Live broadcast fusion element recommendation method and system combined with target group analysis, and medium

Abstract

The application discloses a live broadcast fusion element recommendation method and system combined with target group analysis and a medium, relates to the technical field of live broadcast recommendation, and comprises the following steps: acquiring behavior data of a user in a live broadcast process, dividing the behavior data, forming group stable behavior data and individual instant behavior data, performing target group analysis on the user according to the group stable behavior data, generating a target group analysis result, and generating a group strategy data set; constructing a recommendation strategy space corresponding to the target group according to the group strategy data set, and constraining the selection range of live broadcast fusion elements; outputting user instant behavior state data according to the individual instant behavior data, and dynamically adjusting live broadcast fusion elements in the recommendation strategy space according to the instant behavior state data; and the application solves the problems of unstable recommendation strategy and poor individual adaptability through group stable state analysis and instant behavior adjustment.

Description

Technical Field

[0001] This invention relates to the field of live streaming recommendation technology, and more specifically, to a method, system, and medium for live streaming integrated element recommendation that combines target group analysis. Background Technology

[0002] With the rapid development of mobile internet and short video / live streaming platforms, live streaming content and user scale have experienced explosive growth, making live streaming recommendation systems a crucial technological means to enhance user experience and platform activity. Existing live streaming recommendation technologies primarily rely on users' historical behavior data, interest tags, and recommendation algorithms based on collaborative filtering or content features to push personalized content to users. However, as user behavior becomes increasingly complex, traditional recommendation methods exhibit significant shortcomings in terms of accuracy, real-time performance, and adaptability.

[0003] For example, the invention patent application CN120264033A discloses a live streaming scene-aware recommendation system. The system includes a data acquisition module for real-time collection of multi-source data during the live streaming process; a scene analysis module connected to the data acquisition module that receives and integrates the collected data to analyze the live streaming scene; a user profile building module that tracks user behavior on the live streaming platform in real-time, combining historical viewing data, collection records, consumption records, and interactive behavior during the current live stream to build a dynamic user profile; a recommendation strategy generation module that generates personalized recommendation strategies based on the results of the scene analysis module and the user profiles built by the user profile building module; and a recommendation content push module that pushes the recommended content generated by the recommendation strategy generation module to the user. This achieves personalized and accurate recommendations, effectively improving the user's viewing experience and engagement.

[0004] For example, the invention patent with announcement number CN113207010B discloses a model training method, live streaming recommendation method, device, and program product, involving live streaming recommendation technology. The method includes: acquiring historical live streaming data; determining user attribute information and real-time status based on the historical live streaming data; determining user preference information using a pre-built processing model; using the user preference information to determine a recommendation list; determining user feedback on the recommendation list based on historical live streaming data; determining the model's total expected return using a pre-built evaluation model; and adjusting the processing model and evaluation model based on the model's total expected return and historical live streaming data. The trained processing model is used for personalized recommendations of live streaming videos. The provided solution determines the recommendation list based on user attribute information and preference information, combining user characteristics over a longer period with real-time characteristics to recommend streamers to users. The model is trained based on user feedback to achieve the goal of personalized live streaming video recommendations to users.

[0005] The above-disclosed technical solutions have at least the following technical problems:

[0006] Traditional technical solutions directly model user behavior data based on the entire dataset, failing to distinguish whether the behavior stems from long-term stable preferences or short-term instantaneous fluctuations. This results in recommendation models being highly sensitive to transient noise, user profiles being prone to drift, and the strategy space being unconstrained or weakly constrained during the recommendation strategy generation stage. To address these issues, this invention proposes a solution. Summary of the Invention

[0007] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a live streaming fusion element recommendation method, system, and medium that combines target group analysis, thereby solving the problems of unstable recommendation strategies and poor individual adaptability through group stability analysis and real-time behavior adjustment.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] The live streaming element recommendation method, which incorporates target group analysis, includes: acquiring user behavior data during the live stream and dividing the data into stable group behavior data and individual real-time behavior data; performing target group analysis on users based on the stable group behavior data, generating target group analysis results, and generating a group strategy data set; constructing a recommendation strategy space corresponding to the target group based on the group strategy data set, and constraining the selection range of live streaming element integration; outputting user real-time behavior status data based on individual real-time behavior data, and dynamically adjusting the live streaming element integration within the recommendation strategy space based on the real-time behavior status data; and generating and outputting recommendation results based on the dynamically adjusted live streaming element integration.

[0010] In a preferred embodiment, the acquisition of user behavior data during live streaming, and the segmentation of the behavior data into group stable behavior data and individual real-time behavior data, are specifically as follows: Behavior data is classified according to behavior type, and behavior time series are constructed based on timestamp information; a first time window and a second time window are set in the behavior time series; behavior evolution data is output within the first and second time windows, the behavior evolution data including behavior duration ratio, behavior occurrence density, and behavior change slope; the behavior evolution data is input into a preset state space model to generate a behavior stability discrimination vector, and it is determined whether the current behavior is in a steady-state interval; if the behavior stability discrimination vector continuously falls within the steady-state interval within the first time window, the corresponding behavior data is marked as group stable behavior data; if the behavior stability discrimination vector significantly deviates from the steady-state interval of the first time window within the second time window, and the deviation exceeds a preset threshold, the corresponding behavior data is marked as individual real-time behavior data.

