A social user recommendation method, apparatus and device

Abstract

The social user recommendation method provided in the application comprises the following steps: obtaining a candidate user set to be recommended to a target user for the target user; calculating a recommendation score of each candidate user; wherein the recommendation score of each candidate user is used to represent the preference degree of the target user to the candidate user; sorting each candidate user in the candidate user set in descending order of the recommendation score to obtain a first sorting result; calculating a retention probability increment of each candidate user; wherein the retention probability increment of each candidate user represents the influence degree of the operation of recommending the candidate user to the target user on the retention probability of the candidate user; sorting each candidate user in the candidate user set in descending order of the retention probability increment to obtain a second sorting result; and performing recommendation according to the first sorting result and the second sorting result. The social user recommendation method, device and equipment provided in the application can improve the experience of the two-sided users.

Description

Technical Field

[0001] This application relates to the field of computer technology, and more particularly to a social user recommendation method, apparatus, and device. Background Technology

[0002] With the development of internet technology, users often need a lot of time to filter the users they like when using the internet for social interaction. To improve user experience, recommendation technology has emerged.

[0003] Currently, the mainstream recommendation technology is mainly based on similarity, that is, recommending strangers with similar interests to users. However, this recommendation method cannot meet users' personalized needs, nor does it take into account the social experience of the recommended users, resulting in a low experience for both sides. Summary of the Invention

[0004] In view of this, this application provides a social user recommendation method, apparatus, and device to provide a recommendation method that simultaneously considers the user experience on both sides. Specifically, this application is implemented through the following technical solution:

[0005] The first aspect of this application provides a social user recommendation method, the method comprising:

[0006] For the target user, obtain a set of candidate users to be recommended to the target user;

[0007] Calculate a recommendation score for each candidate user; wherein, the recommendation score for each candidate user is used to characterize the degree of liking the target user for the candidate user;

[0008] The candidate users in the candidate user set are sorted in descending order of their recommendation scores to obtain a first sorting result.

[0009] Calculate the retention probability increment for each candidate user; wherein, the retention probability increment for each candidate user represents the degree of influence of the operation of recommending the candidate user to the target user on the retention probability of the candidate user;

[0010] The candidate users in the candidate user set are sorted in descending order of retention probability increment to obtain a second sorting result.

[0011] Recommendations are made based on the first and second sorting results.

[0012] A second aspect of this application provides a social user recommendation device, the device comprising an acquisition module, a calculation module, a sorting module, and a recommendation module; wherein...

[0013] The acquisition module is used to acquire a set of candidate users to be recommended to the target user;

[0014] The calculation module is used to calculate a recommendation score for each candidate user; wherein, the recommendation score for each candidate user is used to characterize the degree of liking the target user for the candidate user;

[0015] The sorting module is used to sort each candidate user in the candidate user set according to the recommendation score from high to low, and obtain a first sorting result.

[0016] The calculation module is further configured to calculate the retention probability increment of each candidate user; wherein, the retention probability increment of each candidate user represents the degree of influence of the operation of recommending the candidate user to the target user on the retention probability of the candidate user;

[0017] The sorting module is further configured to sort each candidate user in the candidate user set in descending order of retention probability increment, and obtain a second sorting result.

[0018] The recommendation module is used to make recommendations based on the first ranking result and the second ranking result.

[0019] A third aspect of this application provides a social user recommendation device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of any of the methods provided in the first aspect of this application.

[0020] A fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods provided in the first aspect of this application.

[0021] The social user recommendation method, apparatus, and device provided in this application, targeting a specific user, obtain a set of candidate users to be recommended to the target user, calculate the recommendation score and retention probability increment of each candidate user, then sort the candidate users in the set according to the recommendation score from high to low to obtain a first sorting result, and then sort the candidate users in the set according to the retention probability increment from high to low to obtain a second sorting result, and finally make recommendations based on the first and second sorting results. In this way, by obtaining the first sorting result based on the recommendation score and the second sorting result based on the retention probability increment, and then making recommendations based on these two sorting results, it simultaneously considers the personalized needs of the target user and the retention probability increment of the candidate users, thus taking into account both the preferences of the target user and the feelings of the candidate users, improving the experience of both sides. Attached Figure Description

[0022] Figure 1 A flowchart of an embodiment of the social user recommendation method provided in this application;

[0023] Figure 2 A flowchart of Embodiment 2 of the social user recommendation method provided in this application;

[0024] Figure 3 This is a schematic diagram of the structure of a first model shown in an exemplary embodiment of this application;

[0025] Figure 4 A flowchart of Embodiment 3 of the social user recommendation method provided in this application;

[0026] Figure 5 A flowchart of Embodiment 4 of the social user recommendation method provided in this application;

[0027] Figure 6 A flowchart of Embodiment 5 of the social user recommendation method provided in this application;

[0028] Figure 7 A flowchart of Embodiment Six of the Social User Recommendation Method Provided in this Application;

[0029] Figure 8 This is a hardware structure diagram of the social user recommendation device in the social user recommendation device of this application;

[0030] Figure 9 This is a schematic diagram of the structure of a first embodiment of the social user recommendation device provided in this application. Detailed Implementation

[0031] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0032] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0033] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0034] This application provides a social user recommendation method, apparatus, and device to provide a recommendation method that simultaneously considers the user experience on both sides.

