Recall processing method and device, electronic equipment and storage medium

Abstract

The application discloses a recall processing method and device, electronic equipment and a storage medium. The method comprises the following steps: fusing a social feature of a target object, an object feature of the target object, an object feature of a lost friend, and an interaction feature between the target object and the lost friend to obtain a fused feature; predicting, by a gate network, weights of a plurality of expert networks in an invitation prediction task and an invitation acceptance prediction task according to the fused feature; fusing expert processing features output by the plurality of expert networks in the invitation prediction task according to the weights in the invitation prediction task to obtain a first target feature; fusing expert processing features output by the plurality of expert networks in the invitation acceptance prediction task according to the weights in the invitation acceptance prediction task to obtain a second target feature; and performing recall processing on the lost friend according to the first target feature and the second target feature. The application can guarantee the accuracy of the recall processing on the lost friend and improve the recall effect of the lost friend.

Description

Technical Field

[0001] This application relates to the field of computer technology, and more specifically, to a recall processing method, apparatus, electronic device, and storage medium. Background Technology

[0002] In gaming applications, it's common practice to push notifications to active players about their lost friends, allowing active players to invite those lost friends back to participate in the game or its activities. However, this technology is not very effective in recalling lost friends because the process of active players inviting lost friends requires communication between the server and the player, which wastes server communication resources. Summary of the Invention

[0003] In view of the above problems, this application proposes a recall processing method, apparatus, electronic device and storage medium to solve the problem of poor recall effect of lost friends in related technologies.

[0004] Firstly, a recall processing method is provided, comprising: fusing the social characteristics of a target object, the object characteristics of the target object, the object characteristics of the target object's churned friends, and the interaction characteristics between the target object and the churned friends to obtain a fused feature; using a gating network to predict the weights of multiple expert networks on an invitation prediction task and an invitation acceptance prediction task, respectively, based on the fused feature; fusing the expert processing features output by the multiple expert networks according to their respective weights on the invitation prediction task to obtain a first target feature; the expert processing feature is obtained by processing the fused feature by the corresponding expert network; fusing the expert processing features output by the multiple expert networks according to their respective weights on the invitation acceptance prediction task to obtain a second target feature; and performing recall processing on the churned friends based on the first target feature and the second target feature.

[0005] Secondly, a recall processing device is provided, comprising: a first fusion module, used to fuse the social characteristics of a target object, the object characteristics of the target object, the object characteristics of the target object's lost friends, and the interaction characteristics between the target object and the lost friends to obtain fused features; a task weight prediction module, used by a gating network to predict the weights of multiple expert networks on invitation prediction tasks and invitation acceptance prediction tasks respectively, based on the fused features; a second fusion module, used to fuse the expert processing features output by the multiple expert networks respectively according to their respective weights on the invitation prediction task to obtain a first target feature; the expert processing feature is obtained by the corresponding expert network processing the fused features; a third fusion module, used to fuse the expert processing features output by the multiple expert networks respectively according to their respective weights on the invitation acceptance prediction task to obtain a second target feature; and a recall processing module, used to perform recall processing on the lost friends based on the first target feature and the second target feature.

[0006] In some embodiments, the recall processing module includes: a first probability mapping unit, configured to perform probability mapping on the first target feature to obtain a first probability that the target object invites the churned friend back; a second probability mapping unit, configured to perform probability mapping on the second target feature to obtain a second probability that the churned friend accepts the return invitation from the target object; and a recall processing unit, configured to perform recall processing on the churned friend based on the first probability and the second probability.

[0007] In some embodiments, the social features of the target object are obtained by aggregating the object features of multiple friends of the target object in the target friend social graph, and the multiple friends include the churned friends; the recall processing unit includes: a target probability determination unit, used to multiply the first probability and the second probability to obtain a target probability; an update unit, used to update the social features according to the target probability to obtain target social features; a re-prediction unit, used to re-predict a third probability that the target object invites the churned friends back, and a fourth probability that the churned friends accept the return invitation from the target object, based on the target social features, the object features of the target object, the object features of the churned friends, and the interaction features between the target object and the churned friends; a target recall weight determination unit, used to use the product of the third probability and the fourth probability as the target recall weight of the churned friends relative to the target object; and a processing unit, used to perform recall processing on the churned friends according to the recall weight of the churned friends relative to the target object.

[0008] In some embodiments, the social features include a first social feature; the first social feature is obtained by averaging the object features of multiple friends of the target object in the target friend social graph; the update unit is configured to: weight the object features of multiple friends of the target object in the target friend social graph according to the target probability and a preset weight to obtain a second social feature; wherein the target probability is used as the weighting coefficient of the object features of the churned friends, and the preset weight is used as the weighting coefficient of the object features of the non-churned friends among the multiple friends; update the first social feature in the social features to the second social feature to obtain the target social feature.

[0009] In some embodiments, the social features further include a third social feature; the third social feature is obtained by aggregating the object features of multiple friends of the target object in the target friend social graph according to the social relevance between each friend and the target object.

[0010] In some embodiments, the re-prediction unit is configured to: fuse the target social features, the object features of the target object, the object features of the churned friend, and the interaction features between the target object and the churned friend to obtain reference fused features; have a gating network predict reference weights for multiple expert networks on the invitation prediction task and the acceptance invitation prediction task, respectively, based on the reference fused features; have the multiple expert networks process the reference fused features to obtain reference expert processing features output by each expert network; fuse the reference expert processing features output by the multiple experts according to their respective reference weights on the invitation prediction task to obtain a third target feature; fuse the reference expert processing features output by the multiple experts according to their respective reference weights on the acceptance invitation prediction task to obtain a fourth target feature; perform probability mapping on the third target feature to obtain a third probability that the target object invites the churned friend back; and perform probability mapping on the fourth target feature to obtain a fourth probability that the churned friend accepts the return invitation from the target object.

[0011] In some embodiments, the processing unit is configured to: sort multiple churned friends of the target object in descending order of target recall weight to obtain a churned friend ranking; add the top N churned friends in the churned friend ranking to the churned friend recommendation list of the target object; where N is a positive integer; and send the churned friend recommendation list to the client where the target object is located.

[0012] In some embodiments, the first fusion module includes: a splicing unit, configured to splice the social features of the target object, the object features of the target object, the object features of the target object's churned friends, and the interaction features between the target object and the churned friends to obtain spliced ​​features; a cross-processing unit, configured to perform multi-head attention cross-processing on the spliced ​​features to obtain cross-fusion features; and a residual processing unit, configured to perform residual processing on the spliced ​​features and the cross-fusion features to obtain the fused features.

[0013] In other embodiments, the recall processing module includes: a reference recall weight determination unit, configured to multiply the first probability and the second probability to obtain a reference recall weight of the churned friend relative to the target object; a reference ranking determination unit, configured to rank multiple churned friends of the target object in descending order of reference recall weight to obtain a reference ranking; a churned friend recommendation list determination unit, configured to add the top N churned friends in the reference ranking to the churned friend recommendation list of the target object; where N is a positive integer; and a sending unit, configured to send the churned friend recommendation list to the client where the target object is located.

[0014] In some embodiments, the recall processing apparatus further includes a training module, configured to: acquire multiple training samples; wherein, a training sample includes a sample object, a reference sample object, a first label, and a second label, the first label indicating whether the sample object invites the reference sample object to return, and the second label indicating whether the reference sample object accepts the return invitation from the sample object; fuse the social features of the sample object, the object features of the sample object, the object features of the reference sample object, and the interaction features between the sample object and the reference sample object to obtain sample fusion features; the gating network predicts the weights of the multiple expert networks on the invitation prediction task and the acceptance prediction task respectively based on the sample fusion features; and according to the respective invitation... The weights for the prediction task are used to fuse the sample expert features output by the multiple expert networks according to the sample fusion features, resulting in a first sample target feature. The weights for the invitation acceptance prediction task are then used to fuse the sample expert features output by the multiple expert networks, resulting in a second sample target feature. Based on the first sample target feature, a first sample probability is predicted that the sample object will invite the reference sample object back. Based on the second sample target feature, a second sample probability is predicted that the reference sample object will accept the sample object's back invitation. A target loss is calculated based on the first sample probability, the first label, the second sample probability, and the second label. Based on the target loss, the parameters of at least the gating network and the multiple expert networks are adjusted until a first training termination condition is met.

[0015] In some embodiments, the recall processing apparatus further includes: a target friend social graph extraction module, used to acquire a target friend social graph constructed based on the social relationships between objects; a graph encoding processing module, used to perform graph encoding processing on the target friend social graph using a graph neural network to obtain a graph encoding result; the graph encoding result includes object features of objects represented by each node in the target friend social graph; a feature aggregation module, used to aggregate the object features of multiple friends representing the target object in the graph encoding result to obtain the social features of the target object; and an object feature acquisition module, used to acquire the object features of the target object from the graph encoding result, and acquire the object features of the target object's lost friends.

[0016] Thirdly, an electronic device is provided, comprising: a processor; and a memory storing computer-readable instructions, wherein when the computer-readable instructions are executed by the processor, the recall processing method described above is implemented.

[0017] Fourthly, a computer-readable storage medium is provided, on which computer-readable instructions are stored, which, when executed by a processor, implement the recall processing method described above.

