Recommendation list generation method and apparatus, device, and medium

Abstract

This application relates to a method, apparatus, and medium for generating a recommendation list. The method includes: responding to a room recommendation event for a target user; recalling multiple candidate room sets based on preset recall rules; determining extraction and filling rules for candidate rooms from each candidate room set based on the target user's lifecycle stage; determining fixed display positions in the recommendation list template according to preset operational display rules; setting the position mask of the fixed display positions to a locked state; and filling candidate rooms that meet the operational display rules into the fixed display positions; and extracting corresponding candidate rooms from multiple candidate room sets based on the extraction and filling rules, and filling them into non-fixed display positions in the recommendation list template where the position mask is not locked, thereby obtaining the target recommendation list. This application achieves a balance between the certainty of operational goals and the diversity of algorithmic recommendations, taking into account user experience, content ecosystem, and business needs.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method, apparatus, device, and medium for generating a recommendation list. Background Technology

[0002] In highly interactive application scenarios such as voice-based social networking and online entertainment, real-time recommendation is a core technology for improving user stickiness and platform activity. A typical recommendation process includes stages such as recall, sorting, and reordering, aiming to generate personalized recommendation lists for users from massive amounts of candidate content.

[0003] Currently, recommendation list generation methods for real-time interactive scenarios such as live streaming rooms and voice chat rooms mainly focus on predicting user interests through models such as collaborative filtering and deep learning, thereby estimating and ranking candidate content based on click-through rates or interaction rates. In other words, existing recommendation list generation methods in the recommendation process typically prioritize optimizing a single objective, such as user interest matching or content popularity, during the recall and coarse-ranking stages. Consequently, these recommendation algorithms over-recommend popular rooms, resulting in new streamers or high-quality but low-popularity content lacking exposure opportunities, which is detrimental to the long-term healthy development of the content ecosystem.

[0004] Furthermore, during platform operation, it is often necessary to fix specific rooms in a designated position in the list, such as the top position. However, existing methods for generating recommendation lists cannot systematically coordinate the filling and deduplication of the rest of the list after the position is reserved, resulting in the uncertainty of achieving operational goals. Summary of the Invention

[0005] The primary objective of this application is to address at least one of the aforementioned problems by providing a method, apparatus, device, or medium for generating a recommendation list.

[0006] To achieve the various objectives of this application, the following technical solution is adopted: A method for generating a recommendation list, provided for one of the purposes of this application, includes the following steps: In response to room recommendation events targeting specific users, multiple candidate room sets are recalled based on various preset recall rules. Based on the life cycle stage of the target user, the extraction and filling rules for candidate rooms are determined from each candidate room set; According to the preset operation display rules, the fixed display position in the recommendation list template is determined, the position mask of the fixed display position is set to a locked state, and the candidate rooms that meet the operation display rules are filled into the fixed display position. Based on the extraction and filling rules, corresponding candidate rooms are extracted from the multiple candidate room sets and filled into the non-fixed display positions in the recommendation list template where the position mask is not locked, thus obtaining the target recommendation list.

[0007] A recommendation list generation apparatus, proposed to meet one of the purposes of this application, includes: The event response module is configured to respond to room recommendation events targeting specific users, and recall multiple candidate room sets based on various preset recall rules. The rule determination module is configured to determine the extraction and filling rules for candidate rooms from each candidate room set based on the life cycle stage of the target user. The mask generation module is configured to determine a fixed display position in the recommendation list template according to preset operation display rules, set the position mask of the fixed display position to a locked state, and fill the candidate rooms that meet the operation display rules into the fixed display position. The recommendation list generation module is configured to extract corresponding candidate rooms from the multiple candidate room sets based on the extraction and filling rules, and fill them into the non-fixed display positions in the recommendation list template where the position mask is not locked, thereby obtaining the target recommendation list.

[0008] In another aspect, a computer device provided for one of the purposes of this application includes a central processing unit and a memory, the central processing unit being configured to invoke and run a computer program stored in the memory to perform the steps of the recommendation list generation method described in this application.

[0009] In another aspect, a computer-readable storage medium is provided to suit another purpose of this application, which stores, in the form of computer-readable instructions, a computer program implemented according to the aforementioned recommendation list generation method, which, when invoked by a computer, performs the steps included in the corresponding method.

[0010] Compared with existing technologies, the advantages of this application are as follows: This application overcomes the problems of focusing on a single recommendation target and insufficient diversity of content ecosystem by obtaining a recommendation recall set containing multiple candidate room sets determined based on various preset recall rules, and determining corresponding extraction and filling rules in conjunction with the target user's lifecycle stage. Specifically, multiple recall rules ensure the diversity of the candidate room set, covering different types such as popular content and personalized matching content, thus enriching the composition of the recommendation recall set at the source. On this basis, corresponding extraction and filling rules are determined based on the user's lifecycle stage, and candidate rooms are extracted from the candidate room set to fill the recommendation list template according to these rules. The implementation of this technology makes the final target recommendation list not only personalized but also adaptive, dynamically optimizing the composition and proportion of content as the user's lifecycle progresses, thereby satisfying the user's immediate interests while also considering the long-term healthy development of the platform's content ecosystem.

[0011] Furthermore, this application introduces a mechanism to determine fixed display positions and location masks based on preset operational display rules, and prioritizes filling designated content into fixed display positions. In the early stages of generating the recommendation list, a clear distinction is made between fixed display positions locked by operational needs and non-fixed display positions filled by candidate rooms extracted using extraction and filling rules. Operational display rules can precisely specify specific candidate rooms and their mandatory display positions in the recommendation list, and by generating location masks to lock these fixed display positions, it ensures that the content specified by operations can definitively occupy the target positions, unaffected by subsequent sorting and extraction processes. This reliably achieves the operational exposure goals and creatively solves the problem of dynamic list reorganization and content conflicts after fixed display positions are occupied. Attached Figure Description

[0012] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 A flowchart illustrating a typical embodiment of the recommendation list generation method of this application; Figure 2 A schematic block diagram of the device for generating the recommendation list for this application; Figure 3 This is a schematic diagram of the structure of a computer device used in this application. Detailed Implementation

[0013] This application discloses a method for generating a recommendation list, which can be deployed as a software module or service on the server side of a recommendation system. It responds to requests from clients and generates and returns the final recommendation list. This method can be applied to various content recommendation scenarios, and is particularly suitable for real-time interactive content platforms with high requirements for personalization, timeliness, and operational controllability, such as voice chat, live video streaming, and online social applications.

[0014] In a typical embodiment of this application, when a user refreshes the application's homepage or pulls down to refresh, a voice chat room recommendation event can be triggered. Responding to this event, the system obtains the identity identifier of the target user corresponding to the event and queries the user profile based on this identifier to determine the user's current lifecycle stage, such as the new user exploration phase, the active user stable phase, or the dormant user recall phase. Simultaneously, candidate content is retrieved from multiple preset recall channels in parallel. Each recall channel has corresponding recall rules, ultimately forming a recommendation recall set. This recommendation recall set may include: a subset of popular rooms calculated based on real-time online users and interaction volume; a subset of personalized preference rooms derived from user historical behavior using a collaborative filtering algorithm; and a subset of potential rooms recalled from new broadcasters or low-exposure high-quality content pools. Next, based on the determined user lifecycle stage, a preset strategy configuration library is queried. For example, for the new user exploration phase, the corresponding strategy template might stipulate that the generated recommendation list should ensure a high proportion of popular rooms to quickly establish awareness, and mix in a certain proportion of potential rooms for interest exploration. This template is the extraction and filling rule, which clearly defines how many rooms should be extracted from each of the above candidate room sets, and the order in which they should be filled into the recommendation list template.

[0015] In some embodiments, operators can pre-configure display rules through a backend configuration, for example, specifying that a particular official event room must be displayed in the third position of all user recommendation lists today. This rule is parsed, and the third position is marked as a fixed display position in the recommendation list template used in this generation, and this position is locked with a position mask. The specified official event room is then filled in. Afterwards, the remaining positions in the recommendation list template are filled. According to the extraction and filling rules, a specified number of rooms are sequentially extracted from subsets sorted by popularity, personalized preferences, etc. In one embodiment, during the filling process, it is checked whether the extracted rooms are duplicates of rooms already fixed in the operational position. If a duplicate is found, the next room is selected sequentially from that subset to avoid duplication. All rooms are sequentially filled into display positions not locked with position masks, resulting in the final target recommendation list.