[0035] The social user recommendation method, apparatus, and device provided in this application, targeting a specific user, obtain a set of candidate users to be recommended to the target user, calculate the recommendation score and retention probability increment of each candidate user, then sort the candidate users in the set according to the recommendation score from high to low to obtain a first sorting result, and then sort the candidate users in the set according to the retention probability increment from high to low to obtain a second sorting result, and finally make recommendations based on the first and second sorting results. In this way, by obtaining the first sorting result based on the recommendation score and the second sorting result based on the retention probability increment, and then making recommendations based on these two sorting results, it simultaneously considers the personalized needs of the target user and the retention probability increment of the candidate users, thus taking into account both the preferences of the target user and the feelings of the candidate users, improving the experience of both sides.

[0036] The following specific embodiments are given to illustrate the technical solution of this application in detail.

[0037] Figure 1 This is a flowchart of an embodiment of the social user recommendation method provided in this application. Please refer to... Figure 1 The method provided in this embodiment may include:

[0038] S101. For the target user, obtain the set of candidate users to be recommended to the target user.

[0039] Specifically, the candidate user set is a set of users selected from a massive user base according to preset filtering rules. It should be noted that the preset filtering rules are set according to actual needs, and this embodiment will not describe them. For example, in one embodiment, the preset filtering rule could be: registered users are candidate users.

[0040] It should be noted that the social user recommendation method and apparatus provided in this application are applied to a social user recommendation device, which can be a server that interacts with the client.

[0041] Furthermore, the target user can be the user corresponding to the user identifier carried in the currently received recommendation request. Specifically, when a recommendation request is received from the client, a candidate user set can be determined for the target user on the client side.

[0042] For example, in one embodiment, the candidate user set includes 5 users, namely user 1, user 2, ..., user 5.

[0043] S102. Calculate the recommendation score for each candidate user; wherein, the recommendation score for each candidate user is used to characterize the degree of preference of the target user for the candidate user.

[0044] The recommendation score represents the target user's degree of liking for the candidate user; the higher the recommendation score, the greater the target user's degree of liking for the candidate user.

[0045] Specifically, in one possible implementation, for a given candidate user, a similarity value between the target user and the candidate user can be calculated, and this similarity value can then be used as the recommendation score for the candidate user. It should be noted that the specific implementation principles and processes for calculating the similarity value can be found in descriptions in related technologies, and will not be elaborated upon here.

[0046] S103. Sort each candidate user in the candidate user set according to the recommendation score from high to low to obtain the first sorting result.

[0047] Specifically, after obtaining the recommendation scores of each candidate user in the candidate user set, the candidate users corresponding to the recommendation scores are sorted in descending order of recommendation scores, and this order is the first sorting result.

[0048] S104. Calculate the retention probability increment for each candidate user; wherein, the retention probability increment for each candidate user represents the degree of influence of the operation of recommending the candidate user to the target user on the retention probability of the candidate user.

[0049] Specifically, the retention probability increment of each candidate user represents the degree of influence of the operation of recommending the candidate user to the target user on the candidate user's retention probability. In other words, the higher the retention probability increment of a candidate user, the greater the influence of the operation of recommending the candidate user to the target user on the candidate user's retention probability.

[0050] It's important to note that the retention probability of a candidate user refers to the probability that the candidate user will remain in the future. In other words, the retention probability of a candidate user is used to characterize the stickiness and loyalty of the candidate user; a higher retention probability means that the candidate user will remain active and stay in the future.

[0051] Furthermore, in one possible implementation, when calculating the retention probability increment of a candidate user, the retention probability of the user can be estimated based on the candidate user's historical interaction data, and then the retention probability of the candidate user can be predicted after recommending the candidate user to the target user. Finally, the difference between the two retention probabilities is determined as the retention probability increment of the candidate user.

[0052] S105. Sort each candidate user in the candidate user set in descending order of retention probability increment to obtain a second sorting result.

[0053] Specifically, after obtaining the retention probability increment of candidate users, the candidate users are sorted in descending order of retention probability increment, and this order is the second sorting result.

[0054] S106. Make recommendations based on the first sorting result and the second sorting result.

[0055] Specifically, for example, in one embodiment, when making recommendations based on the first ranking result and the second ranking result, the top 10 users in the first ranking result and the top 10 users in the second ranking result can be selected and recommended to the target user.

[0056] For example, in another embodiment, in making recommendations based on the first ranking result and the second ranking result, users who are included in both the first ranking result and the second ranking result can be found from the top 10 users in the first ranking result and the top 10 users in the second ranking result, and then the found users can be recommended to the target users.

[0057] The social user recommendation method provided in this embodiment, targeting a specific user, obtains a set of candidate users to be recommended to the target user, calculates the recommendation score and retention probability increment of each candidate user, and then sorts the candidate users in the set according to the recommendation score from high to low to obtain a first sorting result. Then, it sorts the candidate users in the set according to the retention probability increment from high to low to obtain a second sorting result. Finally, recommendations are made based on both the first and second sorting results. By obtaining the first sorting result based on recommendation score and the second sorting result based on retention probability increment, and then making recommendations based on these two sorting results, the method simultaneously considers the personalized needs of the target user and the retention probability increment of the candidate users, thus balancing the preferences of both the target user and the feelings of the candidate users and improving the experience for both sides.

