terminal 1200 according to an exemplary embodiment. The terminal 1200 can be a portable mobile terminal, such as a smartphone, tablet computer, MP3 player (Moving Picture Experts Group Audio Layer III), MP4 player (Moving Picture Experts Group Audio Layer IV), laptop computer, or desktop computer. The terminal 1200 may also be referred to as a user device, portable terminal, laptop terminal, desktop terminal, or other names.

[0417] Terminal 1200 includes a processor 1201 and a memory 1202.

[0418] Processor 1201 may include one or more processing cores, such as a quad-core processor, an octa-core processor, etc. Processor 1201 may be implemented using at least one hardware form selected from DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). Processor 1201 may also include a main processor and a coprocessor. The main processor, also known as a CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, processor 1201 may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the screen. In some embodiments, processor 1201 may also include an AI (Artificial Intelligence) processor, which is used to handle computational operations related to machine learning.

[0419] The memory 1202 may include one or more computer-readable storage media, which may be non-transitory. The memory 1202 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In some embodiments, the non-transitory computer-readable storage media in the memory 1202 are used to store at least one line of program code, which is executed by the processor 1201 to implement the feature extraction method provided in the method embodiments of this disclosure.

[0420] In some embodiments, the terminal 1200 may also optionally include a peripheral device interface 1203 and at least one peripheral device. The processor 1201, memory 1202, and peripheral device interface 1203 can be connected via a bus or signal line. Each peripheral device can be connected to the peripheral device interface 1203 via a bus, signal line, or circuit board. Specifically, the peripheral device includes at least one of the following: a radio frequency circuit 1204, a display screen 1205, a camera assembly 1206, an audio circuit 1207, a positioning assembly 1208, and a power supply 1209.

[0421] Peripheral device interface 1203 can be used to connect at least one I / O (Input / Output) related peripheral device to processor 1201 and memory 1202. In some embodiments, processor 1201, memory 1202 and peripheral device interface 1203 are integrated on the same chip or circuit board; in some other embodiments, any one or two of processor 1201, memory 1202 and peripheral device interface 1203 can be implemented on separate chips or circuit boards, which is not limited in this embodiment.

[0422] The radio frequency (RF) circuit 1204 is used to receive and transmit RF (Radio Frequency) signals, also known as electromagnetic signals. The RF circuit 1204 communicates with communication networks and other communication devices via electromagnetic signals. The RF circuit 1204 converts electrical signals into electromagnetic signals for transmission, or converts received electromagnetic signals back into electrical signals. Optionally, the RF circuit 1204 includes: an antenna system, an RF transceiver, one or more amplifiers, a tuner, an oscillator, a digital signal processor, a codec chipset, a user identity module card, etc. The RF circuit 1204 can communicate with other terminals via at least one wireless communication protocol. This wireless communication protocol includes, but is not limited to: the World Wide Web, metropolitan area networks, intranets, various generations of mobile communication networks (2G, 3G, 4G, and 5G), wireless local area networks, and / or WiFi (Wireless Fidelity) networks. In some embodiments, the RF circuit 1204 may also include circuitry related to NFC (Near Field Communication), which is not limited in this disclosure.

[0423] Display screen 1205 is used to display a UI (User Interface). This UI may include graphics, text, icons, videos, and any combination thereof. When display screen 1205 is a touch display screen, it also has the ability to collect touch signals on or above its surface. These touch signals can be input as control signals to processor 1201 for processing. In this case, display screen 1205 can also be used to provide virtual buttons and / or a virtual keyboard, also known as soft buttons and / or a soft keyboard. In some embodiments, there may be one display screen 1205, disposed on the front panel of terminal 1200; in other embodiments, there may be at least two display screens, disposed on different surfaces of terminal 1200 or in a folded design; in still other embodiments, display screen 1205 may be a flexible display screen, disposed on a curved or folded surface of terminal 1200. Furthermore, display screen 1205 may also be configured as a non-rectangular, irregular shape, i.e., a non-rectangular screen. The display screen 1205 can be made of materials such as LCD (Liquid Crystal Display) and OLED (Organic Light-Emitting Diode).

[0424] The camera assembly 1206 is used to acquire images or videos. Optionally, the camera assembly 1206 includes a front-facing camera and a rear-facing camera. The front-facing camera is disposed on the front panel of the terminal, and the rear-facing camera is disposed on the back of the terminal. In some embodiments, there are at least two rear-facing cameras, which are any one of a main camera, a depth-sensing camera, a wide-angle camera, and a telephoto camera, to achieve background blurring by fusion of the main camera and the depth-sensing camera, panoramic shooting by fusion of the main camera and the wide-angle camera, VR (Virtual Reality) shooting, or other fusion shooting functions. In some embodiments, the camera assembly 1206 may also include a flash. The flash can be a single-color temperature flash or a dual-color temperature flash. A dual-color temperature flash refers to a combination of a warm light flash and a cool light flash, which can be used for light compensation at different color temperatures.

