Multi-source feature adaptive fusion and dynamic grouping behavior prediction method

Abstract

The application discloses a kind of multi-source feature adaptive fusion and dynamic group's behavior prediction method, the method includes: S1, the multi-source heterogeneous behavior data corresponding to target user is collected, the unified time axis alignment, missing compensation, scale normalization and semantic mapping processing are carried out to the multi-source heterogeneous behavior data, obtain unified time sequence feature set;S2, based on the unified time sequence feature set, respectively extract time sequence response fingerprint for representing behavior response law according to feature source, the present application relates to user behavior prediction technical field.This kind of multi-source feature adaptive fusion and dynamic group's behavior prediction method, based on continuous time window, constructs user group evolution trajectory atlas, can dynamically reflect the migration process of user between different groups, so that prediction model not only considers individual characteristics, but also combines group evolution trend, improves the foresight and stability of prediction.

Description

Technical Field

[0001] This invention relates to the field of behavior prediction, and more specifically, to a behavior prediction method based on multi-source feature adaptive fusion and dynamic grouping. Background Technology

[0002] In today's digital age, user behavior data is characterized by its multi-source heterogeneity, with diverse sources and formats, including but not limited to text behavior records, log behavior records, and event tag records. Furthermore, sampling frequencies and time granularities also vary. While this data contains a wealth of user behavior information, directly utilizing it for analysis and prediction faces numerous challenges.

[0003] Traditional user behavior prediction methods often fail to adequately consider the heterogeneity of multi-source data. The lack of unified and effective processing methods in the data preprocessing stage leads to inconsistent data quality, affecting the accuracy of subsequent analysis. Furthermore, feature extraction struggles to comprehensively and accurately characterize user behavior response patterns, failing to fully uncover the underlying information within the data.

[0004] Most methods for user segmentation use a static approach, which fails to reflect the dynamic changes of user groups over time and ignores the process of group evolution, resulting in discrepancies between prediction results and actual user behavior.

[0005] In terms of multi-source feature fusion, traditional methods usually use fixed weight fusion, which does not take into account the credibility of different feature sources, inter-source conflicts and complementary relationships, and cannot dynamically adjust feature weights according to the actual situation, resulting in the fused features not being able to effectively reflect the user's real behavioral characteristics.

[0006] Furthermore, existing methods lack an effective linkage and update mechanism after obtaining real behavioral feedback. They cannot adjust the user group evolution trajectory, feature source credibility, and fusion weight in a timely manner, making it difficult to continuously adapt to changes in user behavior and limiting prediction accuracy and practicality. Summary of the Invention

[0007] The purpose of this invention is to provide a behavior prediction method based on multi-source feature adaptive fusion and dynamic grouping. This method solves the problems of existing user behavior prediction methods, most of which use static partitioning, which cannot reflect the dynamic changes of user groups over time and ignore the process of group evolution. This results in a deviation between the prediction results and the actual user behavior. Furthermore, the fixed weight fusion method cannot dynamically adjust the feature weights according to the actual situation, which leads to the fused features not effectively reflecting the real user behavior characteristics and failing to meet the usage requirements.

[0008] This invention achieves the above objective through the following technical solution: a behavior prediction method based on multi-source feature adaptive fusion and dynamic grouping, the method comprising:

[0009] S1. Collect multi-source heterogeneous behavior data corresponding to the target user, and perform unified time axis alignment, missing data compensation, scale normalization and semantic mapping on the multi-source heterogeneous behavior data to obtain a unified temporal feature set.

[0010] S2. Based on the unified temporal feature set, extract temporal response fingerprints to characterize the behavior response pattern according to the feature source. The temporal response fingerprints include at least the behavior trigger lag, response duration, response fluctuation, and scene sensitivity.

[0011] S3. Using continuous time windows as analysis units, construct a user group evolution trajectory map based on the time-series response fingerprints of each user, and determine the dynamic group to which the target user belongs and the group migration path based on the user group evolution trajectory map.

[0012] S4. Construct counterfactual perturbation samples for each feature source, input the original sample and the counterfactual perturbation sample into the same prediction inference process, and obtain the difference in prediction response of each feature source before and after perturbation.

[0013] S5. Based on the predicted response differences, determine the confidence parameters, inter-source conflict parameters, and inter-source complementarity parameters of each feature source, and construct the feature source confidence matrix and inter-source complementarity relationship matrix accordingly.

[0014] S6. Input the target user's dynamic group, group migration path, feature source credibility matrix and inter-source complementary relationship matrix into the fusion decision process to generate adaptive fusion weights for each feature source, and perform reliable fusion of multi-source features based on the adaptive fusion weights.

[0015] S7. Input the fused multi-source features and the group evolution trajectory constraints corresponding to the target user into the user behavior prediction model, and output the behavior prediction results of the target user within the target time window.

[0016] S8. After obtaining real behavioral feedback, based on the deviation between the behavior prediction result and the real behavioral feedback, the user group evolution trajectory map, feature source credibility matrix and adaptive fusion weight are updated in a linked manner.

[0017] Furthermore, the process of performing unified timeline alignment, missing data compensation, scale normalization, and semantic mapping on multi-source heterogeneous behavioral data includes:

[0018] Map multi-source heterogeneous behavioral data with different sampling frequencies and time granularities to the same time axis under a unified standard timestamp;

[0019] The missing data is compensated based on the valid data before and after the missing position;

[0020] Convert data with different units and value ranges to a unified numerical range;

[0021] Text behavior records, log behavior records, and event tag records are converted into structured numerical features, thereby forming a unified time-series feature set that is uniform in format, continuous in time sequence, and consistent in semantics.

[0022] Furthermore, the extraction process of the timing response fingerprint includes:

[0023] The behavior trigger hysteresis is determined based on the time series response curves of each feature source to characterize the delay from input of the behavior signal to the generation of an effective response;

[0024] The duration of a response is determined based on the effective response sustainment interval to characterize the duration of a behavioral response;

[0025] The response fluctuation is determined based on the degree of dispersion within the response interval to characterize the stability of the behavioral response;

[0026] The scene sensitivity quantity is determined based on the degree of correlation between the feature response and scene variables under each scene label, so as to characterize the adaptability of the corresponding feature source to changes in the application scene.

[0027] The behavior trigger hysteresis, response duration, response fluctuation, and scene sensitivity are combined to form the temporal response fingerprint of the corresponding feature source.

