An intelligent data analysis method

Abstract

The application relates to an intelligent data analysis method, comprising the following steps: S1. initializing a cognitive state: obtaining a normalized sample of a target data set, and initializing a cognitive state vector for each data feature, the cognitive state vector being a multidimensional vector for dynamically representing individual importance weights of the feature, influence indexes of the feature in a feature correlation network and historical stability degrees of the importance; S2. performing a cognitive evolution cycle: performing at least one round of cognitive evolution cycle, and sequentially performing steps S2a, S2b, S2c and S2d in each round; S2a. local cognitive updating: mapping each sample to a one-dimensional cognitive projection value based on the cognitive state vectors of all the features, dynamically dividing temporary data clusters according to the projection value distribution of all the samples, and updating the individual importance weights in the cognitive state vectors of the features according to the distinguishing abilities of the features in the divided temporary data clusters, and the application realizes more comprehensive intelligent data analysis.

Description

Technical Field

[0001] This invention relates to the field of data analysis, and more specifically to an intelligent data analysis method. Background Technology

[0002] With the development of the digital age, intelligent data analysis has been widely applied in e-commerce, the Internet and other fields. The core requirement is to mine feature correlation patterns from massive and complex data (such as user web browsing behavior data), divide data pattern clusters, and provide reliable decision-making basis to support actual business such as precise operation and risk identification. Existing data analysis methods typically involve core steps such as feature weight setting, clustering, and correlation analysis. However, when faced with real-world scenarios involving expanding data scale, increasing feature dimensions, and dynamic data changes (such as incremental data generation and user feedback iteration), traditional solutions gradually reveal their limitations in adaptability. On the one hand, feature importance assessment often relies on subjective presuppositions or single-dimensional measurements, making it difficult to adapt to the inherent characteristics of the data. On the other hand, local data pattern mining and global feature correlation analysis are often isolated from each other, lacking a co-evolutionary mechanism, resulting in insufficient stability and repeatability of analysis results. At the same time, existing methods generally suffer from the problem of data black boxes, lacking transparency in the feature evolution process and correlation logic, and are difficult to efficiently integrate new data or user feedback. They also lack continuous learning and knowledge updating capabilities, and struggle to accurately identify abnormal behaviors and emerging patterns during online inference. Consequently, they cannot fully meet the high requirements of real-world scenarios for objectivity, robustness, interpretability, and dynamic adaptability in data analysis. Therefore, an intelligent data analysis method is proposed. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides an intelligent data analysis method, comprising the following steps: S1. Initialize cognitive state: Obtain normalized samples of the target dataset and initialize a cognitive state vector for each data feature. The cognitive state vector is a multi-dimensional vector used to dynamically represent the individual importance weight of the feature, its influence index in the feature association network, and the historical stability of its importance. S2. Perform the cognitive evolution cycle: Perform at least one round of cognitive evolution cycle, in which steps S2a, S2b, S2c and S2d are executed sequentially in each round; S2a. Local Cognitive Update: Based on the cognitive state vectors of all current features, each sample is mapped to a one-dimensional cognitive projection value. Temporary data clusters are dynamically divided according to the distribution of projection values ​​of all samples, and the individual importance weights in the cognitive state vectors of each feature are updated according to the distinguishing ability of each feature within the divided temporary data clusters. The mapping process uses an adjustment function to fine-tune the weights, which applies inverse adjustment to the weights based on the influence index of the features. S2b. Global Relationship Distillation: Based on the updated individual importance weights, calculate the nonlinear correlation strength between features to construct a weighted correlation network. By analyzing the topology of this network, extract the influence index of each feature in the global relationship, i.e. update the influence index, and update the corresponding components in its cognitive state vector accordingly. S2c. Consensus Convergence Test: Calculate the overall difference between the current round and the previous round of all feature cognitive state vectors, and determine whether the difference is lower than the preset threshold; S2d. Update stability: Based on the changing trend of the individual importance weight of each feature in multiple iterations, recalculate the stability component in its cognitive state vector. The smoother the changing trend, the higher the stability. S3. Result Generation: If the difference is determined to be lower than the preset threshold in step S2c, the cognitive state is determined to have converged. Based on the converged feature cognitive state vector at this time, the dataset is finally divided and interpreted to generate a stable cluster structure, feature cognitive map and decision basis report; otherwise, the updated cognitive state vector is used to return to step S2 to execute a new round of cognitive evolution cycle.

[0004] Furthermore, step S1 specifically includes: initializing a cognitive state vector for each data feature. The initialization process specifically includes: Calculate the initial individual importance weights : The calculation process is as follows: First, the information entropy and absolute kurtosis of the feature are calculated separately for all samples. Then, these two values ​​are normalized. Finally, the normalized information entropy and absolute kurtosis are weighted and summed according to a preset first weight ratio. The result is the sum of these values. ; Set initial influence metrics Set it to a constant value representing the initial neutral state; Calculate the initial stability The calculation process involves, if historical batch data exists, calculating the coefficient of variation of the estimated individual importance of this feature in each batch of data, and recording this coefficient of variation as... If there is no historical data, set it to a constant value that represents a moderate degree of stability.

[0005] Furthermore, the process of calculating the sample cognitive projection value and updating the individual importance weight in step S2a specifically includes: Sample cognitive projection value Calculation: For the cognitive projection value of the i-th sample in the t-th iteration, the calculation process is as follows: first, the normalized value of the sample on each feature is compared with the current individual importance weight of the corresponding feature. Multiply, then adjust the function The values ​​are scaled, and finally all features are summed to obtain the sample cognitive projection values. The specific process is as follows: ; in, It is the normalized value of the i-th sample on the j-th feature. It is a monotonically decreasing function whose output value is between 0 and 1, which moderately suppresses the direct weight contribution of features with high correlation influence in the projection. Temporary data cluster partitioning: based on the projected values ​​of all samples K temporary data clusters are formed using a density-based clustering algorithm; Individual importance weight Update: For the j-th feature, the updated individual importance weight is calculated by first calculating the discriminative contribution of this feature within each temporary data cluster k. This contribution is obtained by measuring the difference between the distribution of feature values ​​within the temporary data cluster and the global distribution; Then, the contributions of all temporary data clusters are averaged and combined with the current stability of this feature. After smoothing, the weights are finally combined with the weights from the previous round to obtain the individual importance weights. The update formula is expressed as: ; in, This is a preset learning rate parameter used to control the update magnitude.

[0006] Furthermore, the process of updating the influence metric in step S2b specifically includes: Weighted Association Network Construction: This network construction process involves using the maximum information coefficient as a measure of association strength, calculating the nonlinear association strength between any two features, and assigning weights to the current individual importance of the two features. and The geometric mean is used as the weight coefficient of the associated edge, thereby constructing a weighted fully associated network; Network community discovery and centrality calculation: The community discovery algorithm is performed on the above weighted fully interconnected network to identify feature communities with tight internal connections; then, the tight centrality of each feature j within its own community and its betweenness centrality as a bridge connecting different communities are calculated. Influence Metric Update: Influence metrics updated for each feature The calculation process involves first normalizing the compact centrality and betweenness centrality, then merging them according to a preset second weight ratio, and finally mapping them to a reasonable range of values ​​through a saturation function. The update formula is expressed as: ; in, and α and β are the normalized compact centrality and betweenness centrality, respectively, and are the preset fusion weight coefficients.

[0007] Furthermore, in step S2c, the overall difference in the cognitive state vector is calculated. The process is as follows: First, calculate the cosine similarity between the cognitive state vector of each feature in round t+1 and round t. Then, calculate the average of the cosine similarities of all features. Finally, subtract this average from 1 to represent the difference, thus obtaining the overall difference of the cognitive state vector. The specific process is as follows: ; Where n is the total number of features Let j be the cognitive state vector of the j-th feature at the end of the (t+1)-th iteration. Let j be the cognitive state vector of the j-th feature at the end of the t-th iteration. The dot product of two vectors, also known as the inner product. and These are the magnitudes of the two vectors, i.e., their lengths; like If the result is positive, then the convergence is determined. This is the preset convergence threshold.