[0425] The audio circuit 1207 may include a microphone and a speaker. The microphone is used to collect sound waves from the user and the environment, converting the sound waves into electrical signals that are input to the processor 1201 for processing, or input to the radio frequency circuit 1204 for voice communication. For stereo sound acquisition or noise reduction purposes, multiple microphones may be used, each positioned at a different location on the terminal 1200. The microphone may also be an array microphone or an omnidirectional microphone. The speaker is used to convert electrical signals from the processor 1201 or the radio frequency circuit 1204 into sound waves. The speaker may be a conventional diaphragm speaker or a piezoelectric ceramic speaker. When the speaker is a piezoelectric ceramic speaker, it can convert electrical signals not only into audible sound waves but also into inaudible sound waves for purposes such as distance measurement. In some embodiments, the audio circuit 1207 may also include a headphone jack.

[0426] The positioning component 1208 is used to locate the current geographical location of the terminal 1200 in order to enable navigation or LBS (Location Based Service). The positioning component 1208 can be a positioning component based on the US GPS (Global Positioning System), China's BeiDou system, Russia's Granas positioning system, or the European Union's Galileo positioning system.

[0427] Power supply 1209 is used to power the various components in terminal 1200. Power supply 1209 can be AC ​​power, DC power, a disposable battery, or a rechargeable battery. When power supply 1209 includes a rechargeable battery, the rechargeable battery can be a wired rechargeable battery or a wireless rechargeable battery. A wired rechargeable battery is a battery that is charged via a wired line, and a wireless rechargeable battery is a battery that is charged via a wireless coil. The rechargeable battery can also be used to support fast charging technology.

[0428] In some embodiments, the terminal 1200 further includes one or more sensors 1210. The one or more sensors 1210 include, but are not limited to: an accelerometer 1211, a gyroscope 1212, a pressure sensor 1213, a fingerprint sensor 1214, an optical sensor 1215, and a proximity sensor 1216.

[0429] Accelerometer 1211 can detect the magnitude of acceleration along the three coordinate axes of a coordinate system established with terminal 1200. For example, accelerometer 1211 can be used to detect the components of gravitational acceleration along the three coordinate axes. Processor 1201 can control display screen 1205 to display the user interface in either a landscape or portrait view based on the gravitational acceleration signal acquired by accelerometer 1211. Accelerometer 1211 can also be used for games or for acquiring user motion data.

[0430] The gyroscope sensor 1212 can detect the orientation and rotation angle of the terminal 1200. The gyroscope sensor 1212 can work in conjunction with the accelerometer sensor 1211 to collect the user's 3D movements on the terminal 1200. Based on the data collected by the gyroscope sensor 1212, the processor 1201 can perform the following functions: motion sensing (e.g., changing the UI based on the user's tilt), image stabilization during shooting, game control, and inertial navigation.

[0431] The pressure sensor 1213 can be disposed on the side bezel of the terminal 1200 and / or on the lower layer of the display screen 1205. When the pressure sensor 1213 is disposed on the side bezel of the terminal 1200, it can detect the user's grip signal on the terminal 1200, and the processor 1201 can perform left / right hand recognition or quick operation based on the grip signal collected by the pressure sensor 1213. When the pressure sensor 1213 is disposed on the lower layer of the display screen 1205, the processor 1201 can control the operable controls on the UI interface based on the user's pressure operation on the display screen 1205. The operable controls include at least one of button controls, scroll bar controls, icon controls, and menu controls.

[0432] The fingerprint sensor 1214 is used to collect a user's fingerprint. The processor 1201 identifies the user based on the fingerprint collected by the fingerprint sensor 1214, or vice versa. When the user's identity is identified as trusted, the processor 1201 authorizes the user to perform relevant sensitive operations, including unlocking the screen, viewing encrypted information, downloading software, making payments, and changing settings. The fingerprint sensor 1214 can be located on the front, back, or side of the terminal 1200. When the terminal 1200 has physical buttons or a manufacturer's logo, the fingerprint sensor 1214 can be integrated with the physical buttons or manufacturer's logo.

[0433] The optical sensor 1215 is used to collect ambient light intensity. In one embodiment, the processor 1201 can control the display brightness of the display screen 1205 based on the ambient light intensity collected by the optical sensor 1215. Specifically, when the ambient light intensity is high, the display brightness of the display screen 1205 is increased; when the ambient light intensity is low, the display brightness of the display screen 1205 is decreased. In another embodiment, the processor 1201 can also dynamically adjust the shooting parameters of the camera assembly 1206 based on the ambient light intensity collected by the optical sensor 1215.