[0028] Furthermore, the process of constructing the user group evolution trajectory map includes:

[0029] Convert each user's temporal response fingerprint into a fixed-dimensional fingerprint vector;

[0030] Initial group segmentation is completed based on the fingerprint similarity among users;

[0031] By tracking the changes in the location and membership of each group center within a continuous time window, a user group evolution trajectory map reflecting the evolutionary state of the group is formed;

[0032] Based on the target user's current position in the user group evolution trajectory map, its historical dynamic group, and its migration direction in adjacent time windows, the dynamic group to which the target user belongs and the corresponding group migration path are determined.

[0033] Furthermore, the construction process of the counterfactual perturbation sample includes:

[0034] For each feature source, while keeping the information of other feature sources for the same user unchanged, the input value, time segment distribution or scene association item of the feature source is subjected to controlled perturbation to obtain counterfactual perturbation samples that correspond one-to-one with the original samples.

[0035] The original sample and the counterfactual perturbation sample are respectively input into the same prediction inference process to obtain the change in the prediction result before and after the perturbation, and the change is used as the basis for evaluating the strength and stability of the influence of the feature source on the behavior prediction result.

[0036] Furthermore, the construction process of the feature source confidence matrix and the inter-source complementary relationship matrix includes:

[0037] The confidence parameter of a feature source is determined based on the magnitude of the shift in the behavior prediction result caused by multiple counterfactual perturbations.

[0038] The corresponding inter-source conflict parameter is determined based on the degree of deviation between the predicted outputs of any two feature sources on the same user sample.

[0039] The corresponding inter-source complementarity parameter is determined based on the degree of complementarity between any two feature sources in their contribution to behavior prediction.

[0040] The confidence matrix of the feature sources and the inter-source complementarity matrix are formed based on the confidence parameters of all feature sources, the inter-source conflict parameters, and the inter-source complementarity parameters, respectively.

[0041] Furthermore, the process of generating the adaptive fusion weights includes:

[0042] The group confidence information of the dynamic group to which the target user belongs, the path constraint information of the group migration path, the credibility parameters of the corresponding feature sources, and the complementary parameters between sources are all input into the fusion decision process.

[0043] Feature sources with high credibility and high matching degree with the current evolution state of the dynamic group to which the target user belongs are given higher fusion weights, and feature sources with strong complementary ability to other feature sources are given enhanced weights.

[0044] The standardized feature vectors corresponding to each feature source are weighted and combined according to the adaptive fusion weights to obtain the fused multi-source features.

[0045] Furthermore, the user behavior prediction model is used to receive the fused multi-source features and group evolution trajectory constraint information, and output the behavior prediction results of the target user within the target time window;

[0046] The group evolution trajectory constraint information is used to limit the learning range of the user behavior prediction model for the target user's behavior evolution direction, behavior transfer boundary and behavior occurrence probability distribution, so that the behavior prediction result simultaneously satisfies the individual characteristic expression of the target user and the group evolution law of the dynamic group to which it belongs.

[0047] Furthermore, the linked update process includes:

[0048] Based on the deviation between the behavior prediction results and the actual behavior feedback, the corresponding parameters in the feature source credibility matrix are adjusted, and the relevant parameters in the inter-source complementary relationship matrix are corrected simultaneously.

[0049] When the deviation value exceeds a preset tolerance threshold, the degree of suppression of low-confidence feature sources that cause the deviation is increased;

[0050] When the deviation value does not exceed the preset tolerance threshold, the verified and effective fusion relationship is maintained so that the trusted fusion process can continuously adapt to changes in user behavior.

[0051] Furthermore, the linked update of the user group evolution trajectory map includes:

[0052] The group affiliation results of target users in the current time window are re-verified based on real behavioral feedback;

[0053] When behavioral prediction bias indicates a decline in the accuracy of existing group segmentation, adjust the group segmentation threshold and group center update rules, and reconstruct the group migration path of target users in adjacent time windows.

[0054] The updated group affiliation results, group migration paths, feature source credibility matrix, and adaptive fusion weights are re-inputted into the next round of user behavior prediction process to form a closed-loop iterative optimization mechanism.

[0055] The beneficial effects of this invention are as follows:

[0056] 1. By performing unified timeline alignment, missing data compensation, scale normalization, and semantic mapping on multi-source heterogeneous behavioral data, a unified temporal feature set is constructed, which effectively eliminates structural differences between data sources and improves the consistency and usability of feature representation.

[0057] 2. By extracting time-series response fingerprints that include behavior trigger lag, response duration, response fluctuation, and scenario sensitivity, the response patterns of user behavior can be comprehensively characterized from multiple dimensions, thereby significantly improving the precision of behavior modeling.

[0058] 3. Constructing a user group evolution trajectory map based on continuous time windows can dynamically reflect the migration process of users between different groups, enabling the prediction model to not only consider individual characteristics but also combine group evolution trends, thereby improving the foresight and stability of the prediction.

[0059] 4. By constructing counterfactual perturbation samples, the influence of each feature source on the prediction results is quantified, thereby obtaining feature source credibility parameters, inter-source conflict parameters, and complementary parameters, and realizing interpretable modeling of the contribution relationship of multi-source features.

[0060] 5. Based on the feature source credibility matrix and the inter-source complementary relationship matrix, and combined with the user's dynamic group and migration path, dynamic fusion weights are generated to give higher weights to feature sources with high credibility and matching the current behavior state, thereby improving the reliability of the fusion results.

[0061] 6. Inputting the fused multi-source features and population evolution trajectory constraints into the prediction model allows the prediction results to take into account both individual behavioral characteristics and population evolution patterns, effectively reducing noise interference and improving prediction accuracy.

[0062] 7. By introducing a linkage update mechanism based on real behavior feedback, the evolution trajectory map of user groups, the credibility matrix of feature sources, and the fusion weights are dynamically adjusted, enabling the system to continuously adapt to changes in user behavior and achieve long-term stable optimization. Attached Figure Description

[0063] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0064] Figure 1 This is a flowchart illustrating the overall method of the present invention;

[0065] Figure 2 This is a flowchart of the data preprocessing process of the present invention;

[0066] Figure 3 This is a flowchart illustrating the construction of the dynamic population evolution trajectory of this invention. Detailed Implementation

[0067] The present application will now be described in further detail with reference to the accompanying drawings. It should be noted that the following specific embodiments are only used to further illustrate the present application and should not be construed as limiting the scope of protection of the present application. Those skilled in the art can make some non-essential improvements and adjustments to the present application based on the above application content.