[0011] In a preferred embodiment, the step of inputting behavioral evolution data into a preset state space model to generate a behavioral stability discrimination vector and determine whether the current behavior is in a steady-state interval is as follows: The behavioral evolution data corresponding to each time point is combined according to a fixed dimension to form a current behavioral state vector; a state space model for predicting behavioral states is constructed based on historical group behavioral data and the current behavioral state vector; based on historical group behavioral data, the mean and standard deviation of each state vector dimension are calculated, and the steady-state interval of each state feature is output; the current state vector is input into the state space model for prediction to obtain the prediction result of the behavioral state in the next time period, and the prediction result is used as the behavioral stability discrimination vector, while being compared with the steady-state interval.

[0012] In a preferred embodiment, the step of constructing a state-space model for predicting behavioral states based on historical group behavior data and current behavioral state vectors is as follows: Based on the sequence of state vectors at consecutive time points in the historical group behavior data, the changing relationship between state vectors at adjacent time points is analyzed; based on the historical group behavior data, the direction and magnitude of change of different state vector dimensions over time are statistically analyzed to form a state transition structure, thereby constructing a predictive model for behavioral states; the observed behavioral evolution data and corresponding state vectors in the historical group behavior data are correlated and analyzed to determine the mapping relationship between observed behavioral data and state vectors, serving as the observation structure in the state-space model; by fitting the historical group behavior data, the data in the state transition structure and observation structure are determined, enabling the model to output the predicted behavioral state based on the current state vector; during model operation, the behavioral state vector at the current time point is used as the model input, and the behavioral state at the next time point and the current period is predicted based on the constructed state transition structure, thereby achieving the prediction of behavioral trends.

[0013] In a preferred embodiment, the step of performing target group analysis on users based on stable group behavior data, generating target group analysis results, and generating a group strategy data set specifically involves: Based on stable group behavior data, time-aligning stable behavior features that continuously appear in multiple live streaming cycles, and aggregating the stable behavior states of the same user across cycles to form a stable group behavior state set; representing each user behavior state in the stable group behavior state set as a multi-dimensional vector, with each dimension corresponding to a stable behavior feature, forming a behavior state space; mapping the stable group behavior state set to the behavior state space, and analyzing the transition relationship of user behavior states between adjacent time points based on a discrete-time state transition model to obtain distribution characteristics and state transition trends; and identifying the state offset threshold based on the distribution characteristics and state transition trends. The system identifies stable convergence regions within the target group and categorizes users who consistently reside within the same stable convergence region as candidate target groups. An attractor-based decision mechanism identifies user groups within the candidate target groups that exhibit stable convergence characteristics in the behavioral state space. Users converging to the same behavioral attraction domain are grouped into the same target group, resulting in the final target group segmentation. The behavioral state vectors of convergent users within the target group are combined with the corresponding attraction domain boundaries to obtain the target group analysis results. Based on the target group analysis results, a multi-constraint feasible solution space is constructed. Live streaming fusion elements are used as adjustment strategy variables, and the value range of these elements is constrained according to the attraction domain boundaries corresponding to the target group, generating a candidate set of group strategies. Constraint satisfaction filtering is applied to the candidate set of group strategies to form a group strategy dataset.

[0014] In a preferred embodiment, the process of mapping the stable behavioral state set of the group to the behavioral state space and analyzing the transition relationship of user behavioral states between adjacent time points based on a discrete-time state transition model to obtain distribution characteristics and state migration trends is as follows: The behavioral state vectors in the stable behavioral state set are arranged in chronological order to form a discrete-time behavioral state sequence; the discrete-time behavioral state sequence is input into the discrete-time state transition model to analyze the probability of state changes at adjacent time points in the sequence and output the probability of each state vector transitioning to other state vectors, forming a state transition matrix; based on the distribution of behavioral state vectors in the behavioral state space, the distribution density of user behavioral states is statistically analyzed; according to the state transition matrix, the transition probability of each state point is statistically analyzed and combined with the state space coordinates to output a weighted average migration direction; the number of state-based transitions is statistically analyzed based on the state transition matrix to obtain the migration frequency; the migration directions and frequencies of all state points are integrated to form the migration trend of user behavior in the state space; and stable state clustering regions are identified based on the distribution density and migration trend.

[0015] In a preferred embodiment, the step of constructing a recommendation strategy space corresponding to the target group based on the group strategy data set and constraining the selection range of live streaming fusion elements is as follows: Live streaming fusion elements are used as strategy variables; all combinations in the group strategy data set are mapped to a multi-dimensional space, with each dimension corresponding to an initial value range of a live streaming fusion element, forming the recommendation strategy space for the target group; the mapped strategy space is clustered, grouping similar strategy combinations into the same strategy region to form strategy subspaces; for each strategy subspace, the value range of the strategy variables is constrained according to the attraction domain boundary of the target group, eliminating combinations that cause group behavior to deviate from a stable state; the constrained and partitioned strategy subspaces are structurally encapsulated to form the final recommendation strategy space corresponding to the target group.