[0018] Fifthly, a computer program product is provided, including computer instructions, which, when executed by a processor, implement the recall processing method described above.

[0019] Considering the inherent connections between friends in social networks, and the potential influence of an object's actions on its friends' actions, this application not only utilizes the object characteristics of the target object, the object characteristics of the target object's churned friends, and the interaction characteristics between the target object and its churned friends for the invitation prediction and acceptance prediction tasks for the target object's churned friends, but also introduces the target object's social characteristics. These social characteristics reflect the target object's social topology and friend distribution information. Therefore, by combining these four aspects, features reflecting the social links between the target object and its churned friends can be fully extracted to improve the accuracy of the invitation prediction and acceptance prediction tasks, thereby ensuring the accuracy of the recall process for the target object's churned friends. This can improve the recall effect of churned friends and reduce the waste of communication resources during the recall process.

[0020] Furthermore, considering the correlation between the invitation prediction task and the invitation acceptance prediction task, and the commonalities or correlations among the features applicable to both tasks, multiple expert networks are shared. A gating network is used to predict the weights of these expert networks on both tasks. Then, based on the weights of each expert network on the invitation prediction task, the expert-processed features output by each network are fused to obtain a first target feature suitable for the invitation prediction task. Similarly, based on the weights of each network on the invitation acceptance prediction task, the expert-processed features output by each network are fused to obtain a second target feature suitable for the invitation acceptance prediction task. This approach fully utilizes the features extracted by different expert networks while also considering the adaptability of the output features to both the invitation and acceptance prediction tasks, ensuring the accuracy of both the first and second target features and thus improving the accuracy of subsequent friend recall processing. It also eliminates the need to deploy multiple expert networks separately for each task, reducing computational overhead. Attached Figure Description

[0021] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0022] Figure 1 This is a schematic diagram illustrating an application scenario of this application according to an embodiment of this application.

[0023] Figure 2 This is a flowchart illustrating a recall processing method according to an embodiment of this application.

[0024] Figure 3 An example diagram of a target friend social graph is shown.

[0025] Figure 4 This is a flowchart illustrating a recall processing method according to another embodiment of this application.

[0026] Figure 5 An example diagram illustrates the spliced ​​feature obtained by splicing four features together.

[0027] Figure 6 This is a schematic diagram of a display interface according to an embodiment of this application.

[0028] Figure 7 This is a flowchart illustrating step 250 according to an embodiment of this application.

[0029] Figure 8 This is a flowchart illustrating step 730 according to an embodiment of this application.

[0030] Figure 9 This is a flowchart illustrating step 250 according to another embodiment of this application.

[0031] Figure 10 This is a schematic diagram illustrating the comparison between the social characteristics of the target object and the target social characteristics according to an embodiment of this application.

[0032] Figure 11 This is a flowchart illustrating a recall processing method according to an embodiment of this application.

[0033] Figure 12 This is a flowchart illustrating a recall processing method according to another embodiment of this application.

[0034] Figure 13 This is a block diagram of a recall processing apparatus according to an embodiment of this application.

[0035] Figure 14A schematic diagram of the structure of a computer system suitable for implementing the electronic device of the present application is shown. Detailed Implementation

[0036] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0037] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present application without creative effort are within the scope of protection of the present application.

[0038] In the following description, the terms "first" and "second" are used merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first" and "second" may be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0039] In this document, "multiple" refers to two or more. "And / or" describes the relationship between associated objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following associated objects are in an "or" relationship. In the following description, references to "some embodiments or some embodiment methods" describe a subset of all possible embodiments. However, it is understood that "some embodiments" can be the same subset or different subsets of all possible embodiments and can be combined with each other without conflict.

[0040] Figure 1 This is a schematic diagram illustrating an application scenario of this application according to an embodiment of this application. For example... Figure 1 As shown, the application scenario includes a terminal 110 and a server 120 that is communicatively connected to the terminal 110. The server 120 can be used to execute the method provided in this application. The terminal 110 can run an application client. The application is not limited and can be a reading application, a music application, a blog application, a game application, etc.

[0041] like Figure 1As shown, server 120 can fuse the social characteristics of the target object, the object characteristics of the target object, the object characteristics of the target object's churned friends, and the interaction characteristics between the target object and churned friends to obtain fused features. Then, a gating network predicts the weights of multiple expert networks on the invitation prediction task and the invitation acceptance prediction task based on the fused features. Next, according to their respective weights on the invitation prediction task, the expert-processed features output by the multiple expert networks are fused to obtain the first target feature; the expert-processed features are obtained by processing the fused features by the corresponding expert networks. Finally, according to their respective weights on the invitation acceptance prediction task, the expert-processed features output by the multiple expert networks are fused to obtain the second target feature. Based on the first and second target features, churned friends are recalled. That is, based on the first and second target features, a decision is made on whether to add churned friends to the target object's churned friend recommendation list, thereby determining the target object's churned friend recommendation list.

[0042] After determining the list of lost friends of the target object, the server 120 can send the list of lost friends of the target object to the terminal 110 where the target object is located. Correspondingly, the terminal 110 where the target object is located can display the lost friends in the list of lost friends of the target object. In this way, users on the target object's side can invite the displayed lost friends to come back.

[0043] Terminal 110 can be a smartphone, tablet, laptop, desktop computer, wearable device, smart TV, smart home device, vehicle terminal, etc., without specific limitations.

[0044] Server 120 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0045] The implementation details of the technical solutions in the embodiments of this application are described in detail below:

[0046] Figure 2 This is a flowchart illustrating a recall processing method according to an embodiment of this application. The method can be executed by an electronic device, which may be a terminal, server, etc. (Refer to...) Figure 2 As shown, the method includes at least steps 210 to 250, which are described in detail below:

[0047] Step 210: The social features of the target object, the object features of the target object, the object features of the target object's lost friends, and the interaction features between the target object and the lost friends are fused to obtain fused features.

[0048] The target audience refers to the individuals to be recommended to as lost friends. This target audience can be registered users within an application, which can be a game, shopping app, instant messaging app, live streaming app, audio playback app, note-sharing app, etc., without specific limitations. The target audience can be represented by an account within the application.

[0049] The social characteristics of the target object are obtained by aggregating the characteristics of multiple friends of the target object in the target friend social graph. These multiple friends include churned friends and current friends. Churned friends refer to those who haven't interacted with the target object (e.g., likes, comments, messages) for more than a preset period of time, or those who previously had a friend relationship with the target object but have since severed the friendship at least at the time of step 210. Current friends refer to those with whom the target object currently has a friend relationship.

[0050] A target friend social graph is built based on the friendship relationships between objects. It includes nodes and edges connecting those nodes. A node represents an object, and if two nodes are connected by an edge, it means that the objects represented by those two nodes have a friendship relationship, or previously had a friendship relationship. The target friend social graph includes nodes representing the target object, and nodes representing the target object's friends (lost and current friends). Of course, the target friend social graph can also include friends of the target object's friends. A friendship relationship between two objects can be established through mutual friend requests (or mutual following), one object adding another as a friend, or one object following another; there are no specific limitations here.

[0051] Figure 3 An example diagram illustrating the target friend social graph is shown, such as... Figure 3 As shown, if node P0 represents the target object, then the objects represented by nodes P1, P2, P4, P5, and P7 are all friends of the target object. The object represented by node P6 is a friend of both the object represented by node P7 and the object represented by node P5. The object represented by node P4 is also a friend of both the object represented by node P5 and the object represented by node P3.

[0052] In some embodiments, the nodes in the target friend social graph store object information of the objects represented by the nodes. Object information includes, for example, the object's registration time in the application, the object's gender, the object's region, the object's game level, and the object's object tags. The content contained in the object information of objects registered in different applications may be different. For example, in a game application, the object information may also include the object's game badges, game achievements, etc. In a music application, the object information may also include the object's music tags (such as tags indicating the music genres that the object follows).

[0053] In addition, the edges of the target friend social graph store interaction information between the two objects connected by the edges. This interaction information includes, for example, friend relationship establishment information (which may include the establishment time), friend relationship dissolution information, and interaction records between the two objects during the friend relationship period. For example, in a game application, interaction records may include records of participating in the same game, like records (e.g., records of object A liking content posted by object B, records of object A liking object B's game stats), comment records, and so on.

[0054] Next, the target friend social graph, which stores object information in the nodes and interaction information on the edges, is input into a Graph Neural Network (GNN). The GNN performs graph encoding on the target friend social graph to obtain the graph encoding result. The graph encoding result includes the object features of each node in the target friend social graph, the features of the edges connecting two nodes (the edge features can also be understood as the interaction features between the two objects connected by the edge), and the global graph features representing the entire target friend social graph. It is worth mentioning that the object features output by the GNN are features obtained by comprehensively processing the object information of the object and the relationships between the object and other objects (such as direct friend relationships, indirect friend relationships, etc.).

[0055] In some embodiments, based on the object features of the objects represented by each node in the target friend social graph output by the graph neural network, the object features of the target object's friends in the target friend social graph (i.e., the object features of the target object's friends output by the graph neural network) can be aggregated to obtain the social features of the target object.