[0434] The proximity sensor 1216, also known as a distance sensor, is installed on the front panel of the terminal 1200. The proximity sensor 1216 is used to detect the distance between the user and the front of the terminal 1200. In one embodiment, when the proximity sensor 1216 detects that the distance between the user and the front of the terminal 1200 is gradually decreasing, the processor 1201 controls the display screen 1205 to switch from a screen-on state to a screen-off state; when the proximity sensor 1216 detects that the distance between the user and the front of the terminal 1200 is gradually increasing, the processor 1201 controls the display screen 1205 to switch from a screen-off state to a screen-on state.

[0435] Those skilled in the art will understand that Figure 12 The structure shown does not constitute a limitation on terminal 1200 and may include more or fewer components than shown, or combine certain components, or use different component arrangements.

[0436] In some embodiments, the electronic device is provided as a server. Figure 13 This is a structural block diagram of a server according to an exemplary embodiment. The server 1300 can vary significantly due to different configurations or performance. It may include one or more Central Processing Units (CPUs) 1301 and one or more memories 1302. The memories 1302 store at least one line of program code, which is loaded and executed by the processor 1301 to implement the methods provided in the various method embodiments described above. Of course, the server may also have wired or wireless network interfaces, a keyboard, and input / output interfaces for input / output. The server may also include other components for implementing device functions, which will not be elaborated upon here.

[0437] In an exemplary embodiment, a computer-readable storage medium is also provided, which, when executed by a processor of an electronic device, enables the electronic device to perform the steps of the feature extraction method described above. Optionally, the computer-readable storage medium may be a ROM (Read Only Memory), RAM (Random Access Memory), CD-ROM (Compact Disc Read-Only Memory), magnetic tape, floppy disk, and optical data storage device, etc.

[0438] In an exemplary embodiment, a computer program product is also provided, which includes a computer program that is executed by a processor to implement the feature extraction method described above.

[0439] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This disclosure is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the following claims.

[0440] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.

Claims

1. A feature extraction method characterized by, The feature extraction method comprises: obtaining association information, the association information comprising a plurality of objects and an association relationship between any two objects, the plurality of objects comprising an account registered on a server and a multimedia resource published through the server; determining a first object as an initial target object, the first object being any account registered on the server or any multimedia resource published through the server; selecting a next target object satisfying an association condition from the association information based on a current target object until the number of determined target objects reaches a target number, constructing a target object set, the association condition being that the association relationship between the next target object and the current target object is a current determined target association relationship, the target object set comprising the target number of target objects and an association relationship between any two target objects; performing feature extraction on the target object set to obtain an object feature of the first object.

2. The feature extraction method of claim 1, wherein, The selecting a next target object satisfying an association condition from the association information based on a current target object comprises: determining a plurality of types of association relationships of the current target object based on the association information; selecting one type of association relationship from the plurality of types, and determining the selected association relationship as the current target association relationship; selecting an object having an association relationship with the current target object as the current target association relationship from the association information, and determining the selected object as the next target object.

3. The feature extraction method of claim 2, wherein, The selecting an object having an association relationship with the current target object as the current target association relationship from the association information, and determining the selected object as the next target object comprises: selecting a plurality of candidate objects having an association relationship with the current target object as the current target association relationship from the association information; determining any object in the plurality of candidate objects as the next target object.

4. The feature extraction method of claim 1, wherein, The feature extraction method further comprises: dividing the association information into a plurality of association sub-information according to a plurality of types of association relationships contained in the association information, each association sub-information containing a plurality of objects, and different objects in a same association sub-information having a same type of association relationship; The performing feature extraction on the target object set to obtain an object feature of the first object comprises: determining a target association sub-information containing the first object from the plurality of association sub-information; performing feature extraction on the target object set and the target association sub-information to obtain the object feature.

5. The feature extraction method of claim 4, wherein, The target association sub-information is a plurality of, and the performing feature extraction on the target object set and the target association sub-information to obtain the object feature comprises: performing feature extraction on the target object set and the target association sub-information to obtain a first feature of the first object for each target association sub-information. The first features corresponding to the plurality of target association sub-information are weighted based on first weights corresponding to the plurality of target association sub-information, to obtain the object feature, and the first weight corresponding to each target association sub-information represents an influence degree of the association relationship corresponding to the target association sub-information on the object feature of the first object.

6. The feature extraction method of claim 4, wherein, The target association sub-information is multiple, and the object feature is obtained by performing feature extraction on the target object set and the target association sub-information. For each target association sub-information, the first feature of the first object is obtained by performing feature extraction on the target object set and the target association sub-information. The first features corresponding to the plurality of target association sub-information are weighted based on first weights corresponding to the plurality of target association sub-information, to obtain the object feature, and the first weight corresponding to each target association sub-information represents an influence degree of the association relationship corresponding to the target association sub-information on the object feature of the first object. The third feature of the first object is obtained by performing feature extraction on the association information. The second feature and the third feature are spliced to obtain the object feature.

