Account recall method and device, storage medium and electronic device

Abstract

The application discloses a recall method and device of an account, a storage medium and an electronic device. The method comprises the following steps: determining account state information of a target account to be queued; based on the account state information of the target account, acquiring a subset of recalled accounts recalled by each of at least two recall channels for the target account; performing conditional filtering on the acquired subset of recalled accounts according to a queuing strategy of the target account, to obtain a filtered subset of recalled accounts; and merging the filtered subset of recalled accounts, to obtain a result account set for queuing with the target account. The application solves the technical problem of low recall efficiency caused by inaccurate recall results due to the single and one-sided reference dimension of the recall rules.

Description

Technical Field

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

[0002] In Multiplayer Online Battle Arena (MOBA) games, users typically team up with other users to complete game tasks. Currently, in existing MOBA game team recommendation systems, after an inviting account triggers a team invitation request, it usually first recalls a subset of suitable candidate player accounts from a massive pool of player accounts based on preset random recall rules. These candidate accounts are then recommended to the inviting account to facilitate the selection of target accounts for team formation. These random recall rules are usually based on the player account's status, such as its online status and whether it has permission to be invited.

[0003] However, because the account status of the player accounts referenced in setting the above recall rules only passively reflects the team matching status of player accounts, ignoring the player accounts' own subjective willingness to team up, the results of the recall based on the above recall rules often include candidate player accounts that have no habit of teaming up or are not interested in teaming up. When these candidate player accounts without subjective willingness to team up are recommended to the inviting account, even if the inviting account triggers a team invitation, it may receive a response of refusing to team up, resulting in team failure, and thus having to restart the recall process.

[0004] In other words, the recall rules used in the existing account recall methods are too simplistic and one-sided, resulting in inaccurate recall results and thus low recall efficiency.

[0005] There is currently no effective solution to the above problems. Summary of the Invention

[0006] This application provides an account recall method and apparatus, storage medium and electronic device to at least solve the technical problem of low recall efficiency caused by inaccurate recall results due to the overly simplistic and one-sided reference dimensions of the recall rules.

[0007] According to one aspect of the embodiments of this application, an account recall method is provided, comprising: determining the account status information of a target account to be teamed up with; based on the account status information of the target account, obtaining at least two recall channels, each recalling a subset of recalled accounts for the target account, wherein the subset of recalled accounts includes a first type of recalled account subset recalled based on interactive interest tags associated with the account status information, and a second type of recalled account subset recalled based on account relationship features extracted from the account status information; filtering the obtained subset of recalled accounts according to the teaming strategy of the target account to obtain a filtered subset of recalled accounts; and merging the filtered subset of recalled accounts to obtain a set of result accounts for teaming up with the target account.

[0008] According to another aspect of the embodiments of this application, an account recall device is also provided, comprising: a determining unit for determining the account status information of a target account to be teamed up; an acquiring unit for acquiring, based on the account status information of the target account, at least two recall channels each acquiring a subset of recalled accounts for the target account, wherein the subset of recalled accounts includes a first subset of recalled accounts recalled based on interactive interest tags associated with the account status information, and a second subset of recalled accounts recalled based on account relationship features extracted from the account status information; a filtering unit for conditionally filtering the acquired subset of recalled accounts according to the teaming strategy of the target account, to obtain a filtered subset of recalled accounts; and a merging unit for merging the filtered subset of recalled accounts to obtain a set of result accounts for teaming up with the target account.

[0009] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, wherein a computer program is stored in the computer program, and the computer program is configured to execute the account recall method described above when it is run.

[0010] According to another aspect of the embodiments of this application, a computer program product or computer program is provided, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the account recall method described above.

[0011] According to another aspect of the embodiments of this application, an electronic device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to execute the account recall method described above through the computer program.

[0012] In this embodiment, a multi-channel recall fusion approach is adopted. By integrating a first subset of recalled accounts recalled based on interactive interest tags of account status information and a second subset of recalled accounts recalled based on account relationship features of account status information, the obtained subsets of recalled accounts are filtered according to the teaming strategy of the target account to obtain a filtered subset of recalled accounts. The filtered subsets of recalled accounts are then merged to obtain a set of result accounts used to team up with the target account. This achieves the goal of improving the accuracy of teaming recommendations, thereby improving the technical effect of recall efficiency. It also solves the technical problem of low recall efficiency caused by inaccurate recall results due to the overly simplistic and one-sided reference dimensions of the recall rules. Attached Figure Description

[0013] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0014] Figure 1 This is a schematic diagram illustrating the application environment of an optional account recall method according to an embodiment of this application;

[0015] Figure 2 This is a flowchart of an optional account recall method according to an embodiment of this application;

[0016] Figure 3 This is a schematic diagram showing the interface of an optional account recall method according to an embodiment of this application;

[0017] Figure 4 This is a schematic diagram of the model structure of an optional account recall method according to an embodiment of this application;

[0018] Figure 5 This is a schematic diagram of a sample set of an optional account recall method according to an embodiment of this application;

[0019] Figure 6 This is a schematic diagram of an optional account recall method according to an embodiment of this application;

[0020] Figure 7 This is a flowchart of another optional account recall method according to an embodiment of this application;

[0021] Figure 8 This is a schematic diagram of an optional account recall device according to an embodiment of this application;

[0022] Figure 9 This is a schematic diagram of the structure of an optional electronic device according to an embodiment of this application;

[0023] Figure 10 This is an optional computer system architecture block diagram according to an embodiment of this application. Detailed Implementation

[0024] 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 only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0025] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0026] To better understand the technical solutions provided in the embodiments of this application, the key terms involved in the embodiments of this application will be introduced first:

[0027] Team Recommendation: Recommends players on the right end who may be interested to the player on the left end. The player on the left end can send a team invitation. Once the player on the right end accepts the invitation, they can enter the game room on the left end. At this time, the player on the left end can initiate a match search and enter a match.

[0028] Left endpoint, right endpoint: When a player enters the friend recommendation function interface, the game displays a recommendation list for them. This player is called the left endpoint, and the other players included in the recommendation list are called the right endpoint.

[0029] Exposure-to-Match Rate: This is the percentage of players who successfully enter a game after seeing a team-up recommendation message (i.e., exposure), with the player on the left inviting and the player on the right accepting the invitation. Specifically, it's the number of players who were invited to team up through exposure and ultimately joined the team, divided by the number of exposures. A higher value indicates a greater probability that players using this feature will be recommended to join a team.

[0030] Exposure Invitation Rate: The number of times an invitation is invited is divided by the number of times an invitation is made. The higher this value, the greater the probability that a player using this feature will initiate an invitation through a recommendation.

[0031] MMOE structure: includes expert networks, gated networks, and Mixture-of-Experts (MoE), which weights and sums the outputs of multiple expert networks according to specific weights.

[0032] Based on the above technical description, the following is combined with... Figure 1 This paper introduces an account recall system suitable for implementing the account recall method provided in the embodiments of this application. The entity executing the account recall method provided in the embodiments of this application can be, for example, […]. Figure 1 The server 112 shown includes a database 114 and a processing engine 116.

[0033] Server 112 executes S102 to determine the account status information of the target account to be teamed up;

[0034] S104. Based on the account status information of the target account, obtain at least two recall channels for each of the recalled account subsets for the target account. The recalled account subsets include a first type of recalled account subset recalled based on interactive interest tags associated with the account status information, and a second type of recalled account subset recalled based on account relationship features extracted from the account status information.

[0035] S106, According to the team formation strategy of the target account, the obtained subset of recall accounts is filtered by conditions to obtain the filtered subset of recall accounts;

[0036] S108, merge the filtered subset of recalled accounts to obtain the result set of accounts used to team up with the target account.

[0037] S110, server 112 sends the result account set to terminal device 102 via network 110. The team recommendation results can then be displayed on the user device's display interface. Terminal device 102 includes a display 108, a processor 106, and a memory 104.

[0038] Furthermore, it is understood that the embodiments of this application can also be applied to all other related scenarios with "invitation-team-task execution", such as sports competition social platforms, educational platforms for team-based study tours, and e-commerce platforms for group-based group shopping.

[0039] As an optional implementation, in the sports social platform scenario, on the one hand, a tag-based recall strategy is used. This strategy leverages interactive tags (such as teammates who have participated in similar sports competitions) and personal tags (such as sports skill level, active time periods, and preferred sports types) to retrieve a list of right-end users from Redis in real time that have a high historical match rate with the left-end user and share similar sports interests. On the other hand, a multi-task dual-tower recall model is employed. Under the MMOE architecture, this model simultaneously learns sports event invitations and sports match tasks. By calculating the similarity of the feature vectors of the left and right-end users, it further converts the similarity into match probabilities, recalling accounts with a probability value greater than a target threshold. Finally, Elasticsearch (ES) is used to filter the recall results, ensuring that recommended users meet mandatory conditions such as interest matching and geographical proximity. Simultaneously, anti-harassment and exposure strategies are implemented to remove users who do not meet the rules, thereby outputting an accurate and high-quality team recommendation list. This creates a smoother and more personalized sports social experience for users, improves recall efficiency and matching accuracy, and promotes the platform's social activity and user stickiness.