[0008] Furthermore, the result generated in step S3 based on the converged feature cognitive state vector specifically includes: Stable cluster structure generation: Extracting the final individual importance weights of each feature after convergence. As feature weights, calculate the weighted Euclidean distance matrix between samples; The distance matrix is ​​clustered using a hierarchical clustering algorithm, and the optimal number of clusters is automatically determined in the following way: For a set of candidate cluster numbers, for each candidate number N, perform: Divide all samples into N clusters; Calculate the intra-cluster sample density of all clusters under this partition; Calculate the centroid distance separation between all clusters under this partition; The optimal number of clusters is determined by evaluating the inflection point of the evaluation curve composed of compactness and separation under different numbers of candidate clusters. Based on this optimal number of clusters, the updated cluster structure is obtained as the final cluster partitioning result of this round of analysis. Each sample group in this result is defined as a stable cluster. Feature cognitive graph construction: using each feature as a node, and its final influence index The numerical mapping is used to represent the visual size of the nodes, and the network visualization graph is drawn with the final correlation strength between features as the edge weight. At the same time, a cognitive card is generated for each feature to record the complete evolution history of its cognitive state vector. Decision-based report synthesis: For each stable cluster in the final cluster division results, select the top features that contribute the most to its formation as the core feature set, and calculate the contribution ratio of each core feature; through the core feature set of each final stable cluster and its value distribution, summarize and generate readable distinguishing rules; integrate the final cluster division results, cognitive map and distinguishing rules, and output a structured analysis report.

[0009] Furthermore, the method also includes a continuous learning step: S4. Incremental input processing: Perform normalization processing on the new batch of data in the same way as in step S1, or parse the user's feedback tags on the historical analysis results and convert them into cognitive adjustment signals for specific features or the final stable cluster. S5. Cognitive State Adjustment: Based on new data or feedback signals, directly adjust the cognitive state vector of relevant features, specifically by reducing its stability component. The value is used to enhance its plasticity and to weight the individual importance component I. j Make revisions based on new evidence; S6. Rapid Evolution and Knowledge Fusion: Starting with the adjusted set of cognitive state vectors and taking a historical dataset containing new data as input, a cognitive evolution loop with a limited number of rounds is executed. This process employs an increased learning rate parameter. As an accelerated convergence strategy, after the cognitive evolution cycle of this finite number of rounds reaches consensus convergence, based on the finally updated cognitive state vector, the generation process based on the converged feature cognitive state vector in step S3 is executed to obtain the corresponding updated final cluster partitioning result and new rules, which include new discriminative rules. The updated final cluster partitioning result and new rules obtained after evolution are compared and integrated with the original knowledge base of the system to complete the smooth update of the knowledge system.

[0010] Furthermore, the method also includes online inference and early warning steps: S7. Cognitive matching degree calculation: For a new query sample, calculate its cognitive matching degree Mc with each stable cluster c; The specific calculation process is as follows: for each feature j, calculate the similarity between the query sample value and the typical value of the final stable cluster c, and multiply it by the final individual importance weight of that feature. Then multiply by a final impact metric The weighted ordinal difference penalty factor is applied, and finally, all features are summed to obtain the cognitive matching degree. The specific process is as follows: ; in, It queries the value of the j-th feature of a sample. The typical value of the c-th ultimately stable cluster on the j-th feature This indicates the rank of the value within the global distribution of historical samples. It is a similarity calculation function; S8. Intelligent Early Warning and Report Generation: If the query sample matches the highest degree of accuracy among all ultimately stable clusters... If the value falls below the preset abnormal threshold, an abnormal warning will be triggered; If the match degree between the query sample and a previously non-dominant final stable cluster is significantly higher than the statistical baseline of the historical match degree of that cluster, an emerging pattern opportunity warning is triggered. The output includes the most likely final stable cluster affiliation of the query sample and its match degree, the type of warning triggered, a list of key features leading to the match result or warning, and their contribution analysis. Output the most likely cluster affiliation of the query sample and its matching degree, the type of warning triggered, and a list of key features that led to the matching result or warning and their contribution analysis.

[0011] The present invention has the following advantages over the prior art: By dynamically constructing and iteratively updating the cognitive state vector, this method achieves an objective and adaptive assessment of feature importance, overcoming the limitations of traditional methods that rely on subjective weight settings. Furthermore, by combining the co-evolution of local cognition and global relationships, the model can achieve stable and repeatable consensus results, significantly improving the accuracy and robustness of the analysis. Secondly, this method generates a deeply interpretable feature cognitive map and full-process cognitive cards, making complex data patterns and feature associations transparent and traceable, greatly enhancing the insight and credibility of the analysis results. Finally, the solution designs a complete closed loop from offline analysis to online inference and continuous learning. The system can utilize new data or feedback for efficient incremental learning, automatically adapt to data changes, and achieve accurate matching and intelligent early warning for real-time data. This forms a complete intelligent analysis system that is adaptive, evolvable, and has strong decision support capabilities, effectively addressing the key shortcomings of traditional methods in terms of dynamism, interpretability, and sustainability. Attached Figure Description

[0012] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0013] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0014] like Figure 1 As shown, it includes the following steps: S1. Initialize cognitive state: Obtain normalized samples of the target dataset and initialize a cognitive state vector for each data feature. The cognitive state vector is a multi-dimensional vector used to dynamically represent the individual importance weight of the feature, its influence index in the feature association network, and the historical stability of its importance. S2. Perform the cognitive evolution cycle: Perform at least one round of cognitive evolution cycle, in which steps S2a, S2b, S2c and S2d are executed sequentially in each round; S2a. Local Cognitive Update: Based on the cognitive state vectors of all current features, each sample is mapped to a one-dimensional cognitive projection value. Temporary data clusters are dynamically divided according to the distribution of projection values ​​of all samples, and the individual importance weights in the cognitive state vectors of each feature are updated according to the distinguishing ability of each feature within the divided temporary data clusters. The mapping process uses an adjustment function to fine-tune the weights, which applies inverse adjustment to the weights based on the influence index of the features. S2b. Global Relationship Distillation: Based on the updated individual importance weights, calculate the nonlinear correlation strength between features to construct a weighted correlation network. By analyzing the topology of this network, extract the influence index of each feature in the global relationship, i.e. update the influence index, and update the corresponding components in its cognitive state vector accordingly. S2c. Consensus Convergence Test: Calculate the overall difference between the current round and the previous round of all feature cognitive state vectors, and determine whether the difference is lower than the preset threshold; S2d. Update stability: Based on the changing trend of the individual importance weight of each feature in multiple iterations, recalculate the stability component in its cognitive state vector. The smoother the changing trend, the higher the stability. S3. Result Generation: If the difference is determined to be lower than the preset threshold in step S2c, the cognitive state is determined to have converged. Based on the converged feature cognitive state vector at this time, the dataset is finally divided and interpreted to generate a stable cluster structure, feature cognitive map and decision basis report; otherwise, the updated cognitive state vector is used to return to step S2 to execute a new round of cognitive evolution cycle.

[0015] Step S1 specifically includes: initializing a cognitive state vector for each data feature. The initialization process specifically includes: Calculate the initial individual importance weights : The calculation process is as follows: First, the information entropy and absolute kurtosis of the feature are calculated separately for all samples. Then, these two values ​​are normalized. Finally, the normalized information entropy and absolute kurtosis are weighted and summed according to a preset first weight ratio. The result is the sum of these values. ; Set initial influence metrics Set it to a constant value representing the initial neutral state; Calculate the initial stability The calculation process involves, if historical batch data exists, calculating the coefficient of variation of the estimated individual importance of this feature in each batch of data, and recording this coefficient of variation as... If there is no historical data, set it to a constant value representing moderate stability. The initial individual importance weights integrate information entropy (uncertainty in feature distribution) and absolute kurtosis (steepness of feature distribution) to objectively quantify the distinguishing potential of web browsing features and avoid the bias of subjective weight setting in traditional methods. The initial influence index is set as a neutral constant to eliminate initial bias and ensure that each feature in the web browsing data participates fairly in subsequent global relationship analysis, adapting to diverse browsing behavior features (such as duration, number of pages, device type, etc.). The initial stability level is adapted to different scenarios: when historical data is available, the coefficient of variation (reflecting the historical stability of features) is reused; when there is no historical data, a moderate constant is set to balance the batch continuity of web browsing data and the rationality of the first analysis.

[0016] For example, analyzing a "user web browsing behavior dataset" from an e-commerce platform aims to initialize the user's cognitive state of features, laying the foundation for subsequent identification of high-value users. The dataset contains 5 core features: j=1 (average daily browsing time / min), j=2 (average daily number of pages viewed), j=3 (variance of page dwell time), j=4 (whether an ad was clicked, 0=no / 1=yes), and j=5 (browsing device type, 0=mobile phone / 1=tablet / 2=computer), totaling 1000 user samples, with no historical batch data.