7. The feature extraction method according to claim 5 or 6, characterized in that, The type to which the first object belongs is a target type, and the first feature of the first object is obtained by performing feature extraction on the target object set and the target association sub-information for each target association sub-information. For each target association sub-information, the target association sub-information is divided into a plurality of first sub-information, each first sub-information contains the first object and other objects belonging to a reference type, and the other objects and the first object form the association relationship corresponding to the target association sub-information, and the reference types contained in different first sub-information are different, and the reference type is different from the target type. The target object set is divided into a plurality of second sub-information, each second sub-information contains the first object and other objects, and the association relationship between the first object and the other objects includes at least two kinds, and the association relationship between the first object and the other objects in different second sub-information is not completely same. The plurality of first sub-information and the plurality of second sub-information are respectively extracted to obtain the fourth feature corresponding to the plurality of first sub-information and the fifth feature corresponding to the plurality of second sub-information. The fourth feature corresponding to the plurality of first sub-information and the fifth feature corresponding to the plurality of second sub-information are weighted based on the second weight corresponding to the plurality of first sub-information and the second weight corresponding to the plurality of second sub-information, to obtain the first feature of the first object, the second weight corresponding to each first sub-information represents the influence degree of different objects on the object feature of the first object under the same association relationship, and the second weight corresponding to each second sub-information represents the influence degree of different objects on the object feature of the first object under different association relationships.

8. The feature extraction method of claim 7, wherein, The target association sub-information is divided into a plurality of first sub-information for each target association sub-information, including: For each target association sub-information, a plurality of reference types different from the target type are determined; Each time one of the reference types is selected, the first object and other objects satisfying a first selection condition are selected from the target association sub-information to form a first sub-information, until the plurality of reference types are selected, and a plurality of first sub-informations are obtained, wherein the selected objects satisfying the first selection condition means that the objects belong to the reference type and form an association relationship corresponding to the target association sub-information with the first object.

9. The feature extraction method of claim 7, wherein, The division of the target object set into a plurality of second sub-informations includes: determining a plurality of types of association relationships included in the target object set; Each time at least two types of association relationships in the plurality of types are selected, the first object and other objects satisfying a second selection condition are selected from the target object set to form a second sub-information, until there is no at least two types of association relationships in the plurality of types of association relationships that are not selected at the same time, and a plurality of second sub-informations are obtained, wherein the selected objects satisfying the second selection condition means that the association relationship between the selected objects and the first object is the at least two types of association relationships.

10. The feature extraction method of claim 1, wherein, The feature extraction of the target object set to obtain the object feature of the first object includes: calling a feature extraction model to perform feature extraction on the target object set to obtain the object feature of the first object.

11. The feature extraction method of claim 4, wherein, The feature extraction of the target object set and the target association sub-information to obtain the object feature includes: calling a feature extraction model to perform feature extraction on the target object set and the target association sub-information to obtain the object feature.

12. The feature extraction method of claim 11, wherein, The feature extraction model includes a plurality of first feature extraction networks and a second feature extraction network, the target association sub-information is a plurality, and the calling of the feature extraction model to perform feature extraction on the target object set and the target association sub-information to obtain the object feature includes: calling each first feature extraction network to perform feature extraction on the target object set and each target association sub-information to obtain the first feature of the first object; calling the second feature extraction network to perform weighted processing on the first features corresponding to the plurality of target association sub-informations based on first weights corresponding to the plurality of target association sub-informations to obtain the object feature, and the first weight corresponding to each target association sub-information represents the influence degree of the association relationship corresponding to the target association sub-information on the object feature of the first object.

13. The feature extraction method of claim 11, wherein, The feature extraction model includes a plurality of first feature extraction networks, a second feature extraction network and a third feature extraction network, the target association sub-information is a plurality, and the calling of the feature extraction model to perform feature extraction on the target object set and the target association sub-information to obtain the object feature includes: calling each first feature extraction network to perform feature extraction on the target object set and each target association sub-information to obtain the first feature of the first object; The second feature extraction network is called to perform weighted processing on the first features corresponding to the plurality of target association sub-information based on first weights corresponding to the plurality of target association sub-information, to obtain the second feature of the first object, and the first weight corresponding to each target association sub-information represents an influence degree of the association relationship corresponding to the target association sub-information on the object feature of the first object. The third feature extraction network is called to perform feature extraction on the association information, to obtain the third feature of the first object. The second feature extraction network is called to splice the second feature and the third feature, to obtain the object feature.