[0016] Finally, a real-time status check can be performed. For example, it checks whether all rooms in the list are currently live and whether the number of online viewers is higher than the minimum threshold. If a room is determined to be offline, a valid room will be drawn from its original candidate subset to replace it, thus ensuring that every room in the final recommendation list sent to the client is currently available. The resulting recommendation list can simultaneously meet the diverse goals of user experience, content ecosystem, and platform operation.

[0017] Please see Figure 1The method for generating a recommendation list according to this application, in its typical embodiment, includes the following steps: Step S3100: Respond to the room recommendation event for the target user and recall multiple candidate room sets based on preset recall rules; The voice chat room recommendation event can be triggered in the following ways: when a user performs a pull-to-refresh operation on the homepage of the voice chat platform, triggers the homepage loading upon first entering the application, or switches back to the recommendation list page from other functional modules. Furthermore, the triggering of this event is not limited to explicit user-initiated requests, but also includes periodic recommendation updates automatically initiated based on time intervals or business logic. For example, when a user stays on the recommendation page for more than a preset time without generating effective interaction, a silent refresh is automatically triggered to push fresher content.

[0018] After a voice chat room recommendation event is triggered, a corresponding recommendation recall set is obtained. This set of candidate rooms is retrieved through multiple recall channels, each corresponding to a preset recall rule. These channels aim to provide candidate materials for the recommendation list from different content dimensions. For example, a recall rule generated based on popularity values ​​runs a recall channel. First, it accesses the global room pool, which maintains real-time metadata for all currently active voice chat rooms on the platform, including room identifiers, creator information, current online users, cumulative interaction volume, and a comprehensive popularity value calculated based on recent user entry frequency and dwell time. A first popularity threshold can be set to filter out popular rooms with broad appeal, thus constructing a corresponding candidate room set. This threshold can be dynamically adjusted according to the overall platform activity; for example, it can be moderately increased during peak user periods to ensure the quality of popular rooms, and moderately decreased during off-peak periods to maintain the richness of the recommendation list.

[0019] The recall channel, which operates based on recall rules generated from user preference matching, first acquires the target user's historical behavioral data. This data includes records of rooms the user has previously entered, the duration of stay in those rooms, interactive behaviors such as speaking, liking, sending gifts, following anchors, and explicit interest tags. Based on this historical behavioral data, an interaction matrix between the user and the rooms is constructed. A collaborative filtering algorithm is then applied to calculate the user preference matching degree between the target user and each candidate room in the global room pool. Candidate rooms with matching degrees exceeding a preset threshold are then constructed into a corresponding candidate room set.

[0020] The recall channel operates based on a recall rule that considers both quality scores and popularity scores to recall candidate rooms. Specifically, it first accesses a pre-defined newcomer support pool, which includes newly created voice chat rooms on the platform or those that have lacked exposure for a long time but have been deemed promising by the quality assessment model. The quality assessment model comprehensively considers factors such as the room creator's historical performance, the novelty of the content theme, the recent interaction growth rate, and user feedback ratings to calculate a quality score for each room. Simultaneously, a second popularity threshold is set, typically lower than the first popularity threshold, to identify potential rooms that are not yet widely recognized but are of high quality. The candidate room set recalled through this channel ensures that new creators and long-tail content receive a fair exposure opportunity, avoiding content diversity depletion caused by over-concentration of top-tier content in recommendations.

[0021] The recall channel operates based on operational display rules. It first accesses a pre-configured operational resource pool, which is pre-configured by platform operators according to business needs. This pool includes specific rooms that require mandatory exposure, such as official event rooms, exclusive rooms for contracted streamers, or commercial cooperation promotion rooms. These rooms carry clear placement instructions, including target display location, effective time period, and targeted user group characteristics.

[0022] After each channel's recall is completed, a deduplication and merging operation is performed to construct the final recommendation recall set. The deduplication logic performs precise matching based on the room's unique identifier. When the same room is recalled by multiple channels simultaneously, its most complete metadata set is retained.

[0023] Step S3200: Based on the life cycle stage of the target user, determine the extraction and filling rules for extracting candidate rooms from each candidate room set; In one embodiment, a strategy template library is pre-set, which contains multiple basic ratio templates. Each basic ratio template defines different extraction and filling rules. For example, each basic ratio template defines the baseline extraction order and baseline quantity ratio for extracting candidate rooms from each candidate room set. Different basic ratio templates are used to match user needs at different lifecycle stages. For example, for users in the novice exploration phase (registered within a specific timeframe), the corresponding basic content allocation template not only configures the proportion of popular content but also includes newly created voice chat rooms or those lacking exposure but deemed promising by the quality assessment model. For instance, it might stipulate drawing a higher proportion of candidate rooms from the pool of rooms recalled based on popularity, 20% of personalized content from the second pool, and 15% of potential new content from the third pool. For users in the growth and development phase, the proportion of personalized content is increased. For example, the basic content allocation template might appropriately reduce the proportion of popular content to 40% and increase the proportion of personalized content to 35%, while maintaining a reasonable ratio of potential and operational content. This aims to satisfy users' existing interests while guiding them to explore more diverse content types. In other words, the extraction and filling rules differ for different lifecycle stages. These rules can be pre-set by those skilled in the art based on actual data or experience.

[0024] In one embodiment, after matching the corresponding basic ratio template, the real-time behavioral profile features of the target user are further obtained to dynamically fine-tune the benchmark extraction order and benchmark quantity ratio defined in the basic ratio template. For specific fine-tuning steps, please refer to the following detailed implementation method, which will not be elaborated here.

[0025] Step S3300: According to the preset operation display rules, determine the fixed display position in the recommendation list template, set the position mask of the fixed display position to a locked state, and fill the candidate rooms that meet the operation display rules into the fixed display position; Operational display rules can be pre-entered by platform operators through a visual configuration backend, including business promotion needs and content operation strategies within a specific time period. In one embodiment, the core elements extracted from parsing the operational display rules include candidate room identifiers (unique identifiers of specific voice chat rooms that need to be forcibly displayed), fixed display position indexes (specific ranking in the recommendation list template, such as the 3rd, 8th, or 12th position), and effective time and space conditions (the applicable time period and effective geographical information of the rule). The setting of the effective time period supports fine-grained time control, which can be configured to be continuously effective throughout the day, or limited to a specific hourly interval, such as the evening peak period from 8 pm to 11 pm, or for specific dates, such as holidays or anniversary events. The setting of the effective geographical information supports multi-level granularity from country and province to city, so as to adapt to different regional cultural preferences, time zone characteristics, or localized event arrangements. The system obtains the trigger time and trigger location information corresponding to the current voice chat room recommendation event. Through time zone conversion and geocoding parsing, it determines whether the trigger time falls precisely within the effective time period and whether the trigger location has an inclusive or contained relationship with the effective location information. When both of the above judgment results are true, the operation display rule is considered to be currently valid.

[0026] Once the operational display rules are deemed valid, the corresponding positions in the recommendation list template are designated as fixed display positions based on their location indices. The recommendation list template is the structural framework for generating the recommendation list, with a predefined total list length, typically 10, 20, or 50 characters. Specific index positions are marked as fixed display positions within the template; these positions are isolated from the subsequent dynamic filling process and do not participate in the sorting based on the extraction filling rules. Simultaneously, corresponding position masks are generated to lock these fixed display positions. The position mask can be implemented using a bitmap data structure or a Boolean array, where each bit corresponds to a position in the recommendation list template. A value of 1 indicates that the position is locked, and a value of 0 indicates that the position is open for filling by candidate rooms determined based on the extraction filling rules. The generation of the mask ensures that the fixed display positions are not accidentally covered or shifted during subsequent list assembly, maintaining the determinism and stability of the operational display rule execution.

[0027] Furthermore, based on the candidate room identifier, the corresponding candidate room is retrieved from the recommendation recall set, or the complete metadata of the room is obtained by directly querying the operation resource pool. The determined candidate room is then filled into the corresponding fixed display position locked by the location mask. In addition, for cases where multiple operation display rules are in effect at the same time and specify the same fixed display position, a priority arbitration mechanism is introduced. The candidate room for that position is finally locked based on the rule's preset priority level or creation timestamp. The lower priority rule is downgraded to the alternative position or delayed until the next recommendation cycle.

[0028] Step S3400: Based on the extraction and filling rules, extract the corresponding candidate rooms from the multiple candidate room sets and fill them into the non-fixed display positions in the recommendation list template where the position mask is not locked, to obtain the target recommendation list.