[0017] Preset parameters: First weighting ratio: Information entropy weight w1=0.6, kurtosis absolute value weight w2=0.4 (highlighting the priority of feature distribution uncertainty in distinguishing browsing behavior); Initial Influence Metrics (Neutral constant value, ranging from [0,1], ensuring no initial bias); Initial stability (No historical data available, set to a moderately stable constant value, with a range of [0,1]).

[0018] Initial individual importance weights: ; : Normalized information entropy of feature j (reflecting the uncertainty of the distribution of browsing behavior feature values); : The normalized kurtosis absolute value of feature j (reflecting the steepness of the distribution of browsing behavior feature values).

[0019] Initial Influence Metrics: (C is a neutral constant value); Initial stability level: (D is a moderately stable constant value, when there is no historical data).

[0020] Calculate the original information entropy ; The higher the information entropy, the more dispersed the feature value distribution (the more significant the differences in browsing behavior). The Shannon entropy formula is used for calculation (simplified results are shown below): j=1 (average daily browsing time): Sample value distribution [10,15,...,60], moderate dispersion, Ent1=1.8; j=2 (average daily page views): Sample value distribution [2,3,...,15], moderate dispersion, Ent2=1.5; j=3 (variance of page dwell time): The sample value distribution is [5,8,...,30], with the highest dispersion (large fluctuations in user dwell time), Ent3=2.1; j=4 (whether the ad was clicked): 0-1 binary classification, lowest dispersion (only two values), Ent4=0.7; j=5 (Browsing device type): 3 types of discrete values, with low dispersion, Ent5=1.1.

[0021] Calculate the absolute value of the raw kurtosis | Kurt j |: The larger the absolute value of kurtosis, the more concentrated the characteristic values ​​are in a few intervals (the more homogeneous the browsing behavior). j=1 (average daily browsing time): Most users spend 20-40 minutes browsing, |Kurt1|=0.8; j=2 (average daily page views): Most users are concentrated on pages 5-10, |Kurt2|=1.2; j=3 (Variance of page dwell time): User fluctuations are dispersed with no obvious concentrated intervals, |Kurt3|=0.5; j=4 (whether the ad is clicked): 80% of users do not click, highly concentrated, |Kurt4|=2.3; j=5 (Browsing Device Type): 70% of users use mobile phones, high concentration, |Kurt5|=1.6.

[0022] Normalization (mapped to the [0,1] interval): Information entropy normalization: , among which Ent min =0.7 (j=4), Ent max =2.1 (j=3): ; ; ; ; ; Kurtosis absolute value normalization: Kurt min =0.5 (j=3), Kurt max =2.3 (j=4): ; ; ; ; ; Calculate the initial individual importance weights ; Substitute into the formula : ; ; ; ; ; Determine the complete initial cognitive state vector: Based on preset parameters, the initial cognitive state vector for the five features is as follows: j=1: ; j=2: ; j=3: ; j=4: ; j=5: .

[0023] Using the above initialization method, it was objectively identified that the "variance of page dwell time (j=3)" in the web browsing data had the highest initial importance. ), because its browsing fluctuations are the greatest (highest information entropy); while "whether to click on the ad (j=4)" has the lowest initial importance ( This is because its values ​​are highly concentrated (extremely high kurtosis but zero information entropy). The initial influence of all features is unbiased. Consistent stability This provides a fair and accurate starting point for subsequent cognitive evolution cycles (local updates, global distillation), perfectly adapting to the diverse and objective requirements of web browsing behavior data.

[0024] The process of calculating the sample cognitive projection value and updating the individual importance weight in step S2a specifically includes: Sample cognitive projection value Calculation: For the cognitive projection value of the i-th sample in the t-th iteration, the calculation process is as follows: first, the normalized value of the sample on each feature is compared with the current individual importance weight of the corresponding feature. Multiply, then adjust the function The values ​​are scaled, and finally all features are summed to obtain the sample cognitive projection values. The specific process is as follows: ; in, It is the normalized value of the i-th sample on the j-th feature. It is a monotonically decreasing function whose output value is between 0 and 1, which moderately suppresses the direct weight contribution of features with high correlation influence in the projection. Temporary data cluster partitioning: based on the projected values ​​of all samples K temporary data clusters are formed using a density-based clustering algorithm; Individual importance weight Update: For the j-th feature, the updated individual importance weight is calculated by first calculating the discriminative contribution of this feature within each temporary data cluster k. This contribution is obtained by measuring the difference between the distribution of feature values ​​within the temporary data cluster and the global distribution; Then, the contributions of all temporary data clusters are averaged and combined with the current stability of this feature. After smoothing, the weights are finally combined with the weights from the previous round to obtain the individual importance weights. The update formula is expressed as: ; in, The preset learning rate parameter is used to control the update magnitude; By employing a triple weighting of feature normalization values, individual importance weights, and a modulating function, the contribution of key features in web browsing is accurately highlighted, while a monotonically decreasing modulating function suppresses the excessive dominance of high-influence features, avoiding local cognitive biases. Density-based temporary data clustering requires no pre-defined cluster number and adapts to the natural clustering characteristics of web browsing data (such as behavioral clustering of users with different browsing habits), closely aligning with the inherent distribution of the data. The update of individual importance weights integrates "intra-cluster distinguishing contribution + stability smoothing + historical weight iteration," responding to local pattern changes in web browsing data while avoiding excessive weight fluctuations, thus improving the robustness and continuity of the analysis.

[0025] E-commerce user web browsing behavior dataset: 5 core features (j=1: average daily browsing time / min; j=2: average daily number of pages viewed; j=3: variance of page dwell time; j=4: whether an ad was clicked (0=no / 1=yes); j=5: browsing device type (0=mobile phone / 1=tablet / 2=computer)), 1000 samples, no historical data.

[0026] Given the initial cognitive state obtained at t=0: j=1: ; j=2: ; j=3: ; j=4: ; j=5: ; Regulation function: (Monotonically decreasing, output range [0,1]); Learning rate: (Control the magnitude of weight updates) Clustering algorithm: DBSCAN (density-based, parameters eps=0.8, min_samples=5) Differentiating contribution measures: Simplifying the KL divergence formula, ; In the t-th iteration, the j-th feature contributes to the distinction of the k-th temporary data cluster (range [0,1]). The closer the value is to 1, the stronger the feature's effect on distinguishing clusters within that cluster. t represents the iteration number of the current cognitive evolution cycle (e.g., t=0 is the initial round, t=1 is the first iteration). j is the index of the j-th feature in the dataset (such as the index of features like browsing time or ad clicks); k is the kth temporary data cluster partitioned by the density clustering algorithm in the t-th iteration; In the t-th iteration, the mean of all samples of the j-th feature in the k-th temporary data cluster reflects the central tendency of this feature within the cluster; In the t-th iteration, the global mean of the j-th feature in the entire sample reflects the overall central tendency of that feature; In the t-th iteration, the standard deviation of all samples of the j-th feature in the k-th temporary data cluster reflects the degree of dispersion of the feature within the cluster. In the t-th iteration, the global standard deviation of the j-th feature in the full sample reflects the overall dispersion of the feature. Calculate the cognitive projection value of the sample ; ; in, Let t = 0 be the normalized value of the i-th sample in the j-th feature (already mapped to the [0,1] interval), and t = 0 be the initial round.

[0027] Taking the first sample as an example (with clearly defined normalized feature values): X11=0.6 (corresponding to the original browsing time of 35 minutes), X12=0.5 (corresponding to the original number of pages viewed: 8 pages), X13=0.7 (corresponding to the original variance of 20), X14=1 (clicked ad), X15=0.0 (mobile device); Calculate the adjustment function value: because all Therefore ; Substitute into the formula to calculate : ; Full sample calculation: Perform the above logic on each of the 1000 samples to obtain... The value range is 0.32-1.05.