[0068] Example 1:

[0069] Please see Figure 1-3 This invention provides a technical solution: a behavior prediction method based on multi-source feature adaptive fusion and dynamic grouping, the method comprising:

[0070] S1. Collect multi-source heterogeneous behavioral data of target users, and perform time alignment, missing data compensation, scale normalization and semantic mapping on the multi-source heterogeneous behavioral data to generate a unified temporal feature set.

[0071] Among them, multi-source heterogeneous behavioral data refers to user behavior data from different data sources with different structures and characteristics; time alignment unifies multi-source data with different time granularities or different time bases to the same time scale, ensuring that the data is comparable and consistent in the time dimension; missing value compensation uses appropriate methods to fill or estimate missing values ​​in the data; scale normalization transforms data with different dimensions and value ranges to make them have the same scale; semantic mapping unifies the semantic interpretation and transformation of data with different semantic representations but essentially the same or related data from different data sources; and unified temporal feature set is the set of user behavior features with unified time series, complete data, the same scale, and consistent semantics obtained after the above time alignment, missing value compensation, scale normalization, and semantic mapping processes.

[0072] S2. Based on a unified temporal feature set, extract the temporal response fingerprint corresponding to each feature source. The temporal response fingerprint includes at least behavior trigger lag features, impact duration features, fluctuation features, and scene-sensitive features.

[0073] Among them, the temporal response fingerprint describes the response pattern of user behavior to different feature sources of stimulation in the time dimension. It comprehensively reflects the various characteristics of user behavior changes over time and is the basis for subsequent analysis of user behavior patterns and construction of user group evolution trajectory maps. The behavior trigger lag feature refers to the delay between the actual occurrence time of user behavior and the trigger time after being triggered by a certain stimulus or event. The influence duration feature indicates the duration of the influence of a certain stimulus or event on user behavior. The fluctuation feature describes the fluctuation and change of user behavior in the time series. The scene sensitivity feature reflects the sensitivity of user behavior to different scenes. Scenes can include time scenes, such as daytime and nighttime; location scenes, such as home, office, and shopping mall; and social scenes, such as being alone or having a gathering with friends.

[0074] S3. Based on the temporal response fingerprints of each user within a continuous time window, construct a user group evolution trajectory map, and perform dynamic grouping based on the user group evolution trajectory map to obtain the dynamic group to which the target user belongs and its group migration path.

[0075] Among them, the user group evolution trajectory map graphically displays the behavioral changes and group affiliation evolution of different users within a continuous time window. Through this map, the dynamic processes of user group formation, splitting, and merging, as well as the migration paths of individual users between groups, can be intuitively observed; dynamic clustering, based on the temporal response fingerprints of users within a continuous time window, uses clustering algorithms or other group partitioning methods to group users with similar behavioral characteristics and evolutionary trends into the same group. Unlike traditional static clustering, dynamic clustering considers the changes in user behavior over time and can more accurately capture the dynamic characteristics of user groups; group migration path records the trajectory information of users moving from one group to another during the dynamic clustering process. It reflects the changes in user behavior patterns and the evolution of group affiliation, which helps to deeply understand the dynamics and complexity of user behavior;

[0076] S4. Construct counterfactual perturbation samples for each feature source, and input the original samples and counterfactual perturbation samples into the prediction network to obtain the prediction response results of each feature source under different perturbation conditions;

[0077] Counterfactual perturbation samples are sample data generated by making certain hypothetical changes or perturbations to the original samples. In machine learning, counterfactual reasoning is used to explore how the outcome will change if certain conditions or factors change. Prediction networks are machine learning models or neural network structures used to predict user behavior. They receive input sample data, including original samples and counterfactual perturbation samples, and output corresponding prediction response results, i.e., prediction values ​​of future user behavior, by learning patterns and rules in the data. Prediction response results are the prediction values ​​of user behavior output by the prediction network based on the input sample data. Under different perturbation conditions, the prediction response results will be different. By analyzing these differences, we can understand the degree of influence of each feature source on the prediction results.

[0078] S5. Based on the difference in the predicted response of each feature source before and after the disturbance, calculate the confidence parameter, conflict parameter and complementarity parameter of the corresponding feature source, and construct the feature source confidence matrix and the inter-source complementarity matrix.

[0079] The reliability parameter measures the reliability of each feature source in predicting user behavior. By comparing the differences in the prediction responses of each feature source before and after perturbation, the stability and accuracy of the feature source's contribution to the prediction results can be evaluated. The conflict parameter reflects the degree of conflict between different feature sources in the prediction results. When multiple feature sources have significantly different prediction results for the same user behavior, it indicates that there is a conflict between these feature sources. The degree of conflict can be quantified by calculating the conflict parameter. For example, a distance metric or similarity metric can be used to calculate the difference between the prediction results of different feature sources as the value of the conflict parameter. The complementarity parameter represents the complementarity between different feature sources in predicting user behavior. Some feature sources may describe user behavior from different perspectives, and there is a certain complementary relationship between them. Their combined effect can improve the prediction. Accuracy is assessed by analyzing the differences in predicted responses of each feature source before and after perturbation, which reveals their complementary relationships and quantifies the degree of complementarity by calculating complementary parameters. The feature source credibility matrix represents the credibility parameters of each feature source in matrix form, with rows or columns corresponding to different feature sources and element values ​​indicating the credibility of the corresponding feature source. This matrix visually displays the credibility of each feature source, providing a basis for subsequent feature fusion. The inter-source complementarity matrix presents the complementary relationships between different feature sources in matrix form, with element values ​​representing complementary parameters between two corresponding feature sources. This matrix allows understanding of the complementarity between feature sources, enabling the rational utilization of these complementary relationships during feature fusion to improve fusion effectiveness.

[0080] S6. Input the target user's dynamic group, group migration path, feature source credibility matrix, and inter-source complementary relationship matrix into the fusion decision unit to generate adaptive fusion weights for each feature source and complete the reliable fusion of multi-source features.