[0029] The extraction and filling rules specify the exact number, extraction order, and sorting criteria for candidate rooms to be extracted from each candidate room set. Following the order defined by the extraction and filling rules, each candidate room set is accessed sequentially, and the extraction operation is performed. After each extraction, the unique identifier of the candidate room is compared with the identifiers of candidate rooms already filled into the fixed display positions. When a match is detected, the candidate room is determined to be a conflicting candidate room. In this case, the next candidate room is selected from the candidate room set to which the conflicting candidate room belongs, following the same sorting strategy, and conflict detection is performed again until a candidate room that is unique to all candidate rooms in the fixed display positions is found and included in the extraction result set.

[0030] In one embodiment, during the extraction process, a dynamic candidate room buffer pool is maintained to temporarily store candidate rooms extracted from each subset, and preliminary diversity verification is performed on the rooms in the buffer pool. Diversity verification includes detecting the dispersion of topic tags to avoid excessive concentration of rooms extracted in the same batch on a single theme, such as multiple rooms all being game-related; detecting the repetition of streamer identities to prevent multiple rooms by the same streamer or rooms frequently featuring the same streamer from being continuously recommended; and detecting the differentiation of content formats to ensure that rooms with different interactive formats such as voice chat, music performances, and emotional expression are reasonably distributed. When insufficient diversity is detected, a re-extraction mechanism is triggered to select alternative rooms that meet the diversity requirements from the subsequent rankings of the corresponding subset, or to appropriately relax the quantity ratio constraints in the extraction and filling rules to supplement differentiated content from other subsets.

[0031] After extracting all candidate room sets, a preliminary candidate room set is obtained. Then, the non-fixed display positions in the recommended list template are filled. First, the position mask is scanned to identify all open positions with a value of 0. These positions form a queue to be filled according to their index order. Based on the filling priority implicit in the extraction rules, rooms from the candidate room set are sequentially assigned to positions in the queue. The filling priority follows the distribution of user attention heatmaps, assigning the most popular or most relevant rooms to the top of the list, and assigning potential or long-tail content to the back of the list. Rooms with different themes are interspersed between adjacent positions to maintain user browsing freshness.

[0032] In some embodiments, during the filling process, the construction status of the recommendation list is monitored in real time, and multiple constraint checks are performed. These include position continuity checks to ensure that there are no empty slots in the list due to conflict removal or status invalidation; operational rule compatibility checks to ensure that the rooms filled by the algorithm do not have thematic conflicts or competition with the operational content of fixed display positions; and user experience smoothness checks to detect whether there are highly similar cover image styles or title copy between adjacent rooms to avoid user churn due to visual monotony. When a problem is found during the checks, a local rearrangement or replacement strategy is implemented to adjust the allocation of rooms in specific positions, or a better alternative is selected from the remaining reserves of the candidate room set.

[0033] Finally, once all non-fixed display slots have been successfully filled, a target recommendation list is generated. This target recommendation list is a structured data object containing metadata for all rooms arranged in display order.

[0034] As can be seen from the typical embodiments of this application, the technical solution of this application has many advantages, including but not limited to the following aspects: This application overcomes the problems of focusing on a single recommendation target and insufficient diversity of content ecosystem by obtaining a recommendation recall set containing multiple candidate room sets determined based on various preset recall rules, and determining corresponding extraction and filling rules in conjunction with the target user's lifecycle stage. Specifically, multiple recall rules ensure the diversity of the candidate room set, covering different types such as popular content and personalized matching content, thus enriching the composition of the recommendation recall set at the source. On this basis, corresponding extraction and filling rules are determined based on the user's lifecycle stage, and candidate rooms are extracted from the candidate room set to fill the recommendation list template according to these rules. The implementation of this technology makes the final target recommendation list not only personalized but also adaptive, dynamically optimizing the composition and proportion of content as the user's lifecycle progresses, thereby satisfying the user's immediate interests while also considering the long-term healthy development of the platform's content ecosystem.

[0035] Furthermore, this application introduces a mechanism to determine fixed display positions and location masks based on preset operational display rules, and prioritizes filling designated content into fixed display positions. In the early stages of generating the recommendation list, a clear distinction is made between fixed display positions locked by operational needs and non-fixed display positions filled by candidate rooms extracted using extraction and filling rules. Operational display rules can precisely specify specific candidate rooms and their mandatory display positions in the recommendation list, and by generating location masks to lock these fixed display positions, it ensures that the content specified by operations can definitively occupy the target positions, unaffected by subsequent sorting and extraction processes. This reliably achieves the operational exposure goals and creatively solves the problem of dynamic list reorganization and content conflicts after fixed display positions are occupied.

[0036] Based on any embodiment of the method in this application, in response to a voice chat room recommendation event, obtaining the recommendation recall set corresponding to the event includes: Step S3110: Based on the preset popularity recall rules, recall candidate rooms with popularity values ​​greater than the preset first popularity threshold from the preset global room pool to form the first candidate room set; The heat index recall rule is preset by those skilled in the art. It stipulates that candidate rooms with a heat index value greater than a preset first heat index value must be recalled from the preset global room pool. The first heat index value is also preset by those skilled in the art.

[0037] In one embodiment, the popularity value calculation model employs a weighted exponential smoothing algorithm. It comprehensively considers factors such as the real-time online user count, assigning the highest weight to reflect the immediate user aggregation effect; average user dwell time, measuring content attractiveness and stickiness; and interaction density indicators, such as the number of gifts, bullet comments, and likes per unit time, reflecting the room's active atmosphere and depth of participation. The model also introduces a time decay factor, giving higher weight to recent high-activity performance, while appropriately reducing the weight of long-term stable states without fluctuations, preventing "zombie hot topics" from continuously occupying top resources. Finally, after calculating the popularity value of each candidate room based on this model, the popularity value is compared with a preset first popularity threshold. Candidate rooms with popularity values ​​greater than the preset first popularity threshold are selected to form the first candidate room set.

[0038] Step S3120: Based on the preset preference recall rules, determine the user preference matching degree of each candidate room in the global room pool according to the historical behavior data of the target user, and construct the candidate rooms with the user preference matching degree higher than the preset matching degree threshold as the second candidate room set; Historical behavioral data includes the complete trajectory of each room a user has entered, recording the room identifier, entry timestamp, duration of stay, and exit method such as active exit, switching rooms, or app backgrounding; it also includes detailed user interaction behavior within the room, such as the number of times a user spoke and the keywords in the content, sent bullet comments, the type and value of gifts given, and records of likes and follows; and it includes relationship data between the user and the room creator, such as whether the user follows the streamer, the response rate of push notifications when the streamer goes live, and private chat interaction history. This data is stored in a pre-defined distributed data warehouse, organized by user identifier, supporting efficient random access and batch retrieval.

[0039] The execution of preference recall rules uses a user-room interaction matrix as the basic data structure for the algorithm. The rows of this matrix represent user groups, the columns represent candidate rooms in the global room pool, and the matrix elements are the user's explicit or implicit ratings of the room. Explicit ratings are derived from direct user feedback behaviors such as adding to favorites, following, or sending high-value gifts; implicit ratings are inferred from indirect user interactions through behavioral modeling, such as logarithmic transformation of dwell time, exponential weighting of entry frequency, and comprehensive calculation of interaction depth.

[0040] The preference matching degree of each candidate room in the global room pool is calculated, and a set of rooms with a matching degree higher than a threshold is selected. To improve computational efficiency, an approximate nearest neighbor search algorithm is used to quickly locate candidate regions with high matching degrees in a pre-constructed latent vector index space, avoiding scoring all rooms one by one. The selected candidate rooms are sorted in descending order of matching degree, forming a second candidate room set for personalized recommendations.

[0041] Step S3130: Based on the preset strategy recall rules, recall candidate rooms with quality scores greater than the preset score threshold and popularity scores less than the preset second popularity threshold from the preset novice support pool to form a third candidate room set. The pre-set newcomer support pool is used to collect voice chat rooms with high potential but which have not yet gained popularity. The rooms in the pool are sourced from newly registered anchors' first broadcasts, rooms that have been in a low-exposure state for a long time but have recently shown a trend of quality improvement, and specific content that has been marked as potential projects through manual operation review.