[0056] In some embodiments, the social features of the target object can be obtained by calculating the mean (equivalent to mean aggregation) of the object features of the target object's friends in the target friend social graph output by the graph neural network, based on the object features of the objects represented by each node in the target friend social graph. This process can be described by the following formula 1:

[0057]

[0058] x i Indicates the social characteristics of the target object; x j Let represent the object characteristics of the i-th friend of the target object in the target friend social graph; N(i) represents the total number of friends of the target object in the target friend social graph. It is understandable that, since the node representing the target object and the nodes representing the target object's friends are connected by edges in the target social graph, the friends of the target object in the target social graph can be called the first-order neighbors of the target object.

[0059] In other embodiments, the social relevance between the target object and its friends can be calculated based on the interaction records between the target object and its friends. Then, according to the social relevance between the target object and its friends, the object features of the target object's friends in the target friend social graph are aggregated to obtain the target object's social features. For example, the target object's social features can be calculated according to the following formula 2:

[0060]

[0061] Among them, w ij This represents the social relevance between the target object and its j-th friend, where j is a positive integer.

[0062] In some embodiments, the total number of interactions between the target object and other objects can be counted based on the total number of interactions between the target object and each of its friends, and the reference number of interactions between the target object and each of its friends can be counted based on the interaction records between the target object and each of its friends. For the j-th friend of the target object, the ratio between the reference number of interactions between the target object and each friend and the total number of interactions between the target object can be used as the social relevance between the target object and the i-th friend of the target object.

[0063] In some embodiments, the PPR (Personalized Page Rank) algorithm can be used to calculate the social relevance between the target object and its churned friends based on the target friend social graph. This social relevance can also be referred to as the PPR value. Specifically, starting from the node representing the target object (let's call it the first node), a random walk with probability 'a' is performed along the edges of the target friend social graph, maintaining a probability of 1-a and returning to the first node. This process is repeated until it converges to a stable state. From this, the PPR value of each node in the target friend social graph can be calculated. The corresponding PPR value of the target object's churned friends is then used as the social relevance between the target object and its churned friends.

[0064] In some embodiments, the social characteristics of the target object can also be obtained by concatenating the features calculated according to Formula 1 above with the features calculated according to Formula 2; for ease of description, the features calculated according to Formula 1 are referred to as the first social characteristic, and the features calculated according to Formula 2 are referred to as the third social characteristic. The first social characteristic and the second social characteristic are concatenated to obtain the social characteristics of the target object.

[0065] In social networks, there is a certain correlation between the storage of objects and their friends. For example, in gaming applications, a player's activity is influenced by the activity of their friends. The social features of a target object can reflect its social topology and friend distribution information. Furthermore, in this application, based on the target friend social graph, the object features of multiple friends of the target object in the target friend social graph are aggregated to obtain the target object's social features. The idea of ​​decoupled GNN is introduced, and through the pluggable aggregation of friend object features, the scalability of social features can be achieved. For example, when the target object's friends increase, the object features of the newly added friends can be added for aggregation.

[0066] In some embodiments, since the graph encoding result includes the object features of the objects represented by each node in the target friend social graph, the object features of the target object and the object features of the target object's lost friends can be obtained from the graph encoding result.

[0067] In some embodiments, since the graph encoding result includes the features of the edges connecting two nodes in the target friend social graph, the features of the edges representing the nodes of the target object and the churned friends of the target object in the graph encoding result can be used as the interaction features between the target object and the churned friends.

[0068] In some embodiments, interaction information between the target object and its churned friends can also be obtained, and the interaction information can be vectorized to obtain the interaction features between the target object and its churned friends.

[0069] In some embodiments, the social characteristics of the target object, the object characteristics of the target object, the object characteristics of the target object's lost friends, and the interaction characteristics between the target object and the lost friends can be concatenated, and the concatenation result can be used as the fusion feature.

[0070] In other embodiments, a feature fusion network can be used to fuse the social features of the target object, the object features of the target object, the object features of the target object's churned friends, and the interaction features between the target object and churned friends, outputting fused features. The feature fusion network can be a neural network model constructed using fully connected networks, recurrent neural networks, or other neural networks for feature fusion.

[0071] In other embodiments, such as Figure 4 As shown, step 210 includes the following steps 410-430:

[0072] Step 410: The social features of the target object, the object features of the target object, the object features of the target object's lost friends, and the interaction features between the target object and the lost friends are concatenated to obtain the concatenated features.

[0073] Figure 5 An example diagram illustrates the stitched feature obtained by stitching together four features, such as... Figure 5 As shown, the object features of the target object's friends 1 to 4 are aggregated according to Formula 1 above to obtain the first social feature, and the object features of the target object's friends 1 to 4 are aggregated according to Formula 2 above to obtain the third social feature. The first social feature and the third social feature are concatenated to obtain the target object's social feature. Based on this, the target object's social feature, the target object's object features, the object features of churned friends, and the interaction features between the target object and churned friends are concatenated to obtain the concatenated feature.

[0074] Step 420: Perform multi-head attention cross-processing on the spliced ​​features to obtain cross-fused features.

[0075] In this embodiment, the spliced ​​features are mapped to multiple sub-feature spaces through a multi-head attention mechanism. The interaction patterns between different features can be captured in different sub-feature spaces. Then, the features in different sub-feature spaces are stacked to obtain higher-order cross-fusion features.

[0076] Since the social characteristics of the target object, the object characteristics of the target object, the object characteristics of churned friends, and the interaction characteristics between the target object and churned friends all involve features across multiple dimensions, the concatenated feature is also a multidimensional feature. We will now focus on the m-th feature e within the concatenated feature. m Let's take an example to illustrate the process of multi-head attention cross-processing. Correspondingly, the cross-fusion features include the sub-cross-fusion features corresponding to each dimension of the spliced ​​features.

[0077] Under the h-th attention head, the m-th dimension feature e m The matrix is ​​transformed into a vector representation in the sub-feature space corresponding to the h-th attention head through linear multiplication. For the m-th feature e m In a sub-feature space, there are three vector representations, namely:

[0078]

[0079] in, It refers to the m-th dimension feature e m The query vector in the sub-feature space corresponding to the h-th attention head; This refers to the query weight matrix under the h-th attention head; It refers to the m-th dimension feature e m The key vector in the sub-feature space corresponding to the h-th attention head; This refers to the key weight matrix under the h-th attention head; It refers to the m-th dimension feature e m The value vector in the sub-feature space corresponding to the h-th attention head; It refers to the value weight matrix under the h-th attention head.

[0080] Then, the m-th dimension feature e is calculated using the vector inner product. m Compared with other features in the splicing feature (taking the k-th feature e in the splicing feature as an example), k For example, the similarity ψ in the sub-feature space corresponding to the h-th attention head (where m and k are different integers) h (e m e k ):

[0081]

[0082] It refers to the k-th dimension feature e k The key vector in the sub-feature space corresponding to the h-th attention head.

[0083] Subsequently, normalization is performed using the softmax function to obtain the m-th dimension feature e under the h-th attention head.m With the k-th dimension feature e k Attention scores between

[0084]

[0085] M represents the total dimension of the spliced ​​features.

[0086] Finally, we can calculate the attention score based on the m-th dimension feature e under the h-th attention head. m The attention score between the concatenated features and other features of each dimension will be used to determine the concatenated features excluding the m-th feature e. m The value vectors of other features outside the h-th attention head are weighted to obtain the m-th feature e. m Cross features under the h-th attention head

[0087]

[0088] In this embodiment, a multi-head attention mechanism is employed to create feature subspaces corresponding to different attention heads to learn different feature interactions. By combining the features learned in all sub-feature spaces, the m-th dimension feature e can be obtained. m Corresponding sub-cross-fusion features

[0089]

[0090] in, This indicates the concatenation operator, where H is the total number of attention heads.

[0091] Step 430: Perform residual processing on the splicing features and cross-fusion features to obtain the fused features.

[0092] In this embodiment, to preserve the original information in the splicing features, residual processing can be performed on the splicing features and the cross-fusion features. That is, the cross-fusion features are first input into the next layer of the neural network for processing, and then the features output by the next layer of the neural network are added to the splicing features to obtain the added features. Afterwards, activation processing can be performed on the added features to obtain the fused features. The fused features include the sub-fused features corresponding to each dimension of the splicing features, and similarly, the m-th dimension feature e in the splicing features... m Taking the m-th feature e as an example, m Corresponding sub-fusion features for:

[0093]

[0094] W ResThe weight matrix can be determined through training. The ReLU function represents the activation function in the neural network layer, which can be a fully connected layer; no specific limitation is made here.

[0095] Step 220: The gating network predicts the weights of multiple expert networks on the invitation prediction task and the accept invitation prediction task based on the fusion characteristics.

[0096] Invitation prediction task refers to the prediction task used to predict the probability that a user represented by an object will invite another object to return to the network. The probability that a user represented by an object (let's say object A) will invite another object (let's say object B, and object B is a friend of object A) to return to the network can be visualized as: the probability that the client on the object A side will trigger a return invitation operation for object B.

[0097] Invitation acceptance prediction is a task used to predict the probability that an invited user will accept a callback invitation initiated by the inviting party. Again, taking object A inviting object B as an example, it refers to the probability that users on object B's side will accept a callback invitation sent by object A. Expert networks are feedforward networks responsible for learning specific task features.

