Data product pricing trend prediction method based on transaction behavior clustering analysis

Abstract

The present application relates to a data product pricing trend prediction method based on transaction behavior clustering analysis, aiming to solve the technical problems of complex transaction behavior causal relationship identification, causal mechanism dynamic evolution and trend prediction accuracy improvement. The core scheme is: through transaction behavior clustering and multi-dimensional unsupervised correlation measurement to generate a coarse-grained causal graph, combined with causal stability evaluation, conditional independence test and graph fusion to realize dynamic updating of causal relationship; further using semantic annotation path set, path distillation encoder and overall trend prediction link to form an end-to-end price interval prediction and model dynamic closed-loop optimization. This technology can effectively mine the causal structure of transaction behavior, enhance the stability of causal inference, improve the accuracy of trend prediction, and realize adaptive parameter optimization, thereby improving the intelligent level of market trend analysis.

Description

Technical Field

[0001] This invention relates to the field of data-driven pricing trend prediction and causal reasoning modeling technology, and in particular to a data product pricing trend prediction method based on transaction behavior clustering analysis. Background Technology

[0002] In the field of cross-border data service circulation and pricing, to address the challenges of the non-standard nature of data products and the high complexity of value assessment, current mainstream technologies mostly employ correlation analysis, multi-layer feature fusion, static graph construction, and time-series prediction models to assess price trends. Existing technologies widely rely on the clustering and statistical characteristics of transaction behaviors, such as grouping transaction behaviors using clustering algorithms and using statistical indicators or time-series models (such as LSTM, GRU, and other deep learning methods) to capture price change trends. These methods often focus on the superficial correlation between behavioral patterns and price trends, or use static relationship graphs for feature weighted fusion, lacking causal reasoning mechanisms and failing to accurately distinguish the complex, multi-level causal transmission logic between transaction behaviors and their direct, mediating, or synergistic effects on price trends.

[0003] Representative technical solutions include trend prediction methods based on static transaction behavior graphs. These typically construct a single-layer graph structure and perform correlation weighting or single-point intervention analysis. Because the graph structure lacks dynamic evolution capabilities and fails to distinguish multi-order causal paths, the model's prediction accuracy and robustness are insufficient in scenarios involving market structure changes, policy adjustments, or sudden event interventions. In cross-border data flow scenarios, these shortcomings are further amplified: static graphs struggle to adapt to structural differences in market rules, data standards, and policy environments across different jurisdictions; and the lack of causal explanations in black-box predictions fails to meet the stringent requirements of transparency and verifiability for cross-border collaborative decision-making and cross-border compliance audits. Furthermore, most existing technologies rely on manual causal annotation, large-scale counterfactual intervention modeling, or high-cost multi-agent simulations, making it difficult to achieve large-scale automated, low-complexity causal modeling. Their high computational and optimization costs also struggle to support the real-time response and flexible deployment requirements of cross-border businesses.

[0004] Existing technologies suffer from the following prominent problems: First, correlation-driven trend prediction models cannot automatically identify the causal path between trading behavior and price trends, especially multi-level path transmission relationships, leading to a significant reduction in model responsiveness in policy intervention scenarios. Second, static causal graph structures cannot adapt to changes in market conditions, easily causing prediction failures and a decline in model generalization ability. In cross-border contexts with highly heterogeneous data sources, formats, and regulatory environments, this problem directly makes it difficult for models to establish a unified and robust evaluation framework. Third, reliance on complex annotations and high-cost modeling paradigms (such as adversarial training and multi-agent game theory) reduces the efficiency of model deployment and practical usability. Finally, most existing trend prediction models directly output price values, lacking the ability to explain market semantic trend patterns, failing to meet the causal explanation needs of advanced application scenarios such as demand analysis, risk warning, or policy evaluation, and failing to meet the rigid requirements for causal explanation and decision-making coordination in advanced application scenarios such as demand analysis, risk warning, or policy evaluation in cross-border operations.

[0005] Therefore, in the specific field of cross-border data product pricing, the industry urgently needs a pricing evaluation scheme that can serve as a core component of a "data cross-border circulation technology support platform." This scheme needs to be able to automatically identify multi-level causal paths between transaction behavior and multiple pricing model parameters, possess dynamic evolution capabilities to adapt to rule changes and structural migrations in different markets, achieve modeling with low computational costs, and ultimately output results that combine accurate interval prediction with clear causal semantic interpretation.

[0006] This solution aims to fundamentally overcome the core shortcomings of existing pricing technologies in dealing with cross-border circulation scenarios, such as insufficient dynamic response, weak cross-domain generalization ability, high computing costs, and lack of decision interpretation, and to provide key technical support for building a trustworthy, efficient, and compliant cross-border data pricing infrastructure. Summary of the Invention

[0007] This application provides a data product pricing trend prediction method based on transaction behavior clustering analysis, which aims to solve one of the problems or issues of the existing technology mentioned in the background section above.