[0042] In the calculation of the quality score, a pre-defined quality assessment model is deployed, which employs a multi-dimensional feature fusion and deep learning architecture. The model's input features include creator-level features such as the streamer's account registration duration, past broadcast frequency and stability, historical content violation records, user complaint rate, and depth of fan interaction; interaction-level features such as the distribution of early users' dwell time, analysis of bullet screen sentiment, and the incidence of user-initiated sharing and dissemination; and ecosystem-level features such as the scarcity of this topic type in the overall platform supply, and the size and growth trend of the target audience. The quality assessment model learns the complex mapping relationship between these features and the room's long-term retention rate, creator growth trajectory, and user satisfaction through offline training, outputting a quality score between 0 and 1. A higher score indicates a greater cultivation value and a higher probability of success for the room.

[0043] A preset score threshold is set as the entry threshold for the new user support pool. This threshold is determined through historical data analysis to ensure that the overall quality of rooms in the pool is significantly higher than the global average. Simultaneously, a reverse filtering condition based on popularity value is introduced, with a preset second popularity threshold set. This threshold is typically significantly lower than the first popularity threshold in the popularity recall rule, explicitly excluding high-popularity rooms that have already received widespread exposure from the candidate range for guided recall. The calculation of the popularity value uses the same model architecture as the popularity recall rule. A composite condition search is performed in the new user support pool to locate the set of candidate rooms with a quality score greater than the preset score threshold and a popularity value less than the preset second popularity threshold, forming the third candidate room set.

[0044] Step S3140: Deduplicatively merge the first candidate room set, the second candidate room set, and the third candidate room set to construct the recommendation recall set.

[0045] In one embodiment, unique identifiers for all candidate rooms are first extracted from the first, second, and third candidate room sets, respectively. A room is considered a duplicate room if it appears in two or three candidate room sets simultaneously. Deduplication is then performed according to a preset recall strategy priority fusion rule. This rule allows for priority settings for each candidate room set. When a room belongs to two different candidate room sets, duplicate candidate rooms in the lower-priority set are removed. After deduplication, each candidate room set is used to construct a recommendation recall set, and each candidate room appears only once in the final recommendation recall set.

[0046] Through the above embodiments, a multi-dimensional candidate room recall system is constructed, encompassing popularity recall, preference recall, and strategy recall. This application achieves an organic unity of comprehensiveness, accuracy, and ecological health in the supply of recommended content: Popularity recall ensures immediate coverage of mainstream content on the platform, helping users quickly access the most active social and entertainment scenarios; preference recall achieves deep alignment with individual user needs through collaborative filtering; and strategy recall provides high-potential content with opportunities to break through the cold start, maintaining content diversity and building a sustainable creator growth pipeline. Finally, a deduplication and merging mechanism integrates these into a unified recommendation recall set, laying a solid foundation for subsequent flexible extraction strategy adjustments based on user lifecycle stages, achieving a dynamic balance between immediate traffic value, personalized matching accuracy, and ecological cultivation goals.

[0047] Based on any embodiment of the method in this application, and based on the lifecycle stage of the target user, a selection and filling rule for extracting candidate rooms from each candidate room set is determined, including: Step S3210: Determine the lifecycle stage of the target user based on the registration time, and match the corresponding basic ratio template from the preset strategy template library. The basic ratio template defines the proportion of candidate rooms extracted from each candidate room set and the filling rules for filling the recommendation list template. The strategy template library stores multiple basic matching templates designed for different user lifecycle stages. The strategy template library is retrieved based on the target user's lifecycle stage as the query key. The matching process supports a combination of exact and fuzzy matching for each lifecycle stage. The matching result returns the corresponding basic matching template, which explicitly specifies the baseline extraction order and baseline quantity ratio for candidate rooms from each candidate room set. For example, in one embodiment, for users in the novice exploration phase, the baseline extraction order defined by the corresponding basic matching template specifies that rooms are extracted first from the first candidate room set, i.e., the popularity recall subset, to ensure that the most attractive and popular content is presented at the top of the list, helping new users quickly understand the platform's mainstream value; then, high-matching rooms are extracted from the second candidate room set, i.e., the personalized preference subset, so that even if the new user's historical behavior is limited, the collaborative filtering algorithm can still provide initial personalized exploration based on group similarity; finally, high-quality potential content from the third candidate room set, i.e., the novice support subset, is appropriately introduced to cultivate users' awareness of exploring long-tail content. The baseline quantity ratio is configured as follows: popular content accounts for 60%, personalized content accounts for 25%, and potential content accounts for 15%.

[0048] Step S3220: Obtain the real-time behavioral profile features of the target user and input them into a preset template adjustment model. The template adjustment model outputs adjustment parameters for the basic ratio template. The collection of real-time behavioral profile features covers multiple dimensions, including the browsing trajectory dimension, which records the complete swiping path of the user since entering the recommendation scenario; the interaction response dimension, which captures the user's immediate feedback on the recommended content; and the context dimension, which collects the user's current device environment such as screen size and network type, time context such as weekday evening or weekend afternoon, geographical location such as the city of residence or travel status, and social context such as whether there are friends active in a specific room. The template adjustment model can be implemented based on a large language model. It embeds the real-time behavioral profile features of the target user into the adjustment parameter generation template to obtain adjustment parameter generation instructions. These instructions include placeholders for the real-time behavioral profile features and guiding words that instruct the large language model to generate adjustment parameters in the basic ratio template. Based on these instructions, the large language model outputs adjustment parameters in a specified format. In one embodiment, the output of the adjustment parameters is a structured set of incremental instructions, including percentage adjustments to the baseline quantity ratios, such as increasing the proportion of popular content by 5 percentage points and decreasing the proportion of personalized content by 3 percentage points while keeping the proportion of operational content unchanged; and rearranging the position of the baseline extraction order, such as prioritizing the extraction of personalized content.

[0049] Step S3230: Based on the adjustment parameters, modify the quantity ratio and filling rules defined in the basic ratio template, and determine the modified basic ratio template as the extraction and filling rules.

[0050] Based on the adjustment parameters obtained from the aforementioned steps, the extraction order and quantity ratio in the corresponding basic ratio template can be directly modified to obtain the modified basic ratio template as the extraction and filling rule, which is used to subsequently extract candidate rooms from the candidate room set and fill them into the target recommendation list.

[0051] Through the above embodiments, based on the real-time behavioral profile characteristics of the target user, the baseline extraction order and baseline quantity ratio defined in the basic matching template are fine-tuned, so that the final generated extraction and filling rules can break through the static limitations of the life cycle stage framework and achieve accurate response to the user's immediate intent and context.

[0052] Based on any embodiment of the method in this application, before obtaining the target recommendation list by extracting corresponding candidate rooms from the plurality of candidate room sets according to the extraction and filling rules, and filling them into the non-fixed display positions in the recommendation list template where the position mask is not locked, the following steps are included: Step S2100: Determine the corresponding sorting strategy based on the recall rules corresponding to each candidate room set; After the recommended recall set is generated, it contains multiple candidate room sets. The candidate rooms in each subset are recalled using different recall rules, such as the popularity recall rule, preference recall rule, and strategy recall rule based on the aforementioned specific embodiments. Based on different recall rules, the ranking strategy also differs accordingly. For example: For the first candidate room set, i.e., the popularity recall subset, a strategy based on real-time popularity ranking of candidate rooms is automatically matched based on their recall source identifier. The core of this ranking strategy lies in quantifying the current popularity and traffic attraction of the rooms. The ranking is based on a comprehensive consideration of multiple dimensions of popularity indicators, including real-time online users reflecting the immediate scale of user gathering, recent user entry rate reflecting the room's growth momentum and freshness, average user dwell time measuring content stickiness and attractiveness, and interaction density indicators such as the number of gifts, bullet comments, and likes per unit time reflecting the room's active atmosphere. A comprehensive popularity score is calculated for each candidate room. This score uses a weighted exponential smoothing algorithm, giving higher weight to recent high-activity performance, while introducing a time decay factor to avoid excessive dominance of historical accumulated data. During ranking, candidate rooms are arranged in strict descending order according to their comprehensive popularity score, ensuring that the top positions present the most valuable and popular content at present.

[0053] For the second candidate room set, i.e., the personalized preference subset, a strategy based on user preference matching degree ranking is employed. The core of this ranking strategy lies in quantifying the degree of fit between the rooms and the individual interests of the target users. The ranking is based on the user preference matching degree score calculated during the preceding collaborative filtering recall process. This score integrates the prediction score of the latent semantic model, item-based similarity weighting, user-based social dissemination influence, and content-based topic relevance. Through multi-model integration and normalization, a standardized matching degree index between 0 and 1 is formed. During ranking, candidate rooms are arranged in strict descending order according to user preference matching degree, ensuring that the top positions display personalized content highly aligned with the target user's interests.

