User behavior data mining method and system applied to digital enterprise management

Abstract

The invention provides a user behavior data mining method and system applied to digital enterprise management, and the method comprises the steps: collecting the multi-dimensional behavior data of a target user in a business operation interface, carrying out the multi-modal data analysis of the multi-dimensional behavior data, generating a behavior track feature set with time sequence relevance, and carrying out the mining of the behavior track feature set; training an adaptive time sequence analysis model based on the behavior trajectory feature set, capturing a long and short term dependency relationship in a user behavior mode by the time sequence analysis model through a dynamic window division strategy, generating a potential loss risk prediction index, and constructing an interaction process parameter matrix according to the potential loss risk prediction index; and calling the optimized interaction process parameter matrix to drive a service operation interface to reconstruct, generating an interaction interface adaptive to the current user behavior mode, and iteratively updating the time sequence analysis model through an incremental feedback mechanism in a preset verification period. According to the invention, the comprehensiveness and accuracy of user behavior pattern mining can be improved.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and in particular to a method and system for mining user behavior data applied to digital enterprise management. Background Art

[0002] In the field of digital enterprise management, user behavior data analysis serves as a core means of optimizing business processes and improving user experience. It mainly collects user operation data to build behavior models to support decision optimization. Existing technologies usually build static analysis models based on user behavior characteristics (such as click frequency or dwell time), predict user behavior trends through preset rules or simple time series algorithms, and use fixed parameters to drive interface optimization strategies. However, existing methods have difficulty capturing complex behavioral patterns such as the spatial distribution across interfaces, functional module switching logic, and business request correlation implied in user operations, resulting in significant deviations in the prediction of potential churn risk. At the same time, static models cannot adapt to the dynamic evolution of user behavior patterns, and interface optimization strategies are disconnected from real-time operation feedback, resulting in delayed interface reconstruction and rigid recommendation strategies, which seriously restrict the response efficiency of enterprise digital management systems and the improvement of user experience. In addition, the isolated data collection, model training, and interface optimization links in traditional technical solutions form an open-loop architecture and lack an incremental learning mechanism based on real-time user behavior feedback, further exacerbating the technical defects of insufficient system adaptability. Summary of the Invention

[0003] The present invention provides a user behavior data mining method and system applied to digital enterprise management.

[0004] In the first aspect, an embodiment of the present invention provides a user behavior data mining method for digital enterprise management, the method comprising: collecting multi-dimensional behavior data of target users in a business operation interface; performing multimodal data analysis on the multi-dimensional behavior data to generate a behavior trajectory feature set with time series correlation, wherein different modal data are timestamp synchronized and semantically associated through a cross-modal alignment strategy; training an adaptive time series analysis model based on the behavior trajectory feature set, the time series analysis model captures long-term and short-term dependencies in user behavior patterns through a dynamic window partitioning strategy, and generates potential churn risk prediction indicators; constructing an interaction process parameter matrix based on the potential churn risk prediction indicators; calling the optimized interaction process parameter matrix to drive the business operation interface to reconstruct, generating an interaction interface adapted to the current user behavior pattern, and iteratively updating the time series analysis model through an incremental feedback mechanism within a preset verification cycle.

[0005] In a second aspect, an embodiment of the present invention provides a computer system, comprising: a memory storing a computer program; and a processor for loading the computer program to implement the above-mentioned user behavior data mining method for digital enterprise management.

[0006] The user behavior data mining method for digital enterprise management provided by the present invention uses a cross-modal alignment strategy to achieve a deep fusion of spatiotemporal features and semantic associations through multi-dimensional behavioral data, breaking through the limitations of traditional data analysis and significantly improving the comprehensiveness and accuracy of user behavior pattern mining. The adaptive time series analysis model constructed based on the dynamic window partitioning strategy can intelligently capture the dynamic evolution law of long-term and short-term dependencies in user operations, effectively enhancing the reliability of potential churn risk prediction indicators; by constructing an interactive process parameter matrix containing interface element weight distribution, functional module priority sequence and recommendation strategy adjustment coefficient, it realizes the deep coupling of user behavior characteristics and business scenarios, driving the real-time dynamic optimization of the business operation interface reconstruction strategy; combined with the closed-loop learning link formed by the incremental feedback mechanism, the time series analysis model and interface interaction parameters are continuously iterated and updated to ensure the synchronous adaptation of user behavior pattern evolution and interface optimization process, ultimately reducing the risk of user churn while comprehensively improving the response efficiency and user operation experience of the digital management system, providing reliable support for enterprises to make accurate decisions in complex business scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 This is a flowchart of a user behavior data mining method applied to digital enterprise management provided by an embodiment of the present invention.

[0008] Figure 2 It is a schematic diagram of the composition of a computer system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0009] See also Figure 1 , Figure 1 A flowchart of a user behavior data mining method for digital enterprise management provided by an embodiment of the present invention is provided. The user behavior data mining method for digital enterprise management can be executed by a computer system and includes the following steps:

[0010] Step S100: Collect multi-dimensional behavior data of the target user on the business operation interface, which includes interface click tracks, operation time distribution, function module switching paths and related business data request records.

[0011] An interface click trajectory refers to the path of clicks a user makes on a business operation interface. It reflects the location and sequence of user operations on the interface. For example, on an e-commerce platform's business operation interface, a user sequentially clicks on the product details page, the add to cart button, and the checkout button. These clicks constitute the interface click trajectory. Operation duration distribution refers to the distribution of time spent by users on different operation steps. It reflects the user's focus and level of engagement in each operation. For example, if a user spends 5 minutes on the product details page and 2 minutes on the delivery address page, this is a reflection of operation duration distribution. Functional module switching paths refer to the paths users take between different functional modules on the business operation interface. They reflect the user's operational process and business needs. Related business data request records refer to records of related business data requests made by users to the system during operations. They contain information about the user's business data needs. For example, in a financial system, a user's request to query financial statements for a specific time period is a related business data request record.

[0012] Step S200: performing multimodal data analysis on the multi-dimensional behavior data to generate a behavior trajectory feature set with temporal correlation, wherein different modal data are synchronized with timestamps and semantically associated through a cross-modal alignment strategy.

[0013] Multimodal data parsing refers to the analysis and processing of different types of multidimensional behavioral data to extract useful information. In this step, it is necessary to parse data of four different modalities: interface click trajectory, operation time distribution, function module switching path, and associated business data request record. A behavioral trajectory feature set with temporal correlation refers to the conversion of multidimensional behavioral data into a feature set with time sequence and correlation, which can more comprehensively reflect the user's behavior pattern. The cross-modal alignment strategy aligns data of different modalities in time and semantics through timestamp synchronization and semantic association for better analysis and processing. Timestamp synchronization refers to unifying the timestamps of data of different modalities so that they are comparable in time.

[0014] As an embodiment, the multi-dimensional behavior data includes interface click trajectories, operation duration distribution, function module switching paths, and associated business data request records. Based on this, step S200 performs multimodal data analysis on the multi-dimensional behavior data to generate a behavior trajectory feature set with temporal correlation. Specifically, the following steps S210 to S250 may be included:

[0015] Step S210: extracting the spatial coordinate sequence in the interface click trajectory, and calculating the interface area attention heat map based on the coordinate distribution density.

[0016] The spatial coordinate sequence in an interface click trajectory is the sequence of coordinates of the locations where users click on the business operation interface. It accurately reflects the user's click location on the interface. An interface area attention heat map is a chart used to visualize user attention to different areas of the interface. It uses color depth to indicate the degree of user attention to different areas. Darker areas indicate higher user attention.

[0017] Extracting the spatial coordinate sequence can be achieved by processing the interface click trajectory data. Specifically, the position coordinates of each click operation can be extracted from the click trajectory data and arranged in the order of clicks to obtain a spatial coordinate sequence. In one example, the business operation interface is divided into several small areas, for example, the interface is divided into a 10×10 grid. Then, the number of clicks in each area is counted. The more clicks, the greater the coordinate distribution density of the area. Finally, a color value is assigned to each area based on the coordinate distribution density. The depth of the color value is proportional to the coordinate distribution density, thereby generating an interface area attention heat map.

[0018] Step S220: Analyze the dwell interval characteristics in the operation duration distribution, identify operation clusters and abnormal stagnation nodes consisting of continuous operation events exceeding a frequency threshold, and identify high-frequency operation clusters consisting of continuous events whose operation frequency exceeds a preset threshold within a sliding window.

[0019] The dwell interval feature in the operation duration distribution refers to the time intervals between different operation steps. It can reflect the user's operation rhythm and attention level. For example, if a user pauses for a long time between two adjacent operation steps, it may indicate that the user is thinking or encountering a problem. An operation cluster is a collection of consecutive operation events that exceeds a frequency threshold. It represents a series of operations that the user performs intensively over a period of time. An abnormal stagnation node is a location where the user pauses for an unusually long time during an operation. It may indicate that the user encountered difficulties or there is a problem with the operation process. A high-frequency operation cluster is a cluster of consecutive operation events whose operation frequency exceeds a preset threshold within a sliding window.

[0020] Interval feature analysis can be achieved by analyzing the operation duration distribution data. For example, the intervals between adjacent operation steps can be calculated and statistically analyzed, such as the mean and standard deviation, to extract interval features.

[0021] Step S230: constructing a transition probability matrix of the functional module switching path, calculating the state transition weights between the functional modules, and generating a path prediction sequence based on the hidden Markov model.

[0022] The transition probability matrix for a functional module switching path is a matrix that represents the probability of a user switching from one functional module to another. The rows of the matrix represent the starting functional module, the columns represent the target functional module, and the elements in the matrix represent the transition probability from the starting functional module to the target functional module. The state transition weight refers to the importance of the transition from one functional module to another during the functional module switching process. It can be calculated based on factors such as transition probability and operation frequency. For example, the number of times a user switches between different functional modules is counted to obtain a switching count matrix. Then, the elements of each row of the switching count matrix are divided by the sum of the elements in that row to obtain the transition probability matrix. The state transition weight can be calculated using a weighted calculation based on factors such as transition probability and operation frequency. For example, the state transition weight can be obtained by multiplying the transition probability by the operation frequency.

[0023] Step S240: Map the associated business data request record into a request type encoding vector, and construct a request load evaluation index through the request response delay and the result data volume.

[0024] Mapping associated business data request records to request type encoding vectors means representing different types of associated business data request records in vector form. Request response latency refers to the time it takes from a user initiating a data request to the system returning a response result. Result data volume refers to the size of the response result returned by the system. The request load evaluation metric is used to assess the system's request processing load and reflects the system's pressure and performance when processing requests.

[0025] Step S250: Perform multimodal feature fusion on the attention heat map, dwell interval features, transition probability matrix, and request load evaluation index, and use a time axis alignment strategy to generate a temporal embedding vector containing global context information as a behavior trajectory feature set.

[0026] Multimodal feature fusion refers to the fusion of different types of features to obtain a more comprehensive and valuable feature representation. In this step, four different types of features, namely attention heat map, dwell interval feature, transition probability matrix and request load evaluation index, are fused. The timeline alignment strategy is a method for processing multimodal data, which aligns data of different modalities on the timeline so that they are consistent in time. The time series embedding vector is a vector that embeds time series data, which can convert information with time sequence into a representation in a low-dimensional vector space. Global context information refers to the overall information that includes multi-dimensional behavioral data, which can more comprehensively reflect the user's behavior pattern.

[0027] Various methods can be used to fuse multimodal features, such as feature concatenation and weighted summation. In this step, the attention heatmap, dwell interval features, transition probability matrix, and request load evaluation metric are concatenated to generate a fused feature vector. The process of generating a time series embedding vector using a time axis alignment strategy is as follows: First, the fused feature vectors are arranged in chronological order to form a time series feature sequence. Then, an embedding algorithm (such as Word2Vec or GloVe) is used to convert the time series feature sequence into a time series embedding vector. This time series embedding vector incorporates global contextual information and can more comprehensively reflect user behavior patterns, thus serving as a behavioral trajectory feature set. When fusion is performed on multimodal features, different weights can be assigned to each feature based on its importance. For features of different dimensions, weight assignment can balance their influence in the fusion process. For example, the weight of each feature can be determined based on domain knowledge, experimental verification, or a machine learning algorithm (such as a feature selection algorithm). Each feature is multiplied by its corresponding weight before fusion, allowing features of different dimensions to participate in the fusion process in a reasonable manner. Alternatively, minimum-maximum normalization is used to scale the value range of each feature to the interval [0, 1] before fusion is performed. In other embodiments of the present invention, if similar technical content involving the fusion of features of different dimensions is involved, it should be understood that those skilled in the art can unify the dimensions based on their own known technical knowledge to perform subsequent further feature operations.