[0058] Figure 2 The flowchart is for Embodiment 2 of the social user recommendation method provided in this application. Please refer to... Figure 2 The method provided in this embodiment, based on the above embodiments, calculates the recommendation score for candidate users, and may include:

[0059] S201. For each candidate user, perform feature extraction on the target user and the candidate user to obtain a concatenated feature vector of the target user and the candidate user.

[0060] Specifically, the concatenated feature vector is a vector formed by concatenating the feature vectors of the target user and the candidate users. The feature vector of the target user is obtained based on the target user's historical interaction data, and the feature vector of the candidate user is obtained based on the candidate user's historical interaction data. It should be noted that the historical interaction data includes data generated by at least one of the following interaction behaviors: liking, chatting, and matching. For example, in one embodiment, the historical interaction data includes data generated by liking, chatting, and matching. Furthermore, the historical interaction data may also include attribute information. For example, the historical interaction data may also include user ID, age, gender, etc.

[0061] Furthermore, traditional feature extraction methods or neural network-based feature extraction methods can be used to extract features from historical interaction data. For example, collaborative filtering-based feature extraction methods can be used. Alternatively, a network for feature extraction can be employed to extract features from historical interaction data.

[0062] S202. The concatenated feature vector is input into a pre-trained first model for calculating the predicted probabilities corresponding to multiple first recommendation indicators, so that the first model outputs the predicted probability corresponding to each first recommendation indicator; wherein, the first recommendation indicator includes a liking dimension, a pairing dimension, and a chat dimension.

[0063] Specifically, Figure 3 This is a schematic diagram illustrating the structure of a first model as an exemplary embodiment of this application. Please refer to... Figure 3 ,exist Figure 3 In the example shown, the first model may include a shared expert network and a first task network corresponding to the "like" dimension, a second task network corresponding to the "chat" dimension, and a third task network corresponding to the "pairing" dimension. Each task network includes an expert network, a gating network, and a prediction network. It should be noted that the first task network is used to obtain the predicted probability corresponding to the "like" dimension, the second task network is used to obtain the predicted probability corresponding to the "chat" dimension, and the third task network is used to obtain the predicted probability corresponding to the "pairing" dimension.

[0064] Combination Figure 3 The following describes the specific calculation process for "the first model calculating the predicted probability corresponding to each first recommendation indicator":

[0065] Specifically, the process by which the first model calculates the predicted probability for each first recommendation indicator may include:

[0066] (1) For each task network, based on the spliced ​​feature vector, the exclusive expert vector of the recommendation index corresponding to the task network is obtained through the expert network in the task network, and the weight vector of the recommendation index corresponding to the task network is obtained through the gating network in the task network.

[0067] Specifically, in one embodiment, in conjunction with the above description, the expert network 11 in the first task network is used to obtain the exclusive expert vector A1 corresponding to the liking dimension, and the gating network 12 in the first task network is used to obtain the weight vector B1 corresponding to the liking dimension.

[0068] Similarly, expert network 21 in the second task network is used to obtain the exclusive expert vector A2 corresponding to the chat dimension, and gate network 22 in the second task network is used to obtain the weight vector B2 corresponding to the chat dimension.

[0069] Furthermore, the expert network 31 in the third task network is used to obtain the exclusive expert vector A3 corresponding to the paired dimension, and the gate network 32 in the third task network is used to obtain the weight vector B3 corresponding to the paired dimension.

[0070] It should be noted that, in the process of obtaining exclusive expert vectors based on concatenated feature vectors, the expert network extracts the feature vector corresponding to the recommendation metric from the numerous data points of the concatenated feature vectors. Furthermore, referring to the preceding description, each task network is used to perform a specific prediction task, and the features extracted by each expert network correspond to the corresponding prediction task; in other words, each expert network extracts features from a different perspective.

[0071] For details on how expert networks obtain exclusive expert vectors and how gating networks obtain weight vectors, please refer to the descriptions in relevant technologies; they will not be repeated here.

[0072] (2) The spliced ​​feature vector is extracted by the shared expert network to obtain a globally shared expert vector.

[0073] Please continue to refer to Figure 3 Following the example above, in one embodiment, after obtaining the concatenated feature vector, the concatenated feature vector is input into the first model from the input port. The shared expert network generates a globally shared expert vector based on the concatenated feature vector. For ease of distinction, the shared expert vector is denoted by A0.

[0074] (3) For each recommendation indicator, based on the weight vector corresponding to the recommendation indicator, the exclusive expert vector and the shared expert vector corresponding to the recommendation indicator are weighted to obtain the weighted feature vector corresponding to the recommendation indicator; and the weighted feature vector is predicted by the prediction network to obtain the prediction probability corresponding to the recommendation indicator.

[0075] For details, please continue to refer to... Figure 3 Combining the examples above, the result obtained by weighting the exclusive expert vector and the shared expert vector based on the weight vector is the weighted feature vector, which is represented by C1, C2, and C3.

[0076] Combining the above example, for instance, regarding the "like" dimension, we perform a weighted operation on the exclusive expert vector A1 and the shared expert vector A0 output by the shared expert network to obtain the weighted feature vector C1 corresponding to the "like" dimension.

[0077] Furthermore, after obtaining the weighted feature vector, it is input into the prediction network 13 for prediction, obtaining the predicted probability corresponding to the recommendation index. For example, combining the above example, for the first task network, the predicted probability 1 corresponding to the like dimension is obtained.