[0040] As another alternative implementation, in the scenario of educational platforms where team members perform study tour tasks, interactive tags are used on the one hand. For example, information such as the types of activities that team members have participated in together (such as quiz competitions of the same type), teammates in the competition, etc., are recorded, as well as personal tags, including the types of subjects that members are good at and their willingness to participate in quizzes. On the other hand, a multi-task dual-tower recall model is adopted. Under the MMOE architecture, this model learns both the invitation to the quiz competition and the quiz game task. By calculating the similarity of the user feature vectors at the left and right endpoints, the similarity is further converted into game probability. Accounts with a recall probability value greater than the target threshold are recalled. Finally, the forced filtering conditions are combined to filter and recall the final recommendation list.

[0041] As another optional implementation, in the scenario of e-commerce platforms where group buying is conducted, on the one hand, interactive tags are used, such as user interaction behavior in group buying activities, such as users who have participated in a group buying together (e.g., "group buying friends"), or tags reflecting social relationships and interaction patterns among users, such as the frequency of users sharing and inviting friends during the group buying process, as well as personal tags, including tags reflecting personal intentions, such as users' shopping preferences (e.g., frequently purchased product categories, brand preferences), spending power (e.g., spending range, purchase frequency), and active time periods; on the other hand, a multi-task dual-tower recall model is adopted. Under the MMOE architecture, this model simultaneously learns the group buying invitation task and the activity execution task. By calculating the similarity of user feature vectors at the left and right endpoints, it further converts the similarity into game probability, recalls accounts with a probability value greater than the target threshold, and finally combines forced filtering conditions for filtering to recall the final recommendation list.

[0042] Optionally, in this embodiment, the terminal device can be a terminal device configured with a target client, which may include, but is not limited to, at least one of the following: mobile phone (such as Android phone, iOS phone, etc.), laptop computer, tablet computer, PDA, MID (Mobile Internet Devices), PAD, desktop computer, smart TV, etc. The target client may be a video client, instant messaging client, browser client, educational client, etc. The network may include, but is not limited to, wired network and wireless network, wherein the wired network includes: local area network, metropolitan area network and wide area network, and the wireless network includes: Bluetooth, WIFI and other networks that enable wireless communication. The server may be a single server, a server cluster composed of multiple servers, or a cloud server. The above is only an example, and no limitation is made in this embodiment.

[0043] As an example, the terminal device may include, but is not limited to, mobile phones, computers, intelligent voice interaction devices, smart home appliances, vehicle terminals, and aircraft. The embodiments of this invention can be applied to various scenarios, including but not limited to digital humans, virtual humans, games, virtual reality, and extended reality (XR).

[0044] For example, a server can also be 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, CDN, and big data and artificial intelligence platforms.

[0045] Based on the above description of the account recall system, the account recall scheme provided in this application embodiment will be described in more detail below with reference to the accompanying drawings.

[0046] Please see Figure 2 , Figure 2 This application provides an account recall method, which can be executed by a computer device. Specifically, the computer device can be... Figure 1 Device 112 in the account recall system shown. (Example) Figure 2 As shown, the account recall plan may include the following S202-S208:

[0047] S202, Determine the account status information of the target account to be teamed up;

[0048] Optionally, in this embodiment, the target account can be a user account in an interactive scenario with a complete invitation-team-match process; the account status information may specifically include, but is not limited to, the user's online status, game status, social preferences, social network, rank, and other information.

[0049] It should be noted that the information collected in this application (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for display, data used for analysis, etc.) are information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of this data all comply with relevant laws, regulations, and standards, necessary confidentiality measures have been taken, and they do not violate public order and good morals. Corresponding access points are provided for users to choose to authorize or refuse. For example, interfaces are set up between this system and relevant users or organizations, providing users with corresponding access points to choose to agree to or refuse automated decision-making results; if the user chooses to refuse, the process proceeds to the expert decision-making stage.

[0050] S204, Based on the account status information of the target account, obtain at least two recall channels for each of the recalled account subsets for the target account, wherein the recalled account subsets include a first type of recalled account subset recalled based on interactive interest tags associated with the account status information, and a second type of recalled account subset recalled based on account relationship features extracted from the account status information.

[0051] Optionally, in this embodiment, the aforementioned recall channels include, but are not limited to, recall channels based on preference tag similarity, model recommendation, spatiotemporal location, and task achievement. For example, in the preference tag similarity recall channel, based on user interests, sports preferences, and other preference tags, the similarity of tag vectors can be calculated to recall users on the right end who share similar interests with the user on the left end. In the model recommendation recall channel, a multi-task dual-tower model can be used to learn user behavior patterns and team formation intentions, and recall can be performed by calculating the similarity of user embedded representations. The model can capture users' deep-seated preferences and interaction patterns, even in the case of preference tags. Even when good tags don't perfectly match, suitable partners can still be recommended based on the user's historical behavior and potential interests. In the spatiotemporal location-based recall channel, users in geographically proximate locations with matching free time can be recalled based on the user's current geographic location and active time preferences. This effectively recalls potential team partners who best meet the geographical and temporal requirements, reducing team failure rates due to location or time mismatches. In the task achievement-based recall channel, users with similar achievements or complementary skills to the user on the left endpoint can be recalled based on their activity achievements and competition records, improving the matching degree of the recalled team accounts and thus increasing recall efficiency.

[0052] The aforementioned interactive interest tags are used to capture historical interaction patterns and shared interests between the target account (i.e., the user at the left endpoint) and the recalled account (i.e., the user at the right endpoint), thereby recalling a subset of accounts with good interaction records and similar interests with the target account. Specifically, these could be tags for shared game wins, shared friends, similar game strategies, etc. The aforementioned account relationship features are used to quantify and characterize the social network connections and interaction intimacy between the target account and the recalled account, assess the similarity or complementarity between player accounts, and then determine the probability of a match between account pairs based on similarity. Specifically, these could be relationship features extracted from the player's account status information by the model that reflect the potential connection or matching degree between player accounts, such as invitation relationship features and match relationship features.

[0053] As an optional implementation, the first subset of recalled accounts based on the interactive interest tags associated with account status information can be identified by analyzing the account status information of the target account to find player accounts with common interests or good interaction records. The second subset of recalled accounts based on the account relationship features extracted from the account status information can be identified by using a deep learning model, specifically a multi-task dual-tower model, to extract deeper account relationship features from the account status information of the target account, such as the willingness to invite to team up and the conditions for successful matches. The recalled player accounts that may have a high degree of matching with the target account are the second subset of recalled accounts.

[0054] S206, According to the team formation strategy of the target account, the obtained subset of recall accounts is filtered by conditions to obtain the filtered subset of recall accounts;

[0055] Optionally, in this embodiment, the aforementioned team-up strategy includes, but is not limited to, being determined based on rank preference, team size, online time requirements, etc., thereby recalling a set of accounts that meet the specific needs of the target player account. It is understood that different team-up strategies can be applied to different target accounts. The aforementioned conditional filtering process can be preprocessing, such as storing online status and rank data in Redis. By updating user data in Redis in real time, the system can directly query the filtered subset of recalled accounts, or use FAISS to create feature vector indexes for player accounts, etc. During the recall stage, only online player accounts that meet the team-up strategy are queried, thereby reducing unnecessary account screening and computational overhead.

[0056] It should be noted that different filtering strategies can be applied to the recalled subsets obtained from different channels. For example, for the recalled account subsets obtained through tags of player account historical behavior or attribute preferences (such as mutual friends, game records, rank preferences, etc.), the corresponding filtering requirements based on the team strategy can be designed to be online, rank difference within a certain range, and more than 5 mutual friends, etc. For the account subsets recalled through the dual-tower multi-task model, the corresponding filtering requirements based on the team strategy can be designed to be high account activity, no cheating behavior, and high recent win rate, etc. This is only an example, and the embodiments of this application do not impose any limitations on it.

[0057] S208, merge the filtered subset of recalled accounts to obtain the result set of accounts used to team up with the target account.

[0058] Optionally, in this embodiment, merging the filtered subset of recalled accounts integrates the most suitable accounts obtained from different channels to form a final, comprehensive recommendation list. During the merging process, deduplication can be performed. Specifically, the recommendation weight of duplicate accounts can be increased in the recall results, or the ranking order of duplicate accounts can be improved, resulting in a final set of accounts with priority order, as shown in the diagram. Figure 3 As shown, the recommended list on the right includes four inviteable accounts. Further details about the inviteable accounts can be viewed, such as their rank, number of likes, and existing heroes.

[0059] The above-described implementation method of this application adopts a multi-channel recall fusion approach. By integrating a first-type recall account subset recalled based on interactive interest tags of account status information and a second-type recall account subset recalled based on account relationship features of account status information, the obtained recall account subsets are filtered according to the target account's teaming strategy to obtain a filtered recall account subset. Potential recommendable accounts are selected from multiple perspectives, enhancing the targeting and diversity of recommendations. The filtered recall account subsets are then merged to obtain a result account set for teaming with the target account, thereby improving the accuracy of teaming recommendations and achieving the technical effect of improving recall efficiency. This solves the technical problem of low recall efficiency caused by inaccurate recall results due to the overly singular and one-sided reference dimensions of the recall rules.