[0008] The data product pricing trend prediction method based on transaction behavior clustering analysis provided in this application specifically includes: S1: Obtain the data of each behavior cluster output by the transaction behavior clustering module, and generate a coarse-grained transaction behavior causal graph based on unsupervised indicators; S2: Based on the coarse-grained causal graph of trading behavior and the pre-built causal stability evaluator, a sliding window is used to monitor the consistency of the predictive contribution of each directed edge to price trend changes in a continuous period, and when the contribution fluctuation exceeds a set threshold, weight decay is triggered to generate the identifier of the subgraph region to be reconstructed. S3: Perform causal edge identification operation on the identifier of the subgraph region to be reconstructed, update the local graph structure connection relationship, and generate a multi-level transaction behavior causal graph. S4: Define the causal path order based on the shortest number of hops from any behavior node to the price trend response node in the multi-order transaction behavior causal graph, identify typical paths and assign differentiated semantic labels to generate a causal path set; S5: Construct a path distillation encoder based on the causal path set, map the original transaction sequence to a path activation probability distribution, and generate a path activation probability distribution vector through expert rule combination in the teacher-student architecture and KL divergence constraint training of a lightweight graph neural network structure. S6: Based on the path activation probability distribution vector, calculate the activation intensity sequence of each order path, and combine it with the preset path effect mapping table to map different path combinations into trend pattern label sequences. S7: Based on the trend pattern label sequence and historical price level data, perform interval regression calculation to generate price trend interval prediction results.

[0009] S8: Using the deviation between the price trend range prediction results and the actual market feedback data, the contribution fluctuation threshold of the causal stability evaluator and the mapping weight of the path effect mapping table are corrected in reverse to complete the dynamic closed-loop optimization of the model parameters.

[0010] The data product pricing trend prediction method based on transaction behavior clustering analysis provided in this application has the following beneficial effects: (1) In response to the common problems in existing transaction behavior modeling, such as static causal structure, feature fusion relying on correlation rather than real transmission mechanism, and lack of interpretable path support for trend prediction, this solution constructs a three-dimensional unsupervised index system based on temporal adjacency, order coupling strength and subject overlap to automatically generate an initial transaction behavior causal graph. It also introduces a sliding window-driven causal stability evaluator to dynamically monitor the consistency fluctuation of the contribution of each directed edge to price trend prediction. It only triggers the conditional independence test and edge structure rediscovery of the local subgraph region when significant variation is detected. This achieves incremental evolution of the graph structure instead of costly global reconstruction. It can adapt to the structural differences in data standards, transaction rules and policy environments of different countries / regions in the context of cross-border data flow, and realize the rapid adaptation and integration of localized models without global reconstruction.

[0011] (2) To further enhance the semantic expression and logical transparency of trend prediction, this scheme innovatively defines the concept of "causal path order". The system identifies typical path patterns such as first-order direct influence, second-order mediating transmission and third-order synergistic amplification from behavioral nodes to price responses, and assigns them semantic labels with economic meaning, such as "high-frequency bidding → short-term premium" and "cross-platform arbitrage + policy response → long-term inflection point". It can generate readable causal chains such as "due to cross-border bulk purchasing behavior (increasing the weight of the cost method) → triggering changes in market supply and demand expectations (enhancing the sensitivity of the market method) → jointly pushing up the expected net value of the income method". On this basis, a path distillation encoder is designed to map the original transaction sequence into a multi-order path activation probability distribution. A teacher-student architecture is adopted to achieve knowledge transfer: the teacher model is guided by the path distribution constructed by the aforementioned expert rules, while the student model is a lightweight graph neural network structure. Its output distribution is kept consistent with the teacher's through KL divergence constraints. This lightweight path distillation and prediction architecture significantly reduces computational and communication resource consumption while meeting the requirements for real-time cross-border data processing and low-latency response. Furthermore, compared to traditional end-to-end black-box models that only output a single feature vector, this solution's output path activation distribution not only enhances the traceability of the model's decision-making process, but its semantic trend labels and traceable causal paths provide a transparent and consistent logical basis for cross-border and cross-departmental collaborative decision-making, greatly facilitating compliance audits and business verification.