[0016] In a preferred embodiment, the step of outputting user real-time behavior state data based on individual real-time behavior data and dynamically adjusting live streaming fusion elements within the recommendation strategy space based on the real-time behavior state data is as follows: Individual real-time behavior data is processed according to behavior type and high-frequency time windows to extract evolution data; the evolution data at each time point is combined into a real-time state vector, and the real-time state vector is compared with historical stable behavior state vectors to output the behavior deviation; the real-time state vector is input into a state space model to output short-term behavior trend predictions; the strategy subspace that best matches the real-time state vector is identified in the recommendation strategy space, and combined with the short-term behavior trend predictions, the strategy subspace corresponding to the current user behavior state is determined, and an initial strategy combination is selected from this strategy subspace; the adjustment direction of the live streaming fusion elements is determined based on the offset direction of the real-time behavior state vector relative to the center of the target group's stable state region; initial adjustment sensitivity is set for each live streaming fusion element, and the adjusted live streaming fusion elements are progressively updated within a continuous time window; the finally adjusted live streaming fusion elements are output, and the user's real-time behavior state vector generated by the adjustment result is fed back to the real-time behavior state update process.

[0017] The system for recommending live streaming elements based on target group analysis includes a data acquisition module, an analysis module, a constraint module, an adjustment module, and an output module, with interconnections between the modules. The data acquisition module acquires user behavior data during the live stream and categorizes this data into stable group behavior data and individual real-time behavior data. The analysis module performs target group analysis on users based on the stable group behavior data, generating target group analysis results and a group strategy data set. The constraint module constructs a recommendation strategy space corresponding to the target group based on the group strategy data set and constrains the selection range of live streaming elements. The adjustment module outputs real-time user behavior status data based on individual real-time behavior data and dynamically adjusts the live streaming elements within the recommendation strategy space based on this real-time behavior status data. The output module generates and outputs recommendation results based on the dynamically adjusted live streaming elements.

[0018] The technical effects and advantages of this invention, which combines target group analysis with a live streaming element recommendation method, system, and medium, are as follows:

[0019] 1. This invention effectively solves the problem of the difficulty in distinguishing between long-term group preferences and short-term individual fluctuations in existing live streaming recommendation technologies by performing refined layering processing on user live streaming behavior data, separating the behavior data into stable group behavior data and real-time individual behavior data. By introducing time windows, behavior evolution data, and state-space models, it determines whether behavior is in a steady state, ensuring that group analysis is based solely on statistically stable behavioral characteristics, thereby improving the reliability and noise resistance of subsequent group analysis and strategy generation from the data source.

[0020] 2. This invention constructs a behavioral state space based on stable group behavioral data and combines a discrete-time state transition model with an attractor decision mechanism to achieve structured identification of user target groups. Compared with traditional methods based on static features or simple clustering, this invention not only considers behavioral distribution density but also introduces state transition trends and convergence characteristics, characterizing user groups from a dynamic evolution perspective. This results in stronger temporal consistency and behavioral interpretability in the target group segmentation results, avoiding frequent oscillations in group labels. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the live streaming fusion element recommendation method of the present invention, which combines target group analysis.

[0022] Figure 2 This is a schematic diagram of the system structure of the live streaming fusion element recommendation method that combines target group analysis according to the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Example 1, Figure 1 This invention presents a live streaming fusion element recommendation method that combines target group analysis, including:

[0025] S1: Acquire user behavior data during the live stream, and segment the behavior data to form stable group behavior data and individual real-time behavior data;

[0026] In this embodiment, user behavior data during the live stream is acquired and divided into stable group behavior data and individual real-time behavior data, as detailed below:

[0027] Acquire user behavior data during the live stream and associate each piece of behavior data with user information, behavior type, timestamp information, and behavior attributes (data fields describing specific behavior characteristics include dwell time, number of likes, comment content, number of switches, and number of clicked products).

[0028] The behavioral data is classified according to the type of behavior, and a time series of behavior is constructed based on the timestamp information. The types of behavior include viewing behavior, interactive behavior, and operational behavior.

[0029] In the behavioral time series, a first time window and a second time window are set for behavioral stability analysis and instant change analysis, respectively. The first time window is used to extract the low-frequency components of the behavior, and the second time window is used to extract the high-frequency components of the behavior.

[0030] Behavioral evolution data are output in the first time window and the second time window respectively. The behavioral evolution data includes the behavior duration ratio, behavior occurrence density and behavior change slope.

[0031] Input the behavioral evolution data into the preset state space model to generate a behavioral stability discrimination vector and determine whether the current behavior is in the steady state range;

[0032] If the behavior stability discrimination vector continuously falls within the steady-state interval within the first time window, the corresponding behavior data will be marked as stable behavior data of the group.

[0033] If the behavior stability discrimination vector shifts significantly within the second time window relative to the steady-state interval of the first time window, and the shift exceeds a preset threshold, then the corresponding behavior data will be marked as individual real-time behavior data.

[0034] In this embodiment, behavioral evolution data is input into a preset state space model to generate a behavioral stability discrimination vector, and it is determined whether the current behavior is in a steady-state region, as follows:

[0035] The behavioral evolution data corresponding to each time point is combined according to a fixed dimension to form a state vector that describes the current behavioral state, i.e., the current behavioral state vector.

[0036] Based on historical group behavior data and current behavior state vectors, a state space model for predicting behavior state is constructed. The state space model statistically models the relationship between the state vectors in historical group behavior data and the changes over time, forming a state transition structure that describes the evolution of behavior state. The model also constructs an observation structure by mapping the historical observed behavior data to the state vectors, enabling the model to predict the trend of behavior state changes based on the current behavior state vector.