14. The feature extraction method of claim 13, wherein, Before the second feature extraction network is called to perform weighted processing on the first features corresponding to the plurality of target association sub-information based on first weights corresponding to the plurality of target association sub-information, to obtain the second feature of the first object, the feature extraction method further includes: The second feature extraction network is called to process the plurality of target association sub-information, to obtain the first weights corresponding to the plurality of target association sub-information.

15. The feature extraction method of claim 13, wherein, The first feature extraction network is called to perform feature extraction on the target object set and each target association sub-information, to obtain the first feature of the first object, including: For each target association sub-information, a feature extraction layer in the first feature extraction network corresponding to the target association sub-information is called to divide the target association sub-information into a plurality of first sub-information, each first sub-information containing the first object and other objects belonging to a reference type, and the other objects and the first object form an association relationship corresponding to the target association sub-information, and different first sub-information contains different reference types; The feature extraction layer is called to divide the target object set into a plurality of second sub-information, each second sub-information containing the first object and other objects, and the association relationship between the first object and the other objects includes at least two kinds, and the association relationship between the first object and the other objects in different second sub-information is not completely the same; The feature extraction layer is called to perform feature extraction on the plurality of first sub-information and the plurality of second sub-information, to obtain the fourth features corresponding to the plurality of first sub-information and the fifth features corresponding to the plurality of second sub-information; The splicing layer in the first feature extraction network is called to perform weighted processing on the fourth features corresponding to the plurality of first sub-information and the fifth features corresponding to the plurality of second sub-information based on the second weights corresponding to the plurality of first sub-information and the second weights corresponding to the plurality of second sub-information, to obtain the first feature of the first object, the second weight corresponding to each first sub-information representing an influence degree of different objects on the object feature of the first object under the same association relationship, and the second weight corresponding to each second sub-information representing an influence degree of different objects on the object feature of the first object under different association relationships.

16. The feature extraction method of claim 15, wherein, Before the calling the concatenation layer in the first feature extraction network, the feature extraction method further comprises: calling the attention layer in the first feature extraction network, respectively processing the plurality of first sub-information and the plurality of second sub-information, obtaining the second weight corresponding to the plurality of first sub-information and the second weight corresponding to the plurality of second sub-information.

17. The feature extraction method of claim 15, wherein, The calling the feature extraction layer in the first feature extraction network corresponding to the target associated sub-information, the target associated sub-information is divided into a plurality of first sub-information, comprising: calling the feature extraction layer, determining a plurality of reference types different from the target type; calling the feature extraction layer, selecting one of the reference types each time, selecting the first object and other objects satisfying the first selection condition from the target associated sub-information to form a first sub-information, until the plurality of reference types are selected, obtaining a plurality of first sub-information, wherein the selected object satisfying the first selection condition means that the object belongs to the reference type and forms the association relationship corresponding to the target associated sub-information with the first object.

18. The feature extraction method of claim 15, wherein, The calling the feature extraction layer, dividing the target object set into a plurality of second sub-information, comprising: calling the feature extraction layer, determining a plurality of types of association relationships included in the target object set; calling the feature extraction layer, selecting at least two types of association relationships in the plurality of types each time, selecting the first object and other objects satisfying the second selection condition from the target object set to form a second sub-information, until there is no at least two types of association relationships that are not selected at the same time in the plurality of types of association relationships, obtaining a plurality of second sub-information, wherein the selected object satisfying the second selection condition means that the association relationship between the selected object and the first object is the at least two types of association relationships.

19. The feature extraction method according to any one of claims 10-18, characterized in that, The training process of the feature extraction model is as follows: obtaining two sample objects in sample association information, the similarity between the two sample objects is greater than the reference similarity; calling the feature extraction model, respectively obtaining the predicted features corresponding to the two sample objects; using a target function to process the predicted features corresponding to the two sample objects to obtain a loss value; training the feature extraction model based on the loss value.

20. The feature extraction method of claim 19, wherein, After the training of the feature extraction model based on the loss value, the feature extraction method further comprises: in response to the loss value being less than a reference value, ending the training of the feature extraction model.

21. The feature extraction method of claim 1, wherein, The association information is a heterogeneous graph, the heterogeneous graph comprises a plurality of object nodes and a connection line between any two object nodes, an object node represents an object, and a connection line represents an association relationship between objects corresponding to two object nodes connected by the connection line; The next target object satisfying an association condition is selected based on the current target object from the association information until the number of determined target objects reaches a target number, and a target object set is constructed, including: The next target object node satisfying an association condition is selected based on the current target object node from the heterogeneous graph until the number of determined target object nodes reaches the target number, and a target node network is constructed, the association condition refers to that the connection line between the next target object node and the current target object node matches the current determined target association relationship, and the target node network includes the target number of target object nodes and the connection lines between any two target object nodes; The target object set is subjected to feature extraction to obtain the object feature of the first object, including: The target node network is subjected to feature extraction to obtain the object feature.