[0028] Dynamically partition temporary data clusters: Based on all samples The DBSCAN algorithm automatically identifies K=3 temporary data clusters (aligning with natural web browsing behavior patterns): Cluster 1 (High-frequency, short-duration browsing): 320 samples. The value is concentrated between 0.32 and 0.55 (users browse frequently on average every day, but the duration of each single period is short). Cluster 2 (Low-frequency, long-duration browsing): 280 samples. Concentrated between 0.75 and 1.05 (low user browsing frequency but long duration in a single time period); Cluster 3 (Ad-Sensitive Browsing): 400 samples The range is concentrated between 0.58 and 0.72 (indicating high user click-through rate). Calculate intra-cluster distinguishing contribution : Taking feature j=1 (average daily browsing time) as an example, first obtain the global and cluster-specific statistics: Global: ; Within cluster 1: ; Within cluster 2: ; Within cluster 3: ; Substitute into the simplified formula to calculate : ; ; ; Similarly, calculate all features : j=2 (number of pages viewed): ; j=3 (variance of dwell time): ; j=4 (Whether to click the ad): ; j=5 (Browse device type): ; Update individual importance weights : ; Taking j=1 as an example, it is known that K=3: ; ; ; Similarly, calculate all features : j=2: ; j=3: ; j=4: ; j=5: ; Through the above process, a local cognitive update was completed from t=0 to t=1: the web browsing samples were accurately divided into three temporary data clusters that closely match user behavior. The individual importance weights of each feature were adaptively adjusted according to its intra-cluster discrimination ability (e.g., j=4 had the highest discrimination contribution in the ad-sensitive cluster, but did not increase excessively due to stability smoothing). The updated weights both preserved the historical cognitive foundation and responded to local data patterns, laying a precise and robust local cognitive foundation for subsequent global relation distillation and cognitive convergence.

[0029] The process of updating the influence metric in step S2b specifically includes: Weighted Association Network Construction: This network construction process involves using the maximum information coefficient as a measure of association strength, calculating the nonlinear association strength between any two features, and assigning weights to the current individual importance of the two features. and The geometric mean is used as the weight coefficient of the associated edge, thereby constructing a weighted fully associated network; Network community discovery and centrality calculation: The community discovery algorithm is performed on the above weighted fully interconnected network to identify feature communities with tight internal connections; then, the tight centrality of each feature j within its own community and its betweenness centrality as a bridge connecting different communities are calculated. Influence Metric Update: Influence metrics updated for each feature The calculation process involves first normalizing the compact centrality and betweenness centrality, then merging them according to a preset second weight ratio, and finally mapping them to a reasonable range of values ​​through a saturation function. The update formula is expressed as: ; in, and These are the normalized compact centrality and betweenness centrality, respectively, and α and β are preset fusion weight coefficients; The maximum information coefficient (MIC) measures nonlinear correlations between features, accurately capturing complex behavioral relationships in web browsing data (such as the non-proportional relationship between browsing time and ad clicks), overcoming the limitations of traditional linear correlation measures. The weighted association network uses the geometric mean of feature importance as edge weights, highlighting the correlation value of key behavioral features (such as page dwell variance), filtering redundant correlations, and improving the targeting of global relationship analysis. It integrates tight centrality (internal community connection strength) and betweenness centrality (cross-community bridging effect) to comprehensively characterize the global influence of web browsing features, avoiding the one-sidedness of single centrality assessments. The tanh saturation function limits the range of influence indicators, avoiding extreme value interference and ensuring the robustness and comparability of influence assessments for each feature, adapting to the diverse feature attributes of web browsing data.

[0030] For example, the e-commerce user web browsing behavior dataset has 5 core features (j=1: average daily browsing time; j=2: average daily number of pages viewed; j=3: variance of page dwell time; j=4: whether an ad is clicked; j=5: browsing device type), 1000 samples, and no historical data.

[0031] Given the key input (states at t=1 round): Individual importance weight: ; Pre-processing temporary data clusters are divided into 3 clusters (high-frequency short duration, low-frequency long duration, and advertising-sensitive). Second weighting ratio (integration coefficient): α=0.5, β=0.5 (equal emphasis is placed on the role of intra-community connections and cross-community bridges); Community detection algorithm: Louvain algorithm (efficiently identifies closely connected feature communities); Maximum Information Coefficient (MIC) value: set based on web browsing feature logic (range [0,1], the larger the value, the stronger the correlation). Constructing a weighted association network: Core logic: First, calculate the nonlinear correlation strength (MIC) of each pair of features, and then construct a fully weighted network using the geometric mean of feature importance as the edge weight.

[0032] Determine the strength of the nonlinear correlation (MIC) between features: Based on web browsing behavior logic, behavioral features (j1, j2, j3) are closely correlated, while attribute features (j4, j5) are moderately correlated with behavioral features. Specific MIC values ​​are as follows: j1 (view duration) and j2 (number of pages): 0.70 (strong correlation, the longer the duration, the more pages are usually displayed). j1 and j3 (dwell variance): 0.65 (strong correlation, duration fluctuation is related to total duration) J1 and J4 (ad clicks): 0.50 (moderate correlation, duration and click intent are somewhat correlated); j1 and j5 (device type): 0.35 (weak correlation, low correlation between duration and device type); j2 and j3: 0.60 (strong correlation, positive correlation between page count and dwell time fluctuation); j2 and j4: 0.45 (moderate correlation; a larger number of pages may increase the probability of clicking on ads); j2 and j5: 0.30 (weak association, weak association between page count and device type); j3 vs. j4: 0.55 (moderate correlation; users with large fluctuations in dwell time are more likely to click on ads); j3 and j5: 0.40 (moderate correlation, the fluctuation in dwell time is somewhat related to the type of equipment); j4 and j5: 0.65 (strong correlation, device type affects ad click habits); Calculate the associated edge weights (geometric mean): Edge weight formula: (j and m are pairwise feature combinations): Edge weights of j1 and j2: ; Edge weights of j3 and j4: ; Edge weights of j4 and j5: ; The full edge weight logic is as follows: the edge weights between behavioral features are all between 0.394 and 0.415, and the edge weights between attribute and behavioral features are between 0.334 and 0.401, which conforms to the design of "higher weights for key features".

[0033] Constructing a weighted association network: Using five features as nodes and the aforementioned edge weights as connection weights, a fully weighted network is formed: behavioral features (j1, j2, j3) have higher connection weights and constitute the core region of the network; attribute features (j4, j5) are connected to the core region through medium weights.

[0034] Network community discovery and centrality calculation; Community detection (Louvain algorithm): The system automatically identified two closely connected feature communities, which aligns with the logic of web browsing feature classification. Community 1 (Behavioral Characteristic Community): j1, j2, j3 (the edge weights among the three are all >0.394, indicating a close connection, and all reflecting the intensity of user browsing behavior); Community 2 (Attribute Feature Community): j4, j5 (edge ​​weight = 0.335, relatively close connections, both reflecting user attribute-related features); Calculate centrality (closeness centrality CC, betweenness centrality BC): Tight Centrality (CC): Measures the degree of connectivity of a feature within its community. The formula is simplified to "sum of edge weights within the community / (number of nodes in the community - 1)". Community 1 (3 nodes): (j1 connects j2 and j3, and has the highest total edge weight) (j2 connects j1 and j3); (j3 connects j1 and j2); Community 2 (2 nodes): (j4 connects to j5); (j5 connects to j4); Betweenness Centrality (BC): Measures how frequently a feature acts as a "bridge" across communities. The formula simplifies to "sum of cross-community edge weights / sum of all cross-community edge weights". The cross-community edges are only j1-j4, j2-j4, j3-j4, and j3-j5, and the total cross-community edge weight is 0.370 + 0.366 + 0.401 + 0.369 = 1.506; (Only j1-j4 are cross-community edges); (Only j2-j4 are cross-community edges); (J3-J4 and J3-J5 both cross community boundaries, serving a strong bridging function). (Connecting j1, j2, and j3, it serves as the core bridge between the attribute and behavior communities.) (Only j3-j5 are cross-community edges); Update Influence Metrics : ; in, The normalized compact centrality, This represents the normalized betweenness centrality.

[0035] Centrality normalization (mapped to [0,1]): Tight centrality normalization: (j4, j5); ; ; ; ; .

[0036] Betweenness centrality normalization: (j2); ; ; ; ; ; calculate (α=0.5, β=0.5); ; ; ; ; ; Through the global relationship distillation described above, the influence metric update for round t=1 was completed: Feature j3 (variance of page dwell time) has the highest influence. ), because of their close connections within the behavioral community ( ), and is a core bridge across communities ( This aligns with the crucial role of "dwell time fluctuations" in web browsing as a link between behavioral patterns and attribute associations; Feature j4 (ad clicks) has the second greatest impact. Although the connection within the attribute community is weak, it serves as a core bridge across communities. ( ), is a key feature that connects user behavior and attributes; Feature j5 (device type) has the lowest influence. (This is because it has only a weak association with j3 and has no obvious intra- or cross-community value.)

[0037] The updated influence index objectively reflects the global correlation value of web browsing features, providing an accurate global cognitive basis for subsequent cognitive convergence tests and stability updates.