[0037] Based on historical group behavior data, the mean and standard deviation of each state vector dimension are calculated, and the steady-state interval (mean ± standard deviation) of each state feature is output. The steady-state interval is used to characterize the normal fluctuation range of group behavior.

[0038] The current state vector is input into the state space model for prediction to obtain the prediction result of the behavior state in the next time period. The prediction result is used as the behavior stability discrimination vector and compared with the steady state interval.

[0039] In this embodiment, a state-space model for predicting behavioral states is constructed based on historical group behavior data and the current behavior state vector, as follows:

[0040] Based on the state vector sequence of continuous time points in historical group behavior data, the relationship between the changes in state vectors at adjacent time points is analyzed, and the statistical law of the evolution of behavior state over time is extracted as the basic evolution rule of the state space model.

[0041] Based on historical group behavior data, the direction and magnitude of change of different state vector dimensions over time are statistically analyzed to form a state transition structure that describes how the current state evolves to the next state, thereby constructing a predictive model of behavioral state.

[0042] The actual observed behavioral evolution data in historical group behavior data are correlated with the corresponding state vectors to determine the mapping relationship between the observed behavioral data and the state vectors, which serves as the observation structure in the state space model.

[0043] By fitting historical group behavior data, the data in the state transition structure and observation structure are determined, enabling the model to output the predicted behavioral state based on the current state vector.

[0044] During model operation, the behavior state vector at the current time point is used as the model input, and the behavior state at the next time point or the current period is predicted based on the constructed state transition structure, thereby achieving the prediction of behavior trends.

[0045] The predicted behavior stability discriminant vector is compared with the steady-state interval.

[0046] S2, based on stable group behavior data, perform target group analysis on users, generate target group analysis results, and generate a group strategy data set;

[0047] In this embodiment, target group analysis is performed on users based on stable group behavior data, target group analysis results are generated, and a group strategy data set is generated, as follows:

[0048] Based on stable behavior data of the group, the stable behavior characteristics of users that continuously appear in multiple live broadcast cycles are time-aligned, and the stable behavior status of the same user is aggregated across cycles to form a set of stable behavior status of the group. The stable behavior characteristics include average viewing time, viewing frequency, number of likes, number of comments, frequency of sending bullet comments, switching and clicking behavior statistics, which are used to characterize the long-term stable behavior pattern of users.

[0049] Each user behavior state in the set of stable behavior states of the group is represented as a multi-dimensional vector. Each dimension corresponds to a stable behavior feature. Each vector is standardized so that behavior features of different dimensions and frequencies can be compared in the same space to form a behavior state space. Each point in the behavior state space represents the overall behavior state of the user at a certain point in time. The entire space carries the distribution of stable behavior states of all users at different points in time.

[0050] The stable behavioral state set of the group is mapped to the behavioral state space, and the transition relationship of user behavioral state between adjacent time points is analyzed based on the discrete time state transition model to obtain the distribution characteristics and state transition trend of users in the behavioral state space.

[0051] Based on distribution characteristics and state migration trends, stable state clustering regions that exhibit stable convergence characteristics within a preset state offset threshold are identified, and users who have been in the same stable state clustering region for a long time are classified as candidate target groups.

[0052] The attractor decision mechanism identifies user groups with stable convergence characteristics in the behavior state space of candidate target groups, and classifies users who converge to the same behavior attraction domain into the same target group, forming the final target group division result, so as to avoid static grouping based solely on distance or similarity. The attractor decision mechanism is based on the long-term distribution characteristics of user behavior states in the state space and the state transition convergence, and is used to identify whether user behavior states have stably converged to a preset stable state region. The behavior attraction domain is the state cluster region in the state space where user behavior states have remained stable over a long period of time.

[0053] By combining the behavioral state vectors of convergent users in the target group in the behavioral state space with the corresponding attraction domain boundaries, the mean, variance and core state vectors of each stable behavioral feature are statistically analyzed to form the target group analysis results. Among them, the core state vector is the most concentrated and typical long-term behavioral state vector of the target group in the behavioral state space, which is obtained by calculating the mean of each dimension of the behavioral state vector of convergent users.

[0054] Based on the target group analysis results, a multi-constraint feasible solution space is constructed. The live streaming integration elements are used as adjustment strategy variables. The range of values ​​of the live streaming integration elements is constrained according to the attraction domain boundary corresponding to the target group, and a candidate set of group strategies is generated.

[0055] Constraint satisfaction screening is performed on the candidate set of group strategies, eliminating strategy combinations that exceed the stable behavior boundary of the target group and cause deviations in the group's behavior state. The screened strategy combinations are then structured and encapsulated to form a group strategy data set.

[0056] In this embodiment, the stable behavioral state set of the group is mapped to the behavioral state space, and the transition relationship of user behavioral states between adjacent time points is analyzed based on the discrete-time state transition model to obtain the distribution characteristics and state transition trends of users in the behavioral state space, as detailed below:

[0057] Arrange the behavioral state vectors in the set of stable group behavioral states in chronological order to form a discrete-time behavioral state sequence.

[0058] Input the discrete-time behavior state sequence into the discrete-time state transition model, analyze the probability of state change at adjacent time points in the sequence, and output the probability of each state vector transitioning to other state vectors to form a state transition matrix.