[0078] Correspondingly, Task Network 2 outputs the predicted probability 2 for the chat dimension, and Task Network 3 outputs the predicted probability 3 for the pairing dimension.

[0079] S203. Based on the weights pre-set for each of the first recommendation indicators, the predicted probabilities corresponding to each of the first recommendation indicators are weighted, and the processing result obtained by weighting is determined as the recommendation score of the candidate user.

[0080] Specifically, the weights pre-set for each primary recommendation indicator are determined based on actual needs. In this embodiment, the specific values ​​of the pre-set weights for each primary recommendation indicator are not limited. For example, in specific settings, primary recommendation indicators that have a greater impact on users can be assigned higher weights.

[0081] In practice, the recommendation score for each candidate user can be calculated using the following formula:

[0082] F=w1*p(like)+w2*p(chat)+w3*p(match)

[0083] Where F is the recommendation score, p(like) is the predicted probability corresponding to the like dimension, p(chat) is the predicted probability corresponding to the chat dimension, and p(match) is the predicted probability corresponding to the chat dimension; w1, w2 and w3 are the weights corresponding to the like dimension, chat dimension and match dimension, respectively.

[0084] The social user recommendation method provided in this embodiment, for each candidate user, extracts features from both the target user and the candidate user to obtain a concatenated feature vector. This concatenated feature vector is then input into a pre-trained first model for calculating the predicted probabilities corresponding to multiple first recommendation indicators. The first model outputs the predicted probability for each first recommendation indicator. Furthermore, based on pre-set weights for each first recommendation indicator, the predicted probabilities are weighted, and the weighted result is determined as the recommendation score for the candidate user. This provides a method for calculating recommendation scores that considers multiple first recommendation indicators simultaneously. The calculated recommendation score more accurately represents the target user's preferences, and recommendations based on this score can better meet the personalized needs of the target user, thus improving the user experience.

[0085] Figure 4The flowchart for Embodiment 3 of the social user recommendation method provided in this application is shown below. Please refer to... Figure 4 The method provided in this embodiment, based on the above embodiments, includes the following step: calculating the retention probability increment of the candidate user.

[0086] S401. For each candidate user, calculate the first retention probability of the candidate user under a first assumption and the second retention probability of the candidate user under a second assumption; the first assumption is that the candidate user will not be recommended to the target user this time; the second assumption is that the candidate user will be recommended to the target user this time.

[0087] In specific implementation, for each candidate user, the user's first historical interaction data under the first assumption can be obtained. This first historical interaction data may include attribute information, the candidate user's historical interaction data within a first specified time period, and the candidate user's real-time interaction data within a second specified time period. Furthermore, based on the first historical interaction data, the candidate user's first retention probability under the first assumption is estimated.

[0088] It should be noted that the first and second specified time periods are set according to actual needs, and this embodiment does not limit them.

[0089] Furthermore, it can be assumed that the candidate user will be recommended to the target user this time. Based on this assumption, the real-time interaction data in the first historical interaction data is updated to obtain the second historical interaction data of the candidate user under the second assumption. Based on the second historical interaction data, the second retention probability of the candidate user under the second assumption can be estimated.

[0090] S402, The difference between the second retention probability and the first retention probability is determined as the retention probability increment of the candidate user.

[0091] In practice, the retention probability increment can be calculated using the following formula:

[0092] L0 = L2 - L1

[0093] Where L0 is the retention probability increment, L1 is the first retention probability, and L2 is the second retention probability.

[0094] The social user recommendation method provided in this embodiment calculates a first retention probability under a first assumption and a second retention probability under a second assumption for each candidate user. The first assumption is that candidate users will not be recommended to the target user in this instance, while the second assumption is that candidate users will be recommended to the target user. The difference between the second retention probability and the first retention probability is then determined as the candidate user's retention probability increment. This provides a method for calculating the retention probability increment, allowing for recommendations based on this increment. Prioritizing candidate users with higher retention probability increments to the target user improves the user experience for candidate users.

[0095] Figure 5 The flowchart for Embodiment 4 of the social user recommendation method provided in this application is shown below. Please refer to... Figure 5 The method provided in this embodiment, based on the above embodiments, may include the following steps for obtaining the first retention probability and the second retention probability:

[0096] S501. Obtain the first feature vector of the candidate user under the first assumption; wherein, the first feature vector includes an attribute information feature vector, historical feature vectors corresponding to multiple second recommendation indicators, and real-time feature vectors corresponding to the multiple second recommendation indicators; the multiple second recommendation indicators include the multiple first recommendation indicators and the exposure dimension.

[0097] It should be noted that the second recommendation metric includes the dimensions of liking, chatting, matching, and exposure.

[0098] Furthermore, the implementation process of this step may specifically include:

[0099] (1) Obtain the static attribute information of the candidate users.

[0100] Specifically, the static attribute information of candidate users may include the candidate user's ID, age, gender, device information, and geographical location.

[0101] (2) Obtain the historical interaction data of the candidate user within a first specified time period and the real-time interaction data of the candidate user within a second specified time period.

[0102] Specifically, the first specified time period and the second specified time period are set according to actual needs, and they can be the same or different. In this embodiment, the specific duration of the first specified time period and the second specified time period is not limited. For example, in one embodiment, the first specified time period can be one month, and the second specified time period can be a time period shorter than the first specified time period, such as one day.