[0028] Step S300: training an adaptive time series analysis model based on the behavioral trajectory feature set. The time series analysis model captures the long-term and short-term dependencies in the user behavior pattern through a dynamic window partitioning strategy and generates a potential churn risk prediction indicator.

[0029] An adaptive time series analysis model automatically adjusts its parameters and structure based on input data. It can better adapt to different user behavior patterns and data characteristics. A dynamic window partitioning strategy is a method for processing time series data that dynamically adjusts the window size and position based on data characteristics and analysis requirements to capture long-term and short-term dependencies in user behavior patterns. Long-term and short-term dependencies refer to the long-term and short-term correlations between data at different time points in time series data. The potential churn risk prediction indicator is used to predict whether a user is at risk of churn, helping companies take proactive measures to prevent user churn.

[0030] As an implementation method, step S300, training an adaptive time series analysis model based on a behavioral trajectory feature set, may specifically include the following steps S310 to S350:

[0031] Step S310: Input the temporal embedding vector into the bidirectional long short-term memory network, capture the forward and reverse temporal dependencies through the gating mechanism, and generate an initial hidden state sequence.

[0032] Specifically, the time series embedding vector is input into the BiLSTM network in chronological order. At each time step, a gating mechanism calculates the states of the input gate, forget gate, and output gate based on the current input and the hidden state of the previous time step, thereby controlling the inflow, retention, and output of information. This gating mechanism enables the BiLSTM network to effectively capture forward and reverse temporal dependencies. Finally, the outputs of the forward and reverse LSTM layers are concatenated to obtain the initial hidden state sequence.

[0033] Step S320: Connect a multi-head self-attention mechanism layer to the back end of the bidirectional long short-term memory network, calculate the association weights between features at different time steps, and perform attention-weighted aggregation on the initial hidden state sequence to obtain an aggregated feature sequence.

[0034] The association weight refers to the importance of features at different time steps in a multi-head self-attention mechanism. The process of connecting a multi-head self-attention layer to the backend of a bidirectional long short-term memory network is as follows: First, the initial hidden state sequence is input into the multi-head self-attention layer. Within each head, the association weights between features at different time steps in the input sequence are calculated. Specifically, the association weights are calculated by calculating the similarity between the query vector, key vector, and value vector. Then, the input sequence is weighted and summed according to the association weights to obtain the output of each head. Finally, the outputs of multiple heads are concatenated to obtain the aggregated feature sequence.

[0035] Step S330: Use a dynamic convolution kernel group to perform multi-scale feature extraction on the aggregated feature sequence to obtain multi-scale features. The dynamic convolution kernel group adaptively adjusts the convolution kernel size and expansion coefficient according to the current input feature dimension.

[0036] A dynamic convolution kernel group is a set of convolution kernels that can adaptively adjust the size and dilation coefficient of the convolution kernel based on the dimension of the current input feature. Multi-scale feature extraction refers to extracting feature information of different scales from the input feature sequence to more comprehensively describe the characteristics of the data.

[0037] As an implementation method, step S330 uses a dynamic convolution kernel group to perform multi-scale feature extraction on the aggregated feature sequence to obtain multi-scale features. Specifically, the following steps S331 to S335 may be included:

[0038] Step S331: Calculate the convolution kernel size adjustment factor according to the time step and channel dimension of the input feature sequence. The adjustment factor is negatively correlated with the local variance of the feature sequence.

[0039] The time step of the input feature sequence refers to the length of the feature sequence in the time dimension, and the channel dimension refers to the number of feature sequences in the feature dimension. The convolution kernel resize factor is a factor used to adjust the convolution kernel size. It is negatively correlated with the local variance of the feature sequence. Local variance refers to the variance of the feature sequence within a local region and can reflect the degree of fluctuation of the feature sequence.

[0040] The process of calculating the convolution kernel resizing factor based on the time steps and channel dimensions of the input feature sequence is as follows: First, the variance of the input feature sequence in different local regions is calculated. The feature sequence can be divided into several local regions, each containing a number of time steps and channels. Then, the variance of each local region is statistically analyzed to obtain the average of the local variances. Finally, the convolution kernel resizing factor is calculated based on the average of the local variances. The resizing factor is negatively correlated with the local variance. For example, an inverse proportional function can be used to calculate the resizing factor, so that the larger the local variance, the smaller the resizing factor, thereby reducing the size of the convolution kernel.

[0041] Step S332: Generate a basic convolution kernel size set based on the adjustment factor, and allocate a learnable expansion coefficient matrix inside each convolution kernel. The expansion coefficient matrix dynamically adjusts the receptive field range according to the feature differences of adjacent time steps.

[0042] The base convolution kernel size set refers to a set of convolution kernel sizes generated based on the convolution kernel size adjustment factor. A learnable dilation coefficient matrix is ​​a matrix whose elements can be learned and adjusted during training. The dilation coefficient matrix dynamically adjusts the receptive field based on the feature differences between adjacent time steps, thereby better capturing feature information at different scales.

[0043] The process of generating a set of basic convolution kernel sizes based on the adjustment factor is as follows: First, a range of basic convolution kernel sizes is determined based on the convolution kernel size adjustment factor. Then, a set of convolution kernels of different sizes is generated within this range to form a set of basic convolution kernel sizes. The process of assigning a learnable dilation coefficient matrix within each convolution kernel is as follows: A learnable dilation coefficient matrix is ​​assigned to each convolution kernel, and the elements of the matrix are initialized to appropriate values. During training, the elements of the dilation coefficient matrix are dynamically adjusted based on the differences in features between adjacent time steps. Specifically, a neural network can be used to learn the dilation coefficient matrix, with the input being the feature differences between adjacent time steps and the output being the adjusted dilation coefficient matrix. By adjusting the dilation coefficient matrix, the receptive field range of the convolution kernel can be changed, thereby extracting feature information at different scales.

[0044] Step S333: Apply multiple dynamic convolution kernels in parallel to the same input feature sequence to generate feature maps with different scale characteristics.

[0045] A dynamic convolution kernel is one that adaptively adjusts its size and dilation coefficient based on the input feature dimensions. A feature map is the output feature map obtained after a convolution operation, which reflects the characteristic information of the input feature sequence at different scales.

[0046] The process of applying multiple dynamic convolution kernels in parallel to the same input feature sequence is as follows: the input feature sequence is simultaneously fed into multiple dynamic convolution kernels, each with a different size and dilation factor. Each convolution kernel performs a convolution operation on the input feature sequence, generating a feature map. Due to the different sizes and dilation factors of the convolution kernels, the generated feature maps have different scale characteristics. By applying multiple dynamic convolution kernels in parallel, feature information at different scales can be extracted simultaneously, resulting in feature maps with different scale characteristics.

[0047] Step S334: Perform channel attention weighting on the feature map, calculate the importance score of each channel feature and redistribute the channel weight.

[0048] The importance score of each channel feature is the score obtained by evaluating the importance of each channel feature in the feature map. Reassigning channel weights refers to adjusting the channel weights based on the importance scores of each channel feature, giving important channels greater weights. First, a global average pooling operation is performed on each channel of the feature map to compress the feature map of each channel into a scalar value. These scalar values ​​are then input into a fully connected neural network, which outputs an importance score for each channel. Based on the importance score, the channel weights are redistributed. For example, a sigmoid function can be used to map the importance score to the range [0, 1], and this value is then used as the channel weight. Finally, each channel of the feature map is multiplied by the corresponding weight to obtain the feature map after channel attention weighting.

[0049] Step S335: concatenate the weighted feature maps along the channel dimension and perform dimensionality reduction processing through a separable convolutional layer to generate multi-scale features.

[0050] Specifically, the feature maps of multiple channel attention weights are concatenated along the channel dimension, arranging their channels in sequence to form a new feature map. The concatenated feature maps are then convolved using depthwise separable convolution, which performs convolution on each channel separately and then concatenates the results. Next, the output of the depthwise separable convolution is convolved using pointwise convolution, which linearly combines each channel to achieve dimensionality reduction. Finally, the resulting feature map after dimensionality reduction is the multi-scale feature.

[0051] Step S340: Perform a residual connection between the multi-scale features and the original time series embedding vector, and map them into a latent vector space containing an abstract representation of the user behavior pattern through a fully connected layer.

[0052] Specifically, the multi-scale features and the original time series embedding vector are added element-by-element to obtain a residually connected feature vector. This residually connected feature vector is then fed into a fully connected layer, which performs a linear transformation and nonlinear activation on the input vector, mapping it into a low-dimensional vector space. This low-dimensional vector space, known as the latent vector space, contains an abstract representation of user behavior patterns and can be used for subsequent analysis and prediction.

[0053] Step S350: Construct a joint training framework for contrastive learning tasks and prediction tasks in the latent vector space, wherein the contrastive learning task enhances the model's ability to recognize behavioral pattern differences through positive and negative sample pairs, and the prediction task optimizes the accuracy of potential churn risk prediction indicators through the cross-entropy loss function.

[0054] For example, a set of feature vectors of user behavior patterns is first sampled from the latent vector space, positive sample pairs are generated based on the temporal continuity of the feature vectors, and negative sample pairs are generated by random perturbation or pairing of feature vectors in non-adjacent time periods to form a training data set for the contrastive learning task.

[0055] The latent vector space is a low-dimensional vector space that contains abstract representations of user behavior patterns. A feature vector set is a set of feature vectors sampled from the latent vector space, each of which represents a user's behavior pattern. A positive sample pair is a pair of feature vectors with similar behavior patterns, generated based on the temporal continuity of the feature vectors. A negative sample pair is a pair of feature vectors with different behavior patterns, generated by random perturbation or pairing feature vectors from non-adjacent time periods. The training data set for contrastive learning tasks is a dataset consisting of positive and negative sample pairs, used to train contrastive learning models.

[0056] Random sampling can be used to select several feature vectors from the latent vector space to form a feature vector set. For each feature vector in the feature vector set, temporally adjacent feature vectors are selected as positive samples to form a positive sample pair. Feature vectors can be randomly perturbed to generate new feature vectors, which are then paired with the original feature vector to form a negative sample pair. Alternatively, feature vectors from non-adjacent time periods can be paired to form a negative sample pair. Finally, the positive and negative sample pairs are combined to form the training data set for the contrastive learning task.

[0057] Next, the similarity score of the feature vectors in the positive sample pair is calculated and compared with the similarity score of the negative sample pair to generate a contrastive loss value that reflects the consistency of the behavioral pattern. In this step, the similarity score of the feature vectors in the positive sample pair and the similarity score of the feature vectors in the negative sample pair are calculated. The contrastive loss value is a loss value used to measure the difference in similarity between the positive and negative sample pairs, which can reflect the consistency of the behavioral pattern.

[0058] Methods such as cosine similarity and Euclidean distance can be used to calculate the similarity score between two feature vectors. The similarity score of the positive sample pair is then compared with the similarity score of the negative sample pair. A loss function can be used to calculate the comparative loss value. For example, a triplet loss function can be used, with the difference between the similarity scores of the positive sample pair and the negative sample pair as the loss value. This comparative loss value allows the model to learn the differences between positive and negative sample pairs, thereby enhancing the model's ability to identify differences in behavioral patterns.

[0059] Next, the feature vector in the latent vector space is input into the prediction task classifier to obtain the user churn risk prediction result. The cross-entropy loss function is used to calculate the deviation between the prediction result and the true label to generate a prediction loss value. The prediction task classifier is a classifier used to predict user churn risk, classifying the input feature vector as high churn risk or low churn risk. The user churn risk prediction result refers to the prediction task classifier's prediction of user churn risk. The true label refers to the user's actual churn risk label, which can be obtained from historical data or actual observations.

[0060] The contrastive loss and the prediction loss are then fused to generate a joint training loss function. This joint training loss function takes into account the training objectives of both the contrastive learning task and the prediction task. By using this joint training loss function, the model can improve the accuracy of predicting user churn risk while learning from differences in behavioral patterns. For example, the contrastive loss and the prediction loss can be fused using a weighted summation approach.

[0061] The parameters of the time series analysis model are then iteratively updated based on the joint training loss function, simultaneously optimizing the clustering tightness of the behavioral pattern features in the latent vector space and the clarity of the prediction task classification boundary. Based on the gradient of the joint training loss function, the model parameters are continuously adjusted to gradually reduce the joint training loss function. Clustering tightness refers to the degree to which feature vectors with similar behavioral patterns are clustered together in the latent vector space. The clarity of the prediction task classification boundary refers to how clearly the prediction task classifier demarcates high- and low-churn risk regions in the latent vector space.