[0081] The system comprises several components: a fusion decision unit (FDI), a module or algorithm that generates adaptive fusion weights for each feature source based on input information; an adaptive fusion weight unit (AFU), which receives input information such as the target user's dynamic group, group migration path, feature source credibility matrix, and inter-source complementarity matrix; and a fusion decision unit that calculates the weight of each feature source during the fusion process by comprehensively considering these factors and applying specific decision rules or algorithms. The adaptive fusion weights, based on the calculation results of the FDI unit, assign weights to each feature source. These weights are dynamically adjusted according to the target user's dynamic group information, group migration path, and the credibility and complementarity between feature sources, enabling more rational utilization of information from each feature source during multi-source feature fusion, thus improving fusion effectiveness and prediction accuracy. The credible fusion of multi-source features, based on the calculated adaptive fusion weights, weights and fuses the features from each feature source to obtain the fused multi-source features. This fusion method fully considers the credibility and complementarity of each feature source, enabling the fused features to more comprehensively and accurately reflect user behavior information, providing stronger support for subsequent user behavior prediction.

[0082] S7. Input the fused multi-source features and the group evolution trajectory constraints corresponding to the target user into the user behavior prediction model, and output the behavior prediction results of the target user within the preset time window.

[0083] Among them, the group evolution trajectory constraint refers to a constraint condition on the user behavior prediction model based on the evolution trajectory information of the dynamic group to which the target user belongs. It incorporates the dynamic change pattern of the user group into the prediction model, so that when predicting user behavior, the model not only considers the characteristics of the user itself, but also the influence of the evolution trend of the user's group on the user behavior. The user behavior prediction model is a machine learning model used to output the prediction result of the target user's behavior within a preset time window based on the fused multi-source features and the group evolution trajectory constraint. This model predicts the user's future behavior by learning patterns and rules in historical data and combining them with current input information. The preset time window refers to the pre-set time range for predicting user behavior.

[0084] S8. After obtaining real behavioral feedback, calculate the deviation between the predicted result and the real result, and update the user group evolution trajectory map, feature source credibility matrix and adaptive fusion weights in conjunction with the deviation.

[0085] Among these, the "real behavior feedback" refers to the actual behavior data of the target users acquired after the preset time window ends. This data serves as the real result and is used to compare with the prediction results output by the previous prediction model to evaluate the accuracy of the prediction. "Bias" refers to the degree of difference between the prediction result and the real result. Calculating the bias quantifies the accuracy of the prediction model; a smaller bias indicates that the prediction result is closer to reality, and the better the performance of the prediction model. Conversely, a larger bias indicates that the accuracy of the prediction model needs improvement. "Linked updates" is the process of simultaneously adjusting and optimizing the user group evolution trajectory map, feature source credibility matrix, and adaptive fusion weights based on the calculated bias. When there is a large deviation between the prediction result and the real result, it indicates that there may be problems with the current model parameters and input information. Linked updates can dynamically adjust these parameters and information, enabling the model to better adapt to changes in actual data and improve the accuracy and reliability of subsequent predictions.

[0086] It should be noted that, during use, multi-source heterogeneous behavioral data is processed to generate a unified temporal feature set, eliminating data differences and laying the foundation for subsequent analysis. Multiple temporal response fingerprints are extracted to comprehensively characterize user behavior features. A user group evolution trajectory map is constructed and dynamically grouped, accurately grasping the dynamic changes of user groups, clarifying the affiliation and migration path of target users, constructing counterfactual perturbation samples to obtain predicted response results, evaluating the impact of each feature source on prediction, calculating multiple parameters to construct a matrix, quantifying feature source characteristics, and generating adaptive fusion weights through fusion decision units to complete reliable fusion. The advantages of each feature source are fully utilized, and the fused features and group evolution trajectory constraints are input into the prediction model to improve prediction accuracy. Finally, key elements are updated in conjunction with real feedback, enabling the model to continuously adapt to data changes and continuously optimize prediction performance, providing a more reliable and accurate solution for user behavior prediction.

[0087] In one specific embodiment, taking the user behavior prediction scenario of an e-commerce platform as an example, multi-source joint modeling is performed on the target user's clicks, favorites, add-to-cart, order placement, page dwell time, search keywords, activity outreach feedback, and terminal environment changes data. 20,000 active users with complete behavior records over 90 consecutive days are selected as training samples. The multi-source heterogeneous behavioral data includes: in-site browsing behavior data, transaction behavior data, marketing outreach feedback data, search behavior data, and terminal context data, totaling five feature sources. Five feature sources are chosen because they correspond to user interest expression, transaction intent, external stimuli, proactive needs, and scenario constraints, respectively, covering the main information dimensions required for user behavior prediction while avoiding redundant computational burdens caused by too many feature sources.

[0088] In one embodiment, time alignment, missing data compensation, scale normalization, and semantic mapping are performed on multi-source heterogeneous behavioral data, specifically as follows:

[0089] Using a unified standard timestamp as a benchmark, multi-source heterogeneous behavioral data with different collection frequencies and different time dimensions are mapped to the same time axis to achieve strict alignment of the time dimension;

[0090] For data gaps that occur after time alignment, linear interpolation is used to compensate for missing values.

[0091] The compensated data is scaled to map data with different dimensions and different value ranges to the same numerical range, thus eliminating the feature bias caused by the difference in dimensions.

[0092] By semantic mapping, unstructured and semi-structured behavioral data are transformed into structured numerical features that the model can recognize, ultimately forming a unified time-series feature set with a consistent format and complete temporal sequence.

[0093] The calculation process for scale normalization is as follows:

[0094] Let the original behavioral data be The normalized data is ,but:

[0095]

[0096] in, For the minimum value of the behavioral data sequence, The maximum value of the behavioral data sequence;

[0097] Missing data compensation is calculated using linear interpolation:

[0098]

[0099] in, , Given the known data before and after the missing point. , , This corresponds to the timestamp.

[0100] In this embodiment, the unified time granularity is set to 1 hour, meaning all user behavior data is mapped to the same timeline in 1-hour units. This parameter is determined based on the following: when the time granularity is greater than 1 hour, such as 4 hours or 1 day, it easily masks short-term fluctuations in user behavior; when the time granularity is less than 1 hour, such as 5 minutes or 10 minutes, it leads to excessive sparse behavior, increasing the cost of missing data compensation. Therefore, a 1-hour granularity achieves a better balance between sensitivity to behavioral changes and data stability.

[0101] For missing data compensation, linear interpolation is used when the time interval between adjacent valid sampling points does not exceed three time steps. If the missing length exceeds three time steps, linear interpolation is still used, but a confidence flag is added to the compensation result for weight reduction during subsequent confidence calculation. The three time steps threshold is based on the following: in continuous behavior sequences, short-interval missing data usually originates from acquisition jitter, and the linear interpolation error is controllable; after three hours, the user's behavior state may have changed significantly, increasing the uncertainty of direct compensation, thus requiring reduction of its impact in subsequent fusion.