[0103] (3) Extract features from the static attribute information to obtain the attribute information feature vector of the candidate user.

[0104] (4) Extract the features corresponding to the multiple second recommendation indicators from the historical interaction data to obtain the historical feature vectors corresponding to the multiple second recommendation indicators.

[0105] (5) Extract the features corresponding to the multiple second recommendation indicators from the real-time interactive data to obtain the real-time feature vectors corresponding to the multiple second recommendation indicators.

[0106] For the specific implementation methods and principles of feature extraction in steps (3) to (5), please refer to the previous introduction on feature extraction, which will not be repeated here.

[0107] (6) The combination of the attribute information feature vector, the historical interaction feature vector and the real-time interaction feature vector is determined as the first feature vector.

[0108] S502. Update the real-time feature vector corresponding to the exposure dimension, and based on the real-time feature vector and the predicted probability corresponding to each of the first recommendation indicators, update the real-time feature vector corresponding to each of the first recommendation indicators to obtain the updated real-time feature vector.

[0109] Specifically, assuming that candidate users are recommended to the target user this time, when the real-time feature vector corresponding to the exposure dimension is represented by the number of exposures, the real-time feature vector corresponding to the exposure dimension is updated to the current value plus 1.

[0110] Furthermore, after updating the real-time feature vector corresponding to the exposure dimension, the real-time feature vector corresponding to each first recommendation metric is updated. For example, for the like dimension, when the real-time feature vector corresponding to the like dimension is represented by the number of likes, the real-time feature vector corresponding to this dimension is updated to the current value plus the predicted probability corresponding to the like dimension.

[0111] Accordingly, for the matching dimension, when the real-time feature vector corresponding to the matching dimension is represented by the number of matches, the real-time feature vector corresponding to that dimension is updated to the current value plus the predicted probability corresponding to the matching dimension; for the chat dimension, when the real-time feature vector corresponding to the chat dimension is represented by the number of chats, the real-time feature vector corresponding to that dimension is updated to the current value plus the predicted probability corresponding to the chat dimension.

[0112] S503. The combination of the attribute information feature vector, the historical feature vector, and the updated real-time feature vector is determined as the second feature vector of the candidate user under the second assumption.

[0113] S504. The first feature vector and the second feature vector are respectively used as input feature vectors and input into a pre-trained second model for calculating the retention probability, so that the second model outputs the retention probability corresponding to the input feature vector based on the input feature vector.

[0114] Specifically, the second model can output the retention probability based on the input feature vector. For details on the implementation principles and training methods of the second model, please refer to the relevant descriptions in related technologies; they will not be elaborated upon here.

[0115] The social user recommendation method provided in this embodiment obtains the first feature vector of a candidate user under a first assumption, then updates the real-time feature vector corresponding to the exposure dimension. Based on the real-time feature vector and the predicted probability corresponding to each first recommendation indicator, the real-time feature vector corresponding to each first recommendation indicator is updated to obtain the updated real-time feature vector. The combination of the attribute information feature vector, historical feature vector, and updated real-time feature vector is then determined as the second feature vector of the candidate user under the second assumption. Finally, the first and second feature vectors are respectively used as input feature vectors and input into a pre-trained second model for calculating retention probabilities. The second model outputs the retention probability corresponding to the input feature vector based on the input feature vector. In this way, by calculating the retention probability of the candidate user under the first and second assumptions, and then calculating the retention probability increment, recommendations are made based on the retention probability increment, taking into account the candidate user's feelings and improving the candidate user experience.

[0116] Figure 6 The flowchart for Embodiment 5 of the social user recommendation method provided in this application is shown below. Please refer to... Figure 6 The method provided in this embodiment, based on the above embodiments, includes the following steps for making recommendations according to the first sorting result and the second sorting result:

[0117] S601. Calculate the comprehensive sorting result based on the first sorting result and the second sorting result.

[0118] Specifically, for each candidate user, the comprehensive ranking number can be calculated based on the candidate user's first ranking number in the first ranking result, the candidate user's second ranking number in the second ranking result, and the weights pre-set for the first and second ranking results. In other words, for each candidate user, the comprehensive ranking number can be calculated according to the following formula:

[0119] R0 = A1*R1 + A2*R2

[0120] Where R1 and R2 are the first and second sorting numbers, respectively, A1 and A2 are the weights of the first and second sorting results, and R0 is the comprehensive sorting number.

[0121] It should be noted that the weights pre-set for the first and second sorting results are determined according to actual needs, and this embodiment does not limit this. For example, in one embodiment, the weights of the first and second sorting results are 30% and 70%, respectively; in another embodiment, the weights of the first and second sorting results are 50% and 50%, respectively.

[0122] S602. Based on the comprehensive ranking result, select a specified number of candidate users from the candidate user set as recommended users.

[0123] Specifically, the exact value of the specified quantity is set according to actual needs; in this embodiment, the exact value of the specified quantity is not limited. For example, in one embodiment, the specified quantity is 60. In this case, after determining the comprehensive ranking result in this step, the top 60 candidate users are selected as recommended users according to the order of the comprehensive ranking result from front to back.

[0124] S603. Recommend the recommended user to the target user.

[0125] For example, using the example above, the 60 recommended users obtained can be recommended to the target user.