[0062] Use an optimization algorithm (such as stochastic gradient descent, Adam, etc.) to update the model parameters based on the gradient of the joint training loss function. In each iteration, calculate the gradient of the joint training loss function with respect to the model parameters, and then adjust the model parameters based on the direction and magnitude of the gradient. At the same time, optimize the clustering density of behavioral pattern features in the latent vector space and the clarity of the classification boundaries of the prediction task. By continuously adjusting the model parameters, feature vectors with similar behavioral patterns are clustered more closely together in the latent vector space, and the high-churn risk and low-churn risk areas divided by the prediction task classifier in the latent vector space are clearer. This can improve the model's ability to recognize behavioral patterns and the accuracy of predicting user churn risk.

[0063] Finally, after each parameter update, the strategy for generating positive and negative sample pairs is re-evaluated. If the contrastive learning task's contribution to improving the accuracy of the prediction task falls below a preset threshold, the temporal span of the positive sample pairs and the semantic difference between the negative sample pairs are dynamically enhanced. The joint training framework, during model convergence, moves the feature vectors of high-churn risk users away from the low-risk user cluster in the latent vector space, while retaining the similarity in the behavioral patterns of users with the same risk level. After each parameter update, the contribution of the contrastive learning task to improving the accuracy of the prediction task is calculated, comparing the accuracy of the prediction task classifier with and without the contrastive learning task. If the contrastive learning task's contribution to improving the accuracy of the prediction task falls below a preset threshold, this indicates that the current strategy for generating positive and negative sample pairs may not be effective and requires adjustment.

[0064] Step S400: constructing an interaction process parameter matrix based on the potential churn risk prediction index. The interaction process parameter matrix includes the weight distribution of interface elements, the priority sequence of functional modules, and the adjustment coefficient of the recommendation strategy.

[0065] The potential churn risk prediction indicator is used to predict whether a user has a churn risk and can be calculated based on the user's behavior patterns and historical data. The interactive process parameter matrix is ​​a matrix that includes the weight distribution of interface elements, the priority sequence of functional modules, and the recommendation strategy adjustment coefficient. It can be used to optimize the interactive process of the business operation interface and improve the user experience. The interface element weight distribution refers to the weight distribution obtained by evaluating the importance of each element in the business operation interface. It can be used to drive the interface layout rendering. The functional module priority sequence refers to the sequence obtained by sorting the priorities of each functional module in the business operation interface. It can be used for navigation path optimization. The recommendation strategy adjustment coefficient is a coefficient used to adjust the recommended content. It can be used to generate dynamic content push strategies.

[0066] Specifically, step S400 may include the following steps S410 to S450:

[0067] Step S410: extracting user behavior deviation features from potential churn risk prediction indicators. The behavior deviation features include the matching difference results between the interface operation frequency distribution and the historical behavior trajectory, and the abnormal fluctuation coefficient of the functional module stay time.

[0068] The user behavior deviation feature measures the degree of difference between a user's current behavior and their historical behavior. The difference between the user's interface operation frequency distribution and their historical behavior trajectory indicates the degree of match between the user's current interface operation frequency distribution and their historical behavior trajectory, reflecting any changes in user behavior habits. The abnormal fluctuation coefficient of the functional module dwell time indicates the degree of abnormal fluctuation in the duration of user dwell time on different functional modules, reflecting any changes in user interest in different functional modules.

[0069] For example, calculate the matching difference between the interface operation frequency distribution and the historical behavior trajectory. The current interface operation frequency distribution can be compared with the operation frequency distribution of the historical behavior trajectory to calculate the similarity score between them. The lower the similarity score, the greater the matching difference. Next, determine the abnormal fluctuation coefficient of the functional module dwell time. For example, perform a statistical analysis of the user's dwell time on different functional modules and calculate statistics such as the standard deviation or coefficient of variation. These statistics can reflect the degree of abnormal fluctuation of the dwell time. By extracting the user behavior deviation characteristics, we can better understand the changes in user behavior and provide a basis for the subsequent construction of the interaction process parameter matrix.

[0070] Step S420: Based on the interface operation frequency distribution and the matching difference results, the dynamic weight distribution value of each interface element is calculated to generate the interface element weight distribution, wherein the weight value of the area where the matching difference result is lower than the preset difference threshold increases in a positive correlation.

[0071] The dynamic weight distribution of interface elements refers to the weight assigned to each interface element based on the interface operation frequency distribution and the matching difference results. It can reflect the importance of the interface element. The interface element weight distribution can be used to drive interface layout rendering. The preset difference threshold is a pre-set threshold used to determine whether the matching difference between the interface operation frequency distribution and the historical behavior trajectory is within a normal range.

[0072] For example, the number of element clicks and dwell time sequences in the interface operation frequency distribution are analyzed, the spatial coordinate sets and corresponding time distribution features of the operation areas that exceed the frequency threshold are extracted, and an initial weight mapping table is generated. The number of element clicks in the interface operation frequency distribution refers to the number of times the user clicks on each element on the interface, and the dwell time sequence refers to the time series during which the user stays on each element. The frequency threshold is a pre-set threshold used to determine whether the number of clicks on an element exceeds the normal range. The spatial coordinate set of the operation area refers to the set of position coordinates of the operation area that exceeds the frequency threshold. The time distribution feature refers to the temporal distribution feature of the operation area that exceeds the frequency threshold, such as the time interval between operations, the peak period of operations, etc. The initial weight mapping table is a table used to record the initial weight of each interface element, which can provide a basis for subsequent weight allocation.

[0073] Next, based on the initial weight mapping table and the regional division rules of the interface layout, the density distribution gradient of each interface sub-area is calculated to generate a regional density distribution based on spatial aggregation. The initial weight mapping table is a table that records the initial weight of each interface element. The regional division rules of the interface layout refer to the rules for dividing the interface into different sub-areas. The density distribution gradient of the interface sub-area refers to the gradient of change of element weights in different sub-areas of the interface, which can reflect the degree of spatial aggregation of elements. The regional density distribution based on spatial aggregation refers to the regional density distribution generated according to the degree of spatial aggregation of elements, which can be used for subsequent weight allocation and interface layout optimization.

[0074] Then, the matching difference results are decomposed into the difference quantification values ​​of each interface area, and the abnormal operation areas beyond the normal fluctuation range are identified by the preset difference threshold, and the abnormal area space identifier is generated. The matching difference result refers to the matching difference result between the interface operation frequency distribution and the historical behavior trajectory, which can reflect whether the user's operating habits have changed. The difference quantification value of each interface area refers to the quantitative value obtained after decomposing the matching difference result into each interface area, which can more accurately reflect the changes in operating habits of each interface area. The abnormal operation area refers to the interface area where the change in operating habits exceeds the normal range. The abnormal area space identifier is an identifier used to identify the location of the abnormal operation area, which can provide a basis for subsequent weight adjustment and interface optimization.

[0075] Afterwards, dynamic weight allocation is performed based on the regional density distribution and the abnormal area spatial identifier, wherein the initial weight value of the area exceeding the preset density threshold is adjusted through the reverse compensation mechanism of the difference quantization value to generate the dynamic weight allocation value of each interface element. Regional density distribution refers to the regional density distribution generated according to the degree of clustering of elements in space, which can reflect the distribution of elements in different interface areas. Dynamic weight allocation refers to the real-time adjustment of the weight value of each interface element based on the regional density distribution and the abnormal area spatial identifier. The preset density threshold is a pre-set threshold used to determine whether the degree of element clustering in the interface area is too high. The reverse compensation mechanism of the difference quantization value means that when the operating habits of the interface area change abnormally, the initial weight value of the area is adjusted to change in the opposite direction to balance the user's operating habits.

[0076] Specifically, for each interface element, the regional density distribution and the spatial identifier of the abnormal region are obtained based on the interface area in which it is located. Next, a determination is made as to whether the regional density of the region exceeds a preset density threshold. If so, it indicates that the element clustering in the region is too high, and the initial weight value of the region needs to be adjusted. Specifically, an inverse compensation mechanism is used to make adjustments based on the difference quantization value of the region.

[0077] Finally, the dynamic weight distribution values ​​are smoothed across regions to eliminate the weight mutation noise between adjacent interface elements and generate an interface element weight distribution that meets the visual continuity constraint. The final weight value of the interface element weight distribution is determined by the dynamic correction results of the regional density distribution and the spatial identifier of the abnormal area, and the weight distribution process is linked in real time to the differential evolution state of user operation frequency and historical trajectory.

[0078] Cross-region smoothing refers to smoothing the dynamic weight distribution values ​​of adjacent interface elements to eliminate weight mutation noise. Weight mutation noise refers to the sudden changes between the weight values ​​of adjacent interface elements, which will affect the visual continuity of the interface and user experience. The visual continuity constraint means that the weight distribution of interface elements should conform to the principle of visual continuity, that is, the weight values ​​of adjacent interface elements should change gradually rather than suddenly. The final weight value of the interface element weight distribution refers to the interface element weight value obtained after cross-region smoothing, which is determined by the dynamic correction results of the regional density distribution and the abnormal area spatial identifier. The real-time correlation of the difference evolution state of user operation frequency and historical trajectory in the weight distribution process means that in the weight distribution process, it is necessary to consider the difference changes in user operation frequency and historical trajectory in real time, and dynamically adjust the weight value according to the difference evolution state.

[0079] Step S430: Based on the abnormal fluctuation coefficient of the functional module dwell time and the module switching path frequency in the user's historical operation trajectory, a module transition probability graph is constructed, and a functional module priority sequence is generated by sorting the in-degree weights of the nodes in the graph.

[0080] The abnormal fluctuation coefficient of the duration of a functional module's stay refers to the degree of abnormal fluctuation in the length of time a user spends on different functional modules, reflecting changes in user interest in different functional modules. The module switching path frequency in a user's historical operation trajectory refers to the frequency with which users switch between different functional modules, reflecting their operational habits and business needs. The module transition probability graph is a directed weighted graph that represents the probability of a user switching from one functional module to another. The in-degree weight of a node in the graph refers to the sum of the weights of the edges pointing to that node, reflecting the importance and popularity of that functional module.

[0081] Specifically, the temporal fluctuation characteristics in the abnormal fluctuation coefficient of the functional module residence time can be analyzed first, the abnormal module set that exceeds the preset fluctuation threshold can be identified, and an abnormal module identifier and fluctuation intensity association table can be generated. The temporal fluctuation characteristics in the abnormal fluctuation coefficient of the functional module residence time refer to the fluctuation characteristics of the coefficient over time, which can reflect the changes in the user's interest in the functional module over time. The preset fluctuation threshold is a pre-set threshold used to determine whether the fluctuation of the functional module residence time exceeds the normal range. The abnormal module set refers to the set of functional modules whose functional module residence time fluctuation exceeds the preset fluctuation threshold. The abnormal module identifier is an identifier used to identify the abnormal module, and the fluctuation intensity refers to the degree of fluctuation of the functional module residence time of the abnormal module. The abnormal module identifier and fluctuation intensity association table is a table used to record abnormal module identifiers and corresponding fluctuation intensities, which can provide a basis for the subsequent correction of the module transfer path weight matrix.

[0082] Next, we extract the switching path sequence of function modules that were continuously accessed in the user's historical operation trajectory, count the jump frequency and jump direction between modules, and generate a module transfer path weight matrix. The switching path sequence of function modules that were continuously accessed in the user's historical operation trajectory refers to the sequence of function modules that the user accessed in sequence during their historical operation. The jump frequency between modules refers to the number of times the user switched from one function module to another, and the jump direction refers to the direction in which the user switched function modules. The module transfer path weight matrix is ​​a matrix that records the weight of a user switching from one function module to another. The weight can be calculated based on factors such as jump frequency and jump direction.

[0083] For example, historical user operation trajectory data is analyzed and processed to extract information about continuously accessed functional modules, forming a functional module switching path sequence. The functional module switching path sequence is traversed, and the number and direction of jumps between adjacent functional modules are counted. Based on the jump frequency and jump direction, a weight is assigned to each inter-module transition path to form a module transition path weight matrix. For example, the jump frequency can be used as the weight, with higher jump frequency being associated with a greater weight.