[0102] In the scale normalization process, all continuous features are mapped to intervals. The minimum and maximum values ​​are not fixed values ​​from the entire historical period, but rather statistical results within a sliding window of the most recent 30 days are used as the upper and lower bounds for normalization. The rationale for this is that using extreme values ​​from the entire historical period would easily lead to the compression of the current normal data distribution due to abnormal peak values; using a 30-day window can better reflect the true fluctuation range within the current business cycle.

[0103] For example, if the minimum value for page dwell time is 2 seconds and the maximum value is 480 seconds within the last 30 days, the normalized dwell time is calculated as follows:

[0104]

[0105] For semantic mapping, search keywords, page categories, activity tags, and terminal types are converted into structured features. Specifically, search keywords are represented using 128-dimensional word vectors, page categories using 16-dimensional one-hot encoding, activity tags using 12-dimensional multi-hot encoding, and terminal types using 4-dimensional one-hot encoding. This dimensionality setting is based on the following: 128 dimensions can meet the semantic compression requirements of keywords; page categories and terminal types belong to finite discrete sets, and one-hot representation can maintain interpretability; activity tags may coexist in multiple forms, therefore multi-hot encoding better reflects actual business characteristics.

[0106] In one embodiment, based on a unified temporal feature set, temporal response fingerprints are extracted for each type of feature source. The temporal response fingerprints are used to characterize the behavioral response patterns and scene adaptation characteristics of the feature sources over time. The specific extraction and calculation process is as follows:

[0107] Behavior trigger hysteresis is used to characterize the time delay from the input of a behavioral signal to the generation of an effective response. The calculation formula is as follows:

[0108]

[0109] in, For characteristic response function, The preset trigger threshold;

[0110] Trigger threshold The determination rule is based on the historical data of the feature source. The principle is established that the sum of the historical characteristic response mean and three standard deviations is taken, i.e. , The historical mean of the characteristic response, The historical standard deviation of the characteristic response is used to filter out noise interference and identify effective behavioral responses;

[0111] The duration of the impact feature is used to characterize the duration of an effective behavioral response, and the calculation formula is as follows:

[0112]

[0113] This is an indicator function; it takes the value 1 if the condition is met and 0 if the condition is not met.

[0114] Fluctuation characteristics are used to characterize the numerical stability of behavioral response processes, and the calculation formula is as follows:

[0115]

[0116] This is the function for calculating standard deviation;

[0117] Scene-sensitive features are used to characterize the adaptability of feature sources to different application scenarios. The calculation formula is as follows:

[0118]

[0119] in, For the first Scene-related tags The total number of scenes, This is the function for calculating covariance.

[0120] In this embodiment, the trigger threshold for the behavioral trigger lag feature is determined using the historical mean + 3 standard deviation rule. Taking the marketing outreach feedback feature source as an example, the average response value is statistically obtained over the past 30 days. Standard deviation The trigger threshold is:

[0121]

[0122] Therefore, when the marketing response function When the threshold is reached, the state is considered to be in a valid behavior triggering state. The reason for using the mean plus three times the standard deviation is that this threshold can effectively filter out most background noise and retain only statistically significant strong response events, making it suitable for behavior trigger boundary identification.

[0123] In the persistence feature, the effective response threshold is also set to 0.75. When a behavioral response is continuously higher than 0.75, its duration is recorded. For example, if a user's response value is higher than the threshold from the 2nd to the 6th hour after the event exposure, the duration is recorded as 4 hours. This duration is used to characterize the ability of marketing or scenario stimuli to maintain their impact on user behavior.

[0124] The fluctuation characteristics were calculated using the moving standard deviation of the most recent 24 time steps, i.e., using 24 hours as a local stability observation window. The rationale for setting 24 hours as the calculation window is that most user behaviors have diurnal cycle characteristics, and extracting fluctuations on a daily basis can take into account both diurnal patterns and short-term fluctuations.

[0125] In the scenario-sensitive features, user scenarios are divided into four categories: work-hour scenarios, leisure-hour scenarios, promotional activity scenarios, and low-activity nighttime scenarios. (Total number of scenarios) The criteria for determining these four scenario categories are as follows: these categories can cover common business operation cycles, while avoiding instability in covariance estimation caused by overly detailed scenario segmentation. The covariance between each feature source and the scenario label is used to quantify the degree of scenario fit; the larger the covariance, the more suitable the feature source is for prediction in the corresponding scenario.

[0126] In one embodiment, using continuous time windows as analysis units, a user group evolution trajectory map is constructed based on user temporal response fingerprints, and dynamic grouping is achieved, specifically:

[0127] First, the temporal response fingerprint of each user is converted into a feature vector of fixed dimension. Then, the similarity is calculated to measure the similarity of behavioral features between users and the initial grouping is completed based on the similarity. The changes in the center position of the group are tracked within a continuous time window to form an evolutionary trajectory map that reflects the dynamic changes of the group. Based on the map, the dynamic group to which the target user belongs and the migration path between groups are determined.

[0128] The calculation process for user similarity and group center is as follows:

[0129] Vectorize the time-series response fingerprint as Cosine similarity between users:

[0130]

[0131] Clustering was performed based on density-based clustering, with a cluster radius threshold. With minimum sample size Rule Determination: The rule of thumb, combined with business experience, was used to determine the clustering radius threshold. The inflection point where the rate of decrease in clustering error significantly slowed down was used. The minimum sample size was set at a percentage of the total number of target users. This ensures the rationality and stability of group division;

[0132] Group center:

[0133]

[0134] For the group Number of internal users;

[0135] The migration path of a group can be represented as a set of group centers over time:

[0136]

[0137] in, for Always at the center of the group.

[0138] In this embodiment, the temporal response fingerprint of each user is compressed into a 16-dimensional feature vector. The reason for setting it to 16 dimensions is that this dimension is sufficient to accommodate the core statistics of lagging features, continuous features, fluctuating features, and scene-sensitive features, while avoiding distortion of clustering distance due to excessively high dimensionality.

[0139] Dynamic clustering employs density clustering, with an analysis window length of 7 days and a window sliding step of 1 day. These parameters were determined because a 7-day window can cover the complete cyclical behavioral pattern, while a 1-day sliding step can promptly capture population migration changes.