22. The feature extraction method of claim 21, wherein, The feature extraction method further includes: According to a plurality of connection lines contained in the heterogeneous graph, the heterogeneous graph is divided into a plurality of association relationship subgraphs, each association relationship subgraph contains a plurality of object nodes, and the connection lines between different object nodes in the same association relationship subgraph match the same type of association relationship; The target node network is subjected to feature extraction to obtain the object feature, including: A target association relationship subgraph containing the first object node corresponding to the first object is determined from the plurality of association relationship subgraphs; The target node network and the target association relationship subgraph are subjected to feature extraction to obtain the object feature.

23. A feature extraction apparatus characterized by comprising: The feature extraction device includes: An association information acquisition unit configured to perform acquisition of association information, the association information including a plurality of objects and association relationships between any two objects, the plurality of objects including account numbers registered on a server and multimedia resources published through the server; An initial object determination unit configured to perform determination of a first object as an initial target object, the first object being any account number registered on the server or any multimedia resource published through the server; An object set construction unit configured to perform selection of the next target object satisfying an association condition based on the current target object from the association information until the number of determined target objects reaches a target number, and construction of a target object set, the association condition referring to that the association relationship between the next target object and the current target object is the current determined target association relationship, the target object set including the target number of target objects and the association relationships between any two target objects; A feature extraction unit configured to perform feature extraction on the target object set to obtain the object feature of the first object.

24. The feature extraction apparatus according to claim 23, characterized by, The object set construction unit includes: An association relationship determination subunit configured to perform determination of a plurality of types of association relationships of the current target object based on the association information; An association relationship selection subunit configured to perform selection of one type of association relationship from the plurality of types, and determination of the selected association relationship as the current target association relationship; An object determining subunit is configured to perform the following: selecting, from the association information, an object having a target association relationship with the current target object as the current target association relationship; and determining the selected object as the next target object.

25. The feature extraction apparatus according to claim 24, characterized by, The object determining subunit is configured to perform the following: selecting, from the association information, multiple candidate objects having a target association relationship with the current target object as the current target association relationship; and determining any object in the multiple candidate objects as the next target object.

26. The feature extraction apparatus according to claim 23, characterized by The feature extraction apparatus further includes: An association information dividing unit is configured to perform the following: dividing the association information into multiple association sub-information according to multiple types of association relationships contained in the association information, each association sub-information containing multiple objects, and different objects in a same association sub-information having a same type of association relationship; and The feature extraction unit is configured to perform the following: determining, from the multiple association sub-information, a target association sub-information containing the first object; and performing feature extraction on the target object set and the target association sub-information to obtain the object feature.

27. The feature extraction apparatus according to claim 26, characterized by, The target association sub-information is multiple, and the feature extraction unit includes: A first feature extraction subunit is configured to perform the following: performing feature extraction on the target object set and the target association sub-information for each target association sub-information to obtain a first feature of the first object; and An object feature extraction subunit is configured to perform the following: performing weighted processing on the first features corresponding to the multiple target association sub-information based on first weights corresponding to the multiple target association sub-information to obtain the object feature, the first weight corresponding to each target association sub-information representing an influence degree of an association relationship corresponding to the target association sub-information on the object feature of the first object.

28. The feature extraction apparatus according to claim 26, characterized by, The target association sub-information is multiple, and the feature extraction unit includes: A first feature extraction subunit is configured to perform the following: performing feature extraction on the target object set and the target association sub-information for each target association sub-information to obtain a first feature of the first object; and A second feature determination subunit is configured to perform the following: performing weighted processing on the first features corresponding to the multiple target association sub-information based on first weights corresponding to the multiple target association sub-information to obtain a second feature of the first object, the first weight corresponding to each target association sub-information representing an influence degree of an association relationship corresponding to the target association sub-information on the object feature of the first object; and A third feature extraction subunit is configured to perform the following: performing feature extraction on the association information to obtain a third feature of the first object; and An object feature extraction subunit is configured to perform the following: splicing the second feature and the third feature to obtain the object feature.