[0098] Different expert networks focus on extracting domain-specific features from the input, or different expert networks are responsible for learning specific task features, and the knowledge learned by different expert networks is different. In some embodiments, the expert network is a neural network pre-trained on a prediction task similar to an invitation prediction task or an acceptance prediction task. The expert network can be a feedforward neural network (e.g., a multilayer perceptron network), etc., and is not further specifically limited here. The number of expert networks deployed is at least two, and the specific number can be set according to actual needs and the computing power of the device.

[0099] The gating network is used to predict the weights of each expert network on the invitation prediction task and the acceptance prediction task, respectively. The predicted weights represent the degree of contribution of the features output by the expert networks to the corresponding prediction tasks. In some embodiments, the gating network can be a gated recurrent neural network, a feedforward network, etc., without specific limitations.

[0100] In this application, two gating networks can be deployed, referred to as the first gating network and the second gating network, respectively. The first gating network is used to predict the weights of multiple expert networks in the invitation prediction task, and the second gating network is used to predict the weights of multiple expert networks in the acceptance prediction task.

[0101] In this application, a gating network is used to learn the weights of each expert network in the invitation prediction task and the invitation acceptance prediction task. The weight of an expert network in the invitation prediction task determines the contribution of its output features to the invitation prediction task, or in other words, the proportion of its output features used for invitation prediction. Similarly, the weight of an expert network in the invitation acceptance prediction task determines the contribution of its output features to the invitation acceptance prediction task, or in other words, the proportion of its output features used to predict the probability of accepting an invitation. The sum of the weights of all expert networks predicted by the gating network for the same prediction task (invitation prediction or invitation acceptance prediction) is 1.

[0102] Step 230: According to their respective weights on the invitation prediction task, the expert processing features output by multiple expert networks are fused to obtain the first target feature; the expert processing feature is obtained by the corresponding expert network processing the fused feature.

[0103] The fused feature network can be input into multiple expert networks for feature extraction, resulting in expert-processed features output by each expert network. It is understood that, since this application involves two prediction tasks (i.e., an invitation prediction task and an invitation-acceptance prediction task), the deployed multiple expert networks are shared between the two prediction tasks.

[0104] If the i-th expert network in a set of multiple expert networks processes the fusion feature x, the output expert-processed feature is represented as f. i (x), the processing of the fusion feature x by the i-th expert network can be described by the following formula 11:

[0105] f i (x)=ReLU(W i x); (Formula 11)

[0106] W i Let be the parameters of the i-th expert network, which can be determined through training.

[0107] Suppose that the gated network predicts the weights g of the i-th expert network on prediction task a in the two prediction tasks based on the fused feature x. a (x) i :

[0108] g a (x) i =softmax(W ga *x); (Formula 12)

[0109] W ga The parameters of the gating network can be determined through training.

[0110] The first target feature is subsequently used to predict the probability of an invitation prediction task, that is, the first probability of the target object inviting lost friends back, as described below. Assuming task a is an invitation prediction task, the first target feature can be determined according to the following formula 13:

[0111]

[0112] n represents the total number of deployed expert networks, and X1 represents the first target feature.

[0113] Step 240: According to their respective weights in the invitation prediction task, the expert processing features output by multiple expert networks are fused to obtain the second target feature.

[0114] The first target feature is subsequently used to predict the probability of accepting the invitation, i.e., predicting the second probability, as described below, of the churned friends of the target object accepting the target object's return invitation. The calculation of the second target feature can be determined using a process similar to Formula 13 above. It will not be elaborated further here.

[0115] Step 250: Based on the first target characteristics and the second target characteristics, perform recall processing on lost friends.

[0116] The process of recalling churned friends involves deciding whether to add the churned friend to the target user's churned friend recommendation list. For each churned friend of the target user, the process described in steps 210-240 above can be followed to determine whether to add them to the target user's churned friend recommendation list. The target user's churned friend recommendation list is used to display individuals available for churn invitation on the target user's client interface. In other words, the churned friends in the target user's churned friend recommendation list serve as individuals who can be invited back by the user represented by the target user.

[0117] In some embodiments, a user on the target side can invite a lost friend to return by triggering (e.g., clicking) the lost friend displayed on the target's client interface. The target's client then sends an activity invitation message to the client of the lost friend, which is used to invite the lost friend to participate in the activity.

[0118] Figure 6 This is a schematic diagram of a display interface according to an embodiment of the present application, in which a friend display area 610 is shown. Figure 6The friend display area 610 exemplarily shows the avatars and nicknames of three churned friends of the logged-in user of the client, as well as an invitation control 611 displayed for each churned friend. If the user clicks the invitation control 611 displayed for the churned friend with nickname "II," the client can send an activity invitation message to the client where the churned friend with nickname "II" is located. In addition, the friend display area 610 also displays an invitation record viewing control 613. By triggering this control, the user can view historical invitation records. Furthermore, triggering the "More Friends" viewing control 612 shown in the friend display area 610 can display more churned friends that can be invited. Additionally, by triggering the "Refresh" control 614 in the friend display area 610, the churned friends displayed in the friend display area 610 can be replaced with other churned friends. Figure 6 The friend display area shown is merely an example and should not be considered as a limitation on the scope of this application.

[0119] In some embodiments, if the target is an object registered in a game application, a user on the target's side can invite a churned friend back by sending a game invitation to the churned friend's game client triggered by the user on the target's client side. If the target is an object registered in a music application, a user on the target's side can invite a churned friend back by sending a karaoke invitation to the churned friend's client. The form of the back-to-game invitation can vary across different applications.

[0120] Considering the inherent connections between friends in social networks, and the potential influence of an object's actions on its friends' actions, this application not only utilizes the object characteristics of the target object, the object characteristics of the target object's churned friends, and the interaction characteristics between the target object and its churned friends for the invitation prediction and acceptance prediction tasks for the target object's churned friends, but also introduces the target object's social characteristics. These social characteristics reflect the target object's social topology and friend distribution information. Therefore, by combining these four aspects, features reflecting the social links between the target object and its churned friends can be fully extracted to improve the accuracy of the invitation prediction and acceptance prediction tasks, thereby ensuring the accuracy of the recall process for the target object's churned friends and improving the recall effect. This also reduces the waste of server communication resources during the churn recall process.

[0121] Furthermore, considering the correlation between the invitation prediction task and the invitation acceptance prediction task, and the commonalities or correlations among the features applicable to both tasks, multiple expert networks are shared. A gating network is used to predict the weights of each expert network on both tasks. Then, based on the weights of each network on the invitation prediction task, the expert-processed features output by each network are fused to obtain the first target feature suitable for the invitation prediction task. Similarly, based on the weights of each network on the invitation prediction task, the expert-processed features output by each network are fused to obtain the second target feature suitable for the invitation prediction task. This approach fully utilizes the features extracted by different expert networks while also considering the adaptability of the output features to both the invitation and invitation prediction tasks, ensuring the accuracy of both target features and improving the accuracy of subsequent friend recall processing. It also eliminates the need to deploy multiple expert networks separately for each task, reducing computational overhead.

[0122] In some embodiments, such as Figure 7 As shown, step 250 includes:

[0123] Step 710: Perform probability mapping on the first target feature to obtain the first probability of the target object inviting lost friends back.

[0124] The first target feature can be input into the first classification layer, which then performs probability mapping on the first target feature and outputs the first probability that the target object will invite the churned friend back. The first classification layer refers to the classification layer deployed for the invitation prediction task; it can be one or more fully connected layers, without specific limitations here. For example, if task a is an invitation prediction task, the first probability that the target object will invite the churned friend back can be expressed as:

[0125]

[0126] y a The first probability of inviting lost friends back to the target audience; h a This represents the classification layer deployed for task a. When task a is an invitation prediction task, then h... a This is the first classification layer.

[0127] The first probability reflects the likelihood that users on the target side are likely to invite a churned friend back. The higher the first probability, the higher the likelihood that users on the target side are likely to invite a churned friend back. For example, if a user on the target side clicks... Figure 6The higher the likelihood that an invitation control will be displayed for the lost friend in the interface shown.

[0128] Step 720: Perform probability mapping on the second target feature to obtain the second probability that a lost friend will accept the return invitation from the target object.

[0129] Similarly, the second target feature can be input into the second classification layer, which then performs probability mapping on the second target feature, outputting the second probability that the churned friends of the target object will accept the return invitation from the target object. The second classification layer refers to the classification layer deployed for the invitation acceptance prediction task. Likewise, the second classification layer can also be constructed using a fully connected network.

[0130] The second probability reflects the likelihood that the target's churned friends will accept a reconnection invitation initiated by the target. The higher the second probability, the higher the likelihood that the target's churned friends will accept a reconnection invitation initiated by the target. For example, the higher the likelihood that churned friends will accept an activity invitation sent by the client where the target is located.

[0131] Step 730: Based on the first probability and the second probability, perform a recall process for lost friends.

[0132] In some embodiments, such as Figure 8 As shown, step 730 includes:

[0133] Step 810: Multiply the first probability and the second probability to obtain the reference recall weight of the churned friend relative to the target object.