[0060] As an optional implementation, when at least two recall channels include tag-based recall channels, based on the account status information of the target account, a subset of recalled accounts retrieved by each of the at least two recall channels for the target account is obtained, including:

[0061] S1, based on the account status information of the target account, obtain the operation records of the target account's historical interaction operations;

[0062] S2, Based on the statistical results after classifying and statistically analyzing the operation records of historical interaction operations, generate account operation tags for the target account. The account operation tags include at least one of the following: interactive operation tags used to represent the collaborative interaction attributes of the target account, and personalized operation tags used to represent the team-building interest attributes of the target account.

[0063] S3 will retrieve accounts from the account database that match the account operation tags and store them in the subset of recalled accounts obtained under the tag recall channel.

[0064] In step S1 above, the aforementioned historical interaction operation records can be records of player accounts interacting with other player accounts or game elements in the game, such as match history, team invitation and acceptance records, social activity records such as world channel or guild chat records, etc. The historical interaction operation records of the target account can be obtained by setting a log reading task.

[0065] As an optional implementation, the above-mentioned operation record of obtaining the target account's historical interaction operation based on the target account's account status information is, for example, searching the database for all the games the target account has recently participated in based on the online status and rank information in the account status information, and compiling a list of player accounts and results (win, loss, etc.) for each game.

[0066] Furthermore, in step S2 above, the above-mentioned classification statistics include, but are not limited to, statistics based on different operation types (invitation, chat, etc.), different time periods (weekdays, weekends), and different behavioral results (invitation success rate, game win rate, etc.).

[0067] For example, if the target account is player account A, historical data analysis shows that in the past three months, account A sent 200 invitations to guild members on weekends, of which 160 were accepted, resulting in 120 successful team-ups and game starts, achieving a win rate of 60%. On weekdays, account A sent 50 invitations to friends, with a success rate of 80%. Based on these statistics, the following action tags are generated for target account A: Interactive action tags: "Guild Team-Up Expert," "Active Weekend Inviter," "High Success Rate Inviter." Personalized action tags: "Prefers Weekend Challenges," "Close Friend Relationships."

[0068] As an optional implementation method, the account operation tags generated based on the classification statistics results include the interactive operation tags that reflect the collaborative interaction attributes of the target account, such as "high win rate player account" and "active chatterer", which describe the frequency and method of the target account interacting with other player accounts in the game; and the personalized operation tags that represent the target account's personal team-up interests, such as "prefers ranked match team-up" and "night owl player account", which describe the target account's personal intentions and preferences in the game.

[0069] Optionally, based on the types of historical interaction operations mentioned above, the operation records are first divided into different categories, such as team invitations, game participation, chat interaction, friend addition, etc., and then each type of operation record is statistically analyzed. Specifically, this may include calculating the number of team invitations sent by the target account, the number of game wins, the frequency of adding friends, etc. Based on the analysis results, collaborative interaction attribute tags and team interest attribute tags reflecting the target account can be generated. For example, the collaborative interaction attribute tag is "high win rate player account", and the team interest attribute tag is "frequently initiates ranked match teaming in the world channel", etc.

[0070] Optionally, in step S3, accounts that match the account operation tags obtained from the account database are stored in the recall account subset obtained under the tag recall channel. For example, for the tag of personal team-up interest, target accounts with the attribute tag "frequently initiate ranked match team-ups in the world channel" can be recalled from the database with the tags "high acceptance rate of ranked match invitations" and "high activity in the world channel" and stored in the recall account subset of the corresponding tag recall channel.

[0071] Through the above-described implementation methods, the system obtains the historical interaction records of the target account based on the account status information, no longer relying solely on real-time online status or basic game information, but rather deeply exploring the behavioral patterns of player accounts. Then, based on the statistical results of classifying and statistically analyzing the historical interaction records, account operation tags are generated for the target account. Accounts matching the account operation tags obtained from the account database are stored in a subset of recalled accounts obtained through the tag recall channel. By converting the player account's interaction type, frequency, and specific game preferences into quantifiable tags, the system can more comprehensively understand the player account's social needs and gaming style, thereby avoiding the bias that may result from recalling based solely on a single standard (such as rank or online status), and improving the matching degree between the recall results and the player account's actual interests.

[0072] As an optional implementation, accounts matching account operation tags obtained from the account database are stored in a subset of recalled accounts obtained from the tag recall channel, including:

[0073] S1, when the account operation tag is an interactive operation tag, retrieve the operation records of the team combat tasks that the first object account that has already established a collaborative interaction relationship with the target account has participated in from the account database;

[0074] S2, based on the collaborative interaction frequency information indicated by the operation record of the team confrontation task, the first target account is screened to obtain the account that meets the collaborative interaction conditions that match the target account, as the account that matches the account operation tag, and stored in the recall account subset obtained under the tag recall channel.

[0075] Optionally, in step S1 above, the aforementioned collaborative interaction relationship includes, but is not limited to, interactive behaviors such as jointly participating in team matches, completing team tasks, jointly participating in community activities, and mutually viewing personal profiles; the aforementioned team-based competitive tasks may include, but are not limited to, ranked matches, holiday events, and other team-based competitive tasks; the operation records of the team-based competitive tasks participated in by the first object account that has already established a collaborative interaction relationship with the target account may include, for example, the historical team win rate, match type, and team frequency between the first object account and the target account.

[0076] Furthermore, in step S2 above, the aforementioned collaborative interaction conditions can be determined based on the frequency information of collaborative interaction corresponding to interactive behaviors such as win rate between accounts, number of mutual friends, and degree of overlap in online active time periods. Different collaborative interaction conditions can be set for different left-end player accounts, thereby filtering out matching right-end player accounts and improving the accuracy of the recall results.

[0077] For example, the team accounts that teamed up with the target account within a week and had the top ten wins in the above collaborative interaction frequency information can be selected as accounts that meet the collaborative interaction conditions matching the target account. In other words, the results can be truncated after sorting in descending order by the number of games played, and the top K games played are retained as the recall results.

[0078] Alternatively, the results can be sorted in descending order based on the number of mutual friends and then truncated, retaining the top K results based on the number of mutual friends. It should be noted that the sorting results of the top K accounts recalled based on mutual friends can be affected by the level of interaction intimacy. For example, if there is a target account with a high level of intimacy with mutual friend A who matches account 1 in the recall results, the ranking of account 1 can be further improved in the sorting results.

[0079] The implementation methods described in this application focus on interactive operation tags to deeply explore the collaborative interaction history between the target account and the first object account, especially the records of jointly participated team combat missions. Based on the player account relationship based on behavior, a new reference dimension is introduced, namely the frequency of historical collaborative interaction. This allows the recall mechanism to more personally assess the matching degree between player accounts. Subsequently, based on the collaborative interaction frequency information reflected in the collected team combat mission operation records, the first object account is screened to ensure that the accounts ultimately included in the recall account subset have a high interaction preference with the target account, thereby improving the accuracy and efficiency of the recall results.

[0080] As an optional implementation, accounts matching account operation tags obtained from the account database are stored in a subset of recalled accounts obtained from the tag recall channel, including:

[0081] S1. When the account operation tag is a personalized operation tag, determine the second target account that meets the team-up interest condition from the account database. The team-up interest condition includes at least one of the following: having triggered the team-up invitation operation in at least one team-up invitation scenario, having responded to and confirmed the team-up invitation request in at least one team-up invitation scenario, or having the statistical result of accepting team-up invitations reach the target threshold.

[0082] S2, the second target account is treated as an account that matches the account operation tag and stored in the subset of recalled accounts obtained under the tag recall channel.

[0083] In step S1 above, the account database may contain recently active player accounts. The team-up interest conditions include, but are not limited to, conditions triggered by invitation to team-up operation requests and conditions for accepting invitation to team-up operations. Specifically, these may include conditions such as having initiated invitation to team-up operation requests in public community game channels such as the world channel, accepting team-up requests through recruitment halls or other activities, and having an invitation-acceptance rate greater than a certain threshold.

[0084] As an optional implementation, when the account operation tag is "personalized operation," the system filters from the account database player accounts that have proactively initiated invitations in at least two different team-up scenarios within the past month, or accounts that have accepted and confirmed team-up invitations from at least five different players. Furthermore, considering player stability, accounts with a team-up acceptance rate exceeding the average by more than 30% can be further filtered as the second target accounts.

[0085] In step S2 above, as an optional implementation, the second target account is used as an account that matches the account operation tag. For example, an account with a team invitation acceptance rate of more than 80% can be used as an account that matches the personalized operation tag "high personal team willingness". Then, the second target account can be stored in the recall account subset under the corresponding "high personal team willingness" tag channel.

[0086] The implementation methods described in this application focus on personalized operation tags, i.e., the team-building interest attributes of target accounts. Secondary target accounts with a strong tendency to form teams and high response rates are selected from the database. Attention is paid to the proactive behavior of player accounts and their historical invitation acceptance records, increasing the richness and depth of the recall dimensions. This allows the recall mechanism to more accurately target player groups with genuine team-building potential. Furthermore, rigorously screened secondary target accounts that match the personalized operation tags are included in the recall account subset under the tag-based recall channel. This enhances the personalization and effectiveness of the recall results, significantly reduces invalid recalls, and improves overall recall efficiency.