[0059] Based on the distribution of the behavior state vector in the behavior state space, the distribution density of user behavior states is statistically analyzed.

[0060] Based on the state transition matrix, the transition probability of each state point is calculated and combined with the state space coordinates to output the weighted average transition direction to reflect the transition direction of the behavioral state. The number of transitions from the state is also calculated based on the state transition matrix to obtain the transition frequency.

[0061] By integrating the migration directions and frequencies of all state points, a migration trend of user behavior in the state space is formed;

[0062] Based on distribution density and migration trends, identify clusters of users' long-term stable behaviors, i.e., stable state clusters.

[0063] In this embodiment, the attractor decision mechanism is used to identify user groups with stable convergence characteristics in the behavioral state space of the candidate target group, and users who converge to the same behavioral attraction domain are classified into the same target group, forming the final target group classification result, as follows:

[0064] Based on the behavioral state sequence and migration trend of the candidate target group, the convergence characteristics of each user in the potential attraction domain are calculated and compared with the preset convergence threshold. The convergence characteristics include the proportion of the state vector falling into the attraction domain, the average dwell time, and the proportion of the migration direction pointing towards the center of the attraction domain.

[0065] Users whose convergence meets the threshold are grouped into the same target group, and the distribution of the convergent users' behavioral state vectors in the behavioral state space is statistically analyzed. The distribution of the behavioral state vectors includes the mean, variance, maximum value and minimum value of each dimension.

[0066] Based on the distribution of behavioral state vectors, identify dense state regions and output the boundary of dense state regions using the convex hull method, using this boundary as the boundary of the attraction domain.

[0067] The boundary of the behavioral attraction domain is taken as the stable state region of the target group, thus forming the final target group segmentation result.

[0068] In this embodiment, based on the target group analysis results, a multi-constraint feasible solution space is constructed. The live streaming fusion elements are used as adjustment strategy variables, and the value range of the live streaming fusion elements is constrained according to the attraction domain boundary corresponding to the target group, generating a candidate set of group strategies, as follows:

[0069] The live streaming integration elements are used as strategy variables, and an initial value range is set for each variable. The live streaming integration elements include content rhythm, interaction intensity, operation complexity, and the number of recommended products.

[0070] The attraction domain boundary of the target group's stable state region is mapped to the range of policy variable values ​​to form constraints, ensuring that policy combinations do not cause the group's behavior to deviate from the stable state.

[0071] The constraints of all policy variables are combined to form a multi-constraint feasible solution space, with each dimension corresponding to the range of possible values ​​for a policy variable;

[0072] Discretize the solution space within the multi-constraint feasible solution space to generate a list of strategy combinations;

[0073] The feasibility of the generated strategy combination list is verified, combinations that cause the group behavior state to deviate from the stable region are removed, and the verified strategy combinations are structurally encapsulated to form a group strategy candidate set.

[0074] S3, based on the group strategy data set, constructs a recommendation strategy space corresponding to the target group, and constrains the selection range of live streaming integration elements;

[0075] In this embodiment, a recommendation strategy space corresponding to the target group is constructed based on the group strategy data set, and the selection range of live streaming fusion elements is constrained, as follows:

[0076] By using live streaming integration elements as strategy variables, all combinations in the group strategy dataset are mapped to a multi-dimensional space. Each dimension corresponds to the initial value range of a live streaming integration element, thus forming a recommendation strategy space for the target group.

[0077] Cluster the mapped policy space, group similar policies into the same policy region, and form a policy subspace;

[0078] For each strategy subspace, the range of values ​​for strategy variables is constrained according to the attraction domain boundary of the target group, and combinations that cause the group behavior to deviate from the steady state are eliminated. This includes: limiting the upper and lower limits of strategy variables to the range allowed in the steady state region, and eliminating strategy combinations that exceed the allowed range.

[0079] The constrained and partitioned strategy subspace is structured and encapsulated to form the final recommendation strategy space corresponding to the target group.

[0080] In this embodiment, for each strategy subspace, the value range of the strategy variables is constrained according to the attraction domain boundary of the target group, as follows:

[0081] Obtain the attraction domain boundary corresponding to the target group;

[0082] Establish a mapping relationship between each behavioral feature dimension in the behavioral state space and the live streaming fusion elements, so that changes in behavioral state can correspond to the adjustment direction and adjustment range of the live streaming fusion elements.

[0083] For each policy subspace, read the core state vector associated with that policy subspace and determine its positional relationship relative to the boundary of the behavior attraction domain in the behavior state space.

[0084] Based on the distance relationship between the group core state vector and the boundary of the behavior attraction domain, calculate the maximum allowable offset range of each behavioral feature dimension within the policy subspace.

[0085] The maximum allowable offset range of each behavioral feature dimension is mapped to the value variation range of the live streaming fusion element, thereby generating a corresponding strategy variable value range for each live streaming fusion element.

[0086] By using the range of values ​​for strategy variables as constraints, the strategy combinations within the strategy subspace are restricted so that the changes in user behavior caused by the application of the strategy combinations do not exceed the boundary of the attraction domain corresponding to the target group.

[0087] S4 outputs real-time user behavior status data based on individual real-time behavior data, and dynamically adjusts the live streaming integration elements within the recommendation strategy space based on the real-time behavior status data.