[0038] In step S2c, the overall difference in cognitive state vectors is calculated. The process is as follows: First, calculate the cosine similarity between the cognitive state vector of each feature in round t+1 and round t. Then, calculate the average of the cosine similarities of all features. Finally, subtract this average from 1 to represent the difference, thus obtaining the overall difference of the cognitive state vector. The specific process is as follows:

[0039] Where n is the total number of features Let j be the cognitive state vector of the j-th feature at the end of the (t+1)-th iteration. Let j be the cognitive state vector of the j-th feature at the end of the t-th iteration. The dot product of two vectors, also known as the inner product. and These are the magnitudes of the two vectors, i.e., their lengths; like If the result is positive, then the convergence is determined. The preset convergence threshold is used; Cosine similarity is used to measure the difference in cognitive state vectors, accurately capturing web browsing features and avoiding the one-sidedness of single-dimensional judgments. By converting the difference degree through global feature average similarity, the consistency of cognitive evolution across all browsing features is taken into account, ensuring the comprehensiveness and fairness of convergence judgments. A preset quantitative convergence threshold transforms subjective "convergence judgments" into objective numerical comparisons, avoiding manual intervention and adapting to the automated analysis needs of web browsing data. An iterative termination mechanism balances "analysis accuracy" and "computational efficiency," avoiding ineffective iterations, making it particularly suitable for the efficient processing of large-scale web browsing data.

[0040] For example, the e-commerce user web browsing behavior dataset has 5 core features (j=1: average daily browsing time; j=2: average daily number of pages viewed; j=3: variance of page dwell time; j=4: whether an ad is clicked; j=5: browsing device type), 1000 samples, and no historical data.

[0041] Given the key input: Cognitive state vector at t=0 round: ; ; ; Cognitive state vector for round t=1: (σ has not been updated yet; the initial value is still used). ; ; ; ; Preset parameters: Convergence threshold: ε=0.05 (a commonly used and reasonable threshold in the industry; convergence is determined when the value is less than this). Total number of features: n=5; Overall difference in cognitive state vectors.

[0042] Calculate the vector dot product of each feature ( ); Vector dot product formula: ; j=1: 0.420×0.539+0.464×0.5+0.6×0.6≈0.226+0.232+0.36=0.818; j=2: 0.396×0.499+0.407×0.5+0.6×0.6≈0.198+0.203+0.36=0.761; j=3: 0.472×0.600+0.619×0.5+0.6×0.6≈0.283+0.309+0.36=0.952; j=4: 0.338×0.400+0.462×0.5+0.6×0.6≈0.135+0.231+0.36=0.726; j=5: 0.332×0.416+0.002×0.5+0.6×0.6≈0.138+0.001+0.36=0.499; Calculate the vector magnitude of each feature ( ); Modulus formula: ; t=0 wheel module length: ; ; ; t=1 wheel module length: ; ; ; Calculate the cosine similarity of each feature ( ); j=1:0.818 / (0.866×0.950)≈0.818 / 0.823≈0.994; j=2: 0.761 / (0.826×0.927)≈0.761 / 0.766≈0.993; j=3: 0.952 / (0.983×0.985)≈0.952 / 0.968≈0.983; j=4:0.726 / (0.829×0.877)≈0.726 / 0.727≈0.999; j=5:0.499 / (0.686×0.885)≈0.499 / 0.607≈0.822; Calculate global dissimilarity (t=0→t=1 rounds); ; Convergence criteria: The preset convergence threshold ε = 0.05, because It was determined that the cognitive state had converged after t=1 rounds.

[0043] Through the consensus convergence test described above, the cognitive evolution differences from t=0 to t=1 were quantified: among the web browsing features, the cognitive state change was smallest for j=4 (whether an ad was clicked) (similarity 0.999), while the change was largest for j=5 (device type) (similarity 0.822), but the global average difference was only 0.0418, satisfying the convergence condition. This result indicates that after one round of local update and global distillation, the "importance-influence-stability" of each browsing feature has reached a stable consensus, eliminating the need for further iterations and providing a reliable convergence basis for the subsequent generation of a stable cluster structure and feature cognitive map.

[0044] Step S3 generates the result based on the converged feature cognitive state vector, specifically including: Stable cluster structure generation: Extracting the final individual importance weights of each feature after convergence. As feature weights, calculate the weighted Euclidean distance matrix between samples; The distance matrix is ​​clustered using a hierarchical clustering algorithm, and the optimal number of clusters is automatically determined in the following way: For a set of candidate cluster numbers, for each candidate number N, perform: Divide all samples into N clusters; Calculate the intra-cluster sample density of all clusters under this partition; Calculate the centroid distance separation between all clusters under this partition; The optimal number of clusters is determined by evaluating the inflection point of the evaluation curve composed of compactness and separation under different numbers of candidate clusters. Based on this optimal number of clusters, the updated cluster structure is obtained as the final cluster partitioning result of this round of analysis. Each sample group in this result is defined as a stable cluster. Feature cognitive graph construction: using each feature as a node, and its final influence index The numerical mapping is used to represent the visual size of the nodes, and the network visualization graph is drawn with the final correlation strength between features as the edge weight. At the same time, a cognitive card is generated for each feature to record the complete evolution history of its cognitive state vector. Decision-based report synthesis: For each stable cluster in the final clustering results, select the top features that contribute the most to its formation as the core feature set, and calculate the contribution ratio of each core feature; through the core feature sets and their value distributions of each final stable cluster, summarize and generate readable discriminative rules; integrate the final clustering results, cognitive map and discriminative rules, and output a structured analysis report. The stable cluster structure generation integrates weighted Euclidean distance, hierarchical clustering, and inflection point automatic cluster selection. This approach highlights the clustering value of key web browsing features through final importance weights while avoiding subjective biases from manually pre-setting cluster numbers, accurately adapting to the natural distribution of user browsing behavior. The feature cognitive map presents feature influence and correlation strength in a visual format, paired with cognitive cards to trace evolutionary history, making complex web browsing feature relationships transparent and traceable, thus solving the data black box problem. The decision-making basis report focuses on quantifying the contribution of core features and inter-cluster distinguishability rules, generating structured and actionable analytical conclusions that directly support decision-making in scenarios such as e-commerce operations, enhancing the practical value of data analysis.

[0045] For example, the e-commerce user web browsing behavior dataset has 5 core features (j=1: average daily browsing time; j=2: average daily number of pages viewed; j=3: variance of page dwell time; j=4: whether an advertisement is clicked; j=5: browsing device type), with 1000 samples.

[0046] Given the key input (final cognitive state after convergence): Final individual importance weights: ; Final Influence Metrics: ; Final association strength (MIC value) between features: j1-j2=0.70, j1-j3=0.65, j1-j4=0.50, j1-j5=0.35, j2-j3=0.60, j2-j4=0.45, j2-j5=0.30, j3-j4=0.55, j3-j5=0.40, j4-j5=0.65 Number of core features selected: the top two features that contribute the most to each cluster; Evaluation curve metrics: Density (reciprocal of the average distance between samples within a cluster) + Separation (mean distance between centroids between clusters); the inflection point is determined using the "elbow rule". Contribution ratio calculation: The proportion of the feature's distinguishing contribution to the total contribution of all features within the cluster; Generate stable cluster structures: Core logic: Using the final importance weight as the feature weight, calculate the weighted Euclidean distance matrix, and automatically determine the optimal number of clusters after hierarchical clustering.

[0047] Calculate the weighted Euclidean distance between samples; Core formula (distance between sample i and sample k): ; in, , , respectively, are the normalized values ​​of samples i and k in feature j (mapped to the interval [0,1]).

[0048] Taking sample 1 (X1=(0.6,0.5,0.7,1.0,0.0)) and sample 2 (X2=(0.3,0.4,0.2,0.0,0.0)) as examples: ; The weighted Euclidean distance matrix of all 1000 samples was calculated, and the distance range was found to be 0.12-1.83.

[0049] Hierarchical clustering and determination of the optimal number of clusters: Hierarchical clustering: The distance matrix is ​​clustered using the "Ward method" (minimizing intra-cluster variance) to generate a dendrogram; Evaluation curve plotting: Calculate the overall score when the number of clusters K=2, 3, 4, and 5 (score = density × 0.5 + separation × 0.5): K=2: Density = 0.62, Separation = 1.25, Overall Score = 0.935; K=3: Density = 0.78, Separation = 1.10, Overall Score = 0.940 (inflection point, highest score); K=4: Density = 0.85, Separation = 0.82, Overall Score = 0.835; K=5: Density = 0.90, Separation = 0.65, Overall Score = 0.775; The optimal number of clusters is K=3, which perfectly matches the natural pattern of web browsing behavior (high frequency, short duration; low frequency, long duration; and ad-sensitive).