[0134] For a churned friend of a target entity to accept a re-entry invitation, the target entity must first invite that churned friend back. Therefore, the product of the first and second probabilities represents the probability of the target entity inviting a churned friend back, and that churned friend accepting the invitation. In other words, when this event occurs, it indicates a successful re-entry of the churned friend. Thus, multiplying the first and second probabilities yields the churned friend's reference re-entry weight relative to the target entity, representing the probability that the target entity will successfully re-enter that churned friend.

[0135] Step 820: Sort the multiple churned friends of the target object in descending order of reference recall weight to obtain the reference ranking.

[0136] Step 830: Add the top N lost friends from the reference sorting to the lost friend recommendation list of the target object; N is a positive integer.

[0137] Step 840: Send a list of lost friends to the client where the target is located.

[0138] The client containing the target client can display the lost friends in the lost friend recommendation list, which can be used as lost friends for the target to invite back.

[0139] The higher the first probability that a target audience member will invite a churned friend back, the higher the likelihood of that friend being invited. Similarly, the higher the second probability that a churned friend will accept the invitation, the higher the likelihood of that friend accepting. In the above embodiment, the first and second probabilities are multiplied to obtain a reference recall weight for each churned friend relative to the target audience member. Churned friends with higher reference recall weights are then added to the target audience member's churned friend recommendation list. A higher recall weight for a churned friend indicates a higher probability of that friend accepting the invitation and being successfully recalled. Therefore...

[0140] This embodiment can improve the overall success rate of recalling lost friends and enhance the recall effect.

[0141] In other embodiments, churned friends whose recall weight exceeds a weight threshold can be selected from multiple churned friends of the target object and added to the target object's churned friend recommendation list.

[0142] In other embodiments, a secondary screening can be performed on multiple churned friends of the target object based on the first probability and the corresponding second probability of each churned friend. First, based on the first probability of each churned friend of the target object, churned friends whose first probability exceeds a first threshold are added to the candidate set among the multiple churned friends of the target object. Then, churned friends whose second probability exceeds a second threshold are selected from the candidate set and added to the churned friend recommendation list of the target object.

[0143] In other embodiments, the social characteristics of the target object are obtained by aggregating the object characteristics of multiple friends of the target object in the target friend social graph, including churned friends; such as Figure 9 As shown, step 250 includes:

[0144] Step 910: Multiply the first probability and the second probability to obtain the target probability.

[0145] Step 920: Update the social features based on the target probability to obtain the target social features.

[0146] The social features are updated by using the target probability as the aggregation weight for the target object's lost friends, and then aggregating the object features of the target object's multiple friends in the target friend social graph to obtain the new social features of the target object, namely the target social features.

[0147] In some embodiments, the social features include a first social feature; the first social feature is obtained by averaging the object features of multiple friends of the target object in the target friend social graph; step 920 includes:

[0148] Step ①: According to the target probability and preset weight, the object features of the target object in the target friend social graph are weighted to obtain the second social feature; where the target probability is used as the weighting coefficient of the object features of churned friends, and the preset weight is used as the weighting coefficient of the object features of non-churned friends among the multiple friends.

[0149] In this case, the aggregation can be performed according to Formula 15 to obtain the second social feature.

[0150]

[0151] Where the j-th friend of the target is a churned friend, This is equal to the target probability determined for that churned friend; when the j-th friend of the target object is a non-churned friend of the target object... In some embodiments, if the first social feature is determined according to Formula 1 above, the preset weight can be equal to 1. That is, compared with the first social feature, in the process of re-aggregating features, the aggregation weight of the target object's lost friends is updated to the corresponding determined target probability, while the aggregation weight of other non-lost friends remains unchanged.

[0152] Step ②: Update the first social feature in the social features to the second social feature to obtain the target social feature.

[0153] In some embodiments, if the social characteristic of the target object is the first social characteristic, then the second social characteristic is used as the target social characteristic.

[0154] In other embodiments, if the social feature of the target object is the result of concatenating the first social feature and the third social feature, then the result of concatenating the second social feature and the third social feature is taken as the target social feature.

[0155] Figure 10 This is a schematic diagram illustrating the comparison between the social characteristics of the target object and the target social characteristics according to an embodiment of this application, such as... Figure 10As shown, the first social feature is obtained by aggregating the object features of multiple friends of the target object using 1 as the aggregation weight. The third social feature is obtained by aggregating the object features of the target object using the PPR value of the friends as the aggregation weight. The concatenation result of the first and third social features is used as the social feature of the target object.

[0156] Assumption Figure 10 If Friend 1 is a churned friend of the target audience, then the aggregation weight corresponding to Friend 1 is equal to the target probability and the non-churned friends of the target audience (e.g., ...). Figure 10 The aggregation weight of friends 2-4 in the target object is 1. The object features of multiple friends of the target object are aggregated to obtain the second social feature. The concatenation result of the second social feature and the third social feature is used as the target social feature. It can be understood that if the target object's friends include multiple churned friends, and the corresponding first probability and second probability have been predicted in advance for each churned friend, then in step ①, the aggregation weight of each churned friend is replaced with the corresponding target probability.

[0157] Step 930: Based on the target social characteristics, the target object's object characteristics, the churned friend's object characteristics, and the interaction characteristics between the target object and the churned friend, re-predict the third probability of the target object inviting the churned friend back, and re-predict the fourth probability of the churned friend accepting the target object's invitation to return.

[0158] In other words, based on the target social characteristics, the target object's object characteristics, the churned friend's object characteristics, and the interaction characteristics between the target object and the churned friend, the probability of the target object inviting the churned friend back, and the probability of the churned friend accepting the target object's invitation to return, are re-predicted according to the above process. Step 930 may include the following steps A1-A7:

[0159] Step A1 involves fusing the target social features, the target object's object features, the churned friend's object features, and the interaction features between the target object and the churned friend to obtain reference fused features.

[0160] Step A2: The gating network predicts the reference weights of multiple expert networks for the invitation prediction task and the acceptance prediction task, respectively, based on the reference fusion features.

[0161] Step A3 involves multiple expert networks processing the reference fusion features separately to obtain the reference expert-processed features output by each expert network.

[0162] Step A4: According to the reference weights of each expert in the invitation prediction task, the reference expert processing features output by multiple experts are fused to obtain the third target feature.

[0163] Step A5: According to the reference weights of the invitation prediction task, the reference expert processing features output by multiple experts are fused to obtain the fourth target feature.

[0164] Step A6: Perform probability mapping on the third target feature to obtain the third probability of the target object inviting lost friends back.

[0165] Step A7: Perform probability mapping on the fourth target feature to obtain the fourth probability that a lost friend will accept the return invitation from the target object.

[0166] The implementation details of steps A1-A7 above are the same as those in the text above. Figure 2 and Figure 7 The implementation details of the corresponding steps are similar and will not be repeated here.

[0167] Step 940: The product of the third probability and the fourth probability is used as the target recall weight for churned friends relative to the target object.

[0168] Step 950: Recall churned friends according to their recall weight relative to the target audience.

[0169] In some embodiments, step 950 includes: sorting multiple churned friends of the target object in descending order of target recall weight to obtain a churned friend ranking; adding the top N churned friends in the churned friend ranking to the churned friend recommendation list of the target object; where N is a positive integer; and sending the churned friend recommendation list to the client where the target object is located.

[0170] In other embodiments, churned friends whose target recall weight exceeds a weight threshold can be selected from multiple churned friends of the target object and added to the target object's churned friend recommendation list.

[0171] In other embodiments, a secondary screening can be performed on multiple churned friends of the target object based on the third probability and the corresponding fourth probability of each churned friend. First, based on the third probability of each churned friend of the target object, churned friends whose third probability exceeds a first threshold are added to a reference candidate set. Then, churned friends whose fourth probability exceeds a second threshold are selected from the candidate set and added to the recommended list of churned friends of the target object.

[0172] In the above embodiment, the first and second probabilities predicted in the first round are used for self-distillation enhancement. The contribution of the object features of the target object's churned friends to the target social features, the contribution of the object features of the target object's non-churned friends to the target social features, and the contribution of the different churned object features of the target object to the target social features are differentiated. This can provide more effective information for predicting return invitations and accepting return invitations, thus achieving self-distillation enhancement. Compared with the first and second probabilities, the third and fourth probabilities are more accurate, thereby ensuring the accuracy of the subsequent churned friend recommendation list determined for the target object.

[0173] Figure 11 This is a flowchart illustrating a recall processing method according to an embodiment of this application, such as... Figure 11 As shown, the input features are the social features of the target object, the object features of the target object, the interaction features between the target object and churned friends, and the object features of churned friends. These four features are embedded and then concatenated. They are then input into a feature cross-fusion network for feature cross-processing. This feature cross-fusion network performs feature cross-processing according to a multi-head attention mechanism.

[0174] Subsequently, the features output by the feature cross-fusion network are input into the hybrid expert model, which includes a gating network and multiple expert networks, such as expert network 1, expert network 2, ..., expert network n (n is a positive integer). Since this application involves two prediction tasks: an invitation prediction task and an invitation acceptance prediction task, two gating networks can be deployed in the hybrid expert model. The first gating network is used to predict the weights of the multiple expert networks on the invitation prediction task, and the second gating network is used to predict the weights of the multiple expert networks on the invitation acceptance prediction task. Each expert network processes the features output by the feature cross-fusion network, and then, according to the weights of the multiple expert networks on the invitation prediction task, the features output by the multiple expert networks are weighted and fused to obtain the target features suitable for the invitation prediction task (e.g., the first target feature mentioned above); and, according to the weights of the multiple expert networks on the invitation acceptance prediction task, the features output by the multiple expert networks are weighted and fused to obtain the target features suitable for the invitation acceptance prediction task (e.g., the second target feature mentioned above).