[0054] For the third candidate room set, namely the newcomer support subset, a strategy is employed to rank the rooms based on quality scores obtained from a pre-defined quality assessment model. The core of this ranking strategy is to identify high-quality rooms with long-term nurturing value and potential for ecosystem contribution; the ranking is based on the quality scores output by the quality assessment model. During the ranking process, candidate rooms are arranged in strict descending order of quality scores, ensuring that the top positions display the most promising content with the greatest potential for nurturing. For rooms with similar quality scores, priority is given to rooms with a faster recent trend of quality indicator improvement, or rooms with higher creator activity frequency and broadcast stability, in order to identify creators in a rapid growth phase who are most likely to benefit from the support.

[0055] For each candidate room set, a corresponding sorting strategy is matched. For example, the fourth candidate room set in the aforementioned embodiment does not need to be sorted if it does not participate in filling the non-fixed display positions in the target recommendation list.

[0056] Step S2200: Based on the sorting strategy, sort the candidate rooms in the candidate room set in descending order to obtain a sorted candidate room set for subsequent extraction of corresponding candidate rooms based on the extraction and filling rules.

[0057] Based on the sorting strategy matched for each candidate room set in the aforementioned steps, the candidate rooms in each candidate room set are sorted in descending order. After the sorting of each subset is completed in parallel, the sorted candidate room set is encapsulated in a standardized data structure, including an ordered list of room identifiers, the position index of each room in the sorting sequence, and a complete room metadata mapping table, and is passed to the subsequent generation of the recommendation list based on the extraction and filling rules.

[0058] Through the above embodiments, efficient and targeted pre-sorting of each candidate subset is achieved, laying a high-quality and orderly input foundation for subsequent extraction based on extraction and filling rules. Specifically, this method does not apply a uniform sorting standard to all candidate room sets, but rather adaptively matches the most suitable sorting strategy according to their recall source and business objectives. This ensures that each subset achieves its internal optimal order in its respective evaluation dimension, effectively avoiding the problem of potential or personally preferred content being submerged in a sea of ​​popular content and unable to be effectively selected by subsequent rules when using a single indicator (such as global popularity) for mixed sorting.

[0059] Based on any embodiment of the method in this application, according to preset operational display rules, a fixed display position in the recommendation list template is determined, the position mask of the fixed display position is set to a locked state, and candidate rooms that meet the operational display rules are filled into the fixed display position, including: Step S3310: Parse the operation display rules to obtain the effective spatiotemporal conditions, which include the effective time period and the effective geographical information; Operational display rules can be entered and persistently stored by platform operators through a visual rule configuration backend, including content delivery intentions and traffic operation strategies within a specific business cycle. After the rule parsing engine initializes, it sequentially reads the original description of each operational display rule. This description uses a structured data exchange format such as JSON or YAML, or a declarative domain-specific language, containing nested field hierarchies and complex conditional expressions. The engine performs deep parsing, transforming the semantic content of the operational display rules into structured objects in memory, facilitating subsequent matching, judgment, and execution scheduling.

[0060] When parsing the candidate room identifiers, the engine extracts the unique identifier of the target room specified in the rules. This identifier serves as the core indicator for operational deployment and determines which candidate rooms will be forcibly displayed in the target recommendation list.

[0061] The fixed display position index determines the target recommendation list's location information, defining the precise placement of the content within the final generated target recommendation list. The position index can be represented as absolute positions such as the 3rd, 8th, and 12th positions, suitable for scenarios requiring strong deterministic exposure; relative positions such as the first 20% of the list or the first position after the first screen, suitable for scenarios with dynamically changing list lengths; and position sets such as the 2nd or 5th position, suitable for scenarios where operators accept some flexibility. All representations are normalized to absolute position indices, and the percentage of relative position is calculated based on the standard length of the current recommendation list template. For position sets, the optimal available position is selected based on the current list's occupancy.

[0062] The effective time period and effective geographical information defined in the operational display rules determine when and where the rules take effect on users. The effective time period can be parsed in multiple formats, including absolute time intervals (e.g., March 15, 2026, 20:00 to 23:00 for limited-time events), periodic time patterns (e.g., every Friday, 20:00 to 23:00 for fixed sections), and relative time triggers (e.g., within 24 hours of a user's first login for new user-specific operations). These time expressions are converted into a unified time interval representation, and their intersection with the current system time is calculated. The effective geographical information can be parsed at multiple levels of geographic granularity, including country codes for national-level activities, administrative division codes for provincial-level operations, city names for localized promotions, and custom geofences (e.g., latitude and longitude radii for specific business districts or venues). Geographical conditions are normalized into a standard geocoding system and mapped to IP addresses, GPS coordinates, or user registration information, supporting subsequent real-time matching and judgment.

[0063] Finally, the parsed candidate room identifiers, fixed display position location indexes, and effective spatiotemporal conditions are organized in the form of structured rule objects, which can be stored in the operational rule cache of the current recommendation session, waiting to be matched and judged with real-time context parameters.

[0064] Step S3320: Based on the trigger time and triggering region information corresponding to the room recommendation event, determine whether the trigger time is within the effective time period and whether the triggering region information matches the effective region information; The acquisition of the trigger time involves multi-level time base alignment. First, the local system time of the recommendation server is read, which is synchronized with the standard time source via the NTP protocol to ensure millisecond-level accuracy. Then, based on the user's device timezone settings or the geographic timezone inferred from the IP address, the server time is converted into the user's local time representation to support operational strategies based on user-perceived time, such as peak evening hours being relative to the user's local evening rather than the server's location evening. For users traveling across time zones, the device's GPS coordinates or the user's manually set current location are also read to dynamically adjust the timezone mapping. The trigger time is finally encapsulated in a dual format: a Unix timestamp and a structured time representation such as year, month, day, hour, minute, and second, facilitating subsequent interval inclusion judgments and formatted display.

[0065] The acquisition of trigger location information employs a multi-source fusion strategy to address differences in positioning accuracy and availability across various scenarios. The primary source is the user device's GPS coordinates, obtained through client SDK authorization, suitable for mobile applications. The secondary source is IP address geolocation, which maps public IP addresses to country, province, and city levels by querying an IP geolocation database, with accuracy at the kilometer level, suitable for scenarios where location permissions are not enabled or for web-based applications. Supplementary sources include the user's registered residence information, inferred permanent location from historical behavior, and location clues from social relationship chains; these sources are flexibly implemented when real-time location services fail.

[0066] After obtaining the trigger time and trigger location information, for each rule, the engine first performs an inclusion check of the effective time period. This check supports flexible matching of various time representation formats. For absolute time intervals, the engine directly compares the numerical relationship between the trigger timestamp and the start and end timestamps of the interval. For periodic time patterns, the engine decomposes the trigger time into weekday, hour, and minute components and performs pattern matching with the periodic pattern defined in the rule. For example, every Friday from 20:00 to 23:00 requires the weekday component to be equal to 5 and the hour component to be between 20 and 23. For relative time triggers, the engine queries the user profile service to obtain relevant time benchmarks such as the first login timestamp, calculates the time difference, and compares it with the relative interval defined in the rule.

[0067] The matching judgment of the effective geographical information also supports flexible adaptation at multiple granularities. Specifically, it can perform hierarchical inclusion relationship judgment, and prioritize exact matching, such as when the city names are exactly the same; if exact matching fails, it performs inclusion judgment of the superior region, such as when the triggering city belongs to the province specified by the rule; if it still does not match, it performs extended judgment of the neighboring region, such as when the geographical distance between the triggering location and the city specified by the rule is within a preset threshold.

[0068] Finally, the judgment result is output in the form of Boolean value pairs. When the trigger time is true and the triggering region information matches the effective region information, the corresponding operation display rule is considered to be currently valid; when either condition is false, the rule is considered to be currently invalid.

[0069] Step S3330: When the judgment results are all yes, obtain the position index of the fixed display position specified by the operation display rule, and generate the corresponding position mask in the recommendation list template based on the position index corresponding to the fixed display position. The system receives the matching results of the operational display rules from the preceding steps and filters out all valid rules that simultaneously meet the conditions of the trigger time being within the effective time period and the triggering region information matching the effective region information. For these rules, the fixed display position index is extracted and further loaded into the current recommendation list template. Through the mapping of position indexes, the absolute or relative position specified by the operational display rule is converted into a specific index value within the recommendation list template. After determining the fixed display positions, a corresponding position mask is generated to lock these positions, preventing them from being covered or shifted in the subsequent algorithmic dynamic filling process. The position mask data structure uses a compact bitmap representation, where each bit corresponds to a position in the recommendation list template. A bit value of 1 indicates that the position has been locked as a fixed display position, and a bit value of 0 indicates that the position is open for filling by other candidate rooms.