[0050] Final stable cluster partitioning results: Cluster 1 (High Frequency, Short Duration): 320 samples, with the core behavioral characteristics being "normalized daily average browsing time < 0.5 (original < 25 min) + average daily page views < 0.6 (original < 9 pages)", indicating that users primarily browse quickly without a clear purchase intention; Cluster 2 (Low Frequency, Long Duration): 280 samples, with the core behavioral characteristics being "normalized daily average browsing time > 0.7 (original > 40 min) + variance of page dwell time > 0.6 (original > 22)", indicating that users primarily engage in in-depth browsing and comparative decision-making. Cluster 3 (Ad-Sensitive): 400 samples. The core behavioral characteristic is "whether to click on an ad = 1 (click) + page dwell time variance > 0.4 (original > 15)". Users are highly sensitive to ads and are easily converted by marketing content.

[0051] Constructing a feature-based cognitive graph: A visual graph is drawn using features as nodes, influence as node size, and association strength as edge weights, with accompanying cognitive cards to record the evolutionary history.

[0052] Graph visualization rules; Node size: determined by Mapping (values ​​range from 0 to 0.619), the larger the value, the larger the visual size of the node; Edge weight thickness: mapped by MIC value (values ​​range from 0 to 0.70), the larger the value, the thicker the connecting line; Node colors: Behavioral features (j1, j2, j3) are marked in blue, and attribute features (j4, j5) are marked in orange.

[0053] Core presentation of the map: The largest node is j3 (R=0.619), which serves as the global core connection hub. It is connected to j1 (MIC=0.65), j2 (MIC=0.60), and j4 (MIC=0.55) by thick lines, indicating a strong connection. Thickest edge: j1-j2 (MIC=0.70), which intuitively reflects the strong correlation between browsing time and the number of pages; The smallest node is j5 (R=0.002), which is only connected to j3 by a thin line (MIC=0.40), and has a low correlation value.

[0054] Feature recognition cards (taking J3 as an example): The cognitive evolution history of j3 (variance of page dwell time): The initial state (t=0) is I=0.6, R=0.5, ; After iteration, the state (t=1) is I=0.472, R=0.619. ; The core change is that the importance of individuals is smoothly reduced due to the contribution of intra-cluster differentiation, while the influence is significantly improved due to the tightness of intra-community connections and the role of cross-community bridging; the key association is the strong association with j1 and j4, which is the core link connecting user behavior and attribute characteristics.

[0055] Synthetic Decision Basis Report: Core logic: Select core features, calculate contribution ratios, summarize distinguishing rules, and generate a structured analysis report.

[0056] Filtering core features and calculating contribution ratios: Contribution ratio formula: ,in This represents the final distinguishing contribution of feature j within the cluster.

[0057] Cluster 1 (High Frequency, Short Duration): Core Feature 1 is j1 (average daily browsing time), D1=0.45, contributing approximately 41.3%; Core Feature 2 is j2 (average daily page views), D2=0.38, contributing approximately 34.9%. Cluster 2 (Low Frequency, Long Duration): Core Feature 1 is j3 (variance of page dwell time), D3=0.5, contributing approximately 35.2%; Core Feature 2 is j1 (average daily browsing time), D1=0.325, contributing approximately 22.9%. Cluster 3 (Ad-sensitive): Core feature 1 is j4 (whether the ad is clicked), D4=0.85, contributing approximately 57.4%; Core feature 2 is j3 (variance of page dwell time), D3=0.35, contributing approximately 23.6%.

[0058] Inductive distinguishing rules: Distinguishing between cluster 1 and cluster 2: cluster 1's j1 (viewing duration) < 0.5, cluster 2's j1 > 0.7, and cluster 2's j3 (dwelling variance) > 0.6, while cluster 1's j3 < 0.4; Distinguishing between cluster 1 and cluster 3: j4 (ad clicks) = 1 for cluster 3, j4 = 0 for cluster 1, and j3 > 0.4 for cluster 3, while j3 < 0.4 for cluster 1; Distinguishing between cluster 2 and cluster 3: For cluster 2, j1>0.7 and j4=0, while for cluster 3, j1<0.7 and j4=1.

[0059] Core content of structured reports: - Cluster distribution overview: Ad-sensitive users account for the largest proportion (40%), high-frequency short-duration users account for 32%, and low-frequency long-duration users account for 28%; Core operational recommendations: Target cluster 3 users with precise ads and marketing campaigns to improve conversion efficiency; optimize product detail pages and comparison functions for cluster 2 users to meet their in-depth decision-making needs; simplify browsing paths and purchase processes for cluster 1 users to lower the decision-making threshold.

[0060] Through the above-described result generation steps, a complete transformation from a convergent cognitive state to practical analytical results is achieved: a stable cluster structure accurately divides three core user groups; the feature cognitive map intuitively presents the core pivotal role of page dwell time variance; and the decision-making basis report provides directly implementable operational rules. The entire process preserves the objective laws of the data while lowering the decision-making threshold through visualization and structured design, perfectly adapting to the actual business needs of e-commerce platforms for user browsing behavior analysis.

[0061] The method also includes a continuous learning step: S4. Incremental input processing: Perform normalization processing on the new batch of data in the same way as in step S1, or parse the user's feedback tags on the historical analysis results and convert them into cognitive adjustment signals for specific features or the final stable cluster. S5. Cognitive State Adjustment: Based on new data or feedback signals, directly adjust the cognitive state vector of relevant features, specifically by reducing its stability component. The value is used to enhance its plasticity and to weight the individual importance component I. j Make revisions based on new evidence; S6. Rapid Evolution and Knowledge Fusion: Starting with the adjusted set of cognitive state vectors and taking a historical dataset containing new data as input, a cognitive evolution loop with a limited number of rounds is executed. This process employs an increased learning rate parameter. As an accelerated convergence strategy, after the cognitive evolution cycle of this finite number of rounds reaches consensus convergence, based on the finally updated cognitive state vector, the generation process based on the converged feature cognitive state vector in step S3 is executed to obtain the corresponding updated final cluster partitioning result and new rules, which include new discriminative rules. The updated final cluster partitioning result and new rules obtained after evolution are compared and fused with the original knowledge base of the system to complete the smooth update of the knowledge system. It supports dual input of incremental data and user feedback, and can adapt to dynamic changes in web browsing data without retraining. It enhances the plasticity of key features by reducing stability components to avoid cognitive solidification. It balances update speed and result stability with a rapid evolution strategy that increases learning rate and limits the number of rounds. It achieves a smooth transition between old and new rules through a knowledge fusion mechanism, forming an evolutionary web browsing analysis system that greatly improves analysis efficiency and adaptability.

[0062] Using the e-commerce user web browsing behavior dataset: initial 1000 samples, 5 core features (j=1: average daily browsing time; j=2: average daily number of pages viewed; j=3: variance of page dwell time; j=4: whether an ad was clicked; j=5: browsing device type), the final cognitive state after convergence is as follows: Final individual importance weights: ; Final Influence Metrics: ; Final stability level: ; Existing stable clusters: 3 (Cluster 1: high frequency, short duration; Cluster 2: low frequency, long duration; Cluster 3: advertising sensitive). Add scene input Incremental data: 200 new user browsing samples from the past month were added, of which 120 showed "short video ad clicks" (70% of j4=1, higher than the original 40% in cluster 3). The remaining features were normalized and consistent with historical data. User feedback: The operations staff marked "The importance of cluster 3 core feature j4 is underestimated", which was transformed into a cognitive adjustment signal: increase the importance weight of individual j4 and enhance its plasticity.

[0063] Preset parameters: Incremental data normalization: consistent with step S1 (mapped to [0,1]); Learning rate: (Compared to the original) (Increase, accelerate convergence). Evolution round limit: 2 rounds; Stability adjustment rules: j4 The value decreased by 0.2 (from 0.65 to 0.45), other characteristics constant; Weighting adjustment coefficient: The initial adjustment range of the I value of j4 is +0.1, which is further adjusted based on the distribution ratio of the newly added data.

[0064] Incremental input processing: New data normalization: 200 samples underwent normalization consistent with historical data. The typical characteristic value for "short video ad click users" is... ; Feedback signal conversion: quantization ; Merge datasets: The 200 newly added records will be merged with the existing 1,000 records to form 1,200 complete input records.