[0175] Considering that this application involves two prediction tasks, a first classification layer for the invitation prediction task and a second classification layer for the acceptance prediction task are deployed in the classification layer. In this way, the target features applicable to the invitation prediction task are input into the first classification layer for probability prediction to obtain the probability of making a return invitation (e.g., the first probability of the target object inviting a lost friend back as mentioned above); and the target features applicable to the acceptance prediction task are input into the second classification layer for probability prediction to obtain the probability of accepting a return invitation (e.g., the second probability of a lost friend accepting a return invitation from the target object as mentioned above).

[0176] Next, the obtained first and second probabilities are used to update the social features of the target object to obtain the target social features. Self-distillation enhancement is then performed on these features, and the process described above—from the embedding layer to the feature cross-fusion network, the hybrid expert model, and the classification layer—is followed to obtain the third probability of the target object inviting churned friends back, and the fourth probability of churned friends accepting the target object's invitation to return. Subsequently, based on the third and fourth probabilities, a decision can be made on whether to add the invited churned friends to the target object's churned friend recommendation list.

[0177] To ensure the accuracy of the predicted probabilities, the gating network and expert network mentioned above need to be trained in advance. The training process can be as follows: Figure 12 As shown, it includes:

[0178] Step 1210: Obtain multiple training samples; wherein, a training sample includes a sample object, a reference sample object, a first label and a second label, the first label is used to indicate whether the sample object invites the reference sample object to reflow, and the second label is used to indicate whether the reference sample object accepts the sample object's reflow invitation.

[0179] It can retrieve multiple return invitation records and multiple return invitation accepted records. Taking object A and object B as an example, if object B is a lost friend of object A, the return invitation record can be the record generated when object A invites object B to return. The return invitation accepted record can be the record generated when object B accepts the return invitation after receiving it from object A.

[0180] Next, multiple backflow invitation records and multiple backflow accepted records can be grouped together. Backflow invitation records and backflow accepted records corresponding to the same backflow invitation are grouped together to obtain multiple record pairs. Each record pair includes one backflow invitation record and one backflow accepted record. For example, if the backflow invitation record in a record pair is generated when object A invites object B backflow, then the backflow accepted record in that record pair is generated when object B accepts the backflow invitation after receiving it from object A. It's worth noting that since the invited party may not accept the backflow invitation after receiving it, after the above division, there will still be backflow invitation records that do not have a corresponding backflow accepted record.

[0181] Based on this, a positive training sample is generated according to a record pair. That is, the inviter indicated by the return invitation record in a record pair is taken as the sample object, and the invitee indicated by the return invitation record in a record pair is taken as the reference sample object. The first label indicates that the sample object has invited the reference sample object to return; the second label indicates that the reference sample object has accepted the return invitation of the sample object.

[0182] And based on the return invitation records that do not correspond to the return invitation records, negative training samples are generated for the acceptance of invitation prediction task, that is, the inviter indicated in the return invitation record is used as the sample object, and the invitee indicated in the return invitation record is used as the reference sample object, and the first label indicates that the sample object has invited the reference sample object to return; the second label indicates that the reference sample object has not accepted the return invitation of the sample object.

[0183] In addition, negative training samples for the invitation prediction task can be constructed based on the churned friends of each sample object who have not been invited back. That is, the churned friends of the sample object who have not been invited back are used as reference sample objects, and the corresponding first label indicates that the sample object has not invited the reference sample object to come back.

[0184] The positive training samples, negative training samples for the invitation prediction task, and negative training samples for the invitation prediction task are all used as training samples to train the model.

[0185] Step 1220 involves fusing the social features of the sample object, the object features of the sample object, the object features of the reference sample object, and the interaction features between the sample object and the reference sample object to obtain the sample fusion features. The method for determining the social features of the sample object is similar to the process for determining the social features of the target object described above, and will not be repeated here.

[0186] Step 1230: The gating network predicts the weights of multiple expert networks on the invitation prediction task and the acceptance prediction task based on the sample fusion features.

[0187] Step 1240: According to their respective weights in the invitation prediction task, the sample expert features output by multiple expert networks based on the sample fusion features are fused to obtain the first sample target features.

[0188] Step 1250: According to their respective weights in the invitation prediction task, the sample expert features output by multiple expert networks are fused to obtain the second sample target features.

[0189] Step 1260: Based on the target features of the first sample, predict the probability of the first sample object inviting the reference sample object to return.

[0190] Step 1270: Based on the target characteristics of the second sample, predict the probability of the reference sample object accepting the return invitation of the sample object as a second sample.

[0191] The implementation details of steps 1220-1270 are the same as those in the above text. Figure 2 and Figure 7 The implementation details of the corresponding steps are similar and will not be repeated here.

[0192] Step 1280: Calculate the target loss based on the first sample probability, the first label, the second sample probability, and the second label.

[0193] In some embodiments, a first loss for the invitation prediction task can be calculated based on the first sample probability and the corresponding first label, and a second loss for the invitation acceptance prediction task can be calculated based on the second sample probability and the corresponding second label. In some embodiments, the first and second losses can be calculated using a mean error loss function, an absolute value loss function, a cross-entropy loss function, etc. Then, the first and second losses are added together to obtain the target loss.

[0194] For example, if task a is an invitation prediction task, and the loss is calculated using the cross-entropy loss function, then the first sub-loss for each training sample for the invitation prediction task can be calculated according to the following formula 16:

[0195]

[0196] Among them, y ai It represents the first sample probability of the i-th training sample on prediction task a (invitation prediction task), or the mapping value obtained by mapping the first sample probability to 0-1; Let represent the first label of the i-th training sample on prediction task a. Let represent the first sub-loss of the i-th training sample on prediction task a (invitation prediction task).

[0197] Then, the sub-losses of all training samples on the same prediction task can be weighted to obtain the first loss for the invitation prediction task. For example, continuing the example above, the first loss for prediction task a (invitation prediction task) can be determined according to the following formula 17:

[0198]

[0199] w a This represents the loss weight for prediction task a, where K represents the total number of training samples, and L1 is the first loss. In some implementations, the loss weights set for different prediction tasks can be configured as needed; for example, the loss weights for two prediction tasks can both be set to 1, or they can be different.

[0200] Step 1290: Based on the target loss, adjust the parameters of at least the gating network and multiple expert networks until the first training termination condition is met.

[0201] The gradient of the target loss can be calculated, and then the parameters of at least the gating network and multiple expert networks can be adjusted using gradient descent. Additionally, the parameters of the feature cross-fusion network and the classification layers (first and second classification layers) can also be adjusted.

[0202] The first training termination condition can be that the number of iterations reaches a threshold, the target loss converges, or the prediction accuracy for both prediction tasks reaches an accuracy threshold.

[0203] In the training process described above, the gating network and multiple expert networks can simultaneously learn features related to two prediction tasks (invitation prediction task and acceptance prediction task), thus achieving joint training of the two prediction tasks and realizing data augmentation. During training, the feature cross-fusion network uses a multi-head attention mechanism for feature cross-learning, combining different features to capture complex relationships and patterns between features, which can improve the model's prediction accuracy and generalization ability.

[0204] In some embodiments, it is possible to first follow Figure 12 The process shown involves training for a period of time, and then using the currently trained model (e.g.) Figure 11 The model shown is used to predict the first sample probability and the second sample probability for each training sample. The first sample probability and the second sample probability are then used to update the social features of the sample objects in the training samples. The social features in step 1220 are then replaced with the updated social features. The model is then fine-tuned again according to steps 1220-1290 above to achieve self-distillation enhancement of the model and improve the training effect.

[0205] In addition, Figure 11 After training, the model shown was applied to a game application for friend retrieval verification. The proposed solution and related technologies were compared on the invitation prediction and invitation acceptance prediction tasks, testing the ACC (Accuracy) and AUC (Area Under the Curve, used to describe the ROC (Receiver Operating Characteristic) curve). The test results are shown in Tables 1 and 2 below. In Tables 1 and 2, Solution 1 is... Figure 11 The prediction process shown includes a self-distillation enhanced prediction process; Scheme 2 is in Figure 11 The predicted process shown excludes the scheme of self-distillation enhancement; Scheme 3 is in Figure 11 Based on the scheme shown, this scheme does not include social features in the input features and does not perform a self-distillation enhancement process.

[0206]

[0207] Table 1. Comparison of the effects of the invitation prediction task

[0208]

[0209] Table 2. Comparison of the effects of accepting invitations to predict tasks

[0210] As can be seen from Tables 1 and 2 above, using Scheme 1 provided in this application, the AUC index is improved by 7.31% compared to Schemes I and II in related technologies for the invitation prediction task, and by 0.46% compared to Schemes I and II in the invitation acceptance prediction task. This demonstrates that the scheme provided in this application can effectively improve prediction accuracy, thereby improving the subsequent recall effect of churned friends. Furthermore, comparing the test results of Scheme 2 with Scheme 1 shows that introducing the self-distillation enhancement process plays a significant role in improving the prediction effect. Comparing the test results of Scheme 3 with Scheme 2 or Scheme 1 shows that introducing social features also plays a significant role in improving the prediction effect.

