Construction method, resource pushing method and device, equipment, storage medium and program product

Abstract

The invention provides a construction method, a resource pushing method and device, equipment, a storage medium and a program product, and relates to the field of artificial intelligence, the field of digital marketing and the field of intelligent retrieval. The construction method comprises the following steps: acquiring a sample resource feature set related to a sample object; training an object resource recommendation network corresponding to the sample object based on the sample support feature subset to obtain an intermediate object resource recommendation network; processing the sample verification feature subset by utilizing the intermediate object resource recommendation network to obtain object task verification loss; training the initial meta-recommendation network based on the object task verification loss respectively related to the N sample objects to obtain a trained meta-recommendation network; according to the meta-recommendation network, a resource recommendation model is constructed, the resource recommendation model is used for processing resource features related to the target object to obtain an object recommendation index, and the object recommendation index is suitable for determining target resources pushed to the target object.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the field of artificial intelligence, the field of digital marketing, and the field of intelligent retrieval, and more particularly, to a construction method, a resource pushing method, an apparatus, a device, a storage medium, and a program product. BACKGROUND

[0002] With the rapid development of Internet technology, users can conveniently browse resource information such as news and videos through terminal devices such as smart phones, and related Internet platforms can recommend resources to users based on user needs to improve the convenience of users obtaining resources and improve user experience.

[0003] In the process of implementing the present concept, the inventors have found that at least the following problems exist in the related art: the recommended resources are difficult to accurately meet the actual needs of users, especially when the training data of the model is unbalanced and the amount of training data is small, it is difficult to accurately push resources to users who need to obtain the resources, resulting in low accuracy of resource pushing. SUMMARY

[0004] Therefore, the present disclosure provides a resource recommendation model construction method, a resource pushing method, an apparatus, a device, a storage medium, and a program product.

[0005] One aspect of the present disclosure provides a construction method of a resource recommendation model, comprising:

[0006] obtaining a sample resource feature set related to a sample object, the sample resource feature set comprising a sample support feature subset and a sample validation feature subset, the sample object comprising N, N > 1;

[0007] training an object resource recommendation network corresponding to the sample object based on the sample support feature subset, to obtain an intermediate object resource recommendation network;

[0008] processing the sample validation feature subset using the intermediate object resource recommendation network to obtain an object task validation loss;

[0009] training an initial meta recommendation network based on the object task validation loss related to each of the N sample objects, to obtain a trained meta recommendation network, wherein the initial meta recommendation network has the same network structure as the object resource recommendation network; and

[0010] constructing a resource recommendation model according to the meta recommendation network, wherein the resource recommendation model is used to process resource features related to a target object to obtain an object recommendation indicator, and the object recommendation indicator is suitable for determining a target resource to be pushed to the target object.

[0011] According to an embodiment of the present disclosure, training the object resource recommendation network corresponding to the sample object based on the sample support feature subset includes:

[0012] Based on the difference degree detection network, the sample support feature subset related to the sample object is processed to obtain an object task difference degree related to the sample object, and the object task difference degree represents a difference between the object network parameters of the object resource recommendation network and the meta network parameters of the initial meta recommendation network.

[0013] According to the object task difference degree related to the sample object and the current meta network parameters, the current local clustering task network parameters are determined.

[0014] According to the current local clustering task network parameters, parameter migration is performed to obtain the object resource recommendation network.

[0015] The sample support feature subset related to the sample object is processed by using the object resource recommendation network to obtain an object task support loss.

[0016] Based on a gradient descent algorithm, the object resource recommendation network is trained according to the object task support loss.

[0017] Among the N sample objects, M local clustering objects are associated with the same object task difference degree, N>M>1, and the M local clustering objects correspond to the same local clustering network parameters.

[0018] According to an embodiment of the present disclosure, the difference degree detection network includes the Lth detection layer and the L+1th detection layer, and processing the sample support feature subset related to the sample object based on the difference degree detection network includes:

[0019] The L+1th detection layer is used to perform a detection layer operation to process the Lth difference degree detection feature output by the Lth detection layer to obtain the L+1th difference degree detection feature, and the detection layer operation includes:

[0020] The L+1th difference degree between the Lth difference degree detection feature and a detection layer clustering center of the L+1th detection layer is determined, the L+1th detection layer includes K detection layer clustering centers, and K≥1.

[0021] According to the K L+1th difference degrees, a feature update weight of the L+1th detection layer is determined; and

[0022] Based on a neural network algorithm, the Lth difference degree detection feature is processed according to the feature update weight of the L+1th detection layer to obtain the L+1th difference degree detection feature.

[0023] According to an embodiment of the present disclosure, the processing of the sample support feature subset related to the sample object based on the difference degree detection network further includes:

[0024] performing feature fusion on the sample support features in the sample support feature subset related to the sample object, to obtain a sample object task feature; and

[0025] inputting the sample object task feature into the first detection layer of the difference degree detection network, to output a first difference degree detection feature.

[0026] According to an embodiment of the present disclosure, the training of the initial meta-recommendation network based on the object task verification loss related to each of the N sample objects includes:

[0027] determining an object task verification gradient related to the sample object based on the object task verification loss;

[0028] performing gradient fusion on the object task verification gradients corresponding to each of the N sample objects, to obtain a fused task verification gradient; and

[0029] updating the meta-network parameters of the initial meta-recommendation network according to the fused task verification gradient.

[0030] According to an embodiment of the present disclosure, the construction method further includes:

[0031] processing the sample verification feature subset related to the sample object based on a local clustering task network, to obtain a local clustering task verification loss, wherein the local clustering task network is constructed based on current local clustering task network parameters;

[0032] updating the difference degree network parameters of the difference degree detection network according to the local clustering task verification loss based on a gradient descent algorithm.

[0033] According to an embodiment of the present disclosure, the sample support features in the sample support feature subset are obtained by performing feature extraction on sample support data by using an initial encoder, and the sample support data is related to the sample object.

[0034] According to an embodiment of the present disclosure, the construction method further includes:

[0035] updating the network parameters of the initial encoder according to the local clustering task verification loss based on a gradient descent algorithm.

[0036] According to an embodiment of the present disclosure, the processing of the sample verification feature subset by using the intermediate object resource recommendation network includes:

[0037] Determine a positive sample verification feature and a negative sample verification feature from the sample verification feature subset based on the interaction operation relationship between the sample object and the sample resource;

[0038] Determine a first object fusion feature based on the sample object attribute feature related to the sample object and the positive sample verification feature;

[0039] Determine a second object fusion feature based on the sample object attribute feature and the negative sample verification feature;

[0040] Process the first object fusion feature and the second object fusion feature by using the intermediate object resource recommendation network to obtain a first predicted object recommendation index and a second predicted object recommendation index; and

[0041] Process the first predicted object recommendation index and the second predicted object recommendation index according to a loss function to obtain the object task verification loss.

[0042] According to an embodiment of the present disclosure, constructing a resource recommendation model according to the meta recommendation network comprises:

[0043] Determine target local clustering task network parameters related to the target object by using meta recommendation network parameters of the meta recommendation network and an object task difference degree related to the target object;

[0044] Perform parameter migration according to the target local clustering task network parameters to obtain an initial target object resource recommendation network;

[0045] Process a target support feature subset related to the target object by using the initial target object resource recommendation network to obtain a target object task support loss;

[0046] Train the initial target object resource recommendation network according to the target object task support loss based on a gradient descent algorithm to obtain a trained target object resource recommendation network; and

[0047] Determine the trained target object resource recommendation network as the resource recommendation model.

[0048] Another aspect of the present disclosure provides a resource pushing method, comprising:

[0049] Obtain resource features related to a to-be-recommended resource;

[0050] Process the resource features by using a resource recommendation model to obtain an object recommendation index related to the to-be-recommended resource, wherein the resource recommendation model is determined based on the construction method provided by an embodiment of the present disclosure;

[0051] determine a target resource from the to-be-recommended resources based on the object recommendation index; and

[0052] recommend the target resource to the target object.

[0053] Another aspect of the present disclosure provides a construction device of a resource recommendation model, comprising:

[0054] a first obtaining module configured to obtain a sample resource feature set related to a sample object, the sample resource feature set comprising a sample support feature subset and a sample verification feature subset, and the sample object comprising N, N > 1;

[0055] an intermediate object resource recommendation network obtaining module configured to train an object resource recommendation network corresponding to the sample object based on the sample support feature subset, to obtain an intermediate object resource recommendation network;

[0056] an object task verification loss obtaining module configured to process the sample verification feature subset by using the intermediate object resource recommendation network, to obtain an object task verification loss;

[0057] a meta recommendation network obtaining module configured to train an initial meta recommendation network based on object task verification losses respectively related to the N sample objects, to obtain a trained meta recommendation network, wherein the initial meta recommendation network and the object resource recommendation network have the same network structure; and

[0058] a construction module configured to construct a resource recommendation model according to the meta recommendation network, wherein the resource recommendation model is used to process resource features related to a target object to obtain an object recommendation index, and the object recommendation index is applicable to determining a target resource to be pushed to the target object.

[0059] Another aspect of the present disclosure provides a resource pushing device, comprising:

[0060] a second obtaining module configured to obtain resource features related to to-be-recommended resources;

[0061] an object recommendation index obtaining module configured to process the resource features by using a resource recommendation model to obtain an object recommendation index related to the to-be-recommended resources, wherein the resource recommendation model is determined based on a construction method provided by an embodiment of the present disclosure;

[0062] a target resource determining module configured to determine a target resource from the to-be-recommended resources based on the object recommendation index; and

[0063] a pushing module configured to recommend the target resource to the target object.