[0065] Cognitive state adjustment: Stability fine-tuning: j4 Other features Remain unchanged; Individual importance weight adjustment: The adjustment formula is as follows ,in (New data j4=1 ratio), HistRatio(j4)=0.4 (historical data j4=1 ratio), calculated as follows ; Influence metric: Rj will temporarily use the converged value and will be updated in subsequent evolution.

[0066] Rapid evolution and knowledge integration; Initialization of the evolutionary starting point: Fine-tuned cognitive state vector (j4 only adjusted): j1: ; j2: ; j3: ; j4: ; j5: .

[0067] First evolutionary cycle (t=start→t=1): Calculation of cognitive projection values ​​for newly added typical samples (based on newly added typical samples) For example): Adjustment function ,but: ; Temporary data cluster division: DBSCAN automatically identifies 4 temporary data clusters (a new "Short Video Ad Sensitive" cluster, P-value 0.75-0.85). Individual importance weight update (taking j4 as an example): j4's distinguishing contribution in the four clusters The average contribution rates were 0.12, 0.08, 0.92, and 0.85, respectively. Stability smoothing term The updated formula is: ; Global relation distillation, convergence test and stability update: final dissimilarity The situation has not improved.

[0068] Second evolutionary cycle (t=1→t=2): Repeating steps such as local cognitive updating and global relation distillation, the final difference is obtained. ,convergence.

[0069] After the update, the J4 core cognitive state is as follows: .

[0070] Knowledge integration: Cluster structure fusion: The original 3 clusters are retained, and a new "Short Video Ad Sensitive Cluster" (containing 120 new samples) is added. Rule integration: The original Cluster 3 rule has been updated to "high ad clicks + low dwell variance", and the new cluster rule is "high short video ad clicks + medium number of pages"; Knowledge base update: Supplement the fine-tuning and evolutionary history of j4, and retain the core feature relationships.

[0071] Through continuous learning, the system efficiently adapts to the dynamic changes in web browsing data: newly added "short video ad clicks" are accurately identified as new clusters, the importance of j4 features reinforced by user feedback is realized, and there is no need to retrain the entire dataset. The integrated knowledge system retains reliable historical knowledge while absorbing new data patterns, enabling the web browsing analysis system to have adaptive evolution capabilities and adapt to the ever-changing needs of e-commerce user behavior.

[0072] The method also includes online inference and early warning steps: S7. Cognitive matching degree calculation: For a new query sample, calculate its cognitive matching degree Mc with each stable cluster c; The specific calculation process is as follows: for each feature j, calculate the similarity between the query sample value and the typical value of the final stable cluster c, and multiply it by the final individual importance weight of that feature. Then multiply by a final impact metric The weighted ordinal difference penalty factor is applied, and finally, all features are summed to obtain the cognitive matching degree. The specific process is as follows: ; in, It queries the value of the j-th feature of a sample. The typical value of the c-th ultimately stable cluster on the j-th feature This indicates the rank of the value within the global distribution of historical samples. It is a similarity calculation function; S8. Intelligent Early Warning and Report Generation: If the query sample matches the highest degree of accuracy among all ultimately stable clusters... If the value falls below the preset abnormal threshold, an abnormal warning will be triggered; If the match degree between the query sample and a previously non-dominant final stable cluster is significantly higher than the statistical baseline of the historical match degree of that cluster, an emerging pattern opportunity warning is triggered. The output includes the most likely final stable cluster affiliation of the query sample and its match degree, the type of warning triggered, a list of key features leading to the match result or warning, and their contribution analysis. Output the most likely cluster affiliation of the query sample and its matching degree, the type of warning triggered, and a list of key features that led to the matching result or warning and their contribution analysis; Based on the converged feature cognition state, the cognitive matching degree calculation is performed by integrating individual importance weights, similarity, and ordinal difference penalty factors. This accurately quantifies the fit between new samples and stable clusters, supports dual early warnings for abnormal samples and emerging patterns, and adapts to real-time risk identification and opportunity capture in web browsing scenarios. At the same time, it outputs a structured report containing cluster affiliation, early warning type, and key feature contributions, providing accurate and interpretable real-time support for online decision-making and improving the pertinence and practicality of online analysis of web browsing data.

[0073] Known key inputs (final cognitive state after fusion): Final individual importance weights: ; Final Influence Metrics: ; Typical values ​​for each cluster ( (mean of features within the cluster): Cluster 1: ; Cluster 2: ; Cluster 3: ; Cluster 4: ; Historical matching baseline (mean) for each cluster +1 standard deviation ): Cluster 1: Baseline = 0.75; Cluster 2: Baseline = 0.80; Cluster 3: Baseline = 0.87; Cluster 4 (non-dominant cluster, accounting for 10%): Baseline = 0.68 Similarity function: (Values ​​range from 0 to 1, with higher values ​​indicating higher similarity); Abnormal threshold: (An anomaly warning will be triggered if the value is below this level); Emerging pattern identification: The match between the query sample and the non-dominant cluster is greater than the statistical baseline of that cluster, and the proportion of the non-dominant cluster is less than 20%; Rank calculation: rank(x) is the percentile rank (0-100) of x in the global distribution of historical samples; Prepare a query sample; Two typical query samples (normalized feature values) were selected: Sample A (abnormal browsing behavior): (Extremely short duration, very few pages, no ad clicks, computer device); Sample B (Emerging Short Video Advertising Behavior): (Medium duration, multiple pages, high click-through rate for short video ads, mobile devices); Calculate cognitive fit Mc: Core formula (match degree between cluster c and query sample): ; in, To query the value of a sample in feature j, This represents the global percentile ranking of the corresponding value.

[0074] Calculate the cognitive fit of sample A: First, determine the global order of each feature of sample A ( ) and cluster typical value ordinal position ( ): Feature j1: ; Feature j2: ; Feature j3: ; Feature j4: ; Feature j5: ; Calculate the matching degree M1 between sample A and cluster 1: ; Similarly, calculate the matching degree between sample A and other clusters: ; Highest match for sample A: ; Calculate the cognitive fit of sample B: Global order of features in sample B: feature: j1: ; j2: ; j3: ; j4: ; j5: ; Calculate the matching degree between sample B and cluster 4 (non-dominant cluster). : ; Similarly, calculate the matching degree of sample B with other clusters: ; Highest match for sample B: (Cluster 4); Intelligent early warning and report generation: Early warning judgment and reporting for sample A: Warning type: Sample A has the highest matching degree This triggers an anomaly warning; Key feature analysis: j5 (device type) contributed 32% (0.087 / 0.088≈0.989). Due to its large difference in order (90 vs 18-30), its influence was extremely low. The core anomaly was that "extremely short browsing time + no ad clicks + computer device" deviated significantly from the typical behavior of all clusters. Report output: Sample A does not belong to a clear cluster, with a matching degree of 0.088, triggering an anomaly warning. The key anomaly characteristics are the average daily browsing time (contributing 41%) and whether or not ads are clicked (contributing 35%). It is recommended to investigate abnormal account logins or invalid browsing behavior.

[0075] Early warning judgment and reporting for sample B: Warning type: The matching degree between sample B and non-dominant cluster 4 is 0.617 > the baseline of cluster 4 is 0.68; Actual calculation correction: Cluster 4 baseline = 0.68, Sample B (After recalculation), it is higher than the baseline, triggering an early warning of emerging business model opportunities; Key feature analysis: j4 (whether the ad was clicked) contributed 45% (0.325 / 0.723≈0.45), and j2 (number of pages viewed) contributed 28%. The core reason is that "high short video ad clicks + medium number of pages" fits the emerging model of cluster 4 and the matching degree exceeds the baseline. Report output: Sample B most likely belongs to cluster 4 (short video ad sensitive type), with a matching degree of 0.723, triggering an emerging model opportunity warning. The key contributing characteristics are whether the user clicks on the ad (45%) and the average number of pages viewed per day (28%). It is recommended to increase the investment in short video ads for this type of user.