[0084] Next, the abnormal module identifier is mapped to the module transfer path weight matrix, the weight attenuation factor of the abnormal module in the transfer path is calculated, and the module transfer path weight matrix is ​​dynamically modified. Mapping the abnormal module identifier to the module transfer path weight matrix involves associating the abnormal module identifier with the corresponding element in the module transfer path weight matrix to adjust the abnormal module's transfer path weight. The weight attenuation factor of the abnormal module in the transfer path is a factor used to adjust the abnormal module's transfer path weight. It can be calculated based on factors such as the abnormal module's fluctuation intensity. Dynamic modification of the module transfer path weight matrix involves adjusting the abnormal module's transfer path weight in the module transfer path weight matrix based on the weight attenuation factor to reflect changes in user interest in the abnormal module. For example, the abnormal module identifier and fluctuation intensity association table is traversed to find the elements in the module transfer path weight matrix corresponding to the abnormal module and associate them. The weight attenuation factor is calculated based on the fluctuation intensity of the abnormal module; the greater the fluctuation intensity, the smaller the weight attenuation factor.

[0085] Next, a directed weighted graph is constructed based on the modified module transition path weight matrix, with the functional modules as nodes and the modified transition path weights as edge weights, to generate a module transition probability graph. For example, each functional module is abstracted as a node in the graph. Then, for each non-zero element in the modified module transition path weight matrix, a directed edge is added from the starting functional module node to the target functional module node, and the value of this element is used as the edge weight.

[0086] Next, the in-degree weight of each node in the module transition probability graph is calculated. The in-degree weight is a weighted fusion of the edge weights pointing to the node and the fluctuation intensity of the corresponding source node. The in-degree weight is a comprehensive consideration of the weights of the edges pointing to the node. The edge weights pointing to the node reflect the likelihood of other functional modules transferring to the node, while the fluctuation intensity of the corresponding source node reflects the source node's stability and user attention. This weighted fusion of these two factors allows for a more accurate assessment of node importance. For example, for each node in the graph, all edges pointing to the node are traversed. For each edge, its edge weight and the fluctuation intensity of the corresponding source node are obtained. A weight coefficient is assigned to each edge weight and fluctuation intensity. The edge weight is then multiplied by the edge weight coefficient, and the fluctuation intensity is multiplied by the fluctuation intensity coefficient. The two are then added together to determine the edge's contribution to the target node's in-degree weight. The contributions of all edges pointing to the node are then summed to obtain the node's in-degree weight.

[0087] Finally, the nodes are sorted in descending order according to the in-degree comprehensive weight to generate a functional module priority sequence, and the redundancy of the loop path nodes in the sequence is eliminated; among them, the sorting results of the functional modules at the preset priority in the functional module priority sequence simultaneously meet the core hub characteristics of high in-degree weight, low fluctuation anomaly and high-frequency access path.

[0088] Loop path nodes are nodes that form loops in the module transition probability graph. The preset priority is a pre-set priority range. Functional modules within this range should possess the core hub characteristics of high in-degree weight, low volatility, and high-frequency access paths. For example, the combined in-degree weights of all nodes in the module transition probability graph are extracted and sorted in descending order using a sorting algorithm (such as quick sort or merge sort). Based on the sorting results, a priority sequence of functional modules is obtained.

[0089] Step S440: extract the business preference tag set in the user portrait, perform semantic matching on the business preference tags and the functional module priority sequence, calculate the association strength between the recommended content and the target module, and generate the recommendation strategy adjustment coefficient.

[0090] The business preference tag set in the user portrait refers to a tag set that reflects the user's business preferences, extracted by analyzing and mining the user's historical operation data, behavior patterns, etc. These tags can include information such as the business areas, functional modules, product types, etc. that the user is interested in. Semantic matching refers to the semantic matching of business preference tags with modules in the functional module priority sequence to determine the correlation between them. The association strength between recommended content and the target module refers to the degree of correlation between the recommended content and the functional module, which can be calculated based on the results of semantic matching. The recommendation strategy adjustment coefficient is a coefficient used to adjust the recommendation strategy. It can be generated based on the association strength between the recommended content and the target module to optimize the push effect of the recommended content.

[0091] For example, a set of functional module identifiers exceeding a preset access frequency threshold in historical operation records is first extracted from the user profile, and then a set of business preference labels is generated in combination with the preference data actively annotated by the user. The preset access frequency threshold is a pre-set threshold used to filter out functional modules frequently accessed by users. The set of functional module identifiers refers to the set consisting of identifiers of functional modules exceeding the preset access frequency threshold. The business preference label set is a label set composed of the set of functional module identifiers and the preference data actively annotated by the user, which can more comprehensively reflect the user's business preferences.

[0092] Then, the module identifiers in the functional module priority sequence are semantically aligned with the business preference label set, and the context-related word vectors of the module description text and the preference label are extracted to generate a semantic association matrix. The module identifiers in the functional module priority sequence refer to the identifiers of the functional modules sorted by priority, which are used to identify each functional module. Semantic feature alignment refers to the semantic alignment of the module identifiers with the labels in the business preference label set to determine the correlation between them. The module description text refers to the text that describes the function, purpose, etc. of the functional module, and the preference label is the label in the business preference label set. The context-related word vector refers to the vector representation of the associated words extracted by performing context analysis on the module description text and the preference label. The semantic association matrix is ​​a matrix that records the degree of semantic association between the functional module and the business preference label.

[0093] For example, for each module identifier in the functional module priority sequence, the corresponding module description text is searched. Use natural language processing technology to process the module description text and the labels in the business preference label set to extract their semantic features. Align the semantic features of the module description text with the semantic features of the business preference label, and calculate the similarity between them. Use context analysis algorithms (such as dependency syntax analysis, co-occurrence analysis, etc.) to analyze the module description text and preference label to extract the context-related words between them. Convert these related words into vector representations to obtain context-related word vectors. The degree of semantic association between each functional module and each business preference label is used as an element of the matrix to form a semantic association matrix. The rows of the matrix represent functional modules, the columns represent business preference labels, and the values ​​of the elements represent the degree of semantic association between functional modules and business preference labels.

[0094] Afterwards, the matching weight value of each functional module and the business preference label is calculated based on the semantic association matrix, wherein the matching weight of the functional module at the preset priority is adjusted by the product of label coverage and semantic similarity. The matching weight value refers to the weight value used to measure the degree of matching between the functional module and the business preference label, which can be calculated based on the semantic association matrix. Label coverage refers to the proportion of the number of labels related to a functional module in the business preference label set to the total number of labels, which reflects the coverage of the functional module by the business preference label. Semantic similarity refers to the degree of semantic similarity between the functional module and the business preference label, which can be represented by the element value in the semantic association matrix. The functional module at the preset priority refers to the functional module that is within the pre-set priority range in the functional module priority sequence.

[0095] The matching weights are then normalized and combined with the ranking results of the functional module priority sequence to generate a correlation strength distribution between the recommended content and the target module. This correlation strength distribution reflects the recommendation compatibility of different priority modules along the user preference dimension. Normalization (e.g., linear normalization) converts the matching weights to a preset range (e.g., [0, 1]) for comparison and analysis. The correlation strength distribution refers to the distribution of the correlation strength between the recommended content and the target module across modules of different priority levels. It reflects the recommendation compatibility of different priority modules along the user preference dimension.

[0096] Next, recommendation trigger conditions are dynamically set based on the association strength distribution. This can include the following steps: recommendation content display frequency threshold, inter-module recommendation interval, and content presentation format priority, generating a recommendation strategy adjustment coefficient containing multi-dimensional control parameters. The association strength distribution reflects the degree of association between recommended content and modules of different priorities. Recommendation trigger conditions can be dynamically set based on this distribution. The recommendation content display frequency threshold refers to the maximum number of times a recommended content is displayed within a certain time period. It controls the frequency of recommended content and avoids over-recommendations to users. The inter-module recommendation interval refers to the time interval between recommendations in different functional modules. It ensures the rationality and consistency of recommended content. The content presentation format priority refers to the order of precedence of different content presentation formats (such as text, image, video, etc.). It selects the most appropriate content presentation format based on user preferences and operational scenarios. Multi-dimensional control parameters refer to a set of parameters including the recommendation content display frequency threshold, inter-module recommendation interval, and content presentation format priority. The recommendation strategy adjustment coefficient is generated based on these multi-dimensional control parameters and can be used to adjust the recommendation strategy and improve recommendation effectiveness.

[0097] Finally, the recommendation strategy adjustment coefficient is correlated and verified with real-time user operation behavior data. When the correlation strength between the recommended content and the current operation module falls below a preset threshold, a dynamic reallocation mechanism for the recommendation strategy adjustment coefficient is triggered. The generation process of the recommendation strategy adjustment coefficient synchronously correlates the dynamic update status of the functional module priority sequence with the incremental learning results of the user profile to ensure consistency between the recommendation strategy and the evolution of user behavior. Real-time user operation behavior data refers to the user's behavior data during the current operation, including information such as the functional modules accessed, the business data operated, and the buttons clicked. Correlation verification involves correlating the recommendation strategy adjustment coefficient with real-time user operation behavior data to verify its validity. The correlation strength between the recommended content and the current operation module refers to the degree of correlation between the recommended content and the functional module currently being operated. The preset threshold is a pre-set threshold used to determine whether the correlation strength is sufficient. The dynamic reallocation mechanism for the recommendation strategy adjustment coefficient recalculates and reassigns the recommendation strategy adjustment coefficient to adjust the recommendation strategy when the correlation strength between the recommended content and the current operation module falls below the preset threshold. The dynamic update state of the function module priority sequence refers to the state in which the function module priority sequence is updated as user behavior changes. The incremental learning result of the user portrait refers to the update result obtained by continuously learning new user behavior data.

[0098] Step S450: Perform three-dimensional tensor integration on the weight distribution of interface elements, the priority sequence of functional modules, and the adjustment coefficient of recommendation strategy, use orthogonal projection to eliminate cross-interference between dimensions, and generate an interaction process parameter matrix; wherein, the weight distribution of interface elements is used to drive interface layout rendering, the priority sequence of functional modules is used for navigation path optimization, and the recommendation strategy adjustment coefficient is used for dynamic content push strategy generation.

[0099] Three-dimensional tensor integration involves combining data from three different dimensions—the weight distribution of interface elements, the priority sequence of functional modules, and the adjustment coefficient for recommended strategies—into a single three-dimensional tensor. A three-dimensional tensor is a multidimensional array that can simultaneously represent data from three dimensions. For example, the weight distribution of interface elements, the priority sequence of functional modules, and the adjustment coefficient for recommended strategies can be used as the three dimensions of a three-dimensional tensor. For example, the weight distribution of interface elements can be used as the first dimension, the priority sequence of functional modules as the second dimension, and the adjustment coefficient for recommended strategies as the third dimension.

[0100] Step S500: calling the optimized interaction process parameter matrix to drive the business operation interface to reconstruct, generating an interaction interface adapted to the current user behavior pattern, and iteratively updating the timing analysis model through the incremental feedback mechanism within a preset verification cycle.

[0101] The optimized interaction process parameter matrix refers to the interaction process parameter matrix obtained after processing and optimization in the previous steps. It contains information such as the weight distribution of interface elements, the priority sequence of functional modules, and the adjustment coefficient of the recommendation strategy. Business operation interface reconstruction refers to the rearrangement and design of the business operation interface based on the interaction process parameter matrix to improve the interface's usability and user experience. An interaction interface adapted to current user behavior patterns refers to an interaction interface generated based on the user's current behavior patterns, which can better meet user needs and operating habits. The preset verification cycle is a pre-set time period used to verify and evaluate the generated interaction interface. The incremental feedback mechanism is a mechanism for updating models. It continuously collects user feedback information and incrementally updates the model to improve its performance and adaptability. The time series analysis model is a model used to analyze user behavior patterns and predict potential churn risks. It is iteratively updated through the incremental feedback mechanism to enable it to better adapt to changes in user behavior.

[0102] As an implementation method, the above-mentioned interaction process parameter matrix includes the weight distribution of interface elements, the priority sequence of functional modules, and the adjustment coefficient of the recommendation strategy. Based on this, in step S500, the optimized interaction process parameter matrix is ​​called to drive the reconstruction of the business operation interface, which may specifically include the following steps S510 to S550:

[0103] Step S510: parse the weight distribution of interface elements in the interaction process parameter matrix, extract the spatial coordinate set and visual rendering priority parameters of the area exceeding the weight threshold, and generate dynamic focus instructions for the core function control.