[0140] Cluster radius threshold The elbow rule is used to determine the minimum number of samples, which is 0.18 in the dataset of this embodiment. Take the total number of target users ,Right now:

[0141]

[0142] Will The basis for using the minimum sample size is as follows: if the minimum sample size is too low, it is easy to form fragmented small groups; if it is too high, it will mask the real medium-sized user group. Therefore, 1% balances group stability and identification sensitivity.

[0143] In calculating group confidence, the reciprocal of the Euclidean distance from a user to the center of their group is normalized and used as the group confidence score. For example, if a user's distance to the group center is 0.25, the original score is 4.0, and after interval normalization, the confidence score is 0.83. The reciprocal of the distance is used as the basis for confidence because the smaller the distance, the more the user's behavior conforms to the characteristics of the group center, and therefore a higher weight should be given to the group constraint.

[0144] In one embodiment, the reliability of feature sources, the degree of inter-source conflict, and the inter-source complementarity are quantitatively evaluated by the difference in predicted responses before and after counterfactual perturbation, thereby constructing a feature source credibility matrix and an inter-source complementarity matrix. The specific calculation process is as follows:

[0145] The confidence parameter measures the predictive stability and reliability of a single feature source, and is calculated using the following formula:

[0146]

[0147] in, As a feature source No. Predicted values ​​of the second disturbance These are the predicted values ​​for the original samples. This represents the total number of disturbances.

[0148] Number of perturbations Basis for determination: Take The next step is to determine the number of feature sources based on a balance between the number of feature sources and computing resources, ensuring statistical significance without increasing redundant calculations.

[0149] The conflict parameter measures the degree of deviation between the prediction results of two feature sources, and the calculation formula is as follows:

[0150]

[0151] The complementarity parameter measures the information complementarity between two feature sources, and its calculation formula is as follows:

[0152]

[0153] The global maximum conflict value among all feature source pairs;

[0154] Construct a feature source confidence matrix based on the above parameters. Inter-source complementarity matrix , This represents the total number of feature sources.

[0155] In this embodiment, each type of feature source is subjected to 30 counterfactual perturbations. The reason for choosing 30 perturbations is that when the number of perturbations is less than 20, the statistical results fluctuate greatly; when the number of perturbations exceeds 50, the improvement in credibility estimation is limited while the computational cost increases significantly. Therefore, 30 perturbations can achieve a balance between statistical stability and training efficiency.

[0156] The disturbance amplitude is set in three levels, corresponding to the original eigenvalues. The criteria for selecting this third-level disturbance are as follows: In response to slight behavioral fluctuations, In response to general business fluctuations, For more obvious external disturbances, the sensitivity of the feature source to the prediction result can be more completely characterized by three levels of perturbation.

[0157] For example, for a search behavior feature source, if the original prediction value is 0.68 and the average prediction deviation after 30 perturbations is 0.06, then the credibility of this feature source can be calculated using the original formula to obtain a high level of credibility. If the average deviation of another marketing reach feature source reaches 0.15, it indicates that the stability of this feature source is poor, and its weight should be reduced in subsequent fusion.

[0158] The conflict parameter is calculated based on the degree of difference between the prediction results of different feature sources; the complementarity parameter is calculated as "1 minus the relative conflict degree". To ensure the stability of the complementarity parameter distribution, in this embodiment, the global maximum conflict value is taken as the maximum conflict statistical value of all feature source pairs in the current training round, rather than a fixed constant. The rationale is that using a dynamic maximum conflict value can adapt to the changes in feature relationships at different stages of the model, avoiding the distortion of complementarity by a static scale.

[0159] In one embodiment, the fusion decision unit combines user dynamic group attributes, group migration path confidence, feature source credibility, and inter-source complementarity to adaptively allocate fusion weights for each feature source, thereby achieving multi-source feature weighted fusion. The specific calculation process is as follows:

[0160] The adaptive fusion weight comprehensively considers feature source confidence, inter-source complementarity, and group confidence. The calculation formula is as follows:

[0161]

[0162] in, The pre-set balance coefficient is determined according to the following rules:

[0163] Pick Prioritize the fusion of features with high credibility.

[0164] Pick Strengthen the complementarity of information between sources;

[0165] Pick To assist in the introduction of population evolution constraints;

[0166] To the target user group The cluster confidence score is determined by taking the inverse of the distance from the user to the cluster center; the closer the distance, the higher the score. The value is normalized to a certain value. interval;

[0167] The multi-source feature fusion result is obtained by weighted summation:

[0168]

[0169] For the first The standardized feature vectors corresponding to each feature source.

[0170] In this embodiment, the three balancing coefficients in the adaptive fusion weights are respectively:

[0171]

[0172] in, Corresponding to the feature source confidence term, Corresponding source complementary terms, Corresponding to population evolution constraints. The parameters for this set are determined based on: first, on the validation set... , , A grid search was conducted, and considering the business interpretability requirements, a combination was selected that prioritizes credibility, enhances complementarity, and is supplemented by group constraints. The final choice of 0.60 / 0.25 / 0.15 was because this parameter set showed the most balanced performance across accuracy, MAE, and RMSE.

[0173] In this embodiment, the set of example fusion weights obtained after training from the 5 types of feature sources is as follows:

[0174] Site browsing behavior characteristics: 0.27

[0175] Trading behavior characteristics: 0.24

[0176] Marketing reach feedback characteristics: 0.16

[0177] Search behavior characteristics: 0.21

[0178] Terminal context features: 0.12

[0179] The weights mentioned above are not manually assigned, but are calculated jointly from the confidence matrix, the complementarity matrix, and the group confidence score. Browsing and transaction behaviors have higher weights because they are more strongly correlated with the final purchase or conversion behavior; terminal context has lower weights because these features serve more of a supplementary corrective role.

[0180] In one embodiment, the user behavior prediction model is a diffusion autoregressive transformer model, and the specific training steps and parameter settings are as follows:

[0181] The dataset is divided into training, validation and test sets in a 7:2:1 ratio, with a fixed random seed to ensure the reproducibility of the experiment.

[0182] Model initialization: Input dimension, consistent with the fused feature dimension, set to D; Encoder / decoder layer number. Layers; Number of multi-head attention heads: 8; Hidden layer dimensions: 256 / 512; Position encoding: sinusoidal position encoding, maximum sequence length: 512; Number of diffusion steps. Noise scheduling employs a cosine scheduling strategy;

[0183] Training hyperparameter settings: optimizer, AdamW; initial learning rate, 1e-4, using cosine annealing decay; batch size, 32 / 64; maximum number of iterations, 100; early stopping policy, training stops if the validation set loss does not decrease for 10 consecutive iterations.