[0064] Another aspect of the present disclosure provides an electronic device, comprising:

[0065] one or more processors;

[0066] a memory for storing one or more programs,

[0067] wherein the one or more programs, when executed by the one or more processors, cause the one or more processors to implement the method as described above.

[0068] Another aspect of the present disclosure provides a computer-readable storage medium storing computer-executable instructions for implementing the method as described above when executed.

[0069] Another aspect of the present disclosure provides a computer program product comprising computer-executable instructions for implementing the method as described above when executed.

[0070] According to the embodiments of the present disclosure, because of the adopted meta-learning mechanism, by training the object resource recommendation network corresponding to the sample object by utilizing the sample support feature subset, and training the initial meta recommendation network by utilizing the object task validation loss obtained by processing the sample validation feature subset by the trained intermediate object resource recommendation network, the learning ability of the meta recommendation network can be adaptively updated by utilizing the learning ability of the intermediate object resource recommendation network based on the meta-learning mechanism, so as to improve the learning ability of the meta recommendation network for the biased sample resource features, and then make the resource recommendation model constructed according to the meta recommendation network be able to realize the accurate prediction of the recommendation index under the condition of small sample, so as to improve the accuracy of resource pushing and improve the resource retrieval efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0071] The above and other objects, features and advantages of the present disclosure will become more apparent from the following description of embodiments of the present disclosure taken in conjunction with the accompanying drawings, in which:

[0072] Figure 1 An exemplary system architecture to which the resource pushing method and device according to the embodiments of the present disclosure can be applied is schematically shown;

[0073] Figure 2 A flowchart of a construction method of a resource recommendation model according to the embodiments of the present disclosure is schematically shown;

[0074] Figure 3 A schematic diagram of a sample resource feature set and a resource feature set according to the embodiments of the present disclosure is schematically shown;

[0075] Figure 4 A flowchart of processing a sample validation feature subset by an intermediate object resource recommendation network according to the embodiments of the present disclosure is schematically shown;

[0076] Figure 5 A schematic diagram of positive and negative samples according to a comparative example of the present disclosure is shown;

[0077] Figure 6 A schematic diagram of the principle of updating the meta-network parameters according to an embodiment of the present disclosure is shown;

[0078] Figure 7 A flowchart of a resource pushing method according to an embodiment of the present disclosure is shown;

[0079] Figure 8 A block diagram of a device for constructing a resource recommendation model according to an embodiment of the present disclosure is shown;

[0080] Figure 9 A block diagram of a device for constructing a resource recommendation model according to an embodiment of the present disclosure is shown; and

[0081] Figure 10 A block diagram of an electronic device suitable for implementing the method for constructing a resource recommendation model and the resource pushing method according to an embodiment of the present disclosure is shown DETAILED DESCRIPTION

[0082] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. It should be understood, however, that the description which follows is merely exemplary and is not intended to limit the scope of the present disclosure. In the following detailed description of the embodiments of the present disclosure, numerous specific details are set forth in order to provide a thorough understanding of the present disclosure. However, it would be apparent to one skilled in the art that the present disclosure can be practiced without these specific details. In other instances, well-known structures and functions have not been described in detail in order to avoid obscuring aspects of the present disclosure.

[0083] The terms used herein are merely used to describe specific embodiments and are not intended to limit the present disclosure. The terms "include", "comprise", and the like used herein indicate the presence of the described features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0084] All terms used herein, including technical and scientific terms, have the meanings commonly understood by one of ordinary skill in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted as having meanings consistent with the context of the present specification, and should not be interpreted in an idealized or overly formal manner.

[0085] In the case of using expressions similar to "at least one of A, B, and C, etc.", it will be understood that the meaning is generally an "and / or", e.g. "A and / or B and / or C and / or etc." In addition, use of the "and / or" limitation is to be construed as an "even more limiting" alternative (as opposed to using a simple "or" limitation). For example, in the case of "at least one of A and B and / or C" the candidate list of alternatives includes: A only, B only, C only, A and B, A and C, B and C, and / or A and B and C.

[0086] In embodiments of the present disclosure, the collection, updating, analysis, processing, use, transmission, provision, disclosure, storage, etc. of the data involved (for example, including but not limited to user personal information) comply with the relevant legal regulations, are used for legal purposes, and do not violate public order and good customs. In particular, necessary measures have been taken to prevent illegal access to user personal information data, and to maintain user personal information security, network security and national security.

[0087] In embodiments of the present disclosure, the authorization or consent of the user is obtained before the user's personal information is acquired or collected.

[0088] The inventors found that in related recommendation systems, it is difficult to cold start for users with less interaction operations or for to-be-recommended resources, and the accuracy of resource pushing is low. The inventors believe that the reason is that the feature data obtained is sparse, and the effective sample data is less. The way of cold start for to-be-recommended resources in related technologies is difficult to achieve the expected pushing accuracy effect, resulting in poor user resource acquisition experience and low accurate pushing efficiency of to-be-recommended resources.

[0089] Embodiments of the present disclosure provide a resource recommendation model construction method, a resource pushing method, an apparatus, a device, a storage medium and a program product. The resource recommendation model construction method comprises: acquiring a sample resource feature set related to a sample object, the sample resource feature set comprising a sample support feature subset and a sample verification feature subset, the sample object comprising N, N > 1; training an object resource recommendation network corresponding to the sample object based on the sample support feature subset to obtain an intermediate object resource recommendation network; processing the sample verification feature subset by using the intermediate object resource recommendation network to obtain an object task verification loss; training an initial meta recommendation network based on the object task verification loss related to each of the N sample objects to obtain a trained meta recommendation network, wherein the initial meta recommendation network and the object resource recommendation network have the same network structure; and constructing a resource recommendation model according to the meta recommendation network, wherein the resource recommendation model is used to process resource features related to a target object to obtain an object recommendation index, and the object recommendation index is suitable for determining a target resource to be pushed to the target object.

[0090] According to an embodiment of the present disclosure, by training the object resource recommendation network corresponding to the sample object by using the sample support feature subset, training the initial meta recommendation network by using the obtained object task verification loss of the intermediate object resource recommendation network processed by using the trained intermediate object resource recommendation network on the sample verification feature subset, the learning ability of the intermediate object resource recommendation network can be used to adaptively update the network parameters of the meta recommendation network based on the meta learning mechanism, so as to improve the learning ability of the meta recommendation network for the biased sample resource features, and then the resource recommendation model constructed based on the meta recommendation network can realize accurate prediction of the recommendation index under the condition of small sample, so as to improve the accuracy of resource pushing and improve the resource retrieval efficiency.

[0091] The present disclosure also provides a resource pushing method, which comprises: obtaining resource features related to a to-be-recommended resource; processing the resource features by using a resource recommendation model to obtain an object recommendation index related to the to-be-recommended resource, wherein the resource recommendation model is determined based on the construction method of the resource recommendation model provided by the present disclosure; determining a target resource from the to-be-recommended resource based on the object recommendation index; and recommending the target resource to a target object.

[0092] Figure 1 An exemplary system architecture to which the resource pushing method and device according to an embodiment of the present disclosure can be applied is schematically shown. It should be noted that, Figure 1 The system architecture shown is only an example of the system architecture to which the embodiments of the present disclosure can be applied, to help those skilled in the art understand the technical content of the present disclosure, but does not mean that the embodiments of the present disclosure cannot be applied to other devices, systems, environments or scenarios.

[0093] As Figure 1 shown, the system architecture 100 according to this embodiment can include terminal devices 101, 102, 103, a network 104 and a server 105. The network 104 is used as a medium to provide a communication link between the terminal devices 101, 102, 103 and the server 105. The network 104 can include various connection types, such as wired and / or wireless communication links, etc.

[0094] The user can use the terminal devices 101, 102, 103 to interact with the server 105 through the network 104 to receive or send messages, etc. Various communication client applications can be installed on the terminal devices 101, 102, 103, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients and / or social platform software, etc. (only as examples).

[0095] The terminal devices 101, 102, 103 can be various electronic devices with display screens and supporting web browsing, including but not limited to smartphones, tablet computers, laptop computers and desktop computers, etc.

[0096] The server 105 can be a server providing various services, for example, a background management server providing support for a website browsed by a user using the terminal device 101, 102, or 103 (only as an example). The background management server can perform analysis and the like on received user requests and the like, and feed back the processing result (for example, a webpage, information, or data, or the like obtained or generated according to the user request) to the terminal device.

[0097] It should be noted that the resource pushing method provided by the embodiments of the present disclosure can generally be executed by the server 105. Accordingly, the resource pushing apparatus provided by the embodiments of the present disclosure can generally be arranged in the server 105. The resource pushing method provided by the embodiments of the present disclosure can also be executed by a server or a server cluster different from the server 105 and capable of communicating with the terminal device 101, 102, or 103 and / or the server 105. Accordingly, the resource pushing apparatus provided by the embodiments of the present disclosure can also be arranged in a server or a server cluster different from the server 105 and capable of communicating with the terminal device 101, 102, or 103 and / or the server 105. Alternatively, the resource pushing method provided by the embodiments of the present disclosure can also be executed by the terminal device 101, 102, or 103, or by other terminal devices different from the terminal device 101, 102, or 103. Accordingly, the resource pushing apparatus provided by the embodiments of the present disclosure can also be arranged in the terminal device 101, 102, or 103, or in other terminal devices different from the terminal device 101, 102, or 103.