[0087] As an optional implementation, when at least two recall channels include relationship-based recall channels, based on the account status information of the target account, a subset of recalled accounts recalled by each of the at least two recall channels for the target account is obtained, including:

[0088] S1, input the account status information of the target account and the account status information of the candidate accounts to be recalled and matched in the account database into the account recall and matching network. The account recall and matching network includes two team relationship adaptation sub-networks for extracting account relationship features of account status information. The team relationship adaptation sub-network is a neural network model obtained by training network parameters using positive sample data with task relationship labels and negative sample data constructed based on the positive sample data.

[0089] S2, obtain the account call matching degree between the target account and the candidate account output by the account call matching network;

[0090] S3, retrieve candidate accounts whose matching degree is greater than the preset matching degree threshold and store them in the subset of retrieved accounts obtained under the relationship feature retrieval channel.

[0091] Optionally, the aforementioned relationship feature recall channel can be a recall channel that uses a neural network model to calculate the similarity of features extracted from the left and right endpoints, thereby determining a subset of recall accounts that match the target account.

[0092] Specifically, in step S1, the candidate accounts can be accounts from a massive number of online players or accounts of players with shared gaming preferences that have undergone preliminary screening. The aforementioned recall matching network can be a dual-tower recall model, a cross-tower model, etc. Taking the dual-tower recall model as an example... Figure 4 As shown, the model internally includes two team relationship adaptation sub-networks. One sub-network is trained for the target account (the Source Tower corresponding to the left endpoint), and the other is trained for the candidate account (the Destination Tower corresponding to the right endpoint). The aforementioned account relationship features can be invitation relationship features, team relationship features, exposure relationship features, etc.; the aforementioned task relationship labels can be invitation task labels, team task labels, exposure task labels, etc.

[0093] Optionally, the positive sample data corresponding to the invitation task tag can be positive sample player pairs who have completed the invitation interaction operation in the World Recruitment Hall in the past week. The matching negative samples specifically include a first proportion of simple negative sample player pairs constructed by random sampling, and a second proportion of difficult negative sample player pairs constructed based on those who have not completed the invitation interaction operation. In order to approximate the distribution during inference, the second proportion is smaller than the first proportion to avoid the problem of selection bias.

[0094] Further in step S2 above, the account call matching degree between the target account and the candidate account output by the account call matching network is obtained. The account call matching degree can be the probability of successful team formation between account pairs.

[0095] Optionally, in step S3 above, the matching threshold can be dynamically set. For example, when the target account has high activity, the matching threshold can be reduced to recall more accounts, adapting to changes in player account behavior and improving the accuracy of the recall results. A subset of recalled accounts is created for each target account. Candidate accounts whose matching threshold is greater than the preset matching threshold are stored in the relationship feature recall channel, i.e., the subset of recalled accounts obtained based on the recall model outputting a successful team-up probability greater than the preset threshold.

[0096] The relationship feature recall scheme based on deep learning, as described in this application, improves the accuracy and efficiency of the recall results by introducing a team relationship adaptation sub-network for multi-dimensional analysis. This overcomes the limitations of traditional rule-based recall, which suffers from a single reference dimension and poor recall effect, thereby improving the efficiency of recall.

[0097] As an optional implementation, obtaining the account call matching degree between the target account and the candidate account output by the account call matching network includes:

[0098] S1, in the embedded feature transformation layer of the account call-back matching network, the account status information of the target account is transformed into the target account embedded feature vector, and the account status information of the candidate account is transformed into the candidate account embedded feature vector.

[0099] Optionally, the aforementioned embedded feature transformation layer is the first layer in the neural network model, such as... Figure 4 The EmbeddingLayer layer in the middle stores the original state information of the account, i.e., the account data U at the left endpoint. s And the account data U at the right end. t (Such as player account level, online status, game preferences, etc.) are converted into original features. That is, the account status information of the target account and the candidate account are converted into target account embedding feature vector and candidate account embedding feature vector respectively through the Embedding Layer.

[0100] S2, embed the target account into the first pairing relationship adaptation subnetwork of the feature vector input account call-back matching network, and embed the candidate account into the second pairing relationship adaptation subnetwork of the feature vector input account call-back matching network;

[0101] It should be noted that the aforementioned team-matching sub-network can specifically be the MMOE structure in the fine-tuning model. The MMOE structure includes an expert network (Expert) and a gate network (Gate). The Expert is a multi-layer neural network used to extract features from the input data. Different Expert parameters are independent, and each expert network can learn different feature representations, thus providing diverse features for different tasks (such as invitation tasks and game tasks that match the aforementioned task relationship labels). The gate network is another key component in the MMOE architecture. Each task corresponds to a gate network, and the gate network's role is to dynamically select and combine the outputs of the expert networks according to the needs of the current task. The Gate is usually a fully connected layer, with the input being the output of the expert network and the output being a weight vector used to weight and combine the outputs of the expert networks. The Gate parameters for different tasks are independent, meaning that the weights for each task are different. For example, in this embodiment, GateA is used to train the invitation task, and GateB is used to train the game task.

[0102] Optionally, the first set of relationship adaptation subnetworks is used to perform feature processing on the initiator of the invitation task, that is, the target account needs to be embedded into the feature vector and input into the first set of relationship adaptation subnetworks SourceTower in the account call-back matching network. The second set of relationship adaptation subnetworks is used to perform feature processing on the recipient of the invitation task, that is, the candidate account needs to be embedded into the feature vector and input into the second set of relationship adaptation subnetworks DestinationTower in the account call-back matching network.

[0103] S3, obtain the target invitation relationship features and target game relationship features that match the target account output by the first team relationship adaptation subnetwork, and obtain the candidate invitation relationship features and candidate game relationship features that match the candidate account output by the second team relationship adaptation subnetwork.

[0104] Optionally, such as Figure 4 As shown in the first team relationship adaptation subnetwork Source Tower, the output features of the expert network are weighted and summed through the gating network GateA corresponding to the invitation task, and the target invitation relationship features that match the target account are then output. The output features of the expert network are weighted and summed using the gated network GateB corresponding to the game task, and then the target game relationship features that match the target account are output.

[0105] In the second team relationship adaptation subnetwork, the output features of the expert network are weighted and summed through the gate network GateA corresponding to the invitation task to output candidate invitation relationship features that match the candidate accounts. The output features of the expert network are weighted and summed using the GateB gating network corresponding to the game task to output candidate game relationship features that match the candidate accounts.

[0106] S4. Based on the characteristics of the target invitation relationship and the candidate invitation relationship, determine the degree of association of the invitation relationship, and based on the characteristics of the target game relationship and the candidate game relationship, determine the degree of association of the game relationship;

[0107] As an alternative implementation, the degree of association of invitation relationships can be determined by similarity calculation, such as cosine similarity, dot product, Euclidean distance, etc., based on the target invitation relationship features and candidate invitation relationship features. Similarly, the degree of association of game relationships can be determined based on the target game relationship features and candidate game relationship features using the above similarity calculation methods.

[0108] For example, in this embodiment, the aforementioned invitation relationship correlation degree is obtained by calculating the dot product of the target invitation relationship features and the candidate invitation relationship features, i.e., the invitation relationship correlation degree. Similarly, the correlation between game relationships

[0109]

[0110] S5 determines the account call reciprocity matching degree between the target account and the candidate account based on the invitation relationship correlation degree and the game relationship correlation degree.

[0111] As an optional implementation, the invitation relationship correlation and the game relationship correlation are converted into a prediction probability of [0,1]: for example, the invitation prediction probability. Game prediction probability Where σ is the sigmoid function:

[0112] Furthermore, the weight value of the game prediction probability can be set to be greater than the weight value of the invitation prediction probability, and a weighted sum can be performed to obtain the account call return matching degree between the target account and the candidate account.

[0113] The embodiments described in this application transform unstructured state information into structured data representations. Furthermore, through training two sub-networks using deep neural networks, complex features related to invitation and match relationships can be automatically learned and extracted. This allows for a deeper understanding of the relationship characteristics between player accounts, avoiding the inaccuracies caused by matching based solely on a single dimension. Based on feature matching results, the correlation between invitation and match relationships is determined, and the strength of relationships between player accounts is quantitatively assessed. This integrates player accounts' social behavior and gaming habits, thereby more accurately predicting the likelihood of successful invitations and team formations, avoiding inefficient recall, and improving matching efficiency.

[0114] As an optional implementation, before determining the account status information of the target account to be teamed up, the following steps are also included:

[0115] S1, obtain positive sample data configured with task relationship tags, and negative sample data constructed based on the positive sample data. The positive sample data includes multiple independent sample subsets carrying different task relationship tags. The multiple independent sample subsets include an invitation sample subset carrying an invitation task tag and a game sample subset carrying a game task tag.