[0088] In this embodiment, real-time user behavior status data is output based on individual real-time behavior data, and the live streaming fusion elements are dynamically adjusted within the recommendation strategy space based on the real-time behavior status data, as follows:

[0089] Individual real-time behavioral data is processed according to behavioral type and high-frequency time window to extract evolutionary data;

[0090] The evolution data at each time point is combined into an instant state vector, representing the current user behavior state;

[0091] The instantaneous state vector is compared with the historical stable behavior state vector to output the behavior deviation. The historical stable behavior state vector is constructed based on the aforementioned group stable behavior data, which comes from the user's behavior data that was determined to be in a steady state range over multiple live streaming periods.

[0092] Input the instantaneous state vector into the state space model and output a short-term behavioral trend prediction to determine the possible future behavioral state of an individual.

[0093] In the recommendation strategy space, identify the strategy subspace that best matches the immediate state vector, and combine it with short-term behavior trend prediction to determine the strategy subspace corresponding to the current user behavior state, and select an initial strategy combination from the strategy subspace.

[0094] Based on the offset direction of the real-time behavior state vector relative to the center of the stable state region of the target group, the adjustment direction of the live streaming integration elements is determined. Specifically, when the user behavior state deviates outward from the stable state region, the content rhythm, interaction intensity, or number of recommended products are enhanced and adjusted. When the user behavior state converges inward from the stable region, the operation complexity or interaction frequency is moderated and adjusted. This achieves differentiated dynamic adjustment of different live streaming integration elements.

[0095] An initial adjustment sensitivity is set for each live streaming integration element. The initial adjustment sensitivity is determined by the width of the stable state region of the target group. It is used to limit the single adjustment range so that the individual adjustment always falls within the acceptable range corresponding to the stable behavior attraction domain of the target group, and avoids causing the overall shift of the group behavior state due to group behavior instability.

[0096] Within a continuous time window, the adjusted live streaming integration elements are progressively updated, with subsequent adjustments taking the previous adjustment results and the latest real-time behavior state vector as input, to achieve a continuous and smooth evolution of the live streaming integration elements as user behavior changes.

[0097] The final adjusted live streaming integration elements are output, and the real-time behavioral state vectors generated by the users in response to the adjustment results are fed back into the real-time behavioral state update process to drive the next round of strategy subspace selection and element adjustment.

[0098] In this embodiment, the strategy subspace that best matches the immediate state vector is identified in the recommendation strategy space. Combined with short-term behavior trend prediction, the strategy subspace corresponding to the current user behavior state is determined, and an initial strategy combination is selected from this strategy subspace, as follows:

[0099] Each policy subspace in the recommended policy space is associated with its corresponding core state vector and the boundary of the attraction domain.

[0100] After obtaining the user's real-time behavior state vector and its short-term behavior trend prediction results, the real-time behavior state vector is mapped to the behavior state space, and the relative positional relationship of the real-time behavior state vector with respect to each core state vector in the behavior state space is output.

[0101] Based on the relative positional relationship, determine whether the instantaneous behavior state vector falls within the boundary of the attraction domain mapped by the corresponding policy subspace;

[0102] When the instantaneous behavior state vector simultaneously falls into the attraction domain boundary corresponding to a certain strategy subspace, and its short-term behavior trend prediction result indicates that the future state will remain or converge into the attraction domain boundary, then the strategy subspace is determined to be the strategy subspace corresponding to the current user behavior state.

[0103] Within a defined policy subspace, based on the relative position of the instantaneous behavior state vector within that subspace, the policy combination closest to its position or with the same direction of change is selected as the initial policy combination.

[0104] In this embodiment, initial adjustment sensitivities are set for each live streaming fusion element, as follows:

[0105] Based on the boundary of the stable state region corresponding to the target group, the allowable fluctuation range of each behavioral feature dimension in the behavioral state space within the stable state region is determined, and the correspondence between the behavioral feature dimensions and the live streaming fusion elements is established, so that changes in the behavioral state space can be mapped to the adjustment direction of the live streaming fusion elements.

[0106] Based on the maximum permissible offset of each behavioral feature dimension within the stable state region, the maximum permissible change of the corresponding live streaming fusion element in the adjustment dimension is obtained.

[0107] The maximum allowable change is quantified according to a preset ratio and used as the unit adjustment step size of the live streaming fusion element. The unit adjustment step size is set as the initial adjustment sensitivity of the live streaming fusion element.

[0108] By initially adjusting the sensitivity to limit the adjustment range of the integrated elements in a single live broadcast, the adjusted strategy combination always falls within the acceptable range corresponding to the stable state region, thereby avoiding the overall deviation of the group's behavioral state caused by individual instantaneous adjustments.

[0109] S5 generates and outputs recommendation results based on the dynamically adjusted live streaming integration elements;

[0110] In this embodiment, recommendation results are generated and output based on the dynamically adjusted live streaming fusion elements, as detailed below:

[0111] The dynamically adjusted live streaming integration elements are mapped to the preset live streaming recommendation execution template to determine the presentation order of live streaming content, interaction trigger frequency, operation process complexity, and product recommendation display method.

[0112] Based on the integrated elements of live streaming, the current live streaming scenario is generated in real time to form a live streaming recommendation scheme that matches the user's current behavior status.

[0113] The live streaming recommendation scheme is applied to the live streaming process to control the playback pace of live streaming content, the display method of interactive components, and the product recommendation module in real time.