[0098] It should be understood that Figure 1 The number of terminal devices, networks, and servers in the system 100 is only illustrative. According to the needs of implementation, there can be any number of terminal devices, networks, and servers.

[0099] Figure 2 An illustrative flowchart of a method for constructing a resource recommendation model according to an embodiment of the present disclosure is shown.

[0100] As shown in Figure 2 The method for constructing the resource recommendation model includes operations S210-S250.

[0101] In operation S210, a sample resource feature set related to a sample object is obtained.

[0102] According to an embodiment of the present disclosure, the sample resource feature set includes a sample support feature subset and a sample verification feature subset, and the sample object includes N, N>1.

[0103] According to an embodiment of the present disclosure, the sample object can include a user who has performed a resource interaction operation such as a resource-related search operation, a click operation, a purchase operation, and the like in a historical time period. The sample resource can include any type of resource such as an advertisement resource, a news resource, a video resource, a commodity detail page resource, and the like, and the present disclosure does not limit the specific type of the sample resource. The sample resource feature can include feature data related to the sample resource, such as a description text of the resource, an image in the detail page, text, and the like related to the attribute of the sample resource, or can further include feature data related to the number of purchases, the click frequency, and the like related to the interaction operation of the sample resource. The present disclosure does not limit the specific type of the sample resource feature, and a person skilled in the art can select it according to actual needs.

[0104] According to an embodiment of the present disclosure, the sample support feature subset and the sample verification feature subset can be determined from the sample resource feature set based on the time sequence attribute of each sample resource feature, but are not limited thereto. The sample support feature subset and the sample verification feature subset can also be determined based on other manners, and the present disclosure does not limit this.

[0105] In operation S220, an object resource recommendation network corresponding to the sample object is trained based on the sample support feature subset, and an intermediate object resource recommendation network is obtained.

[0106] According to an embodiment of the present disclosure, the object resource recommendation network can be constructed based on any type of deep learning algorithm, for example, can be constructed based on a multilayer perceptron algorithm, but is not limited thereto. The object resource recommendation network can also be constructed based on other types of deep learning algorithms. The present disclosure does not limit the specific algorithm type for constructing the object resource recommendation network, and a person skilled in the art can select it according to actual needs.

[0107] According to an embodiment of the present disclosure, the object resource recommendation network can be used to process the sample support feature subset corresponding to the sample object. The intermediate object resource recommendation network can be trained based on a supervised manner, or can also be trained based on a contrastive learning manner, and the present disclosure does not limit this.

[0108] In operation S230, the sample verification feature subset is processed using the intermediate object resource recommendation network, and an object task verification loss is obtained.

[0109] According to an embodiment of the present disclosure, the N intermediate object resource recommendation networks can correspond to the N sample objects one by one. Each intermediate object resource recommendation network can be used to process the sample verification feature subset of the sample object, so that the object task verification loss corresponding to each sample object can be obtained.

[0110] It should be noted that the output result of the intermediate object resource recommendation network can include a predicted recommendation index related to the sample resource, and the predicted recommendation index obtained by processing the sample verification feature subset based on the intermediate object resource recommendation network can be processed by using a loss function to obtain an object task verification loss.

[0111] In operation S240, the initial meta recommendation network is trained based on the object task verification loss related to each of the N sample objects to obtain a trained meta recommendation network.

[0112] According to an embodiment of the present disclosure, the initial meta recommendation network has the same network structure as the object resource recommendation network. For example, the initial meta recommendation network can have the same number of hidden layers as the object recommendation network, and the initial meta recommendation network and the object resource recommendation network are constructed based on the same deep learning algorithm.

[0113] According to an embodiment of the present disclosure, the N object task verification losses can be aggregated based on a preset aggregation manner, so as to train the initial meta recommendation network based on the obtained aggregated loss. Alternatively, the N object task verification losses can also be processed based on a gradient descent algorithm to generate the gradient of each of the N intermediate object resource recommendation networks, so as to adjust the meta network parameters of the initial meta recommendation network one or more times by fusing the N gradients, and realize the iterative update of the meta recommendation parameters of the initial meta recommendation network.

[0114] In operation S250, a resource recommendation model is constructed based on the meta recommendation network.

[0115] According to an embodiment of the present disclosure, the resource recommendation model is used to process resource features related to a target object to obtain an object recommendation index, and the object recommendation index is suitable for determining a target resource to be pushed to the target object.

[0116] According to an embodiment of the present disclosure, the object recommendation index can represent the probability of the target object performing an interactive operation on the to-be-recommended resource, or can also represent the matching degree between the to-be-recommended resource and the demand of the target object. The object recommendation index can be represented based on a score value, or can also be represented based on a matching level identifier, and the specific manner of representing the object recommendation index is not limited in the present disclosure, and can be selected by a person skilled in the art according to actual needs.

[0117] It should be understood that the target resource can be a resource with a high matching degree with the demand of the target object, or can also be a resource that can be executed by the target object for any type of interactive operation such as a click operation, a collection operation, a purchase operation, and the like. According to the object recommendation index, the target resource is determined and pushed, which can more accurately meet the actual demand of the target object and improve the resource recommendation accuracy. According to the embodiment of the present disclosure, the resource recommendation model can be constructed based on the meta-recommendation network, which can include parameter migration based on the meta-network parameters of the trained meta-recommendation network to obtain the resource recommendation model, or the resource recommendation model can also be constructed by other manners, and the specific manner of constructing the resource recommendation model is not limited by the embodiment of the present disclosure, as long as the meta-recommendation network parameters of the meta-recommendation network are used.

[0118] According to the embodiment of the present disclosure, by training the object resource recommendation network corresponding to the sample object by using the sample support feature subset, and training the initial meta-recommendation network by using the object task validation loss obtained by processing the sample validation feature subset by using the trained intermediate object resource recommendation network, the learning ability of the meta-recommendation network can be adaptively updated based on the meta-learning mechanism by using the learning ability of the intermediate object resource recommendation network, thereby improving the learning ability of the meta-recommendation network for the biased sample resource features, and further enabling the resource recommendation model constructed based on the meta-recommendation network to accurately predict the recommendation index under the condition of small samples, thereby improving the accuracy of resource pushing and improving the resource retrieval efficiency.

[0119] According to the embodiment of the present disclosure, the object recommendation index can include any index type, for example, the index type of the object recommendation index can include at least one of the following: a click type, a transaction type, a collection type, a shopping cart type, a comment type.

[0120] According to the embodiment of the present disclosure, the object recommendation index of the click type can include a click rate index. For example, a click rate index for a preset commodity marketing resource, a news resource.

[0121] According to the embodiment of the present disclosure, the object recommendation index of the transaction type can include a conversion rate index. For example, a purchase probability index for a preset commodity resource.

[0122] According to the embodiment of the present disclosure, the object recommendation index of the collection type can include a probability index of the target object performing a collection operation on a specific resource.

[0123] According to the embodiment of the present disclosure, the object recommendation index of the shopping cart type can include a shopping cart index. For example, a probability index of performing a shopping cart operation on a preset commodity resource.

[0124] According to the embodiment of the present disclosure, the object recommendation index of the comment type can include an index of a sample object performing a comment operation on a specific resource.

[0125] According to an embodiment of the present disclosure, the index type can further include a repurchase type index, a like type index, and any other type of index, and the embodiments of the present disclosure do not limit the specific type of index type, and a person skilled in the art can select according to actual needs, as long as the recommendation accuracy of the resource can be improved.

[0126] According to an embodiment of the present disclosure, the sample support features in the sample support feature subset are obtained by feature extraction on sample support data using an initial encoder, and the sample support data is related to the sample object.

[0127] According to an embodiment of the present disclosure, the initial encoder can include a pre-trained encoder, or can also be constructed based on the initialization encoder parameters generated based on the parameter initialization rule. The initial encoder can also be used to extract features from sample validation data to obtain sample validation features, thereby obtaining a sample validation feature subset.

[0128] It should be noted that the initial encoder can be constructed based on any type of neural network algorithm such as BERT (Bidirectional Encoder Representation from Transformers), SENet (Squeeze-and-Excitation Networks), etc., and the embodiments of the present disclosure do not limit this.

[0129] Figure 3 The schematic diagram of the sample resource feature set and the resource feature set according to an embodiment of the present disclosure is schematically shown.

[0130] As shown in Figure 3 The sample resource features obtained in relation to the sample object can be divided into support features and query features, respectively. For the sample object u1, the sample resource feature set related to the sample object u1 can include a sample support feature subset 311 and a sample validation feature subset 312. Accordingly, Figure 3 The sample support features and sample validation features corresponding to each of the above can correspond to the sample objects u2, u3 and u4, respectively.

[0131] As shown in Figure 3 The resource features corresponding to the target object u1' of the resource to be pushed can also be divided into a target support feature subset 321 and a target validation feature subset 322. Accordingly, the resource features corresponding to the row of the target object u2' can be the target support feature subset and the target validation feature subset of the target object u2'.

[0132] In one example of the present disclosure, the number of sample verification features in the sample verification feature subset can be less than the number of sample support features in the sample support feature subset.