[0104] The weight threshold is a pre-set threshold used to filter out important interface elements. The spatial coordinate set of areas exceeding the weight threshold refers to the set of position coordinates of areas in the interface whose weight values ​​exceed the threshold. The visual rendering priority parameter refers to the parameter used to control the visual rendering order and effect of interface elements. Core function controls refer to controls that provide primary business functions. Dynamic focus instructions refer to instructions used to control the dynamic focus of core function controls. This makes core function controls more prominent in the interface and improves user operation efficiency.

[0105] The process for parsing the weight distribution of UI elements in the interaction process parameter matrix is ​​as follows: The UI element weight distribution data in the interaction process parameter matrix is ​​analyzed and processed to extract the weight value of each UI element. A weight threshold is set to filter out UI elements with weights exceeding the threshold. The spatial coordinates of the areas where these UI elements are located are extracted, along with their associated visual rendering priority parameters.

[0106] For example, based on the spatial coordinate set and visual rendering priority parameters of the area exceeding the weight threshold, a dynamic focus instruction for the core function control is generated. The dynamic focus instruction may include information such as the position, size, color, transparency, and other information of the core function control, as well as parameters such as the time and method of focusing. For example, the core function control can be set to automatically focus when the user enters the interface, or to focus when the user performs an operation. By generating dynamic focus instructions, the core function control can be made more prominent in the interface, attracting the user's attention and improving the user's operating efficiency.

[0107] Step S520: Calculate the access path weight of each module node in the navigation menu according to the sorting result of the functional module priority sequence, and generate the shortest path topology structure based on the weight gradient distribution.

[0108] The sorting result of the functional module priority sequence refers to the sequence obtained by sorting the functional modules from high to low priority. The access path weight of each module node in the navigation menu refers to the weight of the user's access path from the starting node of the navigation menu to each module node. It can reflect the difficulty and importance of the access path. The weight gradient distribution refers to the distribution of the access path weight in the navigation menu, which can reflect the changing trend of the access path weight. The shortest path topology structure refers to the topology structure composed of the shortest access paths from the starting node to each module node in the navigation menu. It can optimize the user's navigation path and improve the user's operating efficiency.

[0109] Specifically, first parse the sorting results of the functional module priority sequence, extract the in-degree weights of the module nodes and the historical statistical values ​​of the transfer frequency between adjacent modules, and generate the initial access path weights of each module node. The sorting result of the functional module priority sequence refers to the sequence obtained after sorting the functional modules from high to low priority. The in-degree weight of a module node refers to the sum of the weights of the edges pointing to the module node, which can reflect the importance of the module node. The historical statistical value of the transfer frequency between adjacent modules refers to the statistical value of the number of times a user transfers between adjacent modules, which can reflect the degree of association between modules. The initial access path weight refers to the access path weight of each module node calculated based on the in-degree weights of the module nodes and the historical statistical values ​​of the transfer frequency between adjacent modules. It provides a basis for subsequent path weight correction and shortest path topology structure generation.

[0110] Then, the initial access path weight is corrected according to the weight difference gradient of adjacent modules in the weight gradient distribution, wherein the module node path weight in the area exceeding the preset gradient change is suppressed from abnormal weight growth by the attenuation factor. The weight gradient distribution refers to the distribution of the access path weight in the navigation menu, which can be obtained by calculating the weight difference between adjacent module nodes. The weight difference gradient of adjacent modules refers to the degree of difference between the access path weights of adjacent module nodes, which can reflect the changing trend of the weight. The initial access path weight refers to the access path weight of each module node generated by the above steps. The preset gradient change is a pre-set threshold used to determine whether the weight difference between adjacent modules is too large. The attenuation factor is a coefficient less than 1, which is used to suppress the abnormal growth of the module node path weight that exceeds the preset gradient change area.

[0111] For example, the weight difference gradient of adjacent modules in the weight gradient distribution is first calculated and compared with the preset gradient change. If the weight difference gradient exceeds the preset gradient change, it indicates that the module node path weight in this area has increased abnormally. For module nodes that exceed the preset gradient change area, their path weight is corrected using an attenuation factor (which can be set according to actual conditions). In this way, the abnormal growth of module node path weight can be suppressed, making the access path weight more reasonable and stable.

[0112] Next, a directed weighted graph structure is constructed based on the revised access path weights, with module nodes as vertices and path weights as edge weights, to generate the initial shortest path topology. For example, each module node is abstracted as a vertex in the graph. Then, for each non-zero element in the revised access path weight matrix, a directed edge is added from the starting module node to the target module node in the graph, with the value of this element serving as the edge weight. In this way, a directed weighted graph structure is constructed, resulting in the initial shortest path topology.

[0113] Afterwards, redundant path verification is performed on the initial shortest path topology to identify jump nodes in the initial shortest path topology that violate the monotonicity of the weight gradient. The reconstructed path topology is obtained by inserting intermediate module nodes or adjusting the edge weight distribution. Redundant path verification refers to checking the paths in the initial shortest path topology and removing unnecessary paths to simplify the topology. Monotonicity of the weight gradient means that the access path weight should show a monotonically increasing or monotonically decreasing trend along the path. Jump nodes refer to nodes in the initial shortest path topology that violate the monotonicity of the weight gradient, which may lead to irrationality of the path. The reconstructed path topology refers to the path topology obtained after redundant path verification and node adjustment, which is more reasonable and efficient.

[0114] For example, all paths in the initial shortest path topology are traversed to check whether the access path weights on the paths satisfy the monotonicity of the weight gradient. If a path violates the monotonicity of the weight gradient, a jump node is identified. Jump nodes can be addressed by inserting intermediate module nodes or adjusting the edge weight distribution. This process results in a reconstructed path topology, improving the rationality and efficiency of the paths.

[0115] Afterwards, the reconstructed path topology is mapped to the guide sequence in the navigation menu, and the visual parameters of the guide sequence match the weight gradient distribution trend of each node on the reconstructed path topology. The reconstructed path topology refers to the path topology obtained after redundant path verification and node adjustment. It shows the shortest access path from the starting node of the navigation menu to each module node. The guide sequence in the navigation menu refers to the sequence used to guide users to perform operations in the navigation menu. It can indicate the user's operation path through visual effects (such as color, size, transparency, etc.). Visual parameters refer to parameters used to control the visual effects of the guide sequence, such as color, size, transparency, etc. The weight gradient distribution trend refers to the changing trend of the access path weight of each node on the reconstructed path topology.

[0116] For example, the position and layout of each module node in the navigation menu are first determined. Then, based on the reconstructed path topology, each module node is assigned a position in the navigation sequence. The visual parameters of the navigation sequence are set based on the weight gradient distribution trend of each node in the reconstructed path topology. In this way, the reconstructed path topology is mapped to the navigation sequence in the navigation menu. The visual parameters of the navigation sequence match the weight gradient distribution trend of each node in the reconstructed path topology, allowing users to more intuitively understand the navigation path and improve operational efficiency.

[0117] Finally, when a user switches modules, the path deviation metric is collected in real time. The access path weights are updated based on the current weight gradient distribution, and the updated results are synchronously fed back into the sorting logic of the functional module priority sequence. The access path weight calculation process uses the weight gradient distribution to dynamically balance the path weights between module nodes, ensuring consistency between the shortest path topology and the evolving user behavior patterns. The path deviation metric refers to the degree of deviation between the actual access path and the path in the shortest path topology when the user switches modules. It can be calculated by calculating the difference between the actual access path length and the shortest path length, or by calculating the degree of difference between nodes on the actual access path and those on the shortest path. The current weight gradient distribution refers to the gradient distribution of the access path weights for each module node in the navigation menu when the user switches modules. Updating the access path weights adjusts the access path weights for each module node in the navigation menu based on the path deviation metric and the current weight gradient distribution. The sorting logic of the functional module priority sequence ranks functional modules based on their importance and user behavior.

[0118] For example, when a user switches modules, the user's actual access path is recorded. The actual access path is compared with the path in the shortest path topology, and a path deviation index is calculated. Based on the path deviation index and the current weight gradient distribution, the access path weights of each module node in the navigation menu are adjusted. The updated access path weights are fed back to the sorting logic of the functional module priority sequence, which reorders the functional modules based on the new access path weights. The access path weight calculation process achieves a dynamic balance of path weights between module nodes through weight gradient distribution, ensuring consistency between the shortest path topology and the evolution of user behavior patterns.

[0119] Step S530: Input the recommendation strategy adjustment coefficient into the pre-trained recommendation rule generator, match the user's real-time operation scenario with historical preference characteristics, and output a personalized recommendation content sequence associated with the current functional module.

[0120] The recommendation strategy adjustment coefficient is a coefficient used to adjust the recommendation strategy. It includes multi-dimensional control parameters such as the recommended content display frequency threshold, the recommendation interval period between modules, and the content presentation form priority. The pre-trained recommendation rule generator is a trained model that can generate recommendation rules based on the input recommendation strategy adjustment coefficient and user-related information. The user's real-time operation scenario refers to the scenario information of the user during the current operation process, including the functional module the user is operating, the time of operation, the operating environment, etc. The historical preference feature refers to the preference information shown by the user during the historical operation process. It can be represented by the business preference label set in the user portrait, etc. The personalized recommendation content sequence refers to the recommended content sequence associated with the current functional module generated based on the user's real-time operation scenario and historical preference features. It can meet the user's personalized needs.

[0121] Step S540: Integrate the dynamic focus instruction, the shortest path topology structure and the personalized recommended content sequence to generate an interface layout rendering configuration file, which includes control positioning coordinates, path guidance marks and spatiotemporal distribution constraints of recommended content plug-ins.

[0122] Dynamic focus instructions refer to instructions used to control the dynamic focus of core function controls. It can make core function controls more prominent in the interface and improve user operation efficiency. The shortest path topology structure refers to the topology structure composed of the shortest access paths from the starting node to each module node in the navigation menu. It can optimize the user's navigation path. Personalized recommendation content sequence refers to the recommended content sequence associated with the current function module generated based on the user's real-time operation scenario and historical preference characteristics. It can meet the user's personalized needs. The interface layout rendering configuration file is a file used to configure interface layout rendering. It contains information such as control positioning coordinates, path guidance marks, and spatiotemporal distribution constraints of recommended content plug-ins, which is used to guide the rendering and display of the interface.

[0123] For example, the position, size, color, and transparency of the core functional controls in the dynamic focus instruction are organized. Next, navigation path information from the shortest path topology, such as node locations and edge weights, is extracted. Next, recommended content information from the personalized recommendation sequence, such as the title, description, and image of the recommended content, is organized. This information is integrated to form a dataset containing control positioning coordinates, path guidance markers, and the spatiotemporal distribution constraints of the recommended content plugins. Based on this integrated dataset, an interface layout rendering configuration file is generated. This configuration file includes control positioning coordinates, which determine the position of core functional controls and other interface elements within the interface; path guidance markers, which indicate the user's navigation path within the navigation menu; and spatiotemporal distribution constraints of the recommended content plugins, which control the display time and location of recommended content within the interface. By generating the interface layout rendering configuration file, the dynamic focus instruction, the shortest path topology, and the personalized recommendation sequence can be combined to optimize the interface layout and display personalized recommended content.

[0124] Step S550: After the target client loads the interface layout rendering configuration file, the user's operation response delay for the core functional controls, the path switching efficiency, and the recommended content adoption rate indicators are collected in real time, and an interface optimization feedback signal is generated and transmitted back to the update queue of the interaction process parameter matrix; wherein, the generation process of the interface layout rendering configuration file synchronously associates the dynamic change trend of the interface element weight distribution with the real-time sorting status of the functional module priority sequence, to ensure the consistency of the interface reconstruction strategy and the evolution of the user behavior pattern.

[0125] The target client refers to the client device running the business operation interface, such as a computer or mobile phone. The interface layout rendering configuration file is a file used to configure the interface layout rendering. It contains information such as control positioning coordinates, path guidance markers, and the spatiotemporal distribution constraints of recommended content plug-ins. The user response latency for core function controls refers to the time it takes for a user to trigger a core function control and for the control to respond. It reflects the responsiveness and usability of the core function control. Path switching efficiency refers to the efficiency of user module switching in the navigation menu. It can be measured by indicators such as the time it takes for users to complete module switching operations and path deviation. The recommended content adoption rate indicator refers to the user's acceptance of recommended content. It can be calculated as the ratio of the number of times a user clicks on and uses a recommended content to the number of times the recommended content is displayed. The interface optimization feedback signal is a signal used to provide feedback on interface optimization. It contains information such as the user response latency for core function controls, path switching efficiency, and recommended content adoption rate. The interaction process parameter matrix update queue is a queue used to store interaction process parameter matrix update instructions. It ensures that updates to the interaction process parameter matrix are performed in order.