[0184] Loss function, total loss ,in To reduce noise loss during diffusion, Cross-entropy loss / mean squared error loss for behavior prediction;

[0185] Model validation involves calculating accuracy, MAE, and RMSE metrics on the validation set after each iteration and saving the optimal model weights.

[0186] In this embodiment, the user behavior prediction model adopts the diffusion autoregressive transformer model, and its parameters are set as follows:

[0187] Input Dimensions The setting is 256, which is determined based on the fact that when the dimension of the fused feature vector is controlled at 256, it can fully preserve the multi-source feature expression, while adapting to the 8-head attention structure and facilitating evenly distributed calculation.

[0188] The encoder has 4 layers and the decoder has 4 layers. The reason for setting the number of layers to 4 is that when the number of layers is less than 3, the extraction of higher-order temporal dependencies is insufficient; when the number of layers exceeds 6, overfitting is likely to occur under the data scale of this embodiment.

[0189] The number of multi-head attention heads is set to 8; this value is determined based on the fact that 8 heads can achieve a balance between feature decomposition capability and computational efficiency.

[0190] The feedforward hidden layer dimension is set to 512; this is based on the fact that it is usually twice the input dimension, which can enhance the non-linear expressive power.

[0191] The maximum sequence length is set to 168, corresponding to 7 consecutive days and one time step per hour. This length is chosen to maintain consistency with the 7-day dynamic clustering window mentioned above, facilitating the sharing of time scales between group constraints and behavior prediction.

[0192] The diffusion steps are set to 1000, and a cosine scheduling strategy is used for noise scheduling. The reason for choosing 1000 steps is that, in behavioral sequence prediction tasks, this number of steps can balance denoising fineness and training cost.

[0193] The optimizer used is AdamW, with an initial learning rate of 1e-4 and a weight decay coefficient of 1e-2. This is because AdamW has good convergence stability in time series model training, and 1e-4 can avoid initial training oscillations.

[0194] The batch size is set to 32; this value is determined based on memory usage and gradient estimation stability.

[0195] The maximum number of training epochs is set to 100, and the early stopping epoch is set to 10. If the validation set loss does not decrease for 10 consecutive epochs, training is stopped. This setting is based on the fact that, given the data scale in this embodiment, the model performance tends to stabilize after 80 epochs, and further training yields little benefit.

[0196] In one embodiment, the prediction bias is calculated based on real behavioral feedback, and the key components of the model are updated in a closed loop based on the bias to achieve continuous iterative optimization of prediction capabilities. Specifically:

[0197] Calculate the mean squared error deviation between the predicted value and the true value, and adjust the clustering threshold of the user group evolution trajectory map, update the feature source confidence matrix, and adjust the adaptive fusion weights in sync with the deviation to ensure that the model adapts to the dynamic changes in user behavior.

[0198] The prediction deviation calculation and parameter update process is as follows:

[0199] Prediction bias:

[0200]

[0201] for Predicted value at any time for Real value at any moment This represents the total number of time steps.

[0202] Learning rate Rule of determination: Initialization is performed using an exponential decay strategy, with the initial value set to... The exponential decreases with each iteration round, balancing update speed and stability;

[0203] Credibility matrix update formula:

[0204]

[0205] Set the learning rate;

[0206] Fusion weight update formula:

[0207]

[0208] The population evolution trajectory map adaptively adjusts the cluster radius and grouping threshold based on the deviation value. The correction rule is: when the deviation is greater than the preset tolerance threshold, the cluster radius is reduced to improve the grouping accuracy; when the deviation is less than the tolerance threshold, the parameters are kept stable.

[0209] In this embodiment, the prediction bias is calculated using mean squared error, and the tolerance threshold is set to 0.08. This threshold is determined by statistically analyzing the MSE distribution of the historical stable period model on the validation set, and then taking the value near the 95th quantile as the upper limit of tolerance, so that most normal fluctuations do not trigger drastic updates, and only obvious drifts are corrected.

[0210] The learning rate uses an exponential decay strategy, with an initial value of:

[0211]

[0212] No. During each update cycle, the learning rate is:

[0213]

[0214] The initial learning rate of 0.05 was chosen because: if the initial value is too large, it will cause oscillations in the confidence matrix and fusion weights; if it is too small, the model will not adapt quickly enough to real feedback. The decay coefficient of 0.98 is used to gradually improve parameter stability in the later stages of training.

[0215] In updating the population evolution trajectory map, if the prediction bias of the current batch is greater than 0.08, the cluster radius will be adjusted. The clustering parameters are reduced by 10%, from 0.18 to 0.162, to improve the accuracy of population segmentation. If the prediction bias is less than 0.03, the clustering parameters are kept unchanged to avoid over-correction. The 10% reduction ratio is based on the principle that this ratio can maintain a balance between correction sensitivity and population structure stability, and will not cause drastic changes in the clustering results due to a single fluctuation.

[0216] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0217] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A behavior prediction method based on multi-source feature adaptive fusion and dynamic grouping, characterized in that, The method includes: S1. Collect multi-source heterogeneous behavior data corresponding to the target user, and perform unified time axis alignment, missing data compensation, scale normalization and semantic mapping on the multi-source heterogeneous behavior data to obtain a unified temporal feature set. S2. Based on the unified temporal feature set, extract temporal response fingerprints to characterize the behavior response pattern according to the feature source. The temporal response fingerprints include at least the behavior trigger lag, response duration, response fluctuation, and scene sensitivity. S3. Using continuous time windows as analysis units, construct a user group evolution trajectory map based on the time-series response fingerprints of each user, and determine the dynamic group to which the target user belongs and the group migration path based on the user group evolution trajectory map. S4. Construct counterfactual perturbation samples for each feature source, input the original sample and the counterfactual perturbation sample into the same prediction inference process, and obtain the difference in prediction response of each feature source before and after perturbation. S5. Based on the predicted response differences, determine the confidence parameters, inter-source conflict parameters, and inter-source complementarity parameters of each feature source, and construct the feature source confidence matrix and inter-source complementarity relationship matrix accordingly. S6. Input the target user's dynamic group, group migration path, feature source credibility matrix and inter-source complementary relationship matrix into the fusion decision process to generate adaptive fusion weights for each feature source, and perform reliable fusion of multi-source features based on the adaptive fusion weights. S7. Input the fused multi-source features and the group evolution trajectory constraints corresponding to the target user into the user behavior prediction model, and output the behavior prediction results of the target user within the target time window. S8. After obtaining real behavioral feedback, based on the deviation between the behavior prediction result and the real behavioral feedback, the user group evolution trajectory map, feature source credibility matrix and adaptive fusion weight are updated in a linked manner.