[0133] According to an embodiment of the present disclosure, training the object resource recommendation network corresponding to the sample object based on the sample support feature subset can include: processing the sample support feature subset related to the sample object based on the difference detection network to obtain an object task difference degree related to the sample object; determining the current local clustering task network parameter according to the object task difference degree related to the sample object and the current meta-network parameter; performing parameter migration according to the current local clustering task network parameter to obtain the object resource recommendation network; processing the sample support feature subset related to the sample object based on the object resource recommendation network to obtain an object task support loss; and training the object resource recommendation network based on the gradient descent algorithm according to the object task support loss.

[0134] According to an embodiment of the present disclosure, the difference detection network can be constructed based on any type of neural network algorithm, for example, can be constructed based on a multi-layer perception algorithm, but is not limited to this, and the difference detection network can also be constructed based on other types of neural network algorithms, and the specific algorithm type for constructing the difference detection network is not limited in the embodiments of the present disclosure.

[0135] According to an embodiment of the present disclosure, the object task difference degree can represent the difference between the object network parameter of the object resource recommendation network and the meta-network parameter of the initial meta-recommendation network.

[0136] According to an embodiment of the present disclosure, the difference detection network can include one or more detection layers, the detection layer can perform difference detection on the hidden features related to the sample support feature subset and the clustering center features of the sample resource features, and the difference detection result obtained can represent the object task difference degree between the sample support feature subset related to the sample object and the global clustering center of the sample features. For each sample object, the difference degree between the current object network parameter and the current meta-network parameter can be represented by the object task difference degree.

[0137] According to an embodiment of the present disclosure, M local clustering objects in N sample objects are associated with the same object task difference degree, N>M>1, and the M local clustering objects correspond to the same local clustering network parameter.

[0138] According to an embodiment of the present disclosure, in a case where M object task difference degrees in N sample objects are the same, or a difference value between the M object task difference degrees is less than or equal to a preset difference degree threshold, a local clustering task network corresponding to the M object task difference degrees can be established in a clustering mapping relationship, and the M sample objects corresponding to the M object task difference degrees can be determined as M local clustering objects.

[0139] According to an embodiment of the present disclosure, determining the current local clustering task network parameter according to the object task difference degree related to the sample object and the current meta network parameter can include processing the object task difference degree and the current meta network parameter based on a fusion function to obtain the current local clustering task network parameter.

[0140] In an example of the present disclosure, the current local clustering task network parameter can be determined based on the following formula (1).

[0141]

[0142] In formula (1), θ k represents the current local clustering task network parameter, o k represents the object task difference degree corresponding to the M local clustering objects, and θ represents the meta network parameter obtained in the last training round, i.e., the current meta network parameter, represents multiplication. θ k represents the current local clustering task network parameter calculated based on formula (1).

[0143] It should be noted that the current meta network parameter can include a meta network parameter obtained through a current round of training operation, for example, a meta network parameter obtained through an i th round of training operation, which can be a meta network parameter updated by performing an (i-1) th round of training operation. The current local clustering task network parameter can be updated in the i th round of training operation, for example, a local clustering task network parameter θ k obtained in the i th round of training operation.

[0144] According to an embodiment of the present disclosure, performing parameter migration according to the current local clustering task network parameter to obtain an object resource recommendation network can include migrating the local clustering task network parameter θ k obtained through the i th round of training operation to the object resource recommendation network corresponding to the M local clustering objects respectively, thereby obtaining M local object resource recommendation networks.

[0145] According to an embodiment of the present disclosure, processing the sample support feature subset related to the sample object by using the local object resource recommendation network to obtain the object task support loss can comprise: for one of the M local clustered objects, processing the sample support feature subset corresponding to the local clustered object by using the local object resource recommendation network corresponding to the local clustered object, so as to obtain the object task support loss corresponding to the local clustered object.

[0146] According to an embodiment of the present disclosure, training the object resource recommendation network based on the object task support loss based on the gradient descent algorithm can comprise: processing the object task support loss corresponding to the local clustered object based on the gradient descent algorithm to obtain the local task gradient corresponding to the local clustered object, and iteratively adjusting the network parameters of the object resource recommendation network corresponding to the local clustered object according to the local task gradient until the object task support loss converges, so as to obtain the trained intermediate object resource recommendation network corresponding to the local clustered object.

[0147] In one example, the network parameters of the object resource recommendation network can be updated based on the following formula (2).

[0148]

[0149] In formula (2), θ k represents the current local clustered task network parameters, and a represents the learning rate related to the local clustered task network, represents the object task support loss, and θ k，u represents the network parameters of the object resource recommendation network obtained by updating in the current round of the object resource recommendation network training operation.

[0150] It should be understood that the object resource recommendation network can be iteratively executed for multiple rounds of object resource recommendation network training operations based on formula (2) until the object task support loss converges.

[0151] According to an embodiment of the present disclosure, processing the sample support feature subset related to the sample object by using the difference degree detection network can comprise: performing feature fusion on the sample support features in the sample support feature subset related to the sample object to obtain a sample object task feature; and inputting the sample object task feature into the first detection layer of the difference degree detection network to output the first difference degree detection feature.

[0152] According to an embodiment of the present disclosure, the feature fusion on the sample support features in the sample support feature subset related to the sample object can be performed based on a fusion algorithm. For example, the feature fusion is performed based on a neural network algorithm such as a pooling network algorithm.

[0153] In one example, the sample support features in the sample support feature subset associated with the sample object can be fused based on the following formula (3).

[0154]

[0155] In formula (3), denotes the sample support feature subset associated with the sample object u, and v denotes the sample support feature, denotes the sample object task feature obtained after feature fusion, so that the sample object task feature can be obtained based on to form a learning task related to the meta-learning mechanism.

[0156] According to an embodiment of the present disclosure, the processing of the sample support feature subset associated with the sample object based on the difference degree detection network can further include: performing a detection layer operation by using the (L+1)th detection layer to process the Lth difference degree detection feature output by the Lth detection layer to obtain the (L+1)th difference degree detection feature, wherein the difference degree detection network includes the Lth detection layer and the (L+1)th detection layer.

[0157] According to an embodiment of the present disclosure, the detection layer operation includes: determining the (L+1)th difference degree between the Lth difference degree detection feature and a detection layer clustering center of the (L+1)th detection layer, the (L+1)th detection layer including K detection layer clustering centers, K≥1; determining a feature update weight of the (L+1)th detection layer according to the K (L+1)th difference degrees; and processing the Lth difference degree detection feature based on a neural network algorithm according to the feature update weight of the (L+1)th detection layer to obtain the (L+1)th difference degree detection feature.

[0158] According to an embodiment of the present disclosure, the difference degree detection feature can represent a hidden feature output by the detection layer, and the detection layer clustering center can represent a clustering center feature of the N sample resource feature sets of the N sample objects, or can also represent a clustering center feature of the N sample support feature subsets of the N sample objects. The detection layer clustering center can be obtained by processing the sample resource features of the N sample resource feature sets based on a clustering algorithm, or can also be trained as a network parameter of the difference degree detection network in the process of training the meta-recommendation network.

[0159] In one example, the difference degree detection network can include a plurality of detection layers, and the Lth difference degree detection feature output by the Lth detection layer is denoted as The (L+1)th detection layer can include K detection layer clustering centers denoted as k l ∈{1,...,K}, and the feature update weight of the (L+1)th detection layer is denoted as The (L+1)th difference degree detection feature can be calculated based on the following formulas (4) and (5)

[0160]

[0161]

[0162] In the formulas (4) and (5), is the network parameter of the L+1th detection layer in the difference detection network. f() can represent the L+1th detection sub-layer in the L+1th detection layer, which is constructed based on a multilayer perceptron algorithm. In this way, the weight distribution between any two detection layers can be normalized using softmax.

[0163] According to an embodiment of the present disclosure, the Lth difference degree detection feature output by the Lth detection layer in the plurality of detection layers may be processed based on an activation function to obtain an object task difference degree o k . For example, the object task difference degree o

[0164]

[0165] In the formula (6), σ() is an activation function. It should be noted that all network parameters in the difference detection network can be represented as

[0166] It should be noted that the English letter "L" involved in the embodiments of the present disclosure can have the same or corresponding meaning as the English lowercase letter "l" in the formula, and the embodiments of the present disclosure will not be repeated here

[0167] Figure 4 A flowchart for processing a sample verification feature subset by using an intermediate object resource recommendation network according to an embodiment of the present disclosure is schematically shown.

[0168] As Figure 4 shown, processing a sample verification feature subset by using an intermediate object resource recommendation network to obtain an object task verification loss can include operations S410 to S450.

[0169] In operation S410, positive sample verification features and negative sample verification features are determined from the sample verification feature subset based on the interaction operation relationship between the sample object and the sample resource.

[0170] In operation S420, a first object fusion feature is determined based on the sample object attribute feature related to the sample object and the positive sample verification feature.

[0171] In operation S430, a second object fusion feature is determined based on the sample object attribute feature and the negative sample verification feature.

[0172] In operation S440, the first object fusion feature and the second object fusion feature are processed by using the intermediate object resource recommendation network to obtain a first predicted object recommendation indicator and a second predicted object recommendation indicator.

[0173] In operation S450, the first predicted object recommendation indicator and the second predicted object recommendation indicator are processed according to a loss function to obtain an object task verification loss.