[0116] S2, mix the positive sample data and negative sample data according to a set ratio to obtain mixed sample data, wherein the set ratio will be adjusted according to the training results of the previous output;

[0117] S3. Input the mixed sample data into the initialized account call-back matching network to train the network parameters until the training loss function value reaches the convergence condition. Each team relationship adaptation subnetwork in the account call-back matching network includes multiple task-gated output structures. Each task-gated output structure is used to output a relationship feature that matches the account.

[0118] In step S1 above, the positive sample data configured with task relationship tags can be player pairs who have historically invited or played against each other. The samples are labeled with specific task relationship tags, which indicate the specific interaction type between the two players, such as "invite" or "play". Negative samples, on the other hand, represent player pairs who have not had corresponding interactions, that is, player combinations that have not invited or played against each other. Specifically, they include a large number of simple negative samples, such as randomly selected players who have not yet interacted, and a small number of difficult negative samples, such as players with a low probability of engaging in invitation or play interactions (e.g., player pairs with significantly different online activity times).

[0119] Further in step S2 above, the positive and negative samples corresponding to each task relationship label are mixed at a set ratio to generate mixed sample data. It should be noted that the set ratio can be dynamically adjusted based on the model's historical training results. For example, the ratio of positive to negative samples can be adjusted based on the ROC curve (Receiver Operating Characteristic curve), which is a curve plotted with the false positive rate on the horizontal axis and the true positive rate on the vertical axis. Specifically, the area under the ROC curve (AUC) can be calculated. It should be noted that the larger the AUC value, the better the model performs in balancing the false positive rate (FPR) and the true positive rate (TPR). Optionally, a 1:5 positive-to-negative sample ratio can be used for initial training. Assuming the model's AUC value is 0.85 after training, if the model's AUC value stabilizes at around 0.92 after multiple iterations, the positive-to-negative sample ratio might be 1:4. For further optimization, the ratio can be adjusted to 1:3.5, and the AUC value can be observed to see if it improves. If the AUC value increases, it indicates that the model has learned more fully from the positive samples; if the AUC value decreases, it indicates that the model may be starting to overfit. At this point, the ratio can be adjusted back to 1:4.5 or 1:5, and the performance change can be observed again.

[0120] Furthermore, the proportion can be adjusted directly based on the False Positive Rate (TPR). For example, if the TPR is below a certain threshold, meaning that the recognition rate of truly positive samples (i.e., player pairs who actually invited or played a game) in the model's prediction is low, it is necessary to increase the amount of positive sample data. It is also understandable that if the model's recognition of invitation tasks is worse than its recognition of game tasks, it can be improved by increasing the proportion of samples carrying the invitation task label.

[0121] like Figure 5As shown, the mixed sample set includes positive samples with category label 1 and negative samples with category label 0 corresponding to the invitation task, and positive samples with category label 1 and negative samples with category label 0 corresponding to the game task. The negative samples include simple negative samples and hard negative samples.

[0122] Optionally, in step S3 above, the convergence condition can be that the calculated loss function value tends to stabilize and no longer decreases significantly. The binary classification cross-entropy loss is calculated as follows:

[0123]

[0124] in, The actual invitation label value of the sample. For the predicted invitation label value of the sample, The actual game label value of the sample. Let be the predicted game label value for the sample, and let the indicator function indicate whether sample i comes from the invitation sample set (otherwise, it comes from the game sample set).

[0125]

[0126] If the traditional multi-task learning method is used to construct samples directly, a large number of "invited but not played" samples will be introduced because the conversion rate of the invitation -> game path is low. For example, the original sample construction method has "invited but not played" samples. In fact, based on the logic of this application, it further includes "invited positive samples" in the invitation task and "game negative samples" in the game task. That is, this application further trains the sample set corresponding to the single task to make the model output more accurate.

[0127] By constructing invitation task samples and game task samples respectively through the embodiments described in this application, the model can be trained in a refined manner on two different tasks. The invitation task model focuses on learning which factors can prompt a player account to issue an invitation, while the game task model focuses on which factors can increase the probability of an invitation being accepted and converted into an actual game. This task decomposition helps the model more accurately capture and distinguish the features of the two different behavioral patterns, thereby providing more targeted results during recall and improving the team formation success rate. Furthermore, independently constructing task samples allows the model to be fully exposed to positive and negative samples from both invitation and game scenarios during training, without being affected by an imbalance in the proportion of a certain type of sample (such as "invitation without game").

[0128] As an optional implementation, in the embedded feature transformation layer of the account call-back matching network, after converting the account status information of the target account into the target account embedded feature vector and the account status information of the candidate accounts into the candidate account embedded feature vector, the following steps are included:

[0129] S1, perform clustering operation on the candidate account embedding feature vectors of multiple candidate accounts to determine at least one candidate account cluster, wherein the similarity between candidate accounts in the candidate account cluster is greater than the target threshold.

[0130] S2, obtain the feature vector corresponding to the cluster center of each candidate account cluster in at least one candidate account cluster, and the vector distance between it and the embedded feature vector of the target account respectively;

[0131] S3. Determine the target candidate account cluster based on the vector distance, wherein the candidate account embedding feature vector of the candidate account in the target candidate account cluster satisfies the target distance condition with the target account embedding feature vector.

[0132] The above processes S1-S3 are described below using a complete implementation method:

[0133] Obtain the embedding representation that matches the target account from the output of the first team relationship adaptation subnetwork corresponding to the left endpoint. The embedding representation can be as follows: Figure 4 The output of the dual-tower multitasking model shown and The splicing and fusion of features includes information such as the target account's team-building interests and willingness, as well as the embedded representation output by the second team-building relationship adaptation subnetwork corresponding to the right endpoint, which matches multiple candidate accounts. The embedded representation can be... and The features are spliced ​​and merged, including information such as the team-building interests and willingness of candidate accounts.

[0134] Optionally, clustering algorithms such as K-means, Mini-Batch K-Means, or other clustering algorithms suitable for large-scale data processing can be used to cluster the embedded representations of all right-endpoint player accounts (i.e., multiple candidate accounts). Specifically, the number of clusters (K value) and a target threshold can be set, where the target threshold is used to measure the similarity of player accounts within a cluster. Through the clustering algorithm, multiple clusters are generated for the aforementioned candidate accounts, and the similarity between candidate accounts in each cluster is greater than the target threshold.

[0135] Furthermore, FAISS is used as the vector index library to construct an index for each candidate account, and each cluster after clustering has a corresponding cluster center vector. The cluster center vector is the average of the embedding representations of all candidate accounts in each cluster, representing the center position of the cluster. The embedding representation of the target account is used as the query vector and input into the FAISS index for an approximate nearest neighbor (ANN) search. The cluster center closest to the embedding representation of the target account (i.e., satisfying the above target distance condition) is found, and the candidate account cluster corresponding to the nearest cluster center is taken as the target candidate account cluster.

[0136] The above implementation method can quickly identify the cluster center with the highest similarity to the embedding vector of the left endpoint player account. Furthermore, in the subsequent precise intra-cluster search, a specific number (e.g., Top K) of high-matching right endpoint player accounts can be further filtered out based on the similarity to the left endpoint. This not only improves the accuracy of the recall results but also significantly reduces the number of vectors that need to be processed to directly calculate the similarity between the left endpoint player account and all right endpoint player accounts in the original dataset.

[0137] As an optional implementation, the obtained subset of recalled accounts is filtered according to the team-up strategy of the target accounts to obtain a filtered subset of recalled accounts, including:

[0138] S1, determine the team strategy configured for the target account;

[0139] S2, obtain the set of team constraints corresponding to the team formation strategy, and filter the obtained subset of recalled accounts according to the set of team constraints to obtain the filtered subset of recalled accounts.

[0140] In step S1 above, the team formation strategy may include, but is not limited to, determining based on conditions such as online status, level restrictions, team formation willingness, and negative feedback.

[0141] Specifically, team-up conditions could include requiring player accounts to be online and idle, rank differences to be within two major ranks, and allowing invitations from strangers. Other measures could include anti-harassment strategies to prevent player accounts from being disturbed excessively, exposure filtering conditions to ensure that previously exposed player accounts do not appear frequently, and negative feedback filtering rules to avoid matching with player accounts that have a history of negative interactions.

[0142] In step S2 above, the set of team formation constraints can be generated based on the team formation strategy. For example, if the team formation strategy indicates that the recent win rate of the right-end player account that teams with the left-end player needs to reach 70%, then the corresponding constraint is a win rate greater than 70%. Then, according to the set of team formation constraints, accounts in the recalled accounts that do not meet the requirement of a win rate greater than 70% are filtered to obtain a filtered subset of recalled accounts.

[0143] As an optional implementation, the system checks whether the recalled right-end player accounts meet team-up requirements such as online status, rank restrictions, and invite permissions. It also adds anti-harassment strategies to prevent target accounts from repeatedly sending invites to the same player account and avoids recommending player accounts that frequently refuse invites or have already been reported. Exposure filtering ensures that the recommendation list does not contain player accounts that have been frequently exposed in a short period, while negative feedback filtering excludes accounts with inappropriate social behavior. Ultimately, the subset of recalled accounts filtered according to these conditions becomes part of the target account recommendation list.