[0126] For example, first parse the abnormal interface element loading delay event in the user operation response delay, identify the control coordinate set and corresponding functional module identifier that exceeds the preset delay trigger threshold, and generate the interface rendering performance bottleneck area mark.

[0127] User operation response delay refers to the time it takes for the user to trigger an interface element operation and for the element to respond. Interface element loading delay abnormal event refers to the abnormal event in which the interface element loading time is too long during the user operation response delay. The preset delay trigger threshold is a pre-set threshold used to determine whether the interface element loading delay is abnormal. The control coordinate set refers to the set consisting of the position coordinates of the interface controls that exceed the preset delay trigger threshold. The corresponding function module identifier refers to the identifier of the function module associated with the interface control that exceeds the preset delay trigger threshold. The interface rendering performance bottleneck area marker is a marker used to mark the rendering performance bottleneck area in the interface, which can provide a basis for subsequent interface optimization.

[0128] Analyze and process the user operation response delay data, and extract the information about the interface element loading delay. Compare the loading delay time of each interface element with the preset delay trigger threshold. When the loading delay time exceeds the preset delay trigger threshold, identify the loading event of the interface element as an interface element loading delay exception event. For each interface element loading delay exception event, extract the control coordinates and corresponding function module identifier of the interface element. Collect these control coordinates and corresponding function module identifiers separately to form a control coordinate set and a corresponding function module identifier set. Generate an interface rendering performance bottleneck area marker based on the control coordinate set and the corresponding function module identifier set. The coordinate points in the control coordinate set can be marked on the interface and annotated with the corresponding function module identifier.

[0129] Then, the path switching efficiency index is extracted from the path jump time and deviation statistics. Combined with the node distribution characteristics of the shortest path topology structure, the matching deviation between the path guidance strategy and the user's actual operation habits is calculated.

[0130] The path switching efficiency index refers to an indicator used to measure the efficiency of users' module switching operations in the navigation menu. It includes statistical values ​​such as path jump time and number of deviations. Path jump time refers to the time it takes for a user to switch from one functional module to another, and the number of deviations refers to the number of times a user deviates from the shortest path in the switching path. The node distribution characteristics of the shortest path topology structure refer to the location, connection relationship and other characteristics of each node in the shortest path topology structure. The path guidance strategy refers to the strategy used to guide users to perform module switching operations in the navigation menu. It can be implemented through path guidance marks and other methods. The matching deviation degree refers to the degree of difference between the path guidance strategy and the user's actual operating habits. It can reflect the effectiveness of the path guidance strategy.

[0131] For example, we analyze and process path switching efficiency indicator data to extract statistics on path jump duration and deviations. We can record the start and end times of each module switching operation to calculate path jump duration. We also record the number of times a user deviates from the shortest path during a switch, generating a deviation statistic. We analyze the node distribution characteristics of the shortest path topology to understand the direction of the shortest path and the node connectivity. We then compare the user's actual operation path with the shortest path topology to calculate the degree of mismatch between the path guidance strategy and the user's actual operating habits. We can use metrics such as path similarity and path deviation distance to measure this mismatch.

[0132] Afterwards, the recommendation content adoption rate indicator is decomposed into the ratio of the number of exposures of recommended content associated with different functional modules to the number of active user interactions, generating a module-level recommendation utility evaluation matrix. The recommendation content adoption rate indicator refers to the user's acceptance of the recommended content. It can be calculated by the ratio of the number of times users click on and use the recommended content to the number of times the recommended content is displayed. Recommended content associated with different functional modules refers to the recommended content associated with each functional module. The number of exposures refers to the number of times the recommended content is displayed on the interface, and the number of active user interactions refers to the number of times users actively click on, use, or perform other actions on the recommended content. The module-level recommendation utility evaluation matrix is ​​a matrix that records the recommendation utility evaluation results of recommended content associated with different functional modules. The rows of the matrix represent functional modules, the columns represent recommended content, and the values ​​of the elements represent the recommendation utility evaluation values ​​of the recommended content associated with the functional modules.

[0133] For example, we analyze and process the recommended content adoption rate indicator data and classify the recommended content by functional module. For each recommended content associated with a functional module, we count its exposure times and the number of active user interactions. We divide the number of active user interactions by the number of exposures to obtain the recommended content adoption rate for the recommended content associated with that functional module. The recommended content adoption rates for each functional module's recommended content are used as elements in a matrix to form a module-level recommendation utility evaluation matrix. The rows of the matrix represent functional modules, the columns represent recommended content, and the values ​​of the elements represent the recommendation utility evaluation values ​​of the recommended content associated with that functional module.

[0134] Afterwards, a multi-dimensional correlation analysis is conducted on the interface rendering performance bottleneck area marking, matching deviation and recommendation utility evaluation matrix to determine the collaborative optimization direction of the interface element weight distribution, functional module priority sequence and recommendation strategy adjustment coefficient in the interaction process parameter matrix.

[0135] Multidimensional correlation analysis refers to the correlation and analysis of data from multiple dimensions to discover the relationships and patterns between the data. In this step, correlation analysis is performed on the data of three dimensions: interface rendering performance bottleneck area marking, matching deviation, and recommendation utility evaluation matrix. The interface rendering performance bottleneck area marking reflects the area with poor rendering performance in the interface, the matching deviation reflects the degree of difference between the path guidance strategy and the user's actual operating habits, and the recommendation utility evaluation matrix reflects the recommendation utility of the recommended content associated with different functional modules. The interaction process parameter matrix contains information such as the weight distribution of interface elements, the priority sequence of functional modules, and the recommendation strategy adjustment coefficient. The collaborative optimization direction refers to the direction of optimizing these parameters to improve the usability of the interface and user experience.

[0136] For example, the data from the interface rendering performance bottleneck area markers, matching deviation, and recommendation utility evaluation matrix are collated and preprocessed to make them comparable. Then, association analysis methods (such as correlation analysis and cluster analysis) are used to analyze these data and identify their correlations. For example, the relationship between the interface rendering performance bottleneck area and matching deviation, and the relationship between matching deviation and the recommendation utility evaluation matrix can be analyzed. Based on the results of the multidimensional association analysis, the optimization direction for the interface element weight distribution, functional module priority sequence, and recommendation strategy adjustment coefficient in the interaction process parameter matrix is ​​determined. For example, if the interface rendering performance bottleneck area is found to be associated with excessive weights of certain interface elements, the weights of these elements can be appropriately reduced; if a large matching deviation is found to be associated with an unreasonable functional module priority sequence, the functional module priority sequence can be adjusted; if the low recommendation utility of certain recommended content in the recommendation utility evaluation matrix is ​​found to be associated with an unreasonable recommendation strategy adjustment coefficient, the recommendation strategy adjustment coefficient can be adjusted. By collaboratively optimizing these parameters, the interface usability and user experience can be improved.

[0137] Then, based on the collaborative optimization direction, parameter matrix update instructions are generated. The instructions include the interface element weight attenuation coefficient, the module priority sorting correction amplitude, and the recommended strategy trigger threshold adjustment amount.

[0138] Collaborative optimization refers to the optimization direction of the interface element weight distribution, functional module priority sequence, and recommendation strategy adjustment coefficient in the interaction process parameter matrix, determined through multi-dimensional correlation analysis of interface rendering performance bottleneck area marking, matching deviation, and recommendation utility evaluation matrix. The parameter matrix update instruction is used to update the interaction process parameter matrix. It contains specific information for adjusting the interface element weight distribution, functional module priority sequence, and recommendation strategy adjustment coefficient. The interface element weight decay coefficient is a coefficient used to adjust the interface element weight. It can proportionally decay the weight of the interface element to optimize the interface layout. The module priority ranking correction amplitude refers to the amplitude of the correction to the functional module priority sequence. It can change the priority ranking of functional modules to improve the rationality of the navigation path. The recommendation strategy trigger threshold adjustment amount refers to the amount of adjustment to the recommendation strategy trigger threshold. It can change the display conditions of recommended content to improve the relevance and effectiveness of the recommended content.

[0139] For example, based on the collaborative optimization direction, the adjustment direction and magnitude of the interface element weight distribution, functional module priority sequence, and recommendation strategy adjustment coefficient are determined. Regarding the interface element weight distribution, based on the identification of interface rendering performance bottleneck areas and the analysis results of the matching deviation degree, the interface elements requiring weight adjustment are identified, and the corresponding interface element weight attenuation coefficients are calculated. For example, if an interface element is located in an area that is a rendering performance bottleneck or does not align with user operating habits, the element's weight can be appropriately reduced by setting an interface element weight attenuation coefficient less than 1. Regarding the functional module priority sequence, based on the analysis results of the matching deviation degree and the recommendation utility evaluation matrix, the functional modules requiring priority ranking adjustments are identified, and the module priority ranking adjustment magnitude is calculated. For example, if a functional module has low recommendation utility or users frequently deviate from its path during operation, the module's priority can be appropriately lowered by setting an appropriate module priority ranking adjustment magnitude. Regarding the recommendation strategy adjustment coefficient, based on the analysis results of the recommendation utility evaluation matrix, the recommendation strategy trigger threshold that requires adjustment is determined, and the adjustment magnitude for the recommendation strategy trigger threshold is calculated. For example, if the adoption rate of recommended content associated with a functional module is low, the module's recommendation strategy trigger threshold can be appropriately lowered by setting an appropriate recommendation strategy trigger threshold adjustment amount. The interface element weight decay coefficient, module priority ranking correction range, and recommendation strategy trigger threshold adjustment amount are integrated together to form a parameter matrix update instruction.

[0140] Then, the parameter matrix update instructions are stored in the update queue in order of priority. When it is detected that the dynamic change amplitude of the interface element weight distribution exceeds a preset threshold, the real-time incremental update of the parameter matrix is ​​triggered.

[0141] The update queue is a queue used to store parameter matrix update instructions, which ensures that update instructions are processed in order. Storing parameter matrix update instructions by priority means sorting the update instructions according to their importance and urgency and storing them in the update queue. This ensures that important and urgent update instructions are processed first. The dynamic change range of the interface element weight distribution refers to the degree of change in the interface element weight distribution over a period of time. It can be measured by calculating statistics such as the rate of change and standard deviation of the interface element weights. The preset threshold is a pre-set threshold used to determine whether the dynamic change range of the interface element weight distribution is too large. Real-time incremental update of the parameter matrix means that when it is detected that the dynamic change range of the interface element weight distribution exceeds the preset threshold, the interaction process parameter matrix is ​​immediately updated to ensure the usability of the interface and user experience.

[0142] For example, a priority is assigned to each parameter matrix update instruction. The priority can be determined based on the impact of the update instruction on interface performance, user experience, and other aspects. For example, adjustment instructions for the bottleneck area of ​​interface rendering performance can be assigned a higher priority, while adjustment instructions for some minor interface elements can be assigned a lower priority. Then, the update instructions are stored in the update queue in descending order of priority. When a new update instruction arrives, it is inserted into the appropriate position in the update queue to ensure the priority order of the queue.

[0143] Monitor the dynamic changes in the weight distribution of interface elements in real time. Statistics such as the rate of change and standard deviation of the weights of interface elements can be calculated regularly and compared with the preset thresholds. When it is detected that the dynamic change exceeds the preset threshold, it means that the weight distribution of interface elements has changed significantly, which may affect the usability of the interface and the user experience. At this time, the update instruction with the highest priority is taken out from the update queue, and the interaction process parameter matrix is ​​incrementally updated in real time. During the update process, the interface element weight distribution, functional module priority sequence and recommendation strategy adjustment coefficient are adjusted accordingly based on the interface element weight attenuation coefficient, module priority sorting correction amplitude and recommendation strategy trigger threshold adjustment amount in the update instruction. Through real-time incremental updates, it is possible to respond to changes in the weight distribution of interface elements in a timely manner to ensure the stability of the interface and user experience.

[0144] Finally, after performing the incremental update, user operation data is collected again to generate a new round of interface optimization feedback signals, forming a closed-loop optimization link; among them, the generation logic of the interface optimization feedback signal synchronously associates the version iteration record of the interface layout rendering configuration file with the evolution trajectory of the user behavior pattern, ensuring the dynamic adaptation of the parameter matrix update strategy and the system performance indicators.