[0174] According to an embodiment of the present disclosure, the positive sample verification feature can include a sample verification feature having an interactive operation relationship with the sample object, for example, a sample resource feature of a sample resource on which the sample object has performed a click operation in a historical time period or has performed a purchase operation. The negative sample verification feature can include a sample verification feature not having an interactive operation relationship with the sample object, for example, a sample resource feature of a sample resource on which the sample object has not performed a click operation in a historical time period or has not performed a purchase operation.

[0175] In one example, the first object fusion feature can be determined based on the following formulas (7) and (8).

[0176] h in = [u; v] (7);

[0177]

[0178] In formulas (7) and (8), h in is the first object fusion feature, is the second object fusion feature, u represents a sample object attribute feature of the sample object, v represents the positive sample verification feature, and v - represents the negative sample verification feature.

[0179] In one example, the first object fusion feature and the second object fusion feature can be processed based on the following formulas (9) and (10) to obtain the first predicted object recommendation indicator and the second predicted object recommendation indicator.

[0180]

[0181]

[0182] In formulas (9) and (10), h o is the first predicted object recommendation indicator, is the second predicted object recommendation indicator, may represent the intermediate object resource recommendation network, θ ku represents the network parameters of the intermediate object resource recommendation network obtained by the current round of training. Ω can represent the network parameters of the initial encoder.

[0183] According to an embodiment of the present disclosure, by constructing a triple <u, v, v - According to formulas (6) to (11), the resource scores (the first predicted object recommendation indicator and the second predicted object recommendation indicator) of the positive sample resource and the negative sample resource can be obtained. The first predicted object recommendation indicator and the second predicted object recommendation indicator can be processed based on a contrast loss function, and then the object task verification loss is obtained.

[0184] In one example, the following formula (11) can be used.

[0185]

[0186] In formula (11), denotes the object task verification loss, denotes a sample verification feature subset, <u, v, v - is a triple obtained by sampling the sample verification feature subset.

[0187] According to an embodiment of the present disclosure, the object resource recommendation network updated according to the local clustering task network parameter can be used to process the sample support feature subset to obtain the object task support loss in a manner similar to operations S310 to S350. The present disclosure will not be repeated here.

[0188] According to an embodiment of the present disclosure, training the initial meta recommendation network based on the object task verification loss related to each of the N sample objects includes: determining an object task verification gradient related to the sample object based on the object task verification loss; performing gradient fusion on the object task verification gradients corresponding to each of the N sample objects to obtain a fused task verification gradient; and updating the meta network parameters of the initial meta recommendation network according to the fused task verification gradient.

[0189] In one example, the object task verification gradient can be determined by deriving the meta network parameter θ based on the following formula (12).

[0190]

[0191] Correspondingly, the meta network parameter θ can be updated in the training operation based on the following formula (13).

[0192]

[0193] In formula (13), θ i-1 denotes the meta network parameter obtained in the i-1th round of training operation, θ i is the meta network parameter obtained after the i-th round of training operation, and β is the learning rate of the meta network parameter.

[0194] According to an embodiment of the present disclosure, the method for constructing the resource recommendation model can further include: processing the sample verification feature subset related to the sample object according to a local clustering task network to obtain a local clustering task verification loss, wherein the local clustering task network is constructed based on the current local clustering task network parameter; and updating the difference degree network parameter of the difference degree detection network according to the local clustering task verification loss based on a gradient descent algorithm.

[0195] According to an embodiment of the present disclosure, during the i-th round of training operation for the meta recommendation network, the local clustering task network can be constructed based on the object task difference degree and the local clustering task network parameter θ k obtained by updating the meta network parameter of the i-1-th round. k The network parameter in formula (6) to (11) can be updated based on a similar calculation method of formula (6) to (11), so as to obtain the local clustering task verification loss related to the sample object.

[0196] According to an embodiment of the present disclosure, the difference degree network parameter θ can be updated in the training operation based on the following formula (14).

[0197]

[0198] In formula (14), B represents a sample object set of N sample objects.

[0199] According to an embodiment of the present disclosure, the method for constructing the resource recommendation model can further include: updating the network parameter of the initial encoder according to the local clustering task verification loss based on a gradient descent algorithm.

[0200] According to an embodiment of the present disclosure, the network parameter Ω of the initial encoder can be updated in the training operation based on the following formula (15).

[0201]

[0202] According to an embodiment of the present disclosure, in the case of convergence of the meta network parameter θ obtained in the i-th round of training operation, the training operation can be ended, and the trained meta recommendation network can be obtained. Wherein, I≥i>1.

[0203] It should be noted that the meta network parameter θ , the difference degree network parameter θ k , and the network parameter Ω of the initial encoder can be updated only once in each round of training operation. The network parameter θ k，uThe object resource recommendation network training operation is performed for multiple rounds, so that the multiple rounds of object resource recommendation network training operations can be nested into one round of training operations for the meta network parameters, and the gradient information of the multiple object resource recommendation networks is aggregated by the local clustering task network to realize adaptive updating of the meta network parameters.

[0204] Figure 5 A schematic diagram of positive samples and negative samples according to the comparative example of the present disclosure is shown.

[0205] As shown in Figure 5 , the column coordinates in the vertical direction of the interaction matrices 510 and 520 are user identifiers, and the row coordinates in the horizontal direction are sample resources. The matrix elements marked with “x” in the interaction matrices 510 and 520 represent that the sample resource and the user identifier have an interaction operation relationship. The matrix elements marked with “x” can be determined as positive samples, and the matrix elements without “x” can be determined as negative samples.

[0206] As shown in Figure 5 , in the case where the total training data contains 9 users and 10 sample resources, the left interaction matrix 510 is adjusted in row / column to obtain the new right interaction matrix 520. The adjustment does not change the interaction operation relationship between the users and the sample resources. It can be found that there is an obvious local clustering effect (low rank of the interaction matrix) among the nine users, and the nine users lack global shared information. In the scenario of a large number of users / resources, the problem of lacking global shared information is more prominent. Therefore, under the condition of sample data distribution lacking shared resources, the intersection between the users and the sample resources is very small, which leads to that the meta parameters in the meta learning strategy cannot extract effective shared features, and the generalization is poor. And the specific task parameters use the poor meta parameters as initialization, and are trained by relying on few samples, which further aggravates the difficulty and instability of learning.

[0207] Figure 6 A schematic diagram of the principle of updating the meta network parameters according to an embodiment of the present disclosure is shown.

[0208] As shown in Figure 6 , according to the method provided by the embodiment of the present disclosure, for the meta network parameters θ, in one round of training operation, the network parameters θ of the converged object resource recommendation network can be obtained by performing multiple rounds of object resource recommendation network training operations. k，u , for example, the network parameters θ k，u1 , θ k，u2 , θ k，u3 , and θ k，u4 of the four sample objects can be obtained. The network parameters θ k，u1 , θ k，u2 , θ k，u3 , and θ k，u4The corresponding gradient directions 601, 602, 603, and 604 are relatively dispersed. The local clustering task network parameters θ can be updated based on the object task difference degree output by the difference detection network and the meta-network parameters. k1 and θ k2 Furthermore, it aggregates the network parameters θ of multiple object resource recommendation networks based on a local clustering task network. k，u1 θ k，u2 θ k，u3 θ k，u4 The gradient direction is determined, and the aggregated gradient update meta-network parameters are generated based on the object task verification loss output by the local clustering task network. This allows the meta-network parameters to be updated quickly in the direction aggregated with multiple gradient update directions based on the meta-learning mechanism, improving the convergence speed of the meta-network parameters and the prediction accuracy of the meta-recommendation network for recommendation metrics, thereby improving the accuracy and efficiency of subsequent resource push. The method provided in this disclosure is particularly suitable for cold start of resources with few samples, improving the accuracy and efficiency of pushing new resources that have not yet been interacted with by users.

[0209] According to embodiments of this disclosure, constructing a resource recommendation model based on a meta-recommendation network includes: determining target local clustering task network parameters related to the target object using the meta-recommendation network parameters and the object task difference degree related to the target object; performing parameter transfer based on the target local clustering task network parameters to obtain an initial target object resource recommendation network; processing a subset of target support features related to the target object using the initial target object resource recommendation network to obtain a target object task support loss; training the initial target object resource recommendation network based on the gradient descent algorithm and the target object task support loss to obtain a trained target object resource recommendation network; and determining the trained target object resource recommendation network as the resource recommendation model.

[0210] According to embodiments of this disclosure, the resource feature set associated with the target object may include a target support feature subset and a target verification feature subset.

[0211] According to embodiments of this disclosure, based on the method provided in the above embodiments, the trained meta-network parameters θ, the encoder network parameters Ω, and the difference detection network network parameters can be obtained. It can acquire target resources and related target resource feature datasets by analyzing the interactive operations of a target object within a time period close to the current time. An encoder, constructed based on the network parameters Ω, encodes the target resource feature dataset and the target object attributes to obtain target object attribute features and the target resource feature set. A subset of target supporting features can be determined from the target resource feature set. and target verification feature subset Among them, the target verification feature subset It can be the resource characteristics of the candidate resources for which the indicator prediction is to be performed.