[0211] The method of this application combines the object characteristics of active players (as the target object mentioned above) and churned friends in game applications, as well as the social characteristics of active players and the interaction characteristics between active players and churned friends, to predict the probability that active players will invite churned friends to return, and the probability that churned friends will accept the return invitation. Then, it recommends churned friends with strong click intentions and high return intentions to active players, thereby achieving the purpose of recalling churned friends in game applications and increasing activity.

[0212] The following describes an apparatus embodiment of this application, which can be used to perform the methods described in the above embodiments of this application. For details not disclosed in the apparatus embodiments of this application, please refer to the method embodiments described in the above embodiments of this application.

[0213] Figure 13 This is a block diagram of a recall processing apparatus according to an embodiment of this application, such as... Figure 13 As shown, the recall processing device includes: a first fusion module 1310, used to fuse the social features of the target object, the object features of the target object, the object features of the target object's lost friends, and the interaction features between the target object and the lost friends to obtain fused features; a task weight prediction module 1320, used by a gating network to predict the weights of multiple expert networks on the invitation prediction task and the invitation acceptance prediction task, respectively, based on the fused features; a second fusion module 1330, used to fuse the expert processing features output by multiple expert networks according to their respective weights on the invitation prediction task to obtain a first target feature; the expert processing features are obtained by the corresponding expert networks processing the fused features; a third fusion module 1340, used to fuse the expert processing features output by multiple expert networks according to their respective weights on the invitation acceptance prediction task to obtain a second target feature; and a recall processing module 1350, used to perform recall processing on lost friends based on the first target feature and the second target feature.

[0214] In some embodiments, the recall processing module 1350 includes: a first probability mapping unit, configured to perform probability mapping on a first target feature to obtain a first probability that the target object invites lost friends to return; a second probability mapping unit, configured to perform probability mapping on a second target feature to obtain a second probability that lost friends accept the return invitation from the target object; and a recall processing unit, configured to perform recall processing on lost friends based on the first probability and the second probability.

[0215] In some embodiments, the social features of the target object are obtained by aggregating the object features of multiple friends of the target object in the target friend social graph, including churned friends; the recall processing unit includes: a target probability determination unit, used to multiply a first probability and a second probability to obtain a target probability; an update unit, used to update the social features according to the target probability to obtain target social features; a re-prediction unit, used to re-predict a third probability of the target object inviting churned friends back, and a fourth probability of churned friends accepting the target object's invitation to return, based on the target social features, the object features of the target object, the object features of churned friends, and the interaction features between the target object and churned friends; a target recall weight determination unit, used to use the product of the third probability and the fourth probability as the target recall weight of churned friends relative to the target object; and a processing unit, used to perform recall processing on churned friends according to the recall weight of churned friends relative to the target object.

[0216] In some embodiments, the social features include a first social feature; the first social feature is obtained by averaging the object features of multiple friends of the target object in the target friend social graph; the update unit is configured to: weight the object features of multiple friends of the target object in the target friend social graph according to the target probability and the preset weight to obtain a second social feature; wherein, the target probability is used as the weighting coefficient of the object features of churned friends, and the preset weight is used as the weighting coefficient of the object features of non-churned friends among the multiple friends; update the first social feature in the social features to the second social feature to obtain the target social feature.

[0217] In some embodiments, the social features also include a third social feature; the third social feature is obtained by aggregating the object features of multiple friends of the target object in the target friend social graph according to the social relevance between each friend and the target object.

[0218] In some embodiments, the re-prediction unit is configured to: fuse target social features, target object features, churned friend object features, and interaction features between the target object and churned friends to obtain reference fused features; use a gating network to predict reference weights of multiple expert networks on the invitation prediction task and the invitation acceptance prediction task, respectively, based on the reference fused features; use multiple expert networks to process the reference fused features to obtain reference expert processing features output by each expert network; fuse the reference expert processing features output by multiple experts according to their respective reference weights on the invitation prediction task to obtain a third target feature; fuse the reference expert processing features output by multiple experts according to their respective reference weights on the invitation acceptance prediction task to obtain a fourth target feature; perform probability mapping on the third target feature to obtain a third probability that the target object invites churned friends back; and perform probability mapping on the fourth target feature to obtain a fourth probability that churned friends accept the return invitation from the target object.

[0219] In some embodiments, the processing unit is configured to: sort multiple churned friends of the target object in descending order of target recall weight to obtain a churned friend ranking; add the top N churned friends in the churned friend ranking to the churned friend recommendation list of the target object; where N is a positive integer; and send the churned friend recommendation list to the client where the target object is located.

[0220] In some embodiments, the first fusion module 1310 includes: a splicing unit, used to splice the social features of the target object, the object features of the target object, the object features of the target object's churned friends, and the interaction features between the target object and the churned friends to obtain spliced ​​features; a cross-processing unit, used to perform multi-head attention cross-processing on the spliced ​​features to obtain cross-fusion features; and a residual processing unit, used to perform residual processing on the spliced ​​features and the cross-fusion features to obtain fused features.

[0221] In other embodiments, the recall processing module 1350 includes: a reference recall weight determination unit, configured to multiply a first probability and a second probability to obtain a reference recall weight of churned friends relative to the target object; a reference ranking determination unit, configured to rank multiple churned friends of the target object in descending order of reference recall weight to obtain a reference ranking; a churned friend recommendation list determination unit, configured to add the top N churned friends in the reference ranking to the churned friend recommendation list of the target object; where N is a positive integer; and a sending unit, configured to send the churned friend recommendation list to the client where the target object is located.

[0222] In some embodiments, the recall processing apparatus further includes a training module, configured to: acquire multiple training samples; wherein, a training sample includes a sample object, a reference sample object, a first label, and a second label, the first label indicating whether the sample object invites the reference sample object to return, and the second label indicating whether the reference sample object accepts the return invitation from the sample object; fuse the social features of the sample object, the object features of the sample object, the object features of the reference sample object, and the interaction features between the sample object and the reference sample object to obtain sample fusion features; and, based on the sample fusion features, a gating network predicts the weights of multiple expert networks on the invitation prediction task and the acceptance prediction task, respectively; and, according to their respective weights... The weights for the invitation prediction task are used to fuse the sample expert features output by multiple expert networks based on the sample fusion features, resulting in the first sample target feature. The weights for the invitation acceptance prediction task are then used to fuse the sample expert features output by multiple expert networks, resulting in the second sample target feature. Based on the first sample target feature, the first sample probability of a sample object inviting a reference sample object back is predicted. Based on the second sample target feature, the second sample probability of a reference sample object accepting a reference sample object's back invitation is predicted. The target loss is calculated based on the first sample probability, the first label, the second sample probability, and the second label. Based on the target loss, the parameters of at least the gating network and multiple expert networks are adjusted until the first training termination condition is met.

[0223] In some embodiments, the recall processing apparatus further includes: a target friend social graph acquisition module, used to acquire a target friend social graph constructed based on the social relationships between objects; a graph encoding processing module, used to perform graph encoding processing on the target friend social graph using a graph neural network to obtain a graph encoding result; the graph encoding result includes the object features of the objects represented by each node in the target friend social graph; a feature aggregation module, used to aggregate the object features of multiple friends representing the target object in the graph encoding result to obtain the social features of the target object; and an object feature acquisition module, used to acquire the object features of the target object from the graph encoding result, and acquire the object features of the target object's lost friends.

[0224] Figure 14 A schematic diagram of a computer system suitable for implementing the embodiments of this application is shown. It should be noted that... Figure 14 The computer system 1400 of the illustrated electronic device is merely an example and should not be construed as limiting the functionality and scope of the embodiments of this application. This electronic device can be used to perform the recall processing method provided in the application.

[0225] like Figure 14As shown, the computer system 1400 includes a Central Processing Unit (CPU) 1401, which can perform various appropriate actions and processes, such as executing the methods described in the above embodiments, based on programs stored in Read-Only Memory (ROM) 1402 or programs loaded from storage portion 1408 into Random Access Memory (RAM) 1403. The RAM 1403 also stores various programs and data required for system operation. The CPU 1401, ROM 1402, and RAM 1403 are interconnected via a bus 1404. An Input / Output (I / O) interface 1405 is also connected to the bus 1404.

[0226] The following components are connected to I / O interface 1405: an input section 1406 including a keyboard, mouse, etc.; an output section 1407 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 1408 including a hard disk, etc.; and a communication section 1409 including a network interface card such as a LAN (Local Area Network) card, modem, etc. The communication section 1409 performs communication processing via a network such as the Internet. A drive 1410 is also connected to I / O interface 1405 as needed. Removable media 1411, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 1410 as needed so that computer programs read from them can be installed into storage section 1408 as needed.

[0227] Specifically, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 1409, and / or installed from removable medium 1411. When the computer program is executed by central processing unit (CPU) 1401, it performs various functions defined in the system of this application.

[0228] It should be noted that the computer-readable medium shown in the embodiments of this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such transmitted data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wired, etc., or any suitable combination thereof.