[0145] For example, the real-time changing trends of the weight distribution of interface elements and the priority sequence of functional modules after incremental updates can be monitored, and the spatial coordinates and associated module identifiers of abnormal fluctuation nodes in the weight gradient distribution can be extracted. The functional module priority sequence is a sequence in which each functional module is sorted according to preset rules, reflecting the importance and usage priority of the functional modules. The weight gradient distribution describes the spatial changes in the weights of interface elements. Abnormal fluctuation nodes refer to nodes in the weight gradient distribution that deviate significantly from the normal change trend. Spatial coordinates are used to locate specific locations on the interface, and associated module identifiers are used to uniquely identify functional modules associated with abnormal fluctuation nodes.

[0146] To monitor the real-time trends in the weight distribution of interface elements and the priority sequence of functional modules, real-time data collection and analysis can be employed. For example, sensors or logging tools can be used to collect real-time data on user interactions with the interface, including clicks, swipes, and other actions. This data can then be processed using data analysis algorithms, such as time series analysis, to determine the real-time changes in the weight distribution of interface elements and the priority sequence of functional modules.

[0147] When extracting the spatial coordinates and associated module identifiers of abnormal fluctuation nodes in the weight gradient distribution, a threshold judgment method can be used. First, calculate the mean and standard deviation of the weight gradient distribution, and then set a threshold. When the weight gradient value of a node exceeds the threshold, it is judged as an abnormal fluctuation node. Then, the spatial coordinates and associated module identifiers of the abnormal fluctuation nodes are obtained through the interface layout information and the functional module association information. For example, assuming that the weight gradient value of an interface element is 10, and the mean value of the weight gradient distribution is 2, the standard deviation is 1, and the threshold is set to 3 times the standard deviation, that is, 6, then the node is judged to be an abnormal fluctuation node. Through the interface layout file, the spatial coordinates of the node can be obtained as (x, y), and then through the functional module association table, its associated module identifier can be determined to be M1.

[0148] Then, the user's operation trajectory data in the new version of the interface is captured, which may include the correction amount of the click frequency distribution of the core function controls, the improvement of the path switching efficiency and the change in the adoption rate of recommended content.

[0149] The correction amount of click frequency distribution refers to the change in the click frequency distribution of the core functional control after the incremental update relative to before the update. The improvement in path switching efficiency refers to the degree of improvement in the efficiency of users switching functional modules in the new version interface relative to the old version. The change in the adoption rate of recommended content refers to the change in the adoption rate of recommended content by users after the incremental update relative to before the update. To calculate the correction amount of click frequency distribution of core functional controls, you can first count the click frequency distribution of the core functional controls before and after the update, and then calculate the difference between the two. For example, the click frequency of the core functional control C1 before the update was 100 times, and after the update it was 120 times, then the correction amount of click frequency distribution is 20 times.

[0150] Then, the operation trajectory data is compared with the historical behavior pattern baseline to identify the behavior pattern offset caused by the parameter matrix update and generate a heat map of the temporal and spatial distribution of the offset.

[0151] The historical behavior pattern baseline refers to a relatively stable pattern of user behavior before the parameter matrix is ​​updated. It serves as a reference standard for measuring changes in user behavior patterns after the parameter matrix is ​​updated. The behavior pattern offset refers to the degree to which user behavior deviates from the historical behavior pattern baseline after the parameter matrix is ​​updated. The offset spatiotemporal distribution heat map is a visual representation that uses color depth to represent the distribution of behavior pattern offsets at different spatial locations and time points.

[0152] When comparing operation trajectory data with historical behavioral pattern baselines, similarity calculation methods can be used. For example, the Dynamic Time Warping (DTW) algorithm can be used to calculate the similarity between the operation trajectory data and the historical behavioral pattern baseline. To identify behavioral pattern deviations caused by parameter matrix updates, causal analysis methods can be used, such as Granger causality tests, to determine whether parameter matrix updates are the cause of behavioral pattern deviations.

[0153] Afterwards, the abnormal areas in the offset spatiotemporal distribution heat map that exceed the preset confidence threshold are extracted, and the rendering parameter adjustment records of the corresponding interface elements and the functional module priority correction history are associated to generate a set of parameter adjustment influencing factors. The abnormal area refers to the area in the offset spatiotemporal distribution heat map where the behavioral pattern offset exceeds the preset confidence threshold. The rendering parameter adjustment record refers to the record of adjusting the rendering parameters of the interface elements during the interface reconstruction process, including the adjustment of parameters such as color, size, and transparency. When extracting the abnormal areas in the offset spatiotemporal distribution heat map that exceed the preset confidence threshold, the threshold screening method can be used. The rendering parameter adjustment records of the corresponding interface elements and the functional module priority correction history can be associated by establishing an association table. In the association table, the rendering parameter adjustment records and the functional module priority correction history of the interface element corresponding to each abnormal area are recorded. The weighted summation method can be used to generate the parameter adjustment influencing factor set.

[0154] Then, the optimization direction correction coefficient of the next round of parameter matrix update instructions is calculated based on the influencing factor set. The correction coefficient is used to suppress unexpected behavior pattern deviations and enhance the target optimization effect. The influencing factor set is generated by the above steps, reflecting the degree of influence of parameter adjustment on user behavior patterns. The optimization direction correction coefficient is a coefficient used to adjust the optimization direction of the next round of parameter matrix update instructions, which can be calculated based on the influencing factor set. Unexpected behavior pattern deviation refers to the changes in user behavior patterns that are not in line with expectations after the parameter matrix is ​​updated. The target optimization effect refers to the optimization goal that is expected to be achieved when updating the parameter matrix, such as improving user operation efficiency, increasing the adoption rate of recommended content, etc.

[0155] Regression analysis can be used to calculate the optimization direction correction coefficient for the next round of parameter matrix update instructions based on the influencing factor set. A regression model is established with the influencing factor set as the independent variable and the unexpected behavior pattern offset as the dependent variable. The regression coefficient of the regression model is used as the optimization direction correction coefficient. When the correction coefficient is positive, the parameter matrix update instructions need to be adjusted in a certain direction to reduce the unexpected behavior pattern offset; when the correction coefficient is negative, the parameter matrix update instructions need to be adjusted in the opposite direction.

[0156] Afterwards, a dynamic learning rate adjustment mechanism is introduced into the closed-loop optimization link to dynamically adjust the parameter matrix update amplitude and frequency based on the execution effect feedback of historical update instructions.

[0157] A closed-loop optimization process involves collecting user operation data, generating interface optimization feedback signals, updating the parameter matrix, and then collecting user operation data again. This process continuously optimizes the interface. A dynamic learning rate adjustment mechanism is introduced into this closed-loop optimization process, using adaptive learning rate algorithms such as Adagrad, Adadelta, and Adam.

[0158] Finally, when it is detected that the user behavior pattern offset is lower than the preset stability threshold after S consecutive updates, a conservative update strategy is switched to, and local parameter correction is only performed on the modules where the interface rendering performance bottleneck area and the path matching deviation exceed the limit, S≥3; among them, the operating status of the closed-loop optimization link is synchronized in real time to the incremental training process of the timing analysis model to ensure the co-evolution consistency of user behavior pattern prediction and interface optimization strategy.

[0159] The preset stability threshold is a pre-set value used to determine whether the user behavior pattern has reached a stable state. The conservative update strategy refers to the use of a more conservative update method after the user behavior pattern reaches a stable state, and only performs local parameter corrections on necessary areas. The interface rendering performance bottleneck area refers to the area where performance bottlenecks occur during the interface rendering process, such as areas with long loading times, freezes, and other problems. When detecting whether the user behavior pattern offset is lower than the preset stability threshold after S consecutive updates, the user behavior pattern offset after each update can be recorded and compared. When the user behavior pattern offset after S consecutive updates is lower than the preset stability threshold, it means that the user behavior pattern has reached a stable state. After switching to the conservative update strategy, local parameter corrections are only performed on modules where the interface rendering performance bottleneck area and the path matching deviation exceed the limit.

[0160] In step S500, the timing analysis model is iteratively updated through an incremental feedback mechanism within a preset verification cycle, which may specifically include the following steps S560 to S5100:

[0161] Step S560: After the interactive interface is deployed, continuously collect user behavior response data and business conversion rate indicators.

[0162] User behavior response data refers to the series of behavioral data generated by users when interacting with the interactive interface, including clicks, swipes, inputs, and other operation data. It reflects users' operating habits and behavioral patterns. The business conversion rate indicator refers to the ratio of the number of users who achieve business goals to the total number of users during business operations. It reflects the conversion effect of the business. Data collection tools such as logging tools and tracking point statistics tools can be used to collect user behavior response data and business conversion rate indicators.

[0163] Step S570: performing a difference comparison between the behavior response data and the historical behavior trajectory feature set to identify the behavior pattern offset after interface optimization.

[0164] The historical behavior trajectory feature set refers to a set of features formed by user operation behaviors before interface optimization. It can be used as a reference standard to measure changes in user behavior patterns after interface optimization. The behavior pattern offset refers to the degree to which user operation behaviors deviate from the historical behavior trajectory feature set after interface optimization.

[0165] When comparing behavioral response data with historical behavioral trajectory feature sets, feature matching can be used. The behavioral response data is extracted as feature vectors, which are then matched with feature vectors in the historical behavioral trajectory feature set. Methods such as cosine similarity and Euclidean distance can be used to measure the similarity between feature vectors. Low similarity indicates a significant shift in the behavioral pattern.

[0166] Step S580: Build a model based on the behavior pattern offset to update the gradient matrix, and use the elastic weight solidification strategy to ensure that the learned key features are not covered.

[0167] The model update gradient matrix is ​​used to update the parameters of the time series analysis model. It can be constructed based on the behavioral pattern offset. The elastic weight hardening strategy is a strategy used to protect learned key features from being overwritten by new data. It penalizes the offset of key parameters by introducing elastic constraints into the loss function.

[0168] The model update gradient matrix is ​​constructed based on the behavioral pattern offset, and a gradient descent algorithm can be used. The behavioral pattern offset is used as the error signal, and the gradient of the model parameters is calculated through the back propagation algorithm, and then the model parameters are updated according to the gradient.

[0169] The elastic weight solidification strategy is used to maintain the learned key features from being overwritten. Specifically, it can include:

[0170] First, calculate the importance score matrix for each layer of the time series analysis model's parameters during historical training. This score matrix is ​​generated based on the cumulative values ​​of the parameter gradient changes. The importance score matrix is ​​used to measure the importance of each layer of the time series analysis model's parameters during historical training. It can be generated using the cumulative values ​​of the parameter gradient changes. The parameter gradient change refers to the magnitude of the parameter gradient during each training session. To calculate the importance score matrix for each layer of the time series analysis model's parameters during historical training, record and accumulate the parameter gradient values ​​during each training session.

[0171] Subsequently, during the incremental training phase, a flexible constraint term is introduced for each parameter. The strength of the constraint term is positively correlated with the corresponding parameter's importance score. A flexible constraint term is a constraint introduced into the loss function that penalizes shifts in key parameters. The strength of the constraint term, which refers to the degree to which the flexible constraint term penalizes parameter shifts, is positively correlated with the corresponding parameter's importance score: the higher the parameter's importance score, the stronger the constraint term.

[0172] Introducing elastic constraints for each parameter during the incremental training phase can add an extra term to the loss function.

[0173] Then, we add a flexible constraint regularization term to the loss function to penalize shifts in key parameters caused by new training data. This term is added to the loss function to penalize shifts in key parameters caused by new training data. By constraining the range of variation of key parameters, it protects already learned key features from being overwritten. Adding a flexible constraint regularization term to the loss function adds a regularization term to the original loss function.

[0174] Next, the optimizer's learning rate allocation strategy is dynamically adjusted, using a decaying learning rate for the first parameter and an adaptive learning rate for the second parameter, where the importance of the first parameter is greater than that of the second. An optimizer refers to an algorithm used to update model parameters, such as stochastic gradient descent (SGD) and Adam. A learning rate allocation strategy is a strategy that assigns different learning rates to different parameters. A decaying learning rate is a strategy that gradually reduces the learning rate as training progresses. An adaptive learning rate is a strategy that automatically adjusts the learning rate based on the gradient changes of the parameters. The first parameter is a parameter of higher importance, and the second parameter is a parameter of lower importance. Dynamic adjustment of the optimizer's learning rate allocation strategy can be performed based on the importance scores of the parameters. For the more important first parameter, a decaying learning rate is used to avoid over-updating and overwriting key features. For the less important second parameter, an adaptive learning rate is used to converge to the optimal solution more quickly.

[0175] Finally, after each training cycle, the importance score matrix is ​​updated and the strength coefficients of the elastic constraints are recalculated. After each training cycle, the importance scores of the parameters change due to updates to the model parameters. Therefore, the importance score matrix needs to be updated and the strength coefficients of the elastic constraints need to be recalculated.