[0212] According to embodiments of this disclosure, a pooling algorithm is used to select a subset of target supporting features. Multiple target support features are pooled to obtain fused target object task features. A trained difference detection network is then used to process the fused target object task features to obtain object task difference related to the target object. The meta-recommendation network parameters θ and the object task difference related to the target object are processed according to formula (1). k The network parameters θ for the target local clustering task are obtained. k Based on the network parameters θ of the target local clustering task k Parameter transfer is performed to obtain an initial target object resource recommendation network, and then the target supporting feature subset is used. Positive and negative sample features are generated from the target object features u, and the target object task support loss is calculated based on formulas (6) to (11). The gradient value is obtained by processing the target object task support loss using the gradient descent algorithm. The network parameters θ of the initial target object resource recommendation network are based on the following formula (16). k，u This allows us to obtain the trained target object source recommendation network.

[0213]

[0214] The trained target object resource recommendation network is defined as the resource recommendation model. A subset of target validation features and target object attribute features related to the target object are sampled. These sampled features are input into the trained resource recommendation model, enabling the prediction of recommendation metrics for resources related to the target validation feature subset, thus obtaining the object recommendation metrics. For example, multiple metric values ​​for each resource to be recommended can be obtained. These resources can then be ranked and filtered based on these metric values, and the filtered target resources can be pushed to the target object according to the filtering results.

[0215] The application provides a recommendation method suitable for user item cold start based on small sample learning. The method can utilize the local aggregation effect of users / items in an extremely sparse scenario, efficiently learn meta parameters through hierarchical clustering method, and improve the recommendation performance. The method only needs to modify the loss function of the model, without modifying the online service system, and has low deployment cost. Extensive experiments are conducted on benchmark datasets, and the results show that the method is more accurate than the most advanced small sample learning method, and maintains competitive and stable performance.

[0216] Embodiments of the present disclosure also provide a resource pushing method.

[0217] Figure 7 A flowchart of the resource pushing method according to an embodiment of the present disclosure is schematically shown.

[0218] As shown in Figure 7 , the resource pushing method includes operation S710 to operation S740.

[0219] In operation S710, resource features related to a to-be-recommended resource are acquired.

[0220] In operation S720, the resource features are processed by using a resource recommendation model to obtain an object recommendation index related to the to-be-recommended resource, wherein the resource recommendation model is determined based on the construction method of the resource recommendation model provided by an embodiment of the present disclosure.

[0221] In operation S730, a target resource is determined from the to-be-recommended resource based on the object recommendation index.

[0222] In operation S740, the target resource is recommended to a target object.

[0223] According to an embodiment of the present disclosure, a plurality of resource features related to a to-be-recommended resource can be divided into a target support feature subset and a target verification feature subset. The target verification feature subset related to a target object and a target object attribute feature of the target object are sampled to obtain a sampling feature, which can be input into a trained resource recommendation model, so that a recommendation index prediction of the to-be-recommended resource related to the target verification feature subset can be performed to obtain an object recommendation index.

[0224] For example, the object recommendation index can be obtained based on a plurality of formulas in formula combination (17).

[0225]

[0226] In formula combination (17), h in represents the spliced target object attribute feature u and target verification feature v. h1 to h n represent the hidden layers of the resource recommendation model containing n hidden layers. σ() represents an activation function. θk，u Ω is a network parameter of the resource recommendation model.

[0227] According to the resource pushing method provided by the embodiment of the present disclosure, cold start recommendation can be performed on the to-be-recommended resource with few interaction sample data. In the case that the sample resource features and interaction operation relations are relatively sparse, the local clustering effect of the user / item is utilized, and hierarchical clustering is performed on the sample resource features and the global sample feature clustering center through the difference detection network, so that efficient meta network parameter learning can be realized, and the object recommendation index prediction accuracy of the resource recommendation model is improved, and the recommendation performance of the resource is improved. According to the construction method provided by the embodiment of the present disclosure, for the cold start of the small sample resource to be recommended, only the loss function of the model needs to be modified, and the parameter structure of the model in the online service system does not need to be modified, so that the training efficiency is high, and the model deployment cost is low. After extensive experiments on the obtained training data set, it is shown that the construction method of the resource recommendation model provided by the embodiment of the present disclosure can more accurately recommend the index prediction for small sample learning, maintains the overall operation efficiency of the resource pushing system, and improves the operation stability.

[0228] Figure 8 A block diagram of a construction device of a resource recommendation model according to an embodiment of the present disclosure is schematically shown.

[0229] As shown in Figure 8 , the construction device 800 of the resource recommendation model includes a first obtaining module 810, an intermediate object resource recommendation network obtaining module 820, an object task verification loss obtaining module 830, a meta recommendation network obtaining module 840, and a construction module 850.

[0230] The first obtaining module 810 is configured to obtain a sample resource feature set related to a sample object, the sample resource feature set including a sample support feature subset and a sample verification feature subset, and the sample object including N, N > 1.

[0231] The intermediate object resource recommendation network obtaining module 820 is configured to train an object resource recommendation network corresponding to the sample object based on the sample support feature subset, to obtain an intermediate object resource recommendation network.

[0232] The object task verification loss obtaining module 830 is configured to process the sample verification feature subset by using the intermediate object resource recommendation network, to obtain an object task verification loss.

[0233] The meta recommendation network obtaining module 840 is configured to train an initial meta recommendation network based on the object task verification loss related to each of the N sample objects, to obtain a trained meta recommendation network, wherein the initial meta recommendation network has the same network structure as the object resource recommendation network.

[0234] The construction module 850 is configured to construct a resource recommendation model according to the meta recommendation network, where the resource recommendation model is configured to process resource features related to a target object to obtain an object recommendation index, and the object recommendation index is suitable for determining a target resource to be pushed to the target object.

[0235] According to an embodiment of the present disclosure, the intermediate object resource recommendation network obtaining module 820 comprises an object task difference degree obtaining sub-module, a local clustering task network parameter obtaining sub-module, an object resource recommendation network obtaining sub-module, an object task support loss obtaining sub-module, and an object resource recommendation network training sub-module.

[0236] The object task difference degree obtaining sub-module is configured to process a sample support feature subset related to a sample object based on a difference degree detection network to obtain an object task difference degree related to the sample object, where the object task difference degree represents a difference between object network parameters of the object resource recommendation network and meta network parameters of the initial meta recommendation network.

[0237] The local clustering task network parameter obtaining sub-module is configured to determine current local clustering task network parameters according to the object task difference degree related to the sample object and the current meta network parameters.

[0238] The object resource recommendation network obtaining sub-module is configured to perform parameter migration according to the current local clustering task network parameters to obtain the object resource recommendation network.

[0239] The object task support loss obtaining sub-module is configured to process the sample support feature subset related to the sample object by using the object resource recommendation network to obtain an object task support loss.

[0240] The object resource recommendation network training sub-module is configured to train the object resource recommendation network based on a gradient descent algorithm according to the object task support loss.

[0241] In the N sample objects, M local clustering objects are associated with the same object task difference degree, N>M>1, and the M local clustering objects correspond to the same local clustering network parameters.

[0242] According to an embodiment of the present disclosure, the object task difference degree obtaining sub-module comprises a first obtaining unit.

[0243] The first obtaining unit is configured to perform a detection layer operation on the Lth difference degree detection feature output by the Lth detection layer by using the (L+1)th detection layer to obtain an (L+1)th difference degree detection feature, wherein the difference degree detection network comprises the Lth detection layer and the (L+1)th detection layer, the detection layer operation comprises: determining an (L+1)th difference degree between the Lth difference degree detection feature and a detection layer clustering center of the (L+1)th detection layer, the (L+1)th detection layer comprises K detection layer clustering centers, and K is greater than or equal to 1; determining a feature update weight of the (L+1)th detection layer according to the K (L+1)th difference degrees; and processing the Lth difference degree detection feature based on a neural network algorithm and according to the feature update weight of the (L+1)th detection layer to obtain the (L+1)th difference degree detection feature.

[0244] According to an embodiment of the present disclosure, the object task difference degree obtaining submodule further comprises a second obtaining unit and a first output unit.

[0245] The second obtaining unit is configured to perform feature fusion on the sample support features in the sample support feature subset related to the sample object to obtain a sample object task feature.

[0246] The first output unit is configured to input the sample object task feature into a first detection layer of the difference degree detection network and output a first difference degree detection feature.

[0247] According to an embodiment of the present disclosure, the meta recommendation network obtaining module 840 comprises an object task verification gradient determining submodule, a fused task verification gradient obtaining submodule, and a second updating submodule.

[0248] The object task verification gradient determining submodule is configured to determine an object task verification gradient related to the sample object based on the object task verification loss.

[0249] The fused task verification gradient obtaining submodule is configured to perform gradient fusion on the object task verification gradients corresponding to the N sample objects respectively to obtain a fused task verification gradient.

[0250] The second updating submodule is configured to update the meta network parameters of the initial meta recommendation network according to the fused task verification gradient.

[0251] According to an embodiment of the present disclosure, the resource recommendation model construction apparatus further comprises a local clustering task verification loss obtaining module and a difference degree network parameter updating module.

[0252] The local clustering task verification loss obtaining module is configured to obtain a local clustering task verification loss by processing a sample verification feature subset related to a sample object according to a local clustering task network, wherein the local clustering task network is constructed based on current local clustering task network parameters.

[0253] The difference degree network parameter updating module is configured to update the difference degree network parameters of the difference degree detection network based on a gradient descent algorithm according to a local clustering task validation loss.

[0254] According to an embodiment of the present disclosure, the sample support features in the sample support feature subset are obtained by feature extraction on sample support data by using the initial encoder, and the sample support data is related to the sample object.

[0255] The resource recommendation model construction apparatus further includes a second updating module.