[0076] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. An intelligent data analysis method, characterized in that, Includes the following steps: S1. Initialize cognitive state: Obtain normalized samples of the target dataset and initialize a cognitive state vector for each data feature. The cognitive state vector is a multi-dimensional vector used to dynamically represent the individual importance weight of the feature, its influence index in the feature association network, and the historical stability of its importance. S2. Perform the cognitive evolution cycle: Perform at least one round of cognitive evolution cycle, in which steps S2a, S2b, S2c and S2d are executed sequentially in each round; S2a. Local Cognitive Update: Based on the cognitive state vectors of all current features, each sample is mapped to a one-dimensional cognitive projection value. Temporary data clusters are dynamically divided according to the distribution of projection values ​​of all samples, and the individual importance weights in the cognitive state vectors of each feature are updated according to the distinguishing ability of each feature within the divided temporary data clusters. The mapping process uses an adjustment function to fine-tune the weights, which applies inverse adjustment to the weights based on the influence index of the features. S2b. Global Relationship Distillation: Based on the updated individual importance weights, calculate the nonlinear correlation strength between features to construct a weighted correlation network. By analyzing the topology of this network, extract the influence index of each feature in the global relationship, i.e. update the influence index, and update the corresponding components in its cognitive state vector accordingly. S2c. Consensus Convergence Test: Calculate the overall difference between the current round and the previous round of all feature cognitive state vectors, and determine whether the difference is lower than the preset threshold; S2d. Update stability: Based on the changing trend of the individual importance weight of each feature in multiple iterations, recalculate the stability component in its cognitive state vector. The smoother the changing trend, the higher the stability. S3. Result Generation: If the difference is determined to be lower than the preset threshold in step S2c, the cognitive state is determined to have converged. Based on the converged feature cognitive state vector at this time, the dataset is finally divided and interpreted to generate a stable cluster structure, feature cognitive map and decision basis report; otherwise, the updated cognitive state vector is used to return to step S2 to execute a new round of cognitive evolution cycle.

2. The intelligent data analysis method according to claim 1, characterized in that: Step S1 specifically includes: initializing a cognitive state vector for each data feature. The initialization process specifically includes: Calculate the initial individual importance weights: The calculation process is as follows: First, the information entropy and absolute kurtosis of the feature are calculated separately for all samples. Then, these two values ​​are normalized. Finally, the normalized information entropy and absolute kurtosis are weighted and summed according to a preset first weight ratio. The result is the sum of these values. ; Set initial influence metrics Set it to a constant value representing the initial neutral state; Calculate the initial stability The calculation process involves, if historical batch data exists, calculating the coefficient of variation of the estimated individual importance of this feature in each batch of data, and recording this coefficient of variation as... If there is no historical data, set it to a constant value that represents a moderate degree of stability.

3. The intelligent data analysis method according to claim 2, characterized in that: The process of calculating the sample cognitive projection value and updating the individual importance weight in step S2a specifically includes: Calculation of cognitive projection value of sample: For the cognitive projection value of the i-th sample in the t-th iteration, the calculation process is as follows: First, the normalized value of the sample on each feature is multiplied by the current individual importance weight of the corresponding feature, then scaled by the adjustment function, and finally summed over all features to obtain the cognitive projection value of the sample. Temporary data cluster partitioning: Based on the projection values ​​of all samples, a density-based clustering algorithm is used to form K temporary data clusters; Individual importance weight update: For the j-th feature, the process of calculating its updated individual importance weight is as follows: First, calculate the distinguishing contribution of the feature in each temporary data cluster. This contribution is obtained by measuring the difference between the distribution of the feature value in the temporary data cluster and the global distribution. Then, the contributions of all temporary data clusters are averaged and smoothed by combining the current stability of the feature. Finally, the weights are weighted and fused with the weights from the previous round to obtain the individual importance weights.

4. The intelligent data analysis method according to claim 1, characterized in that: The process of updating the influence metric in step S2b specifically includes: Weighted Association Network Construction: The construction process involves using the maximum information coefficient as a measure of association strength, calculating the nonlinear association strength between any two features, and using the geometric mean of the sum of the current individual importance weights of the two features as the weight coefficient of the association edge, thereby constructing a weighted fully associated network. Network community discovery and centrality calculation: The community discovery algorithm is performed on the above weighted fully interconnected network to identify feature communities with tight internal connections; then, the tight centrality of each feature j within its own community and its betweenness centrality as a bridge connecting different communities are calculated. Influence index update: The calculation process of the influence index after each feature update is as follows: first, its tight centrality and betweenness centrality are normalized respectively; then, the two are merged according to the preset second weight ratio; and finally, they are mapped to a reasonable value range through a saturation function.

5. The intelligent data analysis method according to claim 1, characterized in that: The process of calculating the overall dissimilarity of the cognitive state vector in step S2c is as follows: First, calculate the cosine similarity between the cognitive state vector of each feature in round t+1 and round t. Then, calculate the average of the cosine similarities of all features. Finally, subtract this average from 1 to represent the dissimilarity, thus obtaining the overall dissimilarity of the cognitive state vector. ; If the overall difference in the cognitive state vector is less than the preset convergence threshold, then convergence is determined.

6. The intelligent data analysis method according to claim 1, characterized in that: Step S3 generates the result based on the converged feature cognitive state vector, specifically including: Stable cluster structure generation: Extracting the final individual importance weights of each feature after convergence. As feature weights, calculate the weighted Euclidean distance matrix between samples; The distance matrix is ​​clustered using a hierarchical clustering algorithm, and the optimal number of clusters is automatically determined in the following way: For a set of candidate cluster numbers, for each candidate number N, perform: Divide all samples into N clusters; Calculate the intra-cluster sample density of all clusters under this partition; Calculate the centroid distance separation between all clusters under this partition; The optimal number of clusters is determined by evaluating the inflection point of the evaluation curve composed of compactness and separation under different numbers of candidate clusters. Based on this optimal number of clusters, the updated cluster structure is obtained as the final cluster partitioning result of this round of analysis. Each sample group in this result is defined as a stable cluster. Feature cognitive graph construction: Each feature is used as a node, and its final influence index value is mapped to the node's visual size. The final correlation strength between features is used as the edge weight to draw a network visualization graph. At the same time, a cognitive card is generated for each feature to record the complete evolution history of its cognitive state vector. Decision-based report synthesis: For each stable cluster in the final cluster division results, select the top features that contribute the most to its formation as the core feature set, and calculate the contribution ratio of each core feature; through the core feature set of each final stable cluster and its value distribution, summarize and generate readable distinguishing rules; integrate the final cluster division results, cognitive map and distinguishing rules, and output a structured analysis report.

7. An intelligent data analysis method according to claim 6, characterized in that: The method also includes a continuous learning step: S4. Incremental input processing: Perform normalization processing on the new batch of data in the same way as in step S1, or parse the user's feedback tags on the historical analysis results and convert them into cognitive adjustment signals for specific features or the final stable cluster. S5. Cognitive State Adjustment: Based on new data or feedback signals, directly adjust the cognitive state vector of relevant features, specifically by reducing the value of its stability component to enhance its plasticity, and by revising its individual importance weight component based on new evidence. S6. Rapid Evolution and Knowledge Fusion: Starting with the adjusted set of cognitive state vectors and taking a historical dataset containing new data as input, a cognitive evolution loop with a limited number of rounds is executed. This process employs an increased learning rate parameter. As a strategy to accelerate convergence; After the cognitive evolution cycle of this limited number of rounds reaches a consensus and converges, based on the final updated cognitive state vector, the generation process based on the converged feature cognitive state vector in step S3 is executed to obtain the corresponding updated final cluster partitioning result and new rules, which include new discriminative rules. The updated final cluster partitioning result and new rules obtained after evolution are compared and integrated with the original knowledge base of the system to complete the smooth update of the knowledge system.

8. An intelligent data analysis method according to claim 5, characterized in that: The method also includes online inference and early warning steps: S7. Cognitive matching degree calculation: For a new query sample, calculate its cognitive matching degree with each stable cluster c; The specific calculation process is as follows: for each feature j, calculate the similarity between the query sample value and the typical value of the final stable cluster c, multiply it by the final individual importance weight of the feature, multiply it by an ordinal difference penalty factor weighted by the final influence index, and finally sum over all features to obtain the cognitive matching degree. S8. Intelligent Early Warning and Report Generation: If the highest matching degree between the query sample and all final stable clusters is lower than the preset anomaly threshold, an anomaly warning will be triggered; If its matching degree with a final stable cluster that is not dominant in historical data is significantly higher than the statistical baseline of the historical matching degree of that cluster, an emerging pattern opportunity warning is triggered. Output the most likely final stable cluster affiliation of the query sample and its matching degree, the type of alert triggered, and a list of key features that led to the matching result or alert, along with their contribution analysis. Output the most likely cluster affiliation of the query sample and its matching degree, the type of warning triggered, and a list of key features that led to the matching result or warning and their contribution analysis.

Images