[0229] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. Each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0230] The units described in the embodiments of this application can be implemented in software or hardware, and the described units can also be located in a processor. The names of these units do not necessarily limit the specific unit itself.

[0231] In another aspect, this application also provides a computer-readable storage medium, which may be included in the electronic device described in the above embodiments; or it may exist independently and not assembled into the electronic device. The aforementioned computer-readable storage medium carries computer-readable instructions that, when executed by a processor, implement the methods in any of the above embodiments.

[0232] In the embodiments of this application, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal. It can be implemented wholly or partially using software, hardware (e.g., processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that functions as a whole.

[0233] According to one aspect of the embodiments of this application, a computer program product is provided, the computer program product including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the methods of any of the above embodiments.

[0234] It should be noted that although several modules or units for the device used to perform actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to the embodiments of this application, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0235] Through the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, touch terminal, or network device, etc.) to execute the method according to the embodiments of this application.

[0236] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the embodiments disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

[0237] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A recall processing method, characterized in that, include: The social characteristics of the target object, the object characteristics of the target object, the object characteristics of the target object's lost friends, and the interaction characteristics between the target object and the lost friends are fused together to obtain the fused characteristics. The gating network predicts the weights of multiple expert networks in the invitation prediction task and the accept invitation prediction task based on the fusion features. According to their respective weights on the invitation prediction task, the expert processing features output by the multiple expert networks are fused to obtain the first target feature; the expert processing feature is obtained by the corresponding expert network processing the fused feature. According to their respective weights on the invitation acceptance prediction task, the expert processing features output by the multiple expert networks are fused to obtain the second target feature; Based on the first target feature and the second target feature, the lost friends are recalled.

2. The method according to claim 1, characterized in that, The step of recalling lost friends based on the first target feature and the second target feature includes: By performing probability mapping on the first target feature, a first probability is obtained that the target object invites the lost friend to return. By performing probability mapping on the second target feature, a second probability is obtained that the churned friend accepts the return invitation from the target object; Based on the first probability and the second probability, the lost friends are recalled.

3. The method according to claim 2, characterized in that, The social features of the target object are obtained by aggregating the object features of multiple friends of the target object in the target friend social graph, and the multiple friends include the lost friends; The step of recalling the lost friends based on the first probability and the second probability includes: Multiply the first probability and the second probability to obtain the target probability; The social features are updated based on the target probability to obtain the target social features; Based on the target social characteristics, the target object's object characteristics, the churned friend's object characteristics, and the interaction characteristics between the target object and the churned friend, a third probability is re-predicted for the target object inviting the churned friend back, and a fourth probability is re-predicted for the churned friend accepting the target object's invitation to return. The product of the third probability and the fourth probability is used as the target recall weight of the lost friend relative to the target object; The lost friends are recalled according to their recall weight relative to the target object.

4. The method according to claim 3, characterized in that, The social features include the first social feature; The first social feature is obtained by averaging the object features of multiple friends of the target object in the target friend social graph; The step of updating the social features based on the target probability to obtain the target social features includes: According to the target probability and preset weight, the object features of the target object in the target friend social graph are weighted to obtain the second social feature; wherein, the target probability is used as the weighting coefficient of the object features of the lost friends, and the preset weight is used as the weighting coefficient of the object features of the non-lost friends among the multiple friends. The first social feature in the social features is updated to the second social feature to obtain the target social feature.

5. The method according to claim 4, characterized in that, The social characteristics also include third social characteristics; The third social feature is obtained by aggregating the object features of multiple friends of the target object in the target friend social graph according to the social relevance between each friend and the target object.

6. The method according to claim 3, characterized in that, The step of re-predicting a third probability that the target object will invite the churned friend back, and a fourth probability that the churned friend will accept the target object's invitation to return, based on the target social characteristics, the target object's object characteristics, the churned friend's object characteristics, and the interaction characteristics between the target object and the churned friend, includes: The target social features, the target object's object features, the churned friend's object features, and the interaction features between the target object and the churned friend are fused together to obtain reference fused features; The gating network predicts the reference weights of multiple expert networks on the invitation prediction task and the invitation-acceptance prediction task, respectively, based on the reference fusion features. The reference fusion features are processed by the multiple expert networks respectively to obtain the reference expert processing features output by each expert network; According to the reference weights respectively on the invitation prediction task, the reference expert processing features output by the multiple experts are fused to obtain the third target feature; According to the reference weights respectively on the invitation prediction task, the reference expert processing features output by the multiple experts are fused to obtain the fourth target feature; By performing probability mapping on the third target feature, a third probability is obtained that the target object invites the lost friend to return. By performing probability mapping on the fourth target feature, a fourth probability is obtained of the churned friend accepting the return invitation from the target object.

7. The method according to claim 3, characterized in that, The step of recalling churned friends according to their recall weight relative to the target object includes: The multiple churned friends of the target object are sorted in descending order of target recall weight to obtain the churned friend ranking. Add the top N churned friends from the churned friend ranking to the churned friend recommendation list of the target object; N is a positive integer; Send the list of lost friends to the client where the target object is located.

8. The method according to any one of claims 1 to 7, characterized in that, The process of fusing the social characteristics of the target object, the object characteristics of the target object, the object characteristics of the target object's churned friends, and the interaction characteristics between the target object and the churned friends to obtain fused characteristics includes: The social features of the target object, the object features of the target object, the object features of the target object's lost friends, and the interaction features between the target object and the lost friends are concatenated to obtain the concatenated features; The spliced ​​features are subjected to multi-head attention cross-processing to obtain cross-fused features; The splicing feature and the cross-fusion feature are subjected to residual processing to obtain the fusion feature.

9. The method according to claim 2, characterized in that, The step of recalling the lost friends based on the first probability and the second probability includes: Multiply the first probability and the second probability to obtain the reference recall weight of the churned friend relative to the target object; The multiple churned friends of the target object are sorted in descending order of reference recall weight to obtain a reference ranking. Add the top N lost friends from the reference sort to the lost friend recommendation list of the target object; N is a positive integer; Send the list of lost friends to the client where the target object is located.

10. The method according to any one of claims 1 to 7 and 9, characterized in that, The method further includes: Multiple training samples are obtained; wherein, a training sample includes a sample object, a reference sample object, a first label and a second label, the first label is used to indicate whether the sample object invites the reference sample object to reflow, and the second label is used to indicate whether the reference sample object accepts the reflow invitation of the sample object; The social features of the sample object, the object features of the sample object, the object features of the reference sample object, and the interaction features between the sample object and the reference sample object are fused to obtain the sample fusion features; The gating network predicts the weights of the multiple expert networks in the invitation prediction task and the invitation acceptance prediction task based on the sample fusion features. According to their respective weights on the invitation prediction task, the sample expert features output by the multiple expert networks based on the sample fusion features are fused to obtain the first sample target features. According to their respective weights on the invitation prediction task, the sample expert features output by the multiple expert networks are fused to obtain the second sample target features. Based on the first sample target features, predict the first sample probability that the sample object will invite the reference sample object to return. Based on the second sample target features, predict the second sample probability that the reference sample object accepts the return invitation from the sample object; Calculate the target loss based on the first sample probability, the first label, the second sample probability, and the second label; Based on the target loss, the parameters of at least the gating network and the plurality of expert networks are adjusted until the first training termination condition is met.

11. The method according to any one of claims 1 to 7 and 9, characterized in that, Before fusing the social characteristics of the target object, the object characteristics of the target object, the object characteristics of the target object's churned friends, and the interaction characteristics between the target object and the churned friends to obtain the fused characteristics, the method further includes: Obtain the target friend social graph constructed based on the social relationships between objects; The target friend social graph is processed by graph encoding using a graph neural network to obtain a graph encoding result; the graph encoding result includes the object features of the objects represented by each node in the target friend social graph; Based on the object features representing multiple friends of the target object in the graph encoding result, the social features of the target object are aggregated to obtain the social features of the target object. The object features of the target object and the object features of the target object's lost friends are obtained from the graph encoding results.

12. A recall processing device, characterized in that, include: The first fusion module is used to fuse the social characteristics of the target object, the object characteristics of the target object, the object characteristics of the target object's lost friends, and the interaction characteristics between the target object and the lost friends to obtain fused characteristics. The task weight prediction module is used by the gating network to predict the weights of multiple expert networks on the invitation prediction task and the acceptance prediction task, respectively, based on the fusion features. The second fusion module is used to fuse the expert processing features output by the multiple expert networks according to their respective weights on the invitation prediction task to obtain a first target feature; the expert processing feature is obtained by the corresponding expert network processing the fused feature; The third fusion module is used to fuse the expert processing features output by the multiple expert networks according to their respective weights on the invitation prediction task to obtain the second target feature. The recall processing module is used to recall the lost friends based on the first target feature and the second target feature.

13. An electronic device, characterized in that, include: processor; A memory storing computer-readable instructions that, when executed by the processor, implement the method as described in any one of claims 1-11.

14. A computer-readable storage medium storing computer-readable instructions thereon, characterized in that, When the computer-readable instructions are executed by a processor, the method as described in any one of claims 1-11 is implemented.

15. A computer program product comprising computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the method of any one of claims 1-11.

Images