[0256] The second updating module is configured to update the network parameters of the initial encoder based on a gradient descent algorithm according to a local clustering task validation loss.

[0257] According to an embodiment of the present disclosure, the object task validation loss obtaining module 830 includes:

[0258] The feature determining submodule is configured to determine the positive sample verification feature and the negative sample verification feature from the sample verification feature subset based on the interactive operation relationship between the sample object and the sample resource.

[0259] The first object fusion feature determining submodule is configured to determine the first object fusion feature based on the sample object attribute feature related to the sample object and the positive sample verification feature.

[0260] The second object fusion feature determining submodule is configured to determine the second object fusion feature based on the sample object attribute feature and the negative sample verification feature.

[0261] The object recommendation indicator determining submodule is configured to process the first object fusion feature and the second object fusion feature by using the intermediate object resource recommendation network to obtain the first predicted object recommendation indicator and the second predicted object recommendation indicator.

[0262] The object task validation loss submodule is configured to process the first predicted object recommendation indicator and the second predicted object recommendation indicator according to a loss function to obtain the object task validation loss.

[0263] According to an embodiment of the present disclosure, the construction module 850 includes:

[0264] The target local clustering task network parameter determining submodule is configured to determine the target local clustering task network parameters related to the target object by using the meta recommendation network parameters of the meta recommendation network and the object task difference degree related to the target object.

[0265] The initial target object resource recommendation network determining submodule is configured to perform parameter migration according to the target local clustering task network parameters to obtain the initial target object resource recommendation network.

[0266] The target object task support loss determination sub-module is configured to process the target support feature subset related to the target object by using the initial target object resource recommendation network, to obtain a target object task support loss.

[0267] The target object resource recommendation network training sub-module is configured to train the initial target object resource recommendation network according to the target object task support loss based on a gradient descent algorithm, to obtain a trained target object resource recommendation network.

[0268] The resource recommendation model determination sub-module is configured to determine the trained target object resource recommendation network as a resource recommendation model.

[0269] Figure 9 A block diagram of a construction apparatus of a resource recommendation model according to an embodiment of the present disclosure is schematically shown.

[0270] As shown in Figure 9 The resource pushing apparatus 900 includes a second acquisition module 910, an object recommendation index obtaining module 920, a target resource determination module 930, and a pushing module 940.

[0271] The second acquisition module 910 is configured to acquire resource features related to a resource to be recommended.

[0272] The object recommendation index obtaining module 920 is configured to process the resource features by using a resource recommendation model to obtain an object recommendation index related to the resource to be recommended, wherein the resource recommendation model is determined based on the construction method of the resource recommendation model provided in the embodiments of the present disclosure.

[0273] The target resource determination module 930 is configured to determine a target resource from the resource to be recommended based on the object recommendation index.

[0274] The pushing module 940 is configured to recommend the target resource to a target object.

[0275] Any of the modules, sub-modules, units, sub-units according to the embodiments of the present disclosure, or at least part of any of them, can be implemented in one module. Any of the modules, sub-modules, units, sub-units according to the embodiments of the present disclosure can be split into multiple modules for implementation. Any of the modules, sub-modules, units, sub-units according to the embodiments of the present disclosure can be implemented at least in part as a hardware circuit, such as a field-programmable gate array (FPGA), a programmable logic array (PLA), a system on chip, a system on board, a system on package, an application-specific integrated circuit (ASIC), or any other reasonable manner of hardware or firmware by integrating or packaging the circuit, or in software, hardware, and firmware in any one of them or in a proper combination of any of them. Alternatively, one or more of the modules, sub-modules, units, sub-units according to the embodiments of the present disclosure can be implemented at least in part as computer program modules that can perform corresponding functions when the computer program modules are run.

[0276] For example, any multiple of the first obtaining module 810, the intermediate object resource recommendation network obtaining module 820, the object task verification loss obtaining module 830, the meta recommendation network obtaining module 840 and the constructing module 850, or any multiple of the second obtaining module 910, the object recommendation index obtaining module 920, the target resource determining module 930 and the pushing module 940 can be combined in one module / unit / sub-unit for implementation, or any one of the modules / units / sub-units can be split into multiple modules / units / sub-units. Alternatively, at least part of the functions of one or more of the modules / units / sub-units can be combined with at least part of the functions of other modules / units / sub-units, and implemented in one module / unit / sub-unit. According to an embodiment of the present disclosure, at least one of the first obtaining module 810, the intermediate object resource recommendation network obtaining module 820, the object task verification loss obtaining module 830, the meta recommendation network obtaining module 840 and the constructing module 850, or at least one of the second obtaining module 910, the object recommendation index obtaining module 920, the target resource determining module 930 and the pushing module 940 can be at least partially implemented as a hardware circuit, such as a field programmable gate array (FPGA), a programmable logic array (PLA), a system on chip, a system on board, a system on package, an application specific integrated circuit (ASIC), or any other reasonable manner of integrating or packaging a circuit, etc. hardware or firmware, or in any one of software, hardware and firmware implementation or in a proper combination of any of them. Alternatively, at least one of the first obtaining module 810, the intermediate object resource recommendation network obtaining module 820, the object task verification loss obtaining module 830, the meta recommendation network obtaining module 840 and the constructing module 850, or at least one of the second obtaining module 910, the object recommendation index obtaining module 920, the target resource determining module 930 and the pushing module 940 can be at least partially implemented as a computer program module which can perform corresponding functions when the computer program module is run.

[0277] It should be noted that the resource recommendation model construction method part in the embodiments of the present disclosure corresponds to the resource recommendation model construction device part in the embodiments of the present disclosure, and the description of the resource recommendation model construction device part is specifically referred to the resource recommendation model construction method part, which will not be repeated here.

[0278] It should be noted that the resource pushing method part in the embodiments of the present disclosure corresponds to the resource pushing device part in the embodiments of the present disclosure, and the description of the resource pushing device part is specifically referred to the resource pushing method part, which will not be repeated here.

[0279] Figure 10The diagram illustrates an electronic device suitable for implementing a method for constructing a resource recommendation model and a resource push method according to embodiments of the present disclosure. Figure 10 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments disclosed herein.

[0280] like Figure 10 As shown, an electronic device 1000 according to an embodiment of the present disclosure includes a processor 1001, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage portion 1008 into a random access memory (RAM) 1003. The processor 1001 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 1001 may also include onboard memory for caching purposes. The processor 1001 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of the present disclosure.

[0281] RAM 1003 stores various programs and data required for the operation of electronic device 1000. Processor 1001, ROM 1002, and RAM 1003 are interconnected via bus 1004. Processor 1001 performs various operations of the method flow according to embodiments of the present disclosure by executing programs in ROM 1002 and / or RAM 1003. It should be noted that the programs may also be stored in one or more memories other than ROM 1002 and RAM 1003. Processor 1001 may also perform various operations of the method flow according to embodiments of the present disclosure by executing programs stored in said one or more memories.

[0282] According to an embodiment of the present disclosure, the electronic device 1000 can further include an input / output (I / O) interface 1005 that is also connected to the bus 1004. The system 1000 can further include one or more of the following components connected to the input / output (I / O) interface 1005: an input part 1006 including, for example, a keyboard and a mouse; an output part 1007 including, for example, a cathode ray tube (CRT), a liquid crystal display (LCD), and a speaker; a storage part 1008 including, for example, a hard disk; and a communication part 1009 including, for example, a LAN card, a modem, and the like. The communication part 1009 performs communication processing via a network such as the Internet. A drive 1010 is also connected to the input / output (I / O) interface 1005 as necessary. A removable medium 1011 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, and the like is mounted on the drive 1010 as necessary, so that a computer program read therefrom is installed in the storage part 1008 as necessary.

[0283] According to an embodiment of the present disclosure, the method flow according to the embodiment of the present disclosure can be implemented as a computer software program. For example, the embodiment of the present disclosure includes a computer program product including a computer program carried on a computer-readable storage medium, the computer program containing program codes for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network by the communication part 1009, and / or installed from the removable medium 1011. When the computer program is executed by the processor 1001, the above-described functions defined in the system implementing the embodiment of the present disclosure are performed. According to an embodiment of the present disclosure, the system, device, apparatus, module, unit, and the like described above can be implemented by computer program modules.

[0284] The present disclosure also provides a computer-readable storage medium, which can be included in the device / apparatus / system described in the above embodiments; or can exist separately without being assembled into the device / apparatus / system. The above computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiment of the present disclosure.

[0285] According to an embodiment of the present disclosure, the computer readable storage medium can be a non-volatile computer readable storage medium. For example, it can include, but is not limited to, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing. In the present disclosure, a computer readable storage medium can be any tangible medium that contains or stores a program, which can be used by or in connection with an instruction execution system, apparatus, or device.

[0286] For example, according to an embodiment of the present disclosure, the computer readable storage medium can include one or more memories of the ROM 1002 and / or the RAM 1003 described above and / or other than the ROM 1002 and the RAM 1003.

[0287] Embodiments of the present disclosure also include a computer program product, which includes a computer program containing program codes for executing the methods provided by the embodiments of the present disclosure, and when the computer program product is run on an electronic device, the program codes are used to make the electronic device implement the methods provided by the embodiments of the present disclosure.

[0288] When the computer program is executed by the processor 1001, the above-mentioned functions defined in the system / apparatus of the embodiments of the present disclosure are performed. According to an embodiment of the present disclosure, the system, apparatus, module, unit, etc. described above can be implemented by computer program modules.