2. The behavior prediction method based on multi-source feature adaptive fusion and dynamic clustering according to claim 1, characterized in that, The process of performing unified timeline alignment, missing data compensation, scale normalization, and semantic mapping on multi-source heterogeneous behavioral data includes: Map multi-source heterogeneous behavioral data with different sampling frequencies and time granularities to the same time axis under a unified standard timestamp; The missing data is compensated based on the valid data before and after the missing position; Convert data with different units and value ranges to a unified numerical range; Text behavior records, log behavior records, and event tag records are converted into structured numerical features, thereby forming a unified time-series feature set that is uniform in format, continuous in time sequence, and consistent in semantics.

3. The behavior prediction method based on multi-source feature adaptive fusion and dynamic grouping according to claim 1, characterized in that, The extraction process of the timing response fingerprint includes: The behavior trigger hysteresis is determined based on the time series response curves of each feature source to characterize the delay from input of the behavior signal to the generation of an effective response; The duration of a response is determined based on the effective response sustainment interval to characterize the duration of a behavioral response; The response fluctuation is determined based on the degree of dispersion within the response interval to characterize the stability of the behavioral response; The scene sensitivity quantity is determined based on the degree of correlation between the feature response and scene variables under each scene label, so as to characterize the adaptability of the corresponding feature source to changes in the application scene. The behavior trigger hysteresis, response duration, response fluctuation, and scene sensitivity are combined to form the temporal response fingerprint of the corresponding feature source.

4. The behavior prediction method for multi-source feature adaptive fusion and dynamic clustering according to claim 1, characterized in that, The process of constructing the user group evolution trajectory map includes: Convert each user's temporal response fingerprint into a fixed-dimensional fingerprint vector; Initial group segmentation is completed based on the fingerprint similarity among users; By tracking the changes in the location and membership of each group center within a continuous time window, a user group evolution trajectory map reflecting the evolutionary state of the group is formed; Based on the target user's current position in the user group evolution trajectory map, its historical dynamic group, and its migration direction in adjacent time windows, the dynamic group to which the target user belongs and the corresponding group migration path are determined.

5. The behavior prediction method for multi-source feature adaptive fusion and dynamic grouping according to claim 3, characterized in that, The process of constructing the counterfactual perturbation sample includes: For each feature source, while keeping the information of other feature sources for the same user unchanged, the input value, time segment distribution or scene association item of the feature source is subjected to controlled perturbation to obtain counterfactual perturbation samples that correspond one-to-one with the original samples. The original sample and the counterfactual perturbation sample are respectively input into the same prediction inference process to obtain the change in the prediction result before and after the perturbation, and the change is used as the basis for evaluating the strength and stability of the influence of the feature source on the behavior prediction result.

6. The behavior prediction method for multi-source feature adaptive fusion and dynamic grouping according to claim 5, characterized in that, The construction process of the feature source confidence matrix and the inter-source complementarity relation matrix includes: The confidence parameter of a feature source is determined based on the magnitude of the shift in the behavior prediction result caused by multiple counterfactual perturbations. The corresponding inter-source conflict parameter is determined based on the degree of deviation between the predicted outputs of any two feature sources on the same user sample. The corresponding inter-source complementarity parameter is determined based on the degree of complementarity between any two feature sources in their contribution to behavior prediction. The confidence matrix of the feature sources and the inter-source complementarity matrix are formed based on the confidence parameters of all feature sources, the inter-source conflict parameters, and the inter-source complementarity parameters, respectively.

7. The behavior prediction method for multi-source feature adaptive fusion and dynamic grouping according to claim 4, characterized in that, The process of generating the adaptive fusion weights includes: The group confidence information of the dynamic group to which the target user belongs, the path constraint information of the group migration path, the credibility parameters of the corresponding feature sources, and the complementary parameters between sources are all input into the fusion decision process. Feature sources with high credibility and high matching degree with the current evolution state of the dynamic group to which the target user belongs are given higher fusion weights, and feature sources with strong complementary ability to other feature sources are given enhanced weights. The standardized feature vectors corresponding to each feature source are weighted and combined according to the adaptive fusion weights to obtain the fused multi-source features.

8. The behavior prediction method for multi-source feature adaptive fusion and dynamic grouping according to claim 7, characterized in that: The user behavior prediction model is used to receive fused multi-source features and population evolution trajectory constraint information, and output the behavior prediction results of the target user within the target time window; The group evolution trajectory constraint information is used to limit the learning range of the user behavior prediction model for the target user's behavior evolution direction, behavior transfer boundary and behavior occurrence probability distribution, so that the behavior prediction result simultaneously satisfies the individual characteristic expression of the target user and the group evolution law of the dynamic group to which it belongs.

9. The behavior prediction method for multi-source feature adaptive fusion and dynamic grouping according to claim 6, characterized in that, The associated update process includes: Adjust the corresponding parameters in the feature source credibility matrix according to the deviation value between the behavior prediction result and the real behavior feedback, and synchronously correct the relevant parameters in the inter-source complementary relationship matrix; When the deviation value exceeds the preset tolerance threshold, increase the suppression degree of the low-credibility feature source that causes the deviation; When the deviation value does not exceed the preset tolerance threshold, maintain the verified and effective fusion relationship to enable the credible fusion process to continuously adapt to the changes in user behavior.

10. The behavior prediction method for multi-source feature adaptive fusion and dynamic grouping according to claim 8, characterized in that, The associated update of the user group evolution trajectory map includes: Re-verify the group membership result of the target user in the current time window according to the real behavior feedback; When the behavior prediction deviation indicates a decrease in the existing group division accuracy, adjust the group division threshold and the group center update rule, and reconstruct the group migration path of the target user in the adjacent time window; Input the updated group membership result, group migration path, feature source credibility matrix and adaptive fusion weight into the next round of user behavior prediction process to form a closed-loop iterative optimization mechanism.

Images