[0070] Step S3340: Based on the candidate room identifier specified by the operation display rule, fill the corresponding candidate room into the corresponding fixed display position.

[0071] For each locked position, i.e., a fixed display slot, a designated candidate room identifier is extracted based on the corresponding operational display rules. Based on this candidate room identifier, the corresponding candidate rooms are retrieved from the recommendation recall set. After determining the candidate rooms, real-time verification and availability assessment of the room status can be performed. Status verification includes: confirming the live broadcast status to verify that the room is currently in an active and accessible state; verifying the number of online users to confirm that the current number of online users is within a reasonable range, neither reaching the upper limit preventing new users from entering nor falling below the minimum effective number, thus affecting the event atmosphere; verifying content compliance by scanning the room's recent content review records to ensure there are no violations or accumulated user complaints; and verifying timeliness, for limited-time event rooms, confirming that the current time is still within the event's validity period.

[0072] The complete metadata of the candidate room that passes the status verification is injected into the corresponding fixed display position of the recommendation list template.

[0073] Through the above embodiments, the operational intent described in natural language or configuration interface is transformed into machine-executable structured instructions by the rule parsing engine. During execution, the trigger time and trigger location information of the target user's recommendation event are matched in real time and accurately with the effective spatiotemporal conditions of each rule, ensuring that the operational content is exposed to users only at the corresponding time and place, avoiding waste of resources and misalignment of user experience.

[0074] Based on any embodiment of the method in this application, and based on the extraction and filling rules, corresponding candidate rooms are extracted from the plurality of candidate room sets and filled into the non-fixed display positions in the recommendation list template where the position mask is not locked, to obtain the target recommendation list, including: Step S3410: Obtain the first identifier of the candidate room extracted based on the extraction and filling rules. When the first identifier is the same as the second identifier of any candidate room in the fixed display position, determine that the corresponding candidate room is a conflicting candidate room. This step involves maintaining a fixed display slot occupancy set, which was generated after the fixed display slots were filled in the previous steps. This set contains unique identifiers for all candidate rooms whose positions have been locked by the operational display rules. This unique identifier is encoded by combining the room creator's user ID and the room creation timestamp, possessing global uniqueness and stability, and accurately identifying a specific voice chat room instance.

[0075] Specifically, when candidate rooms are sequentially extracted from a set of candidate rooms based on extraction and filling rules, the unique identifier of each candidate room is first extracted. This unique identifier is then compared with the fixed display slot occupancy set to check whether the candidate room has already been assigned to a fixed display slot by the operation rules. The comparison process can employ constant-time lookup using hash sets to ensure efficient detection performance even in scenarios involving massive numbers of candidate rooms. When an identifier match is detected, the candidate room is immediately identified as a conflicting candidate room, and the filling operation of that room into the current non-fixed display slot is suspended.

[0076] Step S3420: Determine alternative candidate rooms from the candidate room set to which the conflict candidate room belongs, based on preset alternative rules. Once a candidate room is determined to be a conflict candidate room, the next candidate room that meets the conditions is selected as a candidate room from the original candidate room set to which the conflict candidate room belongs, according to the sorting strategy and selection logic defined by the extraction and filling rules.

[0077] Specifically, the first step is to locate the current position index of the conflict candidate room within its respective candidate room set. Since each candidate room set has been sorted in descending order according to the corresponding sorting strategy in the previous steps, the system then traverses backward based on the current position index, skipping rooms that have been marked as conflict or have been filled in other positions, and selects the next candidate room in the sorted sequence as the initial candidate room.

[0078] After obtaining the initial candidate rooms, a second comparison is performed with the set of fixed display slots. If the comparison result shows that the unique identifier does not exist in the set of occupancy, that is, the candidate room does not conflict with any fixed display slot, then it is confirmed as the final candidate room.

[0079] Step S3430: When the first identifier corresponding to the candidate room is inconsistent with the second identifier, the candidate room replaces the conflicting candidate room and fills it into the corresponding position in the recommendation list template, and finally generates the target recommendation list.

[0080] When the unique identifier of a candidate room is confirmed to be inconsistent with any identifier in the fixed display space occupancy set, the candidate room replaces the conflicting candidate room and is used to fill the corresponding non-fixed display space. As each candidate room set continues to be extracted, conflict detected, candidate selected, and replaced according to the extraction and filling rules, the non-fixed display spaces of the recommendation list template are gradually filled, resulting in the final target recommendation list. The specific target recommendation list filling steps are the same as those in the aforementioned specific embodiment, and will not be repeated here.

[0081] In the above embodiments, the filling of fixed display positions is based on strong business rules, while the filling of non-fixed display positions is based on personalized extraction and filling rules. When the two operate in parallel, there is a risk of content overlap, meaning that the same high-quality room may be forced to the top due to operational goals and also naturally selected due to algorithm ranking. This step, by establishing a fixed display position occupancy set and performing item-by-item comparison, can instantly and accurately identify such duplicate rooms, i.e., conflicting candidate rooms. After identification, a second-best but different candidate room is selected from the same source candidate room set as a replacement. The implementation of this step automatically eliminates duplicate items in the recommendation list without disrupting the original operational layout and algorithm recommendation logic, ensuring that users do not encounter the same room appearing in multiple positions when browsing the list, thus improving the information density and browsing efficiency of the recommendation list.

[0082] Based on any embodiment of the method in this application, and based on the extraction and filling rules, corresponding candidate rooms are extracted from the multiple candidate room sets and filled into the non-fixed display positions in the recommendation list template where the position mask is not locked. After obtaining the target recommendation list, the process includes: Step S4100: Obtain the real-time status data of each candidate room in the target recommendation list. The real-time status data includes one or more of the following: the number of online users, the number of users on the microphone, and the room's locked status. In one embodiment, after the target recommendation list is initially generated, a real-time status query request for all candidate rooms in the target recommendation list is first constructed. This request adopts a batch query mode, encapsulating the unique identifier of each room in the list into a query parameter set, and sending it to the room service cluster or distributed cache system through a high-concurrency asynchronous request channel. The query channel can preferentially access a memory-level distributed cache such as a Redis cluster to obtain high-frequency change data such as the number of online users and the number of users in the microphone with millisecond-level response. For scenarios where the cache is missed or more accurate status is required, the system will retrieve the latest status from the room service's database or real-time message queue.

[0083] Real-time status data collection covers multiple key dimensions. The number of online users reflects the current audience size and social activity of the room, serving as a primary reference indicator for users to determine whether it's worth entering. Accurate counting of this number and its trend over the past five minutes helps identify rooms with sudden drops or abnormal fluctuations in user numbers. The number of users speaking in the microphone reflects the real-time interaction depth within the room. For voice chat rooms, the quantity and quality of speakers directly determine the productivity and attractiveness of the content. Collecting the current list of users speaking in the microphone and their speaking duration helps determine whether the room is in an active discussion state or has fallen silent. The room's locked status is a crucial access control indicator. When a room is manually locked by the host or platform operators, or when automatic risk control mechanisms such as content review are triggered, or when a report is pending investigation, this status prevents new users from entering. Strict monitoring is necessary to avoid recommending inaccessible or invalid rooms to users.

[0084] The acquired real-time status data is organized in a structured form and mapped with the metadata of candidate rooms in the target recommendation list for subsequent invalid room identification.

[0085] Step S4200: When the number of online users is lower than a preset first user threshold, the number of users on the microphone is lower than a preset second user threshold, or the room is locked, the corresponding candidate room is determined to be an invalid room. This step, through preset threshold rules and status flag checks, accurately identifies rooms that, while already included in the recommendation list, are no longer suitable for display to users. This provides a clear decision-making basis for subsequent list updates and content replacements. Specifically, the preset first-number threshold for online users is a core indicator for determining whether a room's social activity meets the standard. This threshold is not a static constant but supports dynamic adaptive adjustment. During peak user hours in the evening, when the overall platform traffic is high, the first-number threshold is appropriately increased to ensure that rooms in the recommendation list have sufficient real-time interactive atmosphere, avoiding the negative impact of sparsely populated rooms on new users' initial experience. During off-peak hours in the early morning, the threshold is appropriately decreased to retain more long-tail content and maintain the richness of the list. The threshold settings also differ for users at different lifecycle stages. New users in the exploration phase are configured with a higher threshold to ensure their certainty of entering highly active rooms, while mature and stable users accept a relatively lenient threshold, allowing them to explore more private or niche room scenarios.