[0114] Output recommendation results generated based on the live streaming recommendation scheme to users, including the display effect of live streaming content, the form of interactive guidance, and the product recommendation results;

[0115] During the recommendation result output process, real-time behavioral data of users in response to the recommendation results is recorded simultaneously, which is used as input data for subsequent real-time behavioral status updates and dynamic adjustments of live streaming integration elements in the next round.

[0116] Example 2, Figure 2 The present invention provides a system for a live streaming fusion element recommendation method that combines target group analysis, comprising a data acquisition module, an analysis module, a constraint module, an adjustment module, and an output module, with connections between the modules;

[0117] The data acquisition module is used to acquire user behavior data during the live broadcast, and to classify the behavior data into stable group behavior data and individual real-time behavior data.

[0118] The analysis module is used to perform target group analysis on users based on stable group behavior data, generate target group analysis results, and generate a group strategy data set.

[0119] The constraint module is used to construct a recommendation strategy space corresponding to the target group based on the group strategy data set, and to constrain the selection range of live streaming integration elements.

[0120] The adjustment module is used to output real-time user behavior status data based on individual real-time behavior data, and to dynamically adjust the live streaming integration elements within the recommendation strategy space based on the real-time behavior status data.

[0121] The output module is used to generate and output recommendation results based on the dynamically adjusted live streaming integration elements.

[0122] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor of an electronic device, can implement the steps in the above-described live streaming fusion element recommendation method incorporating target group analysis.

[0123] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0124] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0125] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0126] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0127] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A live streaming element recommendation method that combines target group analysis, characterized in that, include: Acquire user behavior data during live streaming, and segment the behavior data to form stable group behavior data and individual real-time behavior data; Based on stable group behavior data, target group analysis is performed on users to generate target group analysis results and a group strategy data set. Based on the group strategy data set, a recommendation strategy space corresponding to the target group is constructed, and the selection range of live streaming integration elements is constrained. Based on individual real-time behavior data, output user real-time behavior status data, and dynamically adjust the live streaming integration elements within the recommendation strategy space based on the real-time behavior status data; Recommendation results are generated and output based on the dynamically adjusted live streaming integration elements; The process involves performing target group analysis on users based on stable group behavior data, generating target group analysis results, and generating a group strategy data set, as detailed below: Based on stable behavior data of the group, the stable behavior characteristics of users that continuously appear in multiple live broadcast cycles are time-aligned, and the stable behavior status of the same user is aggregated across cycles to form a set of stable behavior status of the group. Each user behavior state in the set of stable group behavior states is represented as a multi-dimensional vector, with each dimension corresponding to a stable behavior feature, forming a behavior state space. The stable behavioral state set of the group is mapped to the behavioral state space, and the transition relationship of user behavioral state between adjacent time points is analyzed based on the discrete time state transition model to obtain the distribution characteristics and state transition trend. Based on distribution characteristics and state migration trends, stable state clustering regions that exhibit stable convergence characteristics within a preset state offset threshold are identified, and users who have been in the same stable state clustering region for a long time are classified as candidate target groups. The attractor decision mechanism is used to identify user groups with stable convergence characteristics in the behavioral state space of the candidate target group, and users who converge to the same behavioral attraction domain are classified into the same target group to form the final target group classification result. By combining the behavioral state vectors of convergent users in the target group in the behavioral state space with the corresponding attraction domain boundary, the target group analysis results are obtained. Based on the target group analysis results, a multi-constraint feasible solution space is constructed. The live streaming integration elements are used as adjustment strategy variables. The range of values ​​of the live streaming integration elements is constrained according to the attraction domain boundary corresponding to the target group, and a candidate set of group strategies is generated. Constraints are satisfied and the candidate set of group strategies is filtered to form a group strategy data set.

2. The live streaming fusion element recommendation method combining target group analysis according to claim 1, characterized in that, The process of acquiring user behavior data during the live stream and dividing the data into stable group behavior data and real-time individual behavior data is as follows: The behavioral data is categorized according to behavioral type, and behavioral time series are constructed based on timestamp information. Set a first time window and a second time window in the behavioral time series; Behavioral evolution data are output in the first time window and the second time window respectively. The behavioral evolution data includes the behavior duration ratio, behavior occurrence density and behavior change slope. Input the behavioral evolution data into the preset state space model to generate a behavioral stability discrimination vector and determine whether the current behavior is in the steady state range; If the behavior stability discrimination vector continuously falls within the steady-state interval within the first time window, the corresponding behavior data will be marked as stable behavior data of the group. If the behavior stability discrimination vector shifts significantly within the second time window relative to the steady-state interval of the first time window, and the shift exceeds a preset threshold, then the corresponding behavior data will be marked as individual real-time behavior data.

3. The live streaming fusion element recommendation method combining target group analysis according to claim 2, characterized in that, The process of inputting behavioral evolution data into a preset state space model to generate a behavioral stability discrimination vector and determining whether the current behavior is in a steady-state range is as follows: The behavioral evolution data corresponding to each time point is combined according to a fixed dimension to form the current behavioral state vector; Based on the stable behavior data of the group and the current behavior state vector, a state space model for predicting the behavior state is constructed. Based on the stable behavior data of the group, the mean and standard deviation of each state vector dimension are calculated, and the steady-state interval of each state feature is output. The current state vector is input into the state space model for prediction to obtain the prediction result of the behavior state in the next time period. The prediction result is used as the behavior stability discrimination vector and compared with the steady state interval.