[0126] The social user recommendation method provided in this embodiment first calculates a comprehensive ranking result based on a first ranking result and a second ranking result. Then, based on the comprehensive ranking result, a specified number of candidate users are selected from the candidate user set as recommended users, and these recommended users are then recommended to the target user. This recommendation method can simultaneously consider the feelings of both sides of the user base, improving the overall user experience.

[0127] Figure 7 The flowchart for Embodiment Six of the Social User Recommendation Method Provided in this Application is shown below. Please refer to... Figure 7 The method provided in this embodiment may include:

[0128] S701. For the target user, obtain the set of candidate users to be recommended to the target user.

[0129] S702. For each candidate user, perform feature extraction on the target user and the candidate user to obtain a concatenated feature vector of the target user and the candidate user.

[0130] S703. The concatenated feature vector is input into a pre-trained first model for calculating the predicted probabilities corresponding to multiple first recommendation indicators, so that the first model outputs the predicted probability corresponding to each first recommendation indicator; wherein, the first recommendation indicator includes a liking dimension, a pairing dimension, and a chat dimension.

[0131] S704. Based on the weights pre-set for each of the first recommendation indicators, the predicted probabilities corresponding to each of the first recommendation indicators are weighted, and the processing result obtained by weighting is determined as the recommendation score of the candidate user.

[0132] S705. Sort the candidate users in the candidate user set according to the recommendation scores from high to low to obtain the first sorting result.

[0133] S706. For each candidate user, calculate the first retention probability of the candidate user under a first assumption and the second retention probability of the candidate user under a second assumption; the first assumption is that the candidate user will not be recommended to the target user this time; the second assumption is that the candidate user will be recommended to the target user this time.

[0134] It should be noted that in this embodiment, the execution order of steps S706 and S702 is not limited, and steps S706 and S702 can be executed simultaneously.

[0135] S707. The difference between the second retention probability and the first retention probability is determined as the retention probability increment of the candidate user.

[0136] S708. Sort the candidate users in the candidate user set in descending order of retention probability increment to obtain a second sorting result.

[0137] S709. Calculate the comprehensive sorting result based on the first sorting result and the second sorting result.

[0138] S710. Based on the comprehensive ranking result, select a specified number of candidate users from the candidate user set as recommended users.

[0139] S711. Recommend the recommended user to the target user.

[0140] For details on the specific implementation principles and processes of each step, please refer to the descriptions in the previous embodiments; they will not be repeated here.

[0141] The social user recommendation method provided in this embodiment, targeting a target user, obtains a set of candidate users to be recommended to the target user. Then, for each candidate user, features are extracted from both the target user and the candidate users to obtain a concatenated feature vector. This concatenated feature vector is input into a first model, which outputs the predicted probability corresponding to each first recommendation indicator. Then, based on pre-set weights for each first recommendation indicator, the predicted probabilities are weighted to obtain a recommendation score for each candidate user. The candidate users are then sorted in descending order of recommendation scores to obtain a first ranking result. For each candidate user, a first retention probability under a first assumption and a second retention probability under a second assumption are calculated. The difference between the second and first retention probabilities is determined as the retention probability increment for the candidate user. The candidate users are then sorted in descending order of retention probability increments to obtain a second ranking result. Finally, based on the first and second ranking results, a comprehensive ranking result is calculated. According to this comprehensive ranking result, a specified number of candidate users are selected from the candidate user set as recommended users, and these recommended users are then recommended to the target user. In this way, by obtaining a first ranking result based on recommendation scores and a second ranking result based on retention probability increments, and then making recommendations based on these two ranking results, the personalized needs of the target user and the retention probability increments of the candidate users are considered simultaneously. This approach takes into account both the preferences of the target user and the feelings of the candidate users, thus improving the experience for both sides.

[0142] Corresponding to the aforementioned embodiment of a social user recommendation method, this application also provides an embodiment of a social user recommendation device.

[0143] An embodiment of a social user recommendation device disclosed in this application can be applied to a social user recommendation device. The device embodiment can be implemented through software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of the social user recommendation device loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, such as... Figure 8 As shown, Figure 8 This is a hardware structure diagram of the social user recommendation device in this application, except for... Figure 8 In addition to the processor, memory, network interface, and non-volatile memory shown, the social user recommendation device in the embodiment may also include other hardware depending on the actual function of the social user recommendation device, which will not be described in detail here.

[0144] Figure 9 This is a schematic diagram of the structure of a first embodiment of the social user recommendation device provided in this application. Please refer to... Figure 9 The apparatus provided in this embodiment may include an acquisition module 910, a calculation module 920, a sorting module 930, and a recommendation module 940; wherein;

[0145] The acquisition module 910 is used to acquire a set of candidate users to be recommended to the target user for the target user.

[0146] The calculation module 920 is used to calculate the recommendation score for each candidate user; wherein, the recommendation score for each candidate user is used to characterize the degree of liking of the target user for the candidate user;

[0147] The sorting module 930 is used to sort each candidate user in the candidate user set according to the order of recommendation scores from high to low, and obtain a first sorting result.

[0148] The calculation module 920 is further configured to calculate the retention probability increment of each candidate user; wherein, the retention probability increment of each candidate user represents the degree of influence of the operation of recommending the candidate user to the target user on the retention probability of the candidate user;

[0149] The sorting module 930 is further configured to sort each candidate user in the candidate user set in descending order of retention probability increment, and obtain a second sorting result.