[0144] By implementing the methods described in this application and configuring personalized team-up strategies for each target account, the system overcomes the limitations of traditional recall rules that neglect the personalized needs of player accounts. This ensures that the recall results are closer to the player's expectations, improving the accuracy of the recall and user satisfaction. Based on the set of team-up constraints, the initial subset of recalled accounts is refined and filtered to exclude accounts that do not meet the requirements of the team-up strategy. This not only ensures the quality of the recall results and avoids recalling a large number of irrelevant accounts due to a single dimension, but also greatly improves the recall efficiency.

[0145] The present application will be described below with reference to a complete embodiment, such as Figure 6 As shown:

[0146] The first part involves recalling players through tags. These tags can be divided into interaction tags and personal tags. Interaction tag recall involves recalling player accounts that have had specific interactions with the player account on the left-hand side in the past; these accounts are usually more familiar and easier to team up with. Personal tag recall involves recalling player accounts with specific attributes, thereby increasing exposure and match rates by recalling players with a high willingness to team up.

[0147] Interactive tag recall includes recall based on recent match tags and recall based on mutual friend tags. Recent match tag recall uses real-time in-game team data to build a list of teammates who won matches within a certain period. Due to read / write speed limitations, the list is usually sorted in descending order by the number of matches played and then truncated, retaining the top K results. Mutual friend recall uses the in-game friend relationship chain to calculate an offline list of player accounts with mutual friends. Again, due to read / write speed limitations, the list is usually sorted in descending order by the number of mutual friends played and then truncated, retaining the top K results. Both recall lists based on interactive tags are calculated and stored in a Redis database.

[0148] Personal tag recall specifically targets player accounts with a high willingness to team up. Accounts with this tag are more active and have a higher demand for team play, effectively increasing invitation rates and acceptance rates. This tag is derived from the union of various data sources, including the top K player accounts by invitation acceptance rate, player accounts that have initiated team-up requests in the world chat, and player accounts that have teamed up through the recruitment lobby. In other words, the set of accounts with high team-up willingness = {Top K by invitation acceptance rate ∪ Players who initiated team-up requests in the world chat ∪ Players who teamed up through the recruitment lobby}. Further, tags indicating whether player accounts have a high willingness to team up are calculated offline and written into Elasticsearch.

[0149] The second part involves recall using a dual-tower model. The core idea is to use two independent neural networks to embed representations of the left and right endpoints respectively, and then perform recall by calculating the similarity between the feature vectors of the left and right endpoints. The multi-task dual-tower recall model in this application simultaneously utilizes both "team-up" and "invitation" labels to enhance the generalization of the embedded representations.

[0150] This application employs a "mixed sample" approach, constructing separate sample sets for invitation and game tasks and then training them together. For positive samples from invitation and game tasks, corresponding negative samples (a large number of simple negative samples plus a small number of difficult negative samples) are constructed separately. During training, the two sample sets are mixed, with each sample potentially coming from the invitation sample set (corresponding to...). Figure 5 (Middle sample set = 1) or game sample set (corresponding) Figure 5 The sample set is 0, and each sample has only one class label. Only the loss of the task corresponding to that label is trained.

[0151] Specifically, the left and right towers in the dual-tower model use the MMOE structure commonly used in fine-grained ranking models, which can better share label information from different tasks and improve generalization ability. For example... Figure 4 As shown, Expert (expert network) and Gate (gated network) are components of the MMOE network. The left tower outputs the invitation task features that match the left endpoint. Matching the game mission Output the invitation task characteristics that match the right endpoint in the right tower. Matching the game mission

[0152] Calculate the dot product similarity between the left-endpoint embeddings and the right-endpoint embeddings for the invitation task and the game task, respectively. Further transformed into a prediction probability of [0,1]: Invitation prediction probability Game prediction probability Where σ is the sigmoid function:

[0153] It should be noted that the binary classification cross-entropy loss is calculated during training:

[0154]

[0155] in, The actual invitation label value of the sample. For the predicted invitation label value of the sample, The actual game label value of the sample. Let be the predicted game label value for the sample, and let the indicator function indicate whether sample i comes from the invitation sample set (otherwise, it comes from the game sample set). The model training is considered complete when the loss no longer decreases significantly and tends to stabilize.

[0156] In practical applications, the calculated game probability can be used to determine the set of recalled accounts matching the left endpoint. Furthermore, before calculating the similarity between the left and right endpoints, this application can first use an ANN (Approximate Nearest Neighbor) to calculate the top K recall lists for each left endpoint. After batch calculating the embedding representations of all left and right endpoints, the embedding representations of the right endpoints are used to construct a vector index using FAISS. The embedding representations of the left endpoints are then used for querying. The top K most similar right endpoints to the left endpoints are approximated using the ANN algorithm as the recall results, stored in a Redis database, and then similarity and probability calculations are performed, reducing the computational load.

[0157] The third part involves filtering the recall results obtained from each part based on conditions and merging them to obtain multi-path recall results.

[0158] Specifically, the tag data related to the mandatory conditions will be written into the ES index in real time, and it will be used as a condition for search during recommendation recall. It should be noted that the specific settings of the parameters in the filter conditions can be set to different values ​​according to different left endpoints.

[0159] The two recall lists based on interaction tags (recent matches and interacting friends) and the one based on a multi-task dual-tower model (dual-tower model recall) mentioned above are interaction-based recalls. These need to be pre-written into Redis data. During the recall process, the list is also passed as a condition to the search, thus obtaining a recall list that meets the mandatory conditions. Because many real-time mandatory conditions need to be met, the recall list pre-written into Redis is often filtered out, so the number of recall lists generated is often several times larger than the final number needed. In contrast, the tag-based recall (team-up intention recall) uses tag data that can be directly written into an Elasticsearch index, and then combined with mandatory conditions to directly obtain the corresponding number of recall lists.

[0160] As an optional implementation method, the following uses a flowchart. Figure 7 The implementation method of this application is described as follows:

[0161] S702, Account data collection, including but not limited to player rank information, online status, game preferences, historical match records, friend network and chat history, etc.; Based on the collection results, execute S704, Historical interaction tag generation, such as "Recent Matches" and "Mutual Friends"; S706, Personal team-up intention tag generation;

[0162] Furthermore, based on the model training process S708-S710, S708 constructs positive and negative sample sets corresponding to the invitation task label and the game task label, respectively; S708-1 trains the network parameters in the dual-tower model based on the mixed sample set, specifically using the MMOE (Mixture of Experts) architecture, in which the gate network (Gate) and expert network (Expert) can be dynamically adjusted according to the task type of the current sample to ensure that the feature representation and loss function optimization between the two tasks can be effectively handled; S710, when the binary classification cross-entropy loss value tends to stabilize, it is determined that the model training is complete.

[0163] Furthermore, step S710, combined with step S702, executes step S712, where the Embedding Layer performs feature extraction.

[0164] Under the tag-based recall channel, execute S704-1 based on recent game records corresponding to historical interaction tags to recall the top K team accounts with the most recent game wins; execute S704-2 based on social networks to recall the top K team accounts with the most mutual friends. Execute S706-1 based on the feedback behavior of accounts that initiate and receive invitations corresponding to personal team-up intention tags to recall team-up accounts with team-up intentions.

[0165] Under the dual-tower multi-task model recall channel, further execute S714, where the feature data of the left and right endpoints are input into the left and right towers of the dual-tower model, respectively; S716, calculate the similarity between the invitation task and the game task; S718, calculate the success probability of the invitation task and the game task based on the similarity; S720, recall the Top K right endpoint accounts whose success probability is greater than the target threshold.

[0166] Finally, S722 is executed, and the multi-way recall results are filtered and merged according to the team strategy; that is, the recall results of S704-1, S704-2, S706-1 and S720 are filtered, and the player accounts that meet the filtering conditions are merged.

[0167] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0168] According to another aspect of the embodiments of this application, an account recall apparatus for implementing the above-described account recall method is also provided. For example... Figure 8 As shown, the device includes:

[0169] Unit 802 is used to determine the account status information of the target account to be teamed up;

[0170] The acquisition unit 804, based on the account status information of the target account, acquires at least two recall channels for each of the recall account subsets recalled for the target account. The recall account subsets include a first type of recall account subset recalled based on interactive interest tags associated with the account status information, and a second type of recall account subset recalled based on account relationship features extracted from the account status information.

[0171] The filtering unit 806 performs conditional filtering on the obtained subset of recalled accounts according to the team-up strategy of the target account, and obtains the filtered subset of recalled accounts.

[0172] Merging unit 808 merges the filtered subset of recalled accounts to obtain a set of result accounts for teaming up with the target account.

[0173] Optionally, in this embodiment, the acquisition unit 804 includes: a first storage module, configured to acquire operation records of historical interaction operations of the target account based on the account status information of the target account; generate account operation tags for the target account based on the statistical results after classifying and statistically analyzing the operation records of historical interaction operations, wherein the account operation tags include at least one of the following: interaction operation tags for representing the collaborative interaction attributes of the target account, and personalized operation tags for representing the team-building interest attributes of the target account; and store accounts that match the account operation tags obtained from the account database into a subset of recalled accounts obtained under the tag recall channel.