[0012] (3) In the final trend generation stage, the traditional approach of directly regressing price values ​​is abandoned. Instead, the activation intensity sequence of each causal path is predicted first, and then transformed into trend pattern labels with clear market meaning (such as "accelerated upward movement" and "structural reversal") through a preset path effect mapping table. Finally, the trend pattern decoder combines historical price levels to generate interval prediction results with confidence intervals. This semantic trend expression method breaks through the limitation that the prediction results in the existing scheme are difficult for business personnel to understand and verify. It realizes the logical connection from the underlying behavior to the macro trend, and significantly enhances the practicality and credibility of the system. This generation paradigm of "path intensity → pattern label → interval prediction" essentially encapsulates the uncertainty of prediction in an interpretable semantic level. Its output does not depend on the specific numerical distribution of a single market, so it has stronger cross-border market and cross-cycle generalization ability and stability, and is easier to integrate with rule-based compliance audit or risk control processes.

[0013] Therefore, this solution constructs a closed-loop prediction system that requires no frequent parameter tuning, possesses adaptive evolution capabilities, and is highly interpretable. As a core component of the pricing and evaluation capabilities in the data cross-border flow technology support platform, this solution can transform heterogeneous and fragmented transaction data from multiple markets into stable, reliable trend judgments with business semantics. This provides a crucial technical foundation for key application scenarios such as cross-border commodity pricing and supply chain finance risk management, which require the integration of multi-source information for real-time and reliable evaluation. Attached Figure Description

[0014] Figure 1 This is the main flowchart of a data product pricing trend prediction method based on transaction behavior clustering analysis; Figure 2 This is a flowchart illustrating a data product pricing trend prediction method based on transaction behavior clustering analysis; Figure 3 This is a sub-flowchart of a data product pricing trend prediction method based on transaction behavior clustering analysis; Figure 4 This is another sub-flowchart of a data product pricing trend prediction method based on transaction behavior clustering analysis. Detailed Implementation

[0015] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0016] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0017] like Figure 1 As shown, this application provides a data product pricing trend prediction method based on transaction behavior clustering analysis, specifically including: S1: Obtain the data of each behavior cluster output by the transaction behavior clustering module, and generate a coarse-grained transaction behavior causal graph based on unsupervised indicators.

[0018] Specifically, the unsupervised metrics include, but are not limited to, three types of unsupervised metrics: temporal adjacency, order coupling strength, and subject overlap. The coarse-grained transaction behavior causal graph includes, but is not limited to, an initial set of behavior nodes and an initial set of directed edges.

[0019] S2: Based on the coarse-grained causal graph of trading behavior and the pre-built causal stability evaluator, a sliding window is used to monitor the consistency of the predictive contribution of each directed edge to price trend changes in a continuous period, and when the contribution fluctuation exceeds a set threshold, weight decay is triggered to generate the identifier of the subgraph region to be reconstructed.

[0020] S3: Perform causal edge identification operation based on conditional independence test on the identifier of the subgraph region to be reconstructed, update the local graph structure connection relationship, and thus generate a multi-level transaction behavior causal graph with incremental evolution characteristics.

[0021] S4: Define the causal path order based on the shortest number of hops from any behavior node to the price trend response node in the multi-order trading behavior causal graph, automatically identify typical paths, and assign differentiated semantic labels to generate a set of semantically labeled causal paths. The typical paths include, but are not limited to, direct influence, mediating transmission, and synergistic amplification modes.

[0022] S5: Construct a path distillation encoder based on the semantically labeled causal path set, map the original transaction sequence to a path activation probability distribution, and generate a path activation probability distribution vector that implicitly learns causal transmission logic through expert rule combination in the teacher-student architecture and KL divergence constraint training of a lightweight graph neural network structure.

[0023] S6: Input the path activation probability distribution vector into the trend prediction generation process, calculate the activation intensity sequence of each path, and map different path combinations into trend pattern label sequences using a preset path effect mapping table. The trend pattern label sequences include, but are not limited to, accelerated upward movement, plateau oscillation, or structural reversal patterns.

[0024] S7: Based on the trend pattern label sequence and historical price level data, perform interval regression calculation through the trend pattern decoder to generate a price trend interval prediction result with confidence interval.

[0025] S8: Using the deviation between the price trend range prediction results and the actual market feedback data, the contribution fluctuation threshold of the causal stability evaluator and the mapping weight of the path effect mapping table are corrected in reverse to complete the dynamic closed-loop optimization of the model parameters.

[0026] like Figure 2The proposed solution also involves a dynamic adjustment and release process for cross-border data product prices. Specifically, it involves continuously monitoring triggering conditions such as transaction frequency, data quality, and market demand in the cross-border market. When the conditions are met, a price adjustment suggestion is automatically generated, i.e., the price trend range prediction result mentioned above. This suggestion is then submitted to the operator for review, and takes effect at a set time after approval. Once the price adjustment takes effect, the price adjustment information is publicized and relevant users are notified.