[0150] The recommendation module 940 is used to make recommendations based on the first ranking result and the second ranking result.

[0151] The apparatus of this embodiment can be used to perform... Figure 1 The steps of the method embodiment shown are similar in principle and process, and will not be repeated here.

[0152] Optionally, the calculation module 920 is specifically used to calculate, for each candidate user, a first retention probability under a first assumption and a second retention probability under a second assumption; the first assumption is that the candidate user will not be recommended to the target user this time; the second assumption is that the candidate user will be recommended to the target user this time.

[0153] The calculation module 920 is further configured to determine the difference between the second retention probability and the first retention probability as the retention probability increment of the candidate user.

[0154] Optionally, the recommendation module 940 is specifically used to calculate a comprehensive ranking result based on the first ranking result and the second ranking result;

[0155] The recommendation module 940 is further configured to select a specified number of candidate users as recommended users from the candidate user set based on the comprehensive ranking result.

[0156] The recommendation module 940 is also used to recommend the recommended user to the target user.

[0157] Optionally, the calculation module 920 is specifically used to extract features from the target user and the candidate user for each candidate user, and obtain a concatenated feature vector of the target user and the candidate user;

[0158] The calculation module 920 is further configured to input the concatenated feature vector into a pre-trained first model for calculating the predicted probabilities corresponding to multiple first recommendation indicators, so that the first model outputs the predicted probability corresponding to each first recommendation indicator; wherein, the first recommendation indicator includes a liking dimension, a pairing dimension, and a chat dimension;

[0159] The calculation module 920 is further configured to perform weighted processing on the predicted probabilities corresponding to each of the first recommendation indicators according to the weights pre-set for each of the first recommendation indicators, and determine the processing result obtained by weighted processing as the recommendation score of the candidate user.

[0160] Optionally, the calculation module 920 is specifically used to obtain the first feature vector of the candidate user under the first assumption; wherein, the first feature vector includes an attribute information feature vector, historical feature vectors corresponding to multiple second recommendation indicators, and real-time feature vectors corresponding to the multiple second recommendation indicators; the multiple second recommendation indicators include the multiple first recommendation indicators and the exposure dimension;

[0161] The calculation module 920 is also used to update the real-time feature vector corresponding to the exposure dimension, and based on the real-time feature vector and the predicted probability corresponding to each first recommendation indicator, update the real-time feature vector corresponding to each first recommendation indicator to obtain the updated real-time feature vector.

[0162] The calculation module 920 is further configured to determine the combination of the attribute information feature vector, the historical feature vector, and the updated real-time feature vector as the second feature vector of the candidate user under the second assumption condition;

[0163] The calculation module 920 is further configured to input the first feature vector and the second feature vector as input feature vectors into a pre-trained second model for calculating the retention probability, so that the second model outputs the retention probability corresponding to the input feature vector based on the input feature vector.

[0164] Optionally, the calculation module 920 is specifically used to obtain the static attribute information of the candidate user, the historical interaction data of the candidate user within a first specified time period, and the real-time interaction data of the candidate user within a second specified time period.

[0165] The calculation module 920 is also used to extract features from the static attribute information to obtain the attribute information feature vector of the candidate user.

[0166] The calculation module 920 is further configured to extract features corresponding to the plurality of second recommendation indicators from the historical interaction data to obtain historical feature vectors corresponding to the plurality of second recommendation indicators;

[0167] The calculation module 920 is further configured to extract features corresponding to the plurality of second recommendation indicators from the real-time interactive data to obtain real-time feature vectors corresponding to the plurality of second recommendation indicators;

[0168] The calculation module 920 is further configured to determine the combination of the attribute information feature vector, the historical interaction feature vector, and the real-time interaction feature vector as the first feature vector.

[0169] Please continue to refer to Figure 8 This application also provides a social user recommendation device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of any of the methods provided in the first aspect of this application.

[0170] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods provided in this application.

[0171] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0172] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0173] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A social user recommendation method, characterized by, The method comprises: For a target user, a set of candidate users to be recommended to the target user is obtained; A recommendation score of each candidate user is calculated; wherein the recommendation score of each candidate user represents the degree of preference of the target user for the candidate user; wherein the calculation of the recommendation score of each candidate user comprises: For each candidate user, feature extraction is performed on the target user and the candidate user to obtain a spliced feature vector of the target user and the candidate user; The spliced feature vector is input into a pre-trained first model for calculating the prediction probability corresponding to a plurality of first recommendation indicators, so that the first model outputs the prediction probability corresponding to each first recommendation indicator; wherein the first recommendation indicators include like dimension, pairing dimension and chat dimension; According to the weight set in advance for each first recommendation indicator, the prediction probability corresponding to each first recommendation indicator is weighted, and the processing result obtained by weighting is determined as the recommendation score of the candidate user; According to the order of recommendation score from high to low, each candidate user in the set of candidate users is sorted to obtain a first sorting result; The retention probability increment of each candidate user is calculated; wherein the retention probability increment of each candidate user represents the influence degree of the operation of recommending the candidate user to the target user on the retention probability of the candidate user; wherein the calculation of the retention probability increment of each candidate user comprises: For each candidate user, a first retention probability of the candidate user under a first hypothetical condition and a second retention probability of the candidate user under a second hypothetical condition are calculated; the first hypothetical condition is that the candidate user is not recommended to the target user this time; the second hypothetical condition is that the candidate user is recommended to the target user this time; The difference between the second retention probability and the first retention probability is determined as the retention probability increment of the candidate user; According to the order of retention probability increment from high to low, each candidate user in the set of candidate users is sorted to obtain a second sorting result; Recommendation is made according to the first sorting result and the second sorting result.