4. The live streaming fusion element recommendation method combining target group analysis according to claim 3, characterized in that, The state-space model for predicting behavioral states is constructed based on stable group behavior data and the current behavioral state vector, as detailed below: Based on the state vector sequence of continuous time points in the group stability behavior data, analyze the relationship between the changes in state vectors at adjacent time points; Based on stable group behavior data, the direction and magnitude of change of different state vector dimensions over time are statistically analyzed to form a state transition structure, thereby constructing a predictive model of behavioral states. The actual observed behavioral evolution data in the stable behavior data of the group are correlated with the corresponding state vectors to determine the mapping relationship between the observed behavioral data and the state vectors, which serves as the observation structure in the state space model. By fitting the stable behavior data of the group, the data in the state transition structure and observation structure are determined, so that the model can output the predicted behavior state based on the current state vector. During model operation, the behavior state vector at the current time point is used as the model input, and the behavior state at the next time point and the current period is predicted based on the constructed state transition structure, thereby achieving the prediction of behavior trends.

5. The live streaming element recommendation method combining target group analysis according to claim 1, characterized in that, The process involves mapping the stable set of group behavior states to the behavior state space, and analyzing the transition relationship of user behavior states between adjacent time points based on a discrete-time state transition model to obtain distribution characteristics and state transition trends, as detailed below: Arrange the behavioral state vectors in the stable behavioral state set of the group in chronological order to form a discrete-time behavioral state sequence. Input the discrete-time behavior state sequence into the discrete-time state transition model, analyze the probability of state change at adjacent time points in the sequence, and output the probability of each state vector transitioning to other state vectors to form a state transition matrix. Based on the distribution of the behavior state vector in the behavior state space, the distribution density of user behavior states is statistically analyzed. Based on the state transition matrix, the transition probability of each state point is calculated and combined with the state space coordinates to output the weighted average migration direction. The number of transitions from the state is also calculated based on the state transition matrix to obtain the migration frequency. By integrating the migration directions and frequencies of all state points, a migration trend of user behavior in the state space is formed; Based on distribution density and migration trends, identify stable clustering areas.

6. The live streaming fusion element recommendation method combining target group analysis according to claim 1, characterized in that, The process involves constructing a recommendation strategy space corresponding to the target group based on the group strategy data set, and constraining the selection range of live streaming integration elements, as detailed below: By using live streaming integration elements as strategy variables, all combinations in the group strategy data set are mapped to a multi-dimensional space. Each dimension corresponds to the initial value range of a live streaming integration element, forming a recommendation strategy space for the target group. Cluster the mapped policy space, group similar policies into the same policy region, and form a policy subspace; For each strategy subspace, the range of values ​​for strategy variables is constrained by the attraction domain boundary of the target group, and combinations that cause the group behavior to deviate from the stable state are eliminated. The constrained and partitioned strategy subspace is structured and encapsulated to form the final recommendation strategy space corresponding to the target group.

7. The live streaming element recommendation method combining target group analysis according to claim 1, characterized in that, The process involves outputting real-time user behavior status data based on individual real-time behavior data, and dynamically adjusting the live streaming integration elements within the recommendation strategy space based on the real-time behavior status data, as detailed below: Individual real-time behavioral data is processed according to behavioral type and high-frequency time window to extract evolutionary data; The evolution data at each time point is combined into an instant state vector, and the instant state vector is compared with the historical stable behavior state vector to output the behavior deviation. Input the instantaneous state vector into the state space model and output a short-term behavior trend prediction; In the recommendation strategy space, identify the strategy subspace that best matches the immediate state vector, and combine it with short-term behavior trend prediction to determine the strategy subspace corresponding to the current user behavior state, and select an initial strategy combination from the strategy subspace. The adjustment direction of live streaming fusion elements is determined based on the offset direction of the real-time behavior state vector relative to the center of the target group's stable state region. An initial adjustment sensitivity is set for each live streaming fusion element, and the adjusted live streaming fusion elements are progressively updated within a continuous time window; The final adjusted live streaming integration elements are output, and the real-time behavioral state vectors generated by the user's response to the adjustment results are fed back into the real-time behavioral state update process.

8. A system using the live streaming fusion element recommendation method combined with target group analysis as described in any one of claims 1-7, characterized in that, It includes a data acquisition module, an analysis module, a constraint module, an adjustment module, and an output module, and the modules are interconnected. The data acquisition module is used to acquire user behavior data during the live broadcast, and to classify the behavior data into stable group behavior data and individual real-time behavior data. The analysis module is used to perform target group analysis on users based on stable group behavior data, generate target group analysis results, and generate a group strategy data set. The constraint module is used to construct a recommendation strategy space corresponding to the target group based on the group strategy data set, and to constrain the selection range of live streaming integration elements. The adjustment module is used to output real-time user behavior status data based on individual real-time behavior data, and to dynamically adjust the live streaming integration elements within the recommendation strategy space based on the real-time behavior status data. The output module is used to generate and output recommendation results based on the dynamically adjusted live streaming integration elements.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the live streaming fusion element recommendation method that combines target group analysis as described in any one of claims 1 to 7.

Images