[0174] Optionally, in this embodiment, the first storage module includes: a first acquisition module, configured to, when the account operation tag is an interactive operation tag, acquire from the account database the operation records of the first object account that has already established a collaborative interaction relationship with the target account in team-based combat tasks; based on the collaborative interaction frequency information indicated by the operation records of the team-based combat tasks, filter the first object accounts to obtain accounts that meet the collaborative interaction conditions matching the target account, as accounts matching the account operation tag, and store them in the subset of recalled accounts obtained under the tag recall channel.

[0175] Optionally, in this embodiment, the first storage module further includes: a second acquisition module, used to determine a second object account that meets the team-up interest conditions from the account database when the account operation tag is a personalized operation tag, wherein the team-up interest conditions include at least one of the following: having triggered an invitation to team up in at least one invitation to team up scenario, having responded to and confirmed an invitation to team up request in at least one invitation to team up scenario, or having the statistical result of accepting an invitation to team up reach a target threshold; the second object account is used as an account that matches the account operation tag and stored in the subset of recalled accounts obtained under the tag recall channel.

[0176] Optionally, in this embodiment, the acquisition unit 804 further includes: a second storage module, used to input the account status information of the target account and the account status information of the candidate accounts to be recalled and matched in the account database into the account recall matching network, wherein the account recall matching network includes two team relationship adaptation sub-networks for extracting account relationship features of the account status information, and the team relationship adaptation sub-network is a neural network model obtained by training network parameters using positive sample data with task relationship labels and negative sample data constructed based on the positive sample data; acquiring the account recall matching degree between the target account and the candidate accounts output by the account recall matching network; and storing the candidate accounts with an account recall matching degree greater than a preset matching degree threshold in the recall account subset obtained under the relationship feature recall channel.

[0177] Optionally, in this embodiment, the second storage module includes a first determining module, configured to: convert the account status information of the target account into a target account embedding feature vector and the account status information of the candidate account into a candidate account embedding feature vector in the embedding feature transformation layer of the account call-back matching network; input the target account embedding feature vector into the first team relationship adaptation subnetwork of the account call-back matching network and input the candidate account embedding feature vector into the second team relationship adaptation subnetwork of the account call-back matching network; obtain the target invitation relationship features and target game relationship features output by the first team relationship adaptation subnetwork that match the target account, and obtain the candidate invitation relationship features and candidate game relationship features output by the second team relationship adaptation subnetwork that match the candidate account; determine the invitation relationship correlation degree based on the target invitation relationship features and candidate invitation relationship features, and determine the game relationship correlation degree based on the target game relationship features and candidate game relationship features; and determine the account call-back matching degree between the target account and the candidate account based on the invitation relationship correlation degree and the game relationship correlation degree.

[0178] Optionally, in this embodiment, the second storage module further includes a training module, configured to: acquire positive sample data configured with task relationship labels, and negative sample data constructed based on the positive sample data, wherein the positive sample data includes multiple independent sample subsets carrying different task relationship labels, the multiple independent sample subsets including an invitation sample subset carrying an invitation task label, and a game sample subset carrying a game task label; mix the positive sample data and the negative sample data according to a set ratio to obtain mixed sample data, wherein the set ratio will be adjusted according to the training results of the previous output; input the mixed sample data into the initialized account call-back matching network for network parameter training until the training loss function value reaches the convergence condition, wherein each team relationship adaptation subnetwork in the account call-back matching network includes multiple task gating output structures, each task gating output structure is used to output a relationship feature that matches the account.

[0179] Optionally, in this embodiment, the first determining module further performs a clustering operation on the candidate account embedding feature vectors of multiple candidate accounts to determine at least one candidate account cluster, wherein the similarity between candidate accounts in the candidate account cluster is greater than a target threshold; obtains the feature vector corresponding to the cluster center of each candidate account cluster in the at least one candidate account cluster, and the vector distance between it and the target account embedding feature vector; determines the target candidate account cluster based on the vector distance, wherein the candidate account embedding feature vector of the candidate account in the target candidate account cluster satisfies the target distance condition with the target account embedding feature vector.

[0180] Optionally, in this embodiment, the filtering unit 806 is further configured to determine the team-up strategy configured for the target account; obtain the team-up constraint condition set corresponding to the team-up strategy; and perform condition filtering on the obtained subset of recalled accounts according to the team-up constraint condition set to obtain the filtered subset of recalled accounts.

[0181] For specific implementation examples, please refer to the examples shown in the account recall method above. This embodiment will not be repeated here.

[0182] According to another aspect of the embodiments of this application, an electronic device for implementing the above-described account recall method is also provided, the electronic device being... Figure 1 The terminal device or server shown. This embodiment uses the electronic device as a server as an example for illustration. Figure 9 As shown, the electronic device includes a memory 902 and a processor 904. The memory 902 stores a computer program, and the processor 904 is configured to execute the steps of any of the above method embodiments through the computer program.

[0183] Optionally, in this embodiment, the aforementioned electronic device may be located in at least one of a plurality of network devices in a computer network.

[0184] Optionally, in this embodiment, the processor can be configured to perform the following steps via a computer program:

[0185] S1, Determine the account status information of the target account to be teamed up with;

[0186] S2, based on the account status information of the target account, obtain at least two recall channels for each of the recalled account subsets for the target account, wherein the recalled account subsets include a first type of recalled account subset recalled based on interactive interest tags associated with the account status information, and a second type of recalled account subset recalled based on account relationship features extracted from the account status information.

[0187] S3, according to the team formation strategy of the target account, filter the obtained subset of recall accounts according to the conditions to obtain the filtered subset of recall accounts;

[0188] S4 merges the filtered subset of recalled accounts to obtain a set of result accounts for teaming up with the target account.

[0189] Alternatively, as those skilled in the art will understand, Figure 9 The structure shown is for illustrative purposes only. Electronic devices can also be smartphones (such as Android phones, iOS phones, etc.), tablets, PDAs, mobile internet devices (MIDs), PADs, and other terminal devices. Figure 9 This does not limit the structure of the aforementioned electronic devices. For example, the electronic device may also include components that are more... Figure 9 The more or fewer components shown (such as network interfaces, etc.), or having the same Figure 9 The different configurations shown.

[0190] The memory 902 can be used to store software programs and modules, such as the program instructions / modules corresponding to the account recall method and apparatus in this embodiment. The processor 904 executes various functional applications and data processing by running the software programs and modules stored in the memory 902, thereby implementing the aforementioned account recall method. The memory 902 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 902 may further include memory remotely located relative to the processor 904, and these remote memories can be connected to the terminal via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof. Specifically, the memory 902 may be used for information such as account status. As an example, such as... Figure 9 As shown, the memory 902 may include, but is not limited to, the determining unit 802, the obtaining unit 804, the filtering unit 806, and the merging unit 808 in the account recall device. Furthermore, it may include, but is not limited to, other module units in the account recall device, which will not be elaborated upon in this example.

[0191] Optionally, the transmission device 906 described above is used to receive or send data via a network. Specific examples of the network described above may include wired networks and wireless networks. In one example, the transmission device 906 includes a Network Interface Controller (NIC), which can be connected to other network devices and routers via a network cable to communicate with the Internet or a local area network. In another example, the transmission device 906 is a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0192] In addition, the aforementioned electronic device also includes: a display 908 for displaying the recommended team accounts; and a connection bus 910 for connecting the various module components in the aforementioned electronic device.

[0193] In other embodiments, the aforementioned terminal device or server can be a node in a distributed system, wherein the distributed system can be a blockchain system, which is a distributed system formed by connecting multiple nodes through network communication. The nodes can form a point-to-point network, and any form of computing device, such as a server, terminal, or other electronic device, can become a node in the blockchain system by joining this point-to-point network.

[0194] According to one aspect of this application, a computer program product is provided, comprising a computer program / instructions containing program code for performing the methods shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network via communication section 1009, and / or installed from removable media 1011. When the computer program is executed by central processing unit 1001, it performs various functions provided in embodiments of this application.

[0195] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0196] Figure 10 A schematic block diagram of a computer system architecture for implementing an electronic device according to embodiments of the present application is shown.

[0197] It should be noted that, Figure 10 The computer system 1000 of the electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0198] like Figure 10 As shown, the computer system 1000 includes a central processing unit (CPU) 1001, which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) 1002 or programs loaded from storage section 1008 into random access memory (RAM). The RAM 1003 also stores various programs and data required for system operation. The CPU 1001, ROM 1002, and RAM 1003 are interconnected via a bus 1004. An input / output interface 1005 (I / O interface) is also connected to the bus 1004.

[0199] The following components are connected to the input / output interface 1005: an input section 1006 including a keyboard, mouse, etc.; an output section 1007 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 1008 including a hard disk, etc.; and a communication section 1009 including a network interface card such as a local area network card, modem, etc. The communication section 1009 performs communication processing via a network such as the Internet. A drive 1010 is also connected to the input / output interface 1005 as needed. A removable medium 1011, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 1010 as needed so that computer programs read from it can be installed into the storage section 1008 as needed.