29. The feature extraction apparatus according to claim 27 or 28, characterized by, The first object belongs to a target type, and the first feature extraction subunit is configured to perform the following: The target association sub-information is divided into a plurality of first sub-information, each first sub-information contains the first object and other objects belonging to a reference type, and the other objects and the first object constitute an association relationship corresponding to the target association sub-information, and the reference types in different first sub-information are different, and the reference type is different from the target type; The target object set is divided into a plurality of second sub-information, each second sub-information contains the first object and other objects, and the association relationship between the first object and the other objects includes at least two kinds, and the association relationship between the first object and the other objects in different second sub-information is not completely same; Respectively extracting features of the plurality of first sub-information and the plurality of second sub-information, obtaining fourth features corresponding to the plurality of first sub-information and fifth features corresponding to the plurality of second sub-information; Based on the second weight corresponding to the plurality of first sub-information and the second weight corresponding to the plurality of second sub-information, the fourth features corresponding to the plurality of first sub-information and the fifth features corresponding to the plurality of second sub-information are weighted, to obtain the first feature of the first object, the second weight corresponding to each first sub-information represents the influence degree of different objects on the object feature of the first object under the same association relationship, and the second weight corresponding to each second sub-information represents the influence degree of different objects on the object feature of the first object under different association relationships.

30. The feature extraction apparatus according to claim 29, characterized by, The first feature extraction subunit is configured to perform: For each target association sub-information, a plurality of reference types different from the target type are determined; Each time one reference type is selected, the first object and other objects satisfying a first selection condition are selected from the target association sub-information to form a first sub-information, until the plurality of reference types are selected, to obtain a plurality of first sub-information, wherein the selected objects satisfy the first selection condition, that is, the objects belong to the reference type and constitute the association relationship corresponding to the target association sub-information with the first object.

31. The feature extraction apparatus according to claim 29, characterized by, The first feature extraction subunit is configured to perform: Determine a plurality of types of association relationships included in the target object set; Each time at least two types of association relationships in the plurality of types are selected, the first object and other objects satisfying a second selection condition are selected from the target object set to form a second sub-information, until there is no at least two types of association relationships that are not selected at the same time in the plurality of types of association relationships, to obtain a plurality of second sub-information, wherein the selected objects satisfy the second selection condition, that is, the association relationship between the selected objects and the first object is the at least two types of association relationships.

32. The feature extraction apparatus according to claim 23, characterized by, The feature extraction unit is configured to execute the feature extraction model to extract features of the target object set to obtain the object feature of the first object.

33. The feature extraction apparatus according to claim 26, characterized by, The feature extraction unit is configured to execute the feature extraction model to extract features of the target object set and the target association sub-information to obtain the object feature.

34. The feature extraction apparatus according to claim 33, characterized by, The feature extraction model comprises a plurality of first feature extraction networks and a second feature extraction network, the target association sub-information is multiple, the feature extraction unit comprises: A first feature extraction subunit is configured to execute calling each first feature extraction network to perform feature extraction on the target object set and each target association sub-information respectively to obtain first features of the first object. An object feature extraction subunit is configured to execute calling the second feature extraction network to perform weighted processing on the first features corresponding to the multiple target association sub-informations based on first weights corresponding to the multiple target association sub-informations to obtain the object feature, and the first weight corresponding to each target association sub-information represents an influence degree of the association relationship corresponding to the target association sub-information on the object feature of the first object.

35. The feature extraction apparatus according to claim 33, characterized by, The feature extraction model comprises a plurality of first feature extraction networks, a second feature extraction network and a third feature extraction network, the target association sub-information is multiple, and the feature extraction unit comprises: A first feature extraction subunit is configured to execute calling each first feature extraction network to perform feature extraction on the target object set and each target association sub-information respectively to obtain first features of the first object. A second feature determination subunit is configured to execute calling the second feature extraction network to perform weighted processing on the first features corresponding to the multiple target association sub-informations based on first weights corresponding to the multiple target association sub-informations to obtain second features of the first object, and the first weight corresponding to each target association sub-information represents an influence degree of the association relationship corresponding to the target association sub-information on the object feature of the first object. A third feature extraction subunit is configured to execute calling the third feature extraction network to perform feature extraction on the association information to obtain third features of the first object. An object feature extraction subunit is configured to execute calling the second feature extraction network to splice the second features and the third features to obtain the object feature.

36. The feature extraction apparatus according to claim 35, characterized by, The feature extraction unit further comprises: A first weight determination subunit is configured to execute calling the second feature extraction network to process the multiple target association sub-informations to obtain the first weights corresponding to the multiple target association sub-informations.