[0086] The preset second threshold for the number of participants speaking in the microphone focuses on the content productivity dimension of the room. The core value of a voice chat room lies in real-time voice interaction and content production. If there are few speakers in the microphone, even if the number of online participants reaches the threshold, users entering the room will only face a silent audience and will not get the expected interactive experience. In addition, the ratio of participants speaking in the microphone to online participants can be introduced to identify rooms with artificially high online participants but sparse microphone interaction, thus placing them in the high-risk category for invalidation.

[0087] Determining a room's locked status is a hard threshold; when a room is marked as locked, it is considered an invalid room. Scenarios triggering a locked status include the streamer actively setting a private room mode, the platform temporarily freezing rooms for content review, abnormal behavior detected by the system's automatic risk control, and service interruptions due to technical malfunctions. Locked status detection has the highest priority; when a locked flag is detected, the room is immediately marked as invalid, eliminating the need for further threshold comparisons and isolating inaccessible rooms from the recommendation list as quickly as possible.

[0088] An invalid decision is triggered if any one of the following conditions is met: the number of online users is lower than the preset first threshold, the number of users on the microphone is lower than the preset second threshold, or the room is locked.

[0089] Step S4300: Extract candidate rooms whose real-time status data meets preset validity conditions from the candidate room set to which the invalid room belongs, and replace them with the corresponding invalid rooms to update the target recommendation list.

[0090] When a candidate room in the target recommendation list is determined to be invalid, the source and ranking position of the invalid room in its original candidate room set are first located. Since the candidate room sets have been sorted in descending order in the previous steps, the set can be traversed sequentially from the next ranking position of the invalid room in the candidate room set to retrieve alternative candidate rooms whose real-time status data meets preset validity conditions. In one embodiment, the preset validity conditions are the reverse expression of the invalidation determination rule, requiring that the number of online users in the alternative candidate room is higher than or equal to a first threshold, the number of users in the microphone is higher than or equal to a second threshold, and the room's locking status is indicated as unlocked.

[0091] During the sequential retrieval process, real-time status queries and validity condition checks are performed on each candidate room. Rooms that pass the check are marked as candidate rooms. In one embodiment, the retrieval continues to accumulate a specific number of candidate rooms, typically configured to be three times the actual number of replacements needed, to support subsequent optimal selection. After accumulating the candidate room, an optimal selection decision is made, selecting the most suitable replacement room to fill a specific position from multiple candidate rooms that meet the validity conditions. The selection decision comprehensively considers multiple factors, including proximity in sorting position (prioritizing rooms ranked higher in the original subset to maintain content quality), differentiation of content themes (avoiding excessive duplication of theme tags with existing rooms in the list), dispersion of creator attributes (preventing multiple rooms by the same broadcaster from appearing densely), and quality of real-time status (selecting rooms with higher online and microphone users among qualified candidate rooms to optimize user experience). Weight coefficients are assigned to each factor, and a comprehensive score is calculated for each candidate room. The room with the highest score is selected as the final replacement room.

[0092] After the replacement room is determined, the target recommendation list is updated. Finally, the target recommendation list after dynamic replacement and updating can re-enter the quality verification process to verify that all positions have been effectively filled and there are no duplicate rooms.

[0093] Through the above embodiments, after the initial generation of the target recommendation list, by performing high-concurrency batch verification on the real-time status of each candidate room in the list (such as the number of online users, the number of users interacting in the microphone, and the room lock status), rooms that have become invalid due to reasons such as ending the broadcast, cooling down the interaction, or being locked by the operation can be promptly identified and filtered out. This avoids the frustration caused by users clicking into invalid rooms from the source and ensures the expected experience of users every time they click.

[0094] Please see Figure 2This invention provides a recommendation list generation device to meet one of the purposes of this application. It is a functional embodiment of the recommendation list generation method of this application. The device includes an event response module 3100, a rule determination module 3200, a mask generation module 3300, and a recommendation list generation module 3400. The event response module 3100 is configured to respond to room recommendation events targeting a target user and recall multiple candidate room sets based on preset recall rules. The rule determination module 3200 is configured to determine extraction and filling rules for candidate rooms from each candidate room set based on the lifecycle stage of the target user. The mask generation module 3300 is configured to determine fixed display positions in the recommendation list template according to preset operational display rules, set the position mask of the fixed display positions to a locked state, and fill the fixed display positions with candidate rooms that meet the operational display rules. The recommendation list generation module 3400 is configured to extract corresponding candidate rooms from the multiple candidate room sets based on the extraction and filling rules and fill them into non-fixed display positions in the recommendation list template where the position mask is not locked, thus obtaining the target recommendation list.

[0095] Based on any embodiment of the device in this application, the event response module 3100 includes: a first subset composition module, configured to recall candidate rooms with a popularity value greater than a preset first popularity threshold from a preset global room pool based on a preset popularity recall rule, to form a first candidate room set; a second subset composition module, configured to determine the user preference matching degree of each candidate room in the global room pool based on the historical behavior data of the target user according to a preset preference recall rule, and construct a second candidate room set by the candidate rooms with the user preference matching degree higher than a preset matching degree threshold; a third subset composition module, configured to recall candidate rooms with a quality score greater than a preset score threshold and a popularity value less than a preset second popularity threshold from a preset new user support pool based on a preset strategy recall rule, to form a third candidate room set; and a recommendation recall set construction module, configured to deduplicate and merge the first candidate room set, the second candidate room set, and the third candidate room set to construct the recommendation recall set.

[0096] Based on any embodiment of the device in this application, the rule determination module 3200 includes: a template matching module, configured to determine the lifecycle stage of the target user based on the registration time, and match a corresponding basic ratio template from a preset strategy template library, wherein the basic ratio template defines the proportion of candidate rooms extracted from each candidate room set and the filling rules for filling into the recommendation list template; a parameter output module, configured to acquire the real-time behavioral profile features of the target user and input them into a preset template adjustment model, wherein the template adjustment model outputs adjustment parameters for the basic ratio template; and a template correction module, configured to correct the proportion and filling rules defined in the basic ratio template based on the adjustment parameters, and determine the corrected basic ratio template as the extraction and filling rules.

[0097] Based on any embodiment of the apparatus in this application, before the mask generation module 3300, there is a: a strategy matching module, configured to determine a corresponding sorting strategy according to the recall rule corresponding to each candidate room set; and a descending sorting module, configured to sort each candidate room in the candidate room set in descending order based on the sorting strategy, so as to obtain a sorted candidate room set for subsequent extraction of the corresponding candidate room based on the extraction filling rule.

[0098] Based on any embodiment of the device in this application, the mask generation module 3300 includes: a rule parsing module, configured to parse the operation display rule to obtain effective spatiotemporal conditions, the effective spatiotemporal conditions including effective time period and effective geographical information; a condition judgment module, configured to determine whether the trigger time is within the effective time period and whether the trigger geographical information matches the effective geographical information based on the trigger time and trigger geographical information corresponding to the room recommendation event; a condition processing module, configured to obtain the position index of the fixed display position specified by the operation display rule when the judgment results are both yes, and generate a corresponding position mask in the recommendation list template based on the position index corresponding to the fixed display position; and a position filling module, configured to fill the corresponding candidate room into the corresponding fixed display position based on the candidate room identifier specified by the operation display rule.

[0099] Based on any embodiment of the device in this application, the recommendation list generation module 3400 includes: a conflict room determination module, configured to obtain a first identifier of a candidate room extracted based on the extraction and filling rules, and determine that the corresponding candidate room is a conflict candidate room when the first identifier is the same as the second identifier of any candidate room in the fixed display position; an identifier comparison module, configured to determine a candidate room from the candidate room set to which the conflict candidate room belongs based on a preset alternative rule; and a room replacement module, configured to replace the conflict candidate room with the candidate room when the first identifier corresponding to the candidate room is inconsistent with the second identifier, and fill it into the corresponding position of the recommendation list template, and finally generate a target recommendation list.