[0200] Specifically, according to embodiments of this application, the processes described in the various method 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 1009, and / or installed from removable medium 1011. When the computer program is executed by central processing unit 1001, it performs various functions defined in the system of this application.

[0201] According to one aspect of this application, a computer-readable storage medium is provided, from which a processor of a computer device reads computer instructions, and the processor executes the computer instructions, causing the computer device to perform the methods provided in the above-described account recall aspect or various alternative implementations of the account recall aspect.

[0202] Optionally, in this embodiment, the computer-readable storage medium may be configured to store a computer program for performing the following steps:

[0203] S1, Determine the account status information of the target account to be teamed up with;

[0204] S2, based on the account status information of the target account, obtain at least two recall channels for each of the recalled account subsets for the target account, wherein the recalled account subsets include a first type of recalled account subset recalled based on interactive interest tags associated with the account status information, and a second type of recalled account subset recalled based on account relationship features extracted from the account status information.

[0205] S3, according to the team formation strategy of the target account, filter the obtained subset of recall accounts according to the conditions to obtain the filtered subset of recall accounts;

[0206] S4 merges the filtered subset of recalled accounts to obtain a set of result accounts for teaming up with the target account.

[0207] Optionally, in 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, and can be implemented wholly or partially using software, hardware (such as 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 includes the functionality of that module or unit.

[0208] Optionally, in this embodiment, those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing the hardware related to the terminal device. The program can be stored in a computer-readable storage medium, which may include: flash drive, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0209] If the integrated units in the above embodiments are implemented as software functional units and sold or used as independent products, they can be stored in the aforementioned computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause one or more computer devices (which may be personal computers, servers, or network devices, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.

[0210] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0211] In the several embodiments provided in this application, it should be understood that the disclosed client can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, indirect coupling or communication connection between units or modules, and may be electrical or other forms.

[0212] The units described as separate components may or may not be physically separate. 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 units can be selected to achieve the purpose of this embodiment according to actual needs.

[0213] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0214] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for recalling accounts, characterized in that, include: Determine the account status information of the target account to be teamed up with; Based on the account status information of the target account, at least two recall channels are used to retrieve a subset of recalled accounts for the target account. The subset of recalled accounts includes a first subset of recalled accounts retrieved based on interactive interest tags associated with the account status information, and a second subset of recalled accounts retrieved based on account relationship features extracted from the account status information. According to the team formation strategy of the target account, the obtained subset of recalled accounts is filtered by conditions to obtain the filtered subset of recalled accounts. The filtered subset of recalled accounts is merged to obtain a set of result accounts used to team up with the target account.

2. The method according to claim 1, characterized in that, When the at least two recall channels include tag-based recall channels, obtaining the subset of recalled accounts for the target account from each of the at least two recall channels based on the account status information of the target account includes: Based on the account status information of the target account, obtain the operation records of the target account's historical interaction operations; Based on the statistical results of classifying and statistically analyzing the operation records of the historical interaction operations, account operation tags are generated for the target account. The account operation tags include at least one of the following: interactive operation tags used to represent the collaborative interaction attributes of the target account, and personalized operation tags used to represent the team-building interest attributes of the target account. Accounts that match the account operation tags obtained from the account database will be stored in the subset of recalled accounts obtained under the tag recall channel.

3. The method according to claim 2, characterized in that, The step of storing the accounts that match the account operation tags obtained from the account database into the subset of recalled accounts obtained under the tag recall channel includes: When the account operation tag is the interactive operation category tag, retrieve the operation records of the team combat tasks that the first object account, which has already established a collaborative interaction relationship with the target account, has participated in from the account database. Based on the collaborative interaction frequency information indicated by the operation record of the team confrontation task, the first target account is filtered to obtain accounts that meet the collaborative interaction conditions that match the target account. These accounts are then stored in the subset of recalled accounts obtained under the tag recall channel as accounts that match the account operation tag.

4. The method according to claim 2, characterized in that, The step of storing the accounts that match the account operation tags obtained from the account database into the subset of recalled accounts obtained under the tag recall channel includes: When the account operation tag is the personalized operation category tag, a second object account that meets the team-up interest condition is determined from the account database. The team-up interest condition includes at least one of the following: having triggered an invitation to team up in at least one invitation to team up scenario, having responded to and confirmed an invitation to team up request in at least one invitation to team up scenario, or having the statistical result of accepting invitation to team up reach a target threshold. The second target account is designated as the account that matches the account operation tag and stored in the subset of recalled accounts obtained under the tag recall channel.

5. The method according to claim 1, characterized in that, When the at least two recall channels include relationship feature recall channels, obtaining the subset of recalled accounts recalled by each of the at least two recall channels for the target account based on the account status information of the target account includes: The account status information of the target account and the account status information of the candidate accounts to be recalled and matched in the account database are input into the account recall and matching network. The account recall and matching network includes two team relationship adaptation sub-networks for extracting account relationship features of the account status information. The team relationship adaptation sub-network is a neural network model obtained by training network parameters using positive sample data with task relationship labels and negative sample data constructed based on the positive sample data. Obtain the account call matching degree between the target account and the candidate account as output by the account call matching network; Candidate accounts whose matching degree is greater than a preset matching degree threshold are retrieved and stored in the subset of retrieved accounts obtained under the relationship feature retrieval channel.

6. The method according to claim 5, characterized in that, The step of obtaining the account call matching degree between the target account and the candidate account output by the account call matching network includes: In the embedded feature transformation layer of the account call-back matching network, the account status information of the target account is transformed into the target account embedded feature vector, and the account status information of the candidate account is transformed into the candidate account embedded feature vector. The target account embedding feature vector is input into the first team relationship adaptation subnetwork of the account call-back matching network, and the candidate account embedding feature vector is input into the second team relationship adaptation subnetwork of the account call-back matching network. Obtain the target invitation relationship features and target game relationship features that match the target account output by the first team relationship adaptation sub-network, and obtain the candidate invitation relationship features and candidate game relationship features that match the candidate account output by the second team relationship adaptation sub-network. Based on the target invitation relationship features and the candidate invitation relationship features, the invitation relationship correlation degree is determined, and based on the target game relationship features and the candidate game relationship features, the game relationship correlation degree is determined. Based on the invitation relationship correlation degree and the game relationship correlation degree, the account call matching degree between the target account and the candidate account is determined.

7. The method according to claim 5, characterized in that, Before determining the account status information of the target account to be teamed up, the following is also included: Obtain the positive sample data configured with the task relationship tags, and the negative sample data constructed based on the positive sample data, wherein the positive sample data includes multiple independent sample subsets carrying different task relationship tags, the multiple independent sample subsets include an invitation sample subset carrying an invitation task tag, and a game sample subset carrying a game task tag; The positive sample data and the negative sample data are mixed according to a set ratio to obtain mixed sample data, wherein the set ratio will be adjusted according to the training results of the previous output; The mixed sample data is input into the initialized account call-back matching network for network parameter training until the training loss function value reaches the convergence condition. Each of the team relationship adaptation sub-networks in the account call-back matching network includes multiple task-gated output structures, and each task-gated output structure is used to output a relationship feature that matches the account.

8. The method according to claim 6, characterized in that, In the embedded feature transformation layer of the account call-back matching network, after converting the account status information of the target account into a target account embedded feature vector and the account status information of the candidate account into a candidate account embedded feature vector, the process includes: Clustering operation is performed on the candidate account embedding feature vectors of multiple candidate accounts to determine at least one candidate account cluster, wherein the similarity between candidate accounts in the candidate account cluster is greater than a target threshold; Obtain the feature vector corresponding to the cluster center of each candidate account cluster in the at least one candidate account cluster, and the vector distance between it and the embedded feature vector of the target account; The target candidate account cluster is determined based on the vector distance, wherein the candidate account embedding feature vector of the candidate account in the target candidate account cluster satisfies the target distance condition with the target account embedding feature vector.

9. The method according to any one of claims 1 to 8, characterized in that, The process involves filtering the obtained subset of recalled accounts according to the team-up strategy of the target account, resulting in the filtered subset of recalled accounts, which includes: The team-up strategy configured for the target account is determined; Obtain the set of team formation constraints corresponding to the team formation strategy, and filter the obtained subset of recalled accounts according to the set of team formation constraints to obtain the filtered subset of recalled accounts.

10. An account recall device, characterized in that, include: Determine the unit and determine the account status information of the target account to be teamed up; The acquisition unit, based on the account status information of the target account, acquires at least two recall channels, each of which recalls a subset of accounts for the target account. The subset of accounts includes a first subset of accounts recalled based on interactive interest tags associated with the account status information, and a second subset of accounts recalled based on account relationship features extracted from the account status information. The filtering unit performs conditional filtering on the obtained subset of recalled accounts according to the team-up strategy of the target account, and obtains the filtered subset of recalled accounts. The merging unit merges the filtered subset of recalled accounts to obtain a set of result accounts for teaming up with the target account.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein the program is executed by a processor to perform the method described in any one of claims 1 to 9.

12. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 9.

13. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to execute the method described in any one of claims 1 to 9 through the computer program.

Images