2. The method of claim 1, wherein, The recommendation according to the first sorting result and the second sorting result comprises: According to the first sorting result and the second sorting result, a comprehensive sorting result is calculated; According to the comprehensive sorting result, a specified number of candidate users are selected from the set of candidate users as recommended users; The recommended users are recommended to the target user.

3. The method of claim 1, wherein, The calculation of the first retention probability of the candidate user under the first hypothetical condition and the second retention probability of the candidate user under the second hypothetical condition comprises: A first feature vector of the candidate user under the first hypothetical condition is obtained; wherein the first feature vector comprises an attribute information feature vector, a plurality of second recommendation indicator corresponding historical feature vectors and a plurality of second recommendation indicator corresponding real-time feature vectors; the plurality of second recommendation indicators include the plurality of first recommendation indicators and exposure dimension; updating a real-time feature vector corresponding to the exposure dimension, and updating a real-time feature vector corresponding to each of the first recommendation indicators based on the real-time feature vector and the predicted probability corresponding to each of the first recommendation indicators, to obtain an updated real-time feature vector; determining a combination of the attribute information feature vector, the historical feature vector and the updated real-time feature vector as a second feature vector of the candidate user under the second assumed condition; inputting the first feature vector and the second feature vector as input feature vectors into a second model pre-trained for calculating a retention probability, so as to output, by the second model, a retention probability corresponding to the input feature vector based on the input feature vector.

4. The method of claim 3, wherein, The obtaining of the first feature vector of the candidate user under the first assumed condition comprises: obtaining static attribute information of the candidate user, historical interaction data of the candidate user within a first specified time period and real-time interaction data of the candidate user within a second specified time period; performing feature extraction on the static attribute information to obtain an attribute information feature vector of the candidate user; extracting features corresponding to the plurality of second recommendation indicators from the historical interaction data to obtain historical feature vectors corresponding to the plurality of second recommendation indicators; extracting features corresponding to the plurality of second recommendation indicators from the real-time interaction data to obtain real-time feature vectors corresponding to the plurality of second recommendation indicators; determining a combination of the attribute information feature vector, the historical interaction feature vector and the real-time interaction feature vector as the first feature vector.

5. A social user recommendation apparatus, comprising: The device comprises an obtaining module, a calculating module, a sorting module and a recommending module; wherein, The obtaining module is configured to, for a target user, obtain a candidate user set to be recommended to the target user; The calculating module is configured to calculate a recommendation score of each candidate user; wherein, the recommendation score of each candidate user is used to represent a preference degree of the target user for the candidate user; wherein, the calculation of the recommendation score of each candidate user comprises: performing feature extraction on the target user and the candidate user to obtain a spliced feature vector of the target user and the candidate user for each candidate user; inputting the spliced feature vector into a first model pre-trained for calculating predicted probabilities corresponding to a plurality of first recommendation indicators, so as to output, by the first model, a predicted probability corresponding to each of the first recommendation indicators; wherein, the first recommendation indicators comprise a like dimension, a pairing dimension and a chat dimension; performing weighted processing on the predicted probability corresponding to each of the first recommendation indicators according to weights pre-set for each of the first recommendation indicators, and determining a processing result obtained by the weighted processing as the recommendation score of the candidate user; The sorting module is configured to sort each candidate user in the candidate user set in a descending order of recommendation score, to obtain a first sorting result. The computing module is further configured to calculate a retention probability increment of each candidate user, wherein the retention probability increment of each candidate user represents an influence degree of an operation of recommending the candidate user to the target user this time on a retention probability of the candidate user, and wherein the computing of the retention probability increment of each candidate user comprises: For each candidate user, the computing module is configured to calculate a first retention probability of the candidate user under a first hypothetical condition and a second retention probability of the candidate user under a second hypothetical condition, wherein the first hypothetical condition is that the candidate user is not recommended to the target user this time, and the second hypothetical condition is that the candidate user is recommended to the target user this time. The computing module is further configured to determine a difference between the second retention probability and the first retention probability as the retention probability increment of the candidate user. The sorting module is further configured to sort each candidate user in the candidate user set in a descending order of the retention probability increment to obtain a second sorting result. The recommendation module is configured to perform recommendation according to the first sorting result and the second sorting result.

6. The apparatus of claim 5, wherein, The computing module is specifically configured to, for each candidate user, calculate a first retention probability of the candidate user under a first hypothetical condition and a second retention probability of the candidate user under a second hypothetical condition. The first hypothetical condition is that the candidate user is not recommended to the target user this time, and the second hypothetical condition is that the candidate user is recommended to the target user this time. The computing module is further configured to determine a difference between the second retention probability and the first retention probability as the retention probability increment of the candidate user.

7. A social user recommendation device, comprising: The computer program is stored in the memory and executable on the processor, and when the processor executes the program, the steps of the method in any one of claims 1-4 are implemented.

8. A computer readable storage medium having stored thereon a computer program, characterized in that, The program is executable on the processor to implement the steps of the method in any one of claims 1-4.

Images