37. The feature extraction apparatus according to claim 35, characterized by, The first feature extraction subunit is configured to execute: For each target association sub-information, calling a feature extraction layer in the first feature extraction network corresponding to the target association sub-information to divide the target association sub-information into a plurality of first sub-informations, each first sub-information containing the first object and other objects belonging to a reference type, and the other objects and the first object constitute the association relationship corresponding to the target association sub-information, and different first sub-informations contain different reference types. The feature extraction layer is invoked to divide the target object set into a plurality of second sub-information, each second sub-information containing the first object and other objects, and the association relationship between the first object and the other objects includes at least two kinds, and the association relationship between the first object and the other objects in different second sub-information is not completely the same; The feature extraction layer is invoked to perform feature extraction on the plurality of first sub-information and the plurality of second sub-information respectively, to obtain fourth features corresponding to the plurality of first sub-information and fifth features corresponding to the plurality of second sub-information; The concatenation layer in the first feature extraction network is invoked to perform weighted processing on the fourth features corresponding to the plurality of first sub-information and the fifth features corresponding to the plurality of second sub-information based on the second weights corresponding to the plurality of first sub-information and the second weights corresponding to the plurality of second sub-information, to obtain the first feature of the first object, the second weight corresponding to each first sub-information representing the influence degree of different objects on the object feature of the first object under the same association relationship, and the second weight corresponding to each second sub-information representing the influence degree of different objects on the object feature of the first object under different association relationships.

38. The feature extraction apparatus according to claim 37, characterized by, The feature extraction unit further comprises: The second weight determination sub-unit is configured to invoke the attention layer in the first feature extraction network to process the plurality of first sub-information and the plurality of second sub-information respectively to obtain the second weights corresponding to the plurality of first sub-information and the second weights corresponding to the plurality of second sub-information.

39. The feature extraction apparatus according to claim 37, characterized by, The first feature extraction sub-unit is configured to perform: Invoke the feature extraction layer to determine a plurality of reference types different from the target type; The feature extraction layer is invoked to select the first object and other objects satisfying the first selection condition from the target association sub-information to form a first sub-information, until the selection of the plurality of reference types is completed, to obtain a plurality of first sub-information, wherein the selected object satisfies the first selection condition, that is, the object belongs to the reference type and forms the association relationship corresponding to the target association sub-information with the first object.

40. The feature extraction apparatus according to claim 37, characterized by, The first feature extraction sub-unit is configured to perform: The feature extraction layer is invoked to determine a plurality of types of association relationships included in the target object set; The feature extraction layer is invoked to select the first object and other objects satisfying the second selection condition from the target object set to form a second sub-information, until there is no at least two types of association relationships that are not selected at the same time in the plurality of types of association relationships, to obtain a plurality of second sub-information, wherein the selected object satisfies the second selection condition, that is, the association relationship between the selected object and the first object is the at least two types of association relationships.

41. The feature extraction apparatus according to any one of claims 32 to 40, characterized by, The feature extraction device further comprises: The model training unit is configured to execute the two sample objects in the sample association information, and the similarity between the two sample objects is greater than the reference similarity; The model training unit is further configured to execute calling the feature extraction model to obtain the predicted features corresponding to the two sample objects respectively. The model training unit is further configured to execute processing the predicted features corresponding to the two sample objects by using a target function to obtain a loss value. The model training unit is further configured to execute training the feature extraction model based on the loss value.

42. The feature extraction apparatus according to claim 41, characterized by, The model training unit is further configured to execute ending training the feature extraction model in response to the loss value being less than a reference value.

43. The feature extraction apparatus according to claim 23, characterized by, The association information is a heterogeneous graph, the heterogeneous graph includes a plurality of object nodes and connection lines between any two object nodes, one object node represents one object, and one connection line represents an association relationship between objects corresponding to two object nodes connected by the connection line. The object set construction unit is configured to execute selecting a next target object node satisfying an association condition from the heterogeneous graph based on a current target object node until the number of determined target object nodes reaches the target number, to construct a target node network, the association condition refers to that a connection line between the next target object node and the current target object node matches a current determined target association relationship, and the target node network includes the target number of target object nodes and connection lines between any two target object nodes. The feature extraction unit is configured to execute feature extraction on the target node network to obtain the object features.

44. The feature extraction apparatus according to claim 43, characterized by, The feature extraction device further includes: An association information division unit is configured to execute dividing the heterogeneous graph into a plurality of association relationship subgraphs according to a plurality of connection lines included in the heterogeneous graph, each association relationship subgraph includes a plurality of object nodes, and connection lines between different object nodes in the same association relationship subgraph match the same type of association relationship. The feature extraction unit is configured to execute: Determining a target association relationship subgraph containing a first object node corresponding to the first object from the plurality of association relationship subgraphs; Performing feature extraction on the target node network and the target association relationship subgraph to obtain the object features.

45. An electronic device, comprising: The electronic device includes: One or more processors; Memory for storing instructions executable by the one or more processors; The one or more processors are configured to execute the feature extraction method according to any one of claims 1 to 22.

46. A computer-readable storage medium, characterized in that, When the instructions in the computer readable storage medium are executed by the processor of the electronic device, the electronic device can execute the feature extraction method according to any one of claims 1 to 22.

47. A computer program product comprising a computer program, characterised in that, The computer program is executed by the processor to implement the feature extraction method according to any one of claims 1 to 22.

Images