[0289] In one embodiment, the computer program can rely on a tangible storage medium such as an optical storage device, a magnetic storage device, etc. In another embodiment, the computer program can also be transmitted, distributed, and downloaded in the form of a signal over a network medium, and be downloaded and installed through the communication part 1009 and / or installed from the detachable medium 1011. The program codes contained in the computer program can be transmitted by any appropriate network medium, including but not limited to wireless, wired, etc., or any suitable combination of the foregoing.

[0290] According to embodiments of the present disclosure, program code of a computer program provided by embodiments of the present disclosure can be written in any combination of one or more programming languages, and specifically, can be implemented using a high-level procedural and / or object-oriented programming language, and / or an assembly / machine language. Programming languages include, but are not limited to, Java, C++, python, "C" language, or similar programming languages. Program code can execute entirely on a user's computing device, partly on a user device, partly on a remote computing device, or entirely on a remote computing device or server. In the latter scenario, the remote computing device can be connected to the user's computing device through any kind of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computing device, such as through the Internet using an Internet Service Provider.

[0291] The flow diagrams and the block diagrams in the drawings are illustrations of architectures, functionalities, and operations of possible implementations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in the flow diagrams or block diagrams can represent a module, a segment, or a portion of code, which comprises one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions noted in the blocks can occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and / or flow diagrams, and combinations thereof, can be implemented by special purpose hardware-based systems that perform the specified functions or operations, or combinations of special purpose hardware and computer instructions.

[0292] Embodiments of the present disclosure have been described. However, these embodiments are merely intended to illustrate the present disclosure, and are not intended to limit the scope of the present disclosure. Although each of the embodiments is described above separately, this does not mean that the measures in each of the embodiments cannot be used advantageously in combination. The scope of the present disclosure is defined by the appended claims and their equivalents. Without departing from the scope of the present disclosure, those skilled in the art can make various substitutions and modifications, and these substitutions and modifications should all fall within the scope of the present disclosure.

Claims

1. A method for constructing a resource recommendation model, comprising: Obtain a sample resource feature set related to the sample objects. The sample resource feature set includes a sample support feature subset and a sample verification feature subset. The sample objects include N, where N>1. Based on the sample supporting feature subset, train an object resource recommendation network corresponding to the sample object to obtain an intermediate object resource recommendation network; The intermediate object resource recommendation network is used to process the sample verification feature subset to obtain the object task verification loss. Based on the object task verification loss associated with each of the N sample objects, an initial meta-recommendation network is trained to obtain a trained meta-recommendation network, wherein the initial meta-recommendation network has the same network structure as the object resource recommendation network; and A resource recommendation model is constructed based on the meta-recommendation network, wherein the resource recommendation model is used to process resource features related to the target object to obtain object recommendation indicators, and the object recommendation indicators are suitable for determining the target resources to be pushed to the target object.

2. The method according to claim 1, wherein, Training an object resource recommendation network corresponding to the sample objects based on the sample supporting feature subset includes: Based on the difference detection network, a subset of sample support features related to the sample object is processed to obtain the object task difference degree related to the sample object. The object task difference degree characterizes the difference between the object network parameters of the object resource recommendation network and the meta network parameters of the initial meta recommendation network. Based on the object task difference degree associated with the sample object and the current meta-network parameters, determine the current local clustering task network parameters; Based on the current local clustering task network parameters, parameter migration is performed to obtain the object resource recommendation network; The object resource recommendation network is used to process a subset of sample support features related to the sample object to obtain the object task support loss. The object resource recommendation network is trained based on the gradient descent algorithm and the object task support loss. Among the N sample objects, M local cluster objects are associated with the same object task difference degree, N>M>1, and the M local cluster objects correspond to the same local clustering network parameters.

3. The method according to claim 2, wherein, The difference detection network includes an Lth detection layer and an L+1th detection layer; The subset of sample supporting features related to the sample object processed by the difference detection network includes: The L+1th detection layer performs the following detection layer operations to process the Lth difference detection feature output by the Lth detection layer, thereby obtaining the L+1th difference detection feature. The detection layer operations include: Determine the (L+1)th difference between the Lth difference detection feature and the cluster center of the (L+1)th detection layer, wherein the (L+1)th detection layer includes K cluster centers of the detection layer, where K≥1; Based on the K L+1th difference values, determine the feature update weights of the L+1th detection layer; and Based on the neural network algorithm, the Lth difference detection feature is processed by updating the weights according to the features of the (L+1)th detection layer to obtain the (L+1)th difference detection feature.

4. The method according to claim 2, wherein, The processing of the sample support feature subset related to the sample object based on the difference detection network also includes: Feature fusion is performed on the sample support features in the subset of sample support features related to the sample object to obtain the sample object task features; and The task features of the sample object are input into the first detection layer of the difference detection network, and the first difference detection feature is output.

5. The method according to claim 2, wherein, Based on the object task verification loss associated with each of the N sample objects, the initial meta-recommendation network is trained as follows: Based on the object task verification loss, determine the object task verification gradient related to the sample object; Gradient fusion is performed on the object task verification gradients corresponding to each of the N sample objects to obtain the fused task verification gradient; and The meta-network parameters of the initial meta-recommendation network are updated based on the gradient verification of the fusion task.

6. The method according to claim 5, further comprising: The local clustering task verification loss is obtained by processing a subset of sample verification features related to the sample object using the local clustering task network, wherein the local clustering task network is constructed based on the current local clustering task network parameters; Based on the gradient descent algorithm, the difference network parameters of the difference detection network are updated according to the verification loss of the local clustering task.

7. The method according to claim 6, wherein, The sample support features in the sample support feature subset are obtained by feature extraction from the sample support data using the initial encoder, and the sample support data is related to the sample object; The method further includes: Based on the gradient descent algorithm, the network parameters of the initial encoder are updated according to the verification loss of the local clustering task.

8. The method according to claim 1, wherein, The intermediate object resource recommendation network is used to process the sample verification feature subset to obtain the object task verification loss, which includes: Based on the interaction relationship between the sample object and the sample resource, positive sample verification features and negative sample verification features are determined from the sample verification feature subset. Based on the sample object attribute features associated with the sample object and the positive sample verification features, the first object fusion feature is determined; Based on the sample object attribute features and the negative sample verification features, the second object fusion features are determined; The intermediate object resource recommendation network is used to process the first object fusion feature and the second object fusion feature to obtain a first predicted object recommendation index and a second predicted object recommendation index; and The object task verification loss is obtained by processing the first prediction object recommendation index and the second prediction object recommendation index according to the loss function.

9. The method according to claim 1, wherein, Constructing a resource recommendation model based on the meta-recommendation network includes: Using the meta-recommendation network parameters of the meta-recommendation network and the object task difference degree related to the target object, the target local clustering task network parameters related to the target object are determined; Based on the network parameters of the target local clustering task, parameter migration is performed to obtain the initial target object resource recommendation network; The target object resource recommendation network is used to process a subset of target support features related to the target object to obtain the target object task support loss; Based on the gradient descent algorithm, the initial target object resource recommendation network is trained according to the target object task support loss, resulting in the trained target object resource recommendation network; and The trained target object resource recommendation network is determined as the resource recommendation model.

10. A resource push method, comprising: Obtain resource characteristics related to the resources to be recommended; The resource features are processed using a resource recommendation model to obtain object recommendation indicators related to the resource to be recommended, wherein the resource recommendation model is determined based on the method of any one of claims 1 to 9; Based on the object recommendation metrics, target resources are determined from the resources to be recommended; and The target resource is recommended to the target object.

11. An apparatus for constructing a resource recommendation model, comprising: The first acquisition module is used to acquire a sample resource feature set related to the sample objects. The sample resource feature set includes a sample support feature subset and a sample verification feature subset. The sample objects include N, where N>1. The intermediate object resource recommendation network acquisition module is used to train an object resource recommendation network corresponding to the sample object based on the sample supporting feature subset, and obtain the intermediate object resource recommendation network. The object task verification loss acquisition module is used to process the sample verification feature subset using the intermediate object resource recommendation network to obtain the object task verification loss. A meta-recommendation network acquisition module is used to train an initial meta-recommendation network based on the object task verification loss associated with each of the N sample objects, thereby obtaining a trained meta-recommendation network, wherein the initial meta-recommendation network has the same network structure as the object resource recommendation network; and A construction module is used to construct a resource recommendation model based on the meta-recommendation network. The resource recommendation model is used to process resource features related to the target object to obtain object recommendation indicators. The object recommendation indicators are suitable for determining the target resources to be pushed to the target object.

12. A resource delivery device, comprising: The second acquisition module is used to acquire resource features related to the resources to be recommended. An object recommendation index acquisition module is used to process the resource features using a resource recommendation model to obtain object recommendation indices related to the resource to be recommended, wherein the resource recommendation model is determined based on the method of any one of claims 1 to 9; The target resource determination module is used to determine target resources from the resources to be recommended based on the object recommendation metrics; and The push module is used to recommend the target resource to the target object.

13. An electronic device, comprising: One or more processors; Memory, used to store one or more programs. Wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the method of any one of claims 1 to 10.

14. A computer-readable storage medium having stored thereon executable instructions that, when executed by a processor, cause the processor to perform the method of any one of claims 1 to 10.

15. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1 to 10.

Images