[0100] Based on any embodiment of the device in this application, after the recommendation list generation module 3400, it includes: a status data acquisition module, configured to acquire real-time status data of each candidate room in the target recommendation list, the real-time status data including one or more of the following: online users, users in the microphone, and room lock status; an invalid room judgment module, configured to determine that the corresponding candidate room is an invalid room when the online users are lower than a preset first user threshold, the users in the microphone are lower than a preset second user threshold, or the room lock status is characterized as a locked state; and a list update module, configured to extract candidate rooms whose real-time status data meets preset validity conditions from the candidate room set to which the invalid room belongs, and replace them with the corresponding invalid room to update the target recommendation list.

[0101] To address the aforementioned technical problems, embodiments of this application also provide a computer device. For example... Figure 3 The diagram shows the internal structure of a computer device. The computer device includes a processor, a computer-readable storage medium, a memory, and a network interface connected via a system bus. The computer-readable storage medium stores an operating system, a database, and computer-readable instructions. The database may store a sequence of control information. When executed by the processor, the computer-readable instructions enable the processor to implement a recommendation list generation method. The processor provides computational and control capabilities, supporting the operation of the entire computer device. The memory stores computer-readable instructions, which, when executed by the processor, enable the processor to execute the recommendation list generation method of this application. The network interface of the computer device is used for communication with a terminal. Those skilled in the art will understand that… Figure 3The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0102] In this embodiment, the processor is used to execute... Figure 2 The specific functions of each module and its submodules are defined within the system. The memory stores the program code and various data required to execute the aforementioned modules or submodules. The network interface is used for data transmission between the user terminal and the server. In this embodiment, the memory stores the program code and data required to execute all modules / submodules in the recommendation list generation device of this application. The server can call the server's program code and data to execute the functions of all submodules.

[0103] This application also provides a storage medium storing computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the steps of the recommendation list generation method of any embodiment of this application.

[0104] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. This computer program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. The aforementioned storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0105] Those skilled in the art will understand that the steps, measures, and solutions in the various operations, methods, and processes discussed in this application can be alternated, modified, combined, or deleted. Furthermore, other steps, measures, and solutions in the various operations, methods, and processes discussed in this application can also be alternated, modified, rearranged, decomposed, combined, or deleted. Furthermore, steps, measures, and solutions in the prior art that are similar to those in the open-source operations, methods, and processes of this application can also be alternated, modified, rearranged, decomposed, combined, or deleted.

[0106] The above description is only a partial 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 recommendation list generation method characterized by comprising: include: In response to room recommendation events targeting specific users, multiple candidate room sets are recalled based on various preset recall rules. Based on the life cycle stage of the target user, the extraction and filling rules for candidate rooms are determined from each candidate room set; According to the preset operation display rules, the fixed display position in the recommendation list template is determined, the position mask of the fixed display position is set to a locked state, and the candidate rooms that meet the operation display rules are filled into the fixed display position. Based on the extraction and filling rules, corresponding candidate rooms are extracted from the multiple candidate room sets and filled into the non-fixed display positions in the recommendation list template where the position mask is not locked, thus obtaining the target recommendation list.

2. The recommendation list generation method of claim 1, wherein, In response to room recommendation events targeting specific users, multiple candidate room sets are recalled based on various preset recall rules, including: Based on the preset popularity recall rules, candidate rooms with popularity values ​​greater than the preset first popularity threshold are recalled from the preset global room pool to form the first candidate room set; Based on the preset preference recall rules, the user preference matching degree of each candidate room in the global room pool is determined according to the historical behavior data of the target user, and the candidate rooms with the user preference matching degree higher than the preset matching degree threshold are constructed as the second candidate room set. Based on the preset strategy recall rules, candidate rooms with quality scores greater than the preset score threshold and popularity scores less than the preset second popularity threshold are recalled from the preset novice support pool to form a third candidate room set. The first candidate room set, the second candidate room set, and the third candidate room set are deduplicated and merged to construct the recommendation recall set.

3. The recommendation list generation method of claim 1, wherein, Based on the lifecycle stage of the target user, the extraction and filling rules for candidate rooms from each candidate room set are determined, including: The lifecycle stage of the target user is determined based on the registration time, and the corresponding basic ratio template is matched from the preset strategy template library. The basic ratio template defines the proportion of candidate rooms extracted from each candidate room set and the filling rules for filling the recommendation list template. The real-time behavioral profile features of the target user are obtained and input into a preset template adjustment model. The template adjustment model outputs adjustment parameters for the basic ratio template. Based on the adjustment parameters, the quantity ratios and filling rules defined in the basic proportion template are modified, and the modified basic proportion template is determined as the extraction and filling rules.

4. The recommendation list generation method of claim 1, wherein, Based on the extraction and filling rules, corresponding candidate rooms are extracted from the multiple candidate room sets and filled into the non-fixed display positions in the recommendation list template where the position mask is not locked. Before obtaining the target recommendation list, the process includes: Based on the recall rules corresponding to each candidate room set, the corresponding ranking strategy is determined; Based on the sorting strategy, the candidate rooms in the candidate room set are sorted in descending order to obtain a sorted candidate room set for subsequent extraction of corresponding candidate rooms based on the extraction and filling rules.

5. The recommendation list generation method of claim 1, wherein, Based on preset operational display rules, a fixed display position in the recommendation list template is determined, the position mask of the fixed display position is set to a locked state, and candidate rooms that meet the operational display rules are filled into the fixed display position, including: The operational display rules are analyzed to obtain the effective spatiotemporal conditions, which include the effective time period and the effective geographical information. Based on the trigger time and trigger location information corresponding to the room recommendation event, determine whether the trigger time is within the effective time period and whether the trigger location information matches the effective location information; When the judgment results are both yes, obtain the position index of the fixed display position specified by the operation display rule, and generate the corresponding position mask in the recommendation list template based on the position index corresponding to the fixed display position. Based on the candidate room identifiers specified in the operation display rules, the corresponding candidate rooms will be filled into the corresponding fixed display positions.

6. The recommendation list generation method of claim 1, wherein, Based on the extraction and filling rules, corresponding candidate rooms are extracted from the multiple candidate room sets and filled into the non-fixed display positions in the recommendation list template where the position mask is not locked, to obtain the target recommendation list, including: Obtain the first identifier of the candidate room extracted based on the extraction and filling rules. When the first identifier is the same as the second identifier of any candidate room in the fixed display position, determine that the corresponding candidate room is a conflicting candidate room. From the candidate room set to which the conflict candidate room belongs, alternative candidate rooms are determined from the candidate room set based on preset alternative rules; When the first identifier corresponding to the candidate room is inconsistent with the second identifier, the candidate room replaces the conflicting candidate room and fills the corresponding position in the recommendation list template, thus generating the target recommendation list.

7. The method for generating a recommendation list according to claim 1, characterized in that, Based on the extraction and filling rules, corresponding candidate rooms are extracted from the multiple candidate room sets and filled into the non-fixed display positions in the recommendation list template where the position mask is not locked. After obtaining the target recommendation list, the process includes: Obtain real-time status data for each candidate room in the target recommendation list. The real-time status data includes one or more of the following: the number of online users, the number of users on the microphone, and the room's locked status. When the number of online users is lower than a preset first threshold, the number of users on the microphone is lower than a preset second threshold, or the room is locked, the corresponding candidate room is determined to be an invalid room. From the candidate room set to which the invalid room belongs, select alternative candidate rooms whose real-time status data meets the preset validity conditions, and replace them with the corresponding invalid rooms to update the target recommendation list.

8. A recommendation list generation device, characterized in that, include: The event response module is configured to respond to room recommendation events targeting specific users, and recall multiple candidate room sets based on various preset recall rules. The rule determination module is configured to determine the extraction and filling rules for candidate rooms from each candidate room set based on the life cycle stage of the target user. The mask generation module is configured to determine a fixed display position in the recommendation list template according to preset operation display rules, set the position mask of the fixed display position to a locked state, and fill the candidate rooms that meet the operation display rules into the fixed display position. The recommendation list generation module is configured to extract corresponding candidate rooms from the multiple candidate room sets based on the extraction and filling rules, and fill them into the non-fixed display positions in the recommendation list template where the position mask is not locked, thereby obtaining the target recommendation list.

9. A computer device comprising a central processing unit and a memory, characterized in that, The central processing unit is used to invoke and run a computer program stored in the memory to perform the steps of the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores, in the form of computer-readable instructions, a computer program implemented according to any one of claims 1 to 7, which, when invoked by a computer, executes the steps included in the corresponding method.

Images