[0176] Step S590: gradually adjust the parameters of the time series analysis model using a small batch incremental training method, and calculate the model prediction stability index through the online validation set after each update.

[0177] Small-batch incremental training divides the training data into several small batches, using only one small batch at a time to update the model parameters. An online validation set is a dataset used to verify model performance in real time during training. Model prediction stability metrics, such as mean squared error (MSE) and mean absolute error (MAE), are used to measure the stability of model predictions. Using small-batch incremental training to gradually adjust the parameters of a time series analysis model can reduce computational effort and improve training efficiency. After each update, the model prediction stability metrics are calculated using the online validation set to provide a timely understanding of changes in model performance.

[0178] Step S5100: When the fluctuation range of the stability indicator for K consecutive verification cycles is lower than the preset threshold, terminate the incremental update process and lock the current model version, K≥3.

[0179] The preset threshold is a pre-set value used to determine whether the model's stability indicator fluctuation range is within an acceptable range. K consecutive validation cycles refers to the process of performing K consecutive validations. The stability indicator fluctuation range refers to the difference between the maximum and minimum values ​​of the model's predicted stability indicator over K consecutive validation cycles.

[0180] When the stability indicator fluctuation range for K consecutive validation cycles is below the preset threshold, it indicates that the model performance has stabilized. At this time, the incremental update process can be terminated and the current model version can be locked. This ensures the stability and reliability of the model and avoids excessive updates that may cause performance degradation.

[0181] See also Figure 2 , Figure 2 A structural diagram of a computer system provided in an embodiment of the present invention, the computer system at least includes a processor 101, a communication interface 102 and a memory 103. The processor 101, the communication interface 102 and the memory 103 can be connected via a bus or other means. The processor 101 can parse various instructions in the computer system and process various data of the computer system. The communication interface 102 can optionally include a standard wired interface, a wireless interface (such as WI-FI, a mobile communication interface, etc.), which can be used to send and receive data under the control of the processor 101; the communication interface 102 can also be used for the transmission and interaction of data within the computer system. The memory 103 (Memory) is a memory device in the computer system for storing programs and data. The memory 103 provides a storage space, which stores the operating system of the computer system. The processor 101 executes the user behavior data mining method for digital enterprise management provided in the above embodiment of the present invention by running the computer program in the memory 103.

Claims

1. A user behavior data mining method for digital enterprise management, characterized in that: The method comprises: Collect multi-dimensional behavioral data of target users on the business operation interface; Performing multimodal data analysis on the multi-dimensional behavioral data to generate a behavioral trajectory feature set with temporal correlation, wherein different modal data are timestamp synchronized and semantically associated through a cross-modal alignment strategy; Training an adaptive time series analysis model based on the behavioral trajectory feature set, wherein the time series analysis model captures the long-term and short-term dependencies in the user behavior pattern through a dynamic window partitioning strategy and generates a potential churn risk prediction indicator; Constructing an interactive process parameter matrix based on the potential loss risk prediction indicators; The optimized interaction process parameter matrix is ​​called to drive the business operation interface for reconstruction, generate an interaction interface adapted to the current user behavior pattern, and iteratively update the timing analysis model through an incremental feedback mechanism within a preset verification cycle.

2. The method according to claim 1, characterized in that The multi-dimensional behavior data includes interface click trajectories, operation duration distribution, function module switching paths, and associated business data request records. The multi-modal data analysis of the multi-dimensional behavior data generates a behavior trajectory feature set with temporal correlation, including: Extracting the spatial coordinate sequence from the interface click trajectory, and calculating the interface area attention heat map based on the coordinate distribution density; Analyze the dwell interval characteristics in the operation duration distribution to identify operation clusters and abnormal stagnation nodes consisting of continuous operation events exceeding a frequency threshold. The high-frequency operation cluster is composed of continuous events whose operation frequency exceeds a preset threshold within a sliding window. Constructing a transition probability matrix of the functional module switching path, calculating the state transition weights between the functional modules, and generating a path prediction sequence based on a hidden Markov model; Mapping the associated business data request record into a request type encoding vector, and constructing a request load evaluation index based on the request response delay and the result data volume; Multimodal feature fusion is performed on the attention heat map, dwell interval features, transition probability matrix and request load evaluation index, and a time axis alignment strategy is adopted to generate a temporal embedding vector containing global context information as the behavior trajectory feature set.

3. The method according to claim 2, characterized in that The step of training the adaptive time series analysis model based on the behavior trajectory feature set includes: Inputting the temporal embedding vector into a bidirectional long short-term memory network, capturing the forward and reverse temporal dependencies through a gating mechanism, and generating an initial hidden state sequence; Connecting a multi-head self-attention mechanism layer to the back end of the bidirectional long short-term memory network, calculating the association weights between features at different time steps, and performing attention-weighted aggregation on the initial hidden state sequence to obtain an aggregated feature sequence; A dynamic convolution kernel group is used to extract multi-scale features from the aggregated feature sequence to obtain multi-scale features. The dynamic convolution kernel group adaptively adjusts the convolution kernel size and dilation coefficient according to the current input feature dimension. Performing a residual connection between the multi-scale features and the original time series embedding vector, and mapping them into a latent vector space containing an abstract representation of the user behavior pattern through a fully connected layer; A joint training framework for contrastive learning tasks and prediction tasks is constructed in the latent vector space, wherein the contrastive learning task enhances the model's ability to recognize behavioral pattern differences through positive and negative sample pairs, and the prediction task optimizes the accuracy of potential churn risk prediction indicators through a cross-entropy loss function.

4. The method according to claim 3, characterized in that The dynamic convolution kernel group is used to extract multi-scale features from the aggregated feature sequence to obtain multi-scale features, including: Calculate the convolution kernel size adjustment factor based on the time step and channel dimension of the input feature sequence, and the adjustment factor is negatively correlated with the local variance of the feature sequence; Generate a set of basic convolution kernel sizes based on the adjustment factor, and allocate a learnable expansion coefficient matrix within each convolution kernel, wherein the expansion coefficient matrix dynamically adjusts the receptive field range according to the feature difference between adjacent time steps; Apply multiple dynamic convolution kernels in parallel to the same input feature sequence to generate feature maps with different scale characteristics; Performing channel attention weighting on the feature map, calculating the importance score of each channel feature and redistributing the channel weights; The weighted feature maps are concatenated along the channel dimension and subjected to dimensionality reduction processing through a separable convolutional layer to generate the multi-scale features.

5. The method according to claim 1, characterized in that The interaction process parameter matrix includes the weight distribution of interface elements, the priority sequence of functional modules, and the recommended strategy adjustment coefficient. The optimized interaction process parameter matrix is ​​called to drive the business operation interface to be reconstructed, including: Analyzing the weight distribution of interface elements in the interaction process parameter matrix, extracting the spatial coordinate set and visual rendering priority parameters of the area exceeding the weight threshold, and generating dynamic focus instructions for core functional controls; Calculate the access path weight of each module node in the navigation menu according to the sorting result of the functional module priority sequence, and generate the shortest path topology structure based on the weight gradient distribution; Input the recommendation strategy adjustment coefficient into the pre-trained recommendation rule generator, match the user's real-time operation scenario with historical preference characteristics, and output a personalized recommendation content sequence associated with the current functional module; Integrating the dynamic focus instruction, the shortest path topology structure, and the personalized recommendation content sequence to generate an interface layout rendering configuration file; After the target client loads the interface layout rendering configuration file, the user's operation response delay, path switching efficiency and recommended content adoption rate indicators for the core function controls are collected in real time, and an interface optimization feedback signal is generated and sent back to the update queue of the interaction process parameter matrix.

6. The method according to claim 5, characterized in that The step of calculating the access path weight of each module node in the navigation menu according to the sorting result of the functional module priority sequence and generating the shortest path topology structure based on the weight gradient distribution includes: Analyze the sorting results of the functional module priority sequence, extract the in-degree weight of the module node and the historical statistical value of the transfer frequency between adjacent modules, and generate the initial access path weight of each module node; The initial access path weight is modified according to the weight difference gradient of adjacent modules in the weight gradient distribution, wherein the module node path weight exceeding the preset gradient change area is suppressed from abnormal growth by using an attenuation factor; Based on the modified access path weights, a directed weighted graph structure is constructed, with module nodes as vertices and path weights as edge weights, to generate the initial shortest path topology structure; Performing redundant path verification on the initial shortest path topology structure, identifying jump nodes in the initial shortest path topology structure that violate the monotonic change of weight gradient, and obtaining a reconstructed path topology structure by inserting intermediate module nodes or adjusting edge weight distribution; Mapping the reconstructed path topology structure into a guide sequence in the navigation menu, wherein the visual parameters of the guide sequence match the weight gradient distribution trend of each node on the reconstructed path topology structure; When the user performs a module switching operation, the path deviation index is collected in real time, the access path weight is updated based on the current weight gradient distribution, and the updated result is synchronously fed back to the sorting logic of the functional module priority sequence.

7. The method according to claim 1, characterized in that The iterative updating of the timing analysis model through an incremental feedback mechanism within a preset verification cycle includes: After the interactive interface is deployed, continuously collect user behavior response data and business conversion rate indicators; Comparing the behavioral response data with a historical behavioral trajectory feature set to identify behavioral pattern offsets after interface optimization; Building a model based on the behavioral pattern offset to update the gradient matrix, and maintaining the learned key features from being overwritten through an elastic weight solidification strategy; The parameters of the time series analysis model are gradually adjusted using a small batch incremental training method, and the model prediction stability index is calculated through an online validation set after each update; When the fluctuation range of the stability index for K consecutive verification cycles is lower than the preset threshold, the incremental update process is terminated and the current model version is locked, K ≥ 3.

8. The method according to claim 7, characterized in that The elastic weight solidification strategy is used to maintain the learned key features from being overwritten, including: Calculating an importance scoring matrix of parameters of each layer of the time series analysis model during the historical training process, wherein the scoring matrix is ​​generated based on the cumulative value of the parameter gradient change amplitude; In the incremental training phase, an elastic constraint term is introduced for each parameter, and the strength of the constraint term is positively correlated with the importance score of the corresponding parameter; Add elastic constraint regularization terms to the loss function to penalize the shift of key parameters caused by new training data; Dynamically adjust the optimizer's learning rate allocation strategy to use a decaying learning rate for a first parameter and an adaptive learning rate for a second parameter, wherein the first parameter is more important than the second parameter; After each training cycle, the importance score matrix is ​​updated and the strength coefficient of the elastic constraint item is recalculated.

9. The method according to claim 5, characterized in that The generating of the interface optimization feedback signal and transmitting it back to the update queue of the interaction process parameter matrix includes: Analyze abnormal interface element loading delay events in user operation response delays, identify control coordinate sets and corresponding functional module identifiers that exceed the preset delay trigger threshold, and generate interface rendering performance bottleneck area markers; Extracting the path jump time and deviation statistics from the path switching efficiency index, and combining them with the node distribution characteristics of the shortest path topology structure, calculating the matching deviation between the path guidance strategy and the user's actual operation habits; Decompose the recommended content adoption rate indicator into the ratio of the number of exposures of recommended content associated with different functional modules to the number of active user interactions, and generate a module-level recommendation utility evaluation matrix; Performing a multi-dimensional correlation analysis on the interface rendering performance bottleneck area marking, matching deviation, and recommendation utility evaluation matrix to determine the collaborative optimization direction of the interface element weight distribution, functional module priority sequence, and recommendation strategy adjustment coefficient in the interaction process parameter matrix; Generate parameter matrix update instructions based on the collaborative optimization direction, the instructions include the interface element weight attenuation coefficient, the module priority sorting correction amplitude and the recommended strategy trigger threshold adjustment amount; Storing parameter matrix update instructions in the update queue in order of priority, and triggering a real-time incremental update of the parameter matrix when it is detected that the dynamic change amplitude of the interface element weight distribution exceeds a preset threshold; After performing the incremental update, user operation data is collected again to generate a new round of interface optimization feedback signals, forming a closed-loop optimization link; The generation logic of the interface optimization feedback signal synchronously associates the version iteration record of the interface layout rendering configuration file with the evolution trajectory of the user behavior pattern, ensuring the dynamic adaptation of the parameter matrix update strategy and the system performance indicators.

10. A computer system, characterized in that: include: a memory storing a computer program; A processor, configured to load the computer program to implement the user behavior data mining method for digital enterprise management as described in any one of claims 1 to 9.

Images