[0027] Step S1: Obtain the data of each behavior cluster output by the transaction behavior clustering module, and generate a coarse-grained transaction behavior causal graph containing an initial set of behavior nodes and an initial set of directed edges based on three unsupervised indicators: temporal adjacency, order coupling strength, and subject overlap. Specifically, this includes: S1.1: Obtain the data of each behavior cluster output by the transaction behavior clustering module, and perform timestamp alignment and discretization binning on the original transaction record sequence within each behavior cluster to generate a set of behavior cluster time-series feature vectors with uniform time granularity.

[0028] The system uses the transaction behavior clustering module's output data as input to unify the original transaction record sequences in the time dimension, ensuring temporal consistency for subsequent causal analysis. For each transaction record sequence within a behavior cluster, a high-precision timestamp parsing program is invoked to convert the timestamp field to a standardized UTC time format, removing records with abnormal offsets or missing values. Based on the set unified time granularity parameters, alignment operations are performed on the standardized timestamp sequences, mapping the timelines of different transaction records to equidistant time nodes, forming a time-synchronized transaction event sequence. Based on the synchronized transaction event sequence, according to a preset discretization binning strategy, the continuous timeline is divided into multiple fixed-length time periods, and transaction records falling within the same time period are aggregated into the same bin. Feature encoding processing is performed on the transaction records in each time bin, extracting multi-dimensional indicators including transaction amount, transaction quantity, order type, and transaction entity category to construct a feature vector for each bin. The feature vectors of each time bin are combined sequentially according to time order to form a set of behavioral cluster time-series feature vectors with a unified time granularity. Through this structural unification and discretization binning processing method, the original transaction record sequence is transformed into standardized feature data that is comparable on the time scale and can be used to calculate causal indicators such as time adjacency, thereby achieving the time consistency technical effect of subsequent causal relationship measurement.

[0029] For example, the output of a transaction behavior clustering module in an electronic data product market is obtained, containing 5 behavior clusters, each with approximately 2000 original transaction records. The timestamp field in the records is converted to UTC format, and the time granularity parameter is set to 1 hour. Time alignment processing is performed to ensure that all transaction events fall on a unified hour node. The time axis is divided into bins in 1-hour units, resulting in approximately 720 time bins per cluster. For the transaction records within each bin, the total transaction amount, number of transactions, proportion of major order types, and number of participating entity categories are statistically analyzed to form a 4-dimensional feature vector. The 720 4-dimensional feature vectors are combined in chronological order to form the temporal feature vector set for that behavior cluster. Standardization processing is applied to the transaction amount feature. After processing, the temporal feature vector sets of each behavior cluster are completely consistent in terms of time scale and numerical statistical distribution, significantly improving the accuracy of subsequent time lag mutual information calculation and causal path identification.

[0030] S1.2: Based on the set of temporal feature vectors of the behavior clusters, calculate the time lag mutual information value between any two behavior clusters as a time adjacency measure, and count the concurrency frequency and monetary correlation of cross-cluster orders as an order coupling strength index. At the same time, extract the intersection ratio of the identity identifiers of the main participants in the transaction as the main overlap value, thereby generating a relationship matrix between behavior clusters containing a three-dimensional unsupervised correlation measure.

[0031] Using the set of temporal feature vectors of behavioral clusters output in step S1.1 as input, delayed mutual information calculation is performed on the temporal feature vectors corresponding to any two behavioral clusters in the set. The mutual information value is used as the core indicator of temporal adjacency. A sliding delay window mechanism is used to accumulate and average the values ​​at multiple delay steps to reduce the impact of noise, resulting in a stable temporal adjacency numerical matrix. Cross-cluster order concurrency statistics are performed on transaction records of the same behavioral cluster pair at different time periods. By counting the frequency of simultaneous orders and calculating the Pearson correlation coefficient between the transaction amount and the time segments, the correlation between concurrency frequency and amount is obtained and used as elements of the order coupling strength indicator matrix. Based on this, an intersection operation is performed on the set of identity identifiers of the transaction participants of the behavioral cluster pair. The ratio of the number of intersection elements to the total number of participants is used as the subject overlap value to quantify the degree of subject participation crossover. The three dimensions of the relationship matrix are filled with three types of unsupervised indicators to obtain a behavioral cluster relationship matrix containing temporal adjacency values, order coupling strength values, and subject overlap values.

[0032] By generating a relation matrix between behavioral clusters using three-dimensional indicators of delayed mutual information, order coupling strength, and subject overlap, the unified time-granularity feature vector set from the previous step is transformed into quantitative correlation data that can be used for topology construction, thus achieving simultaneous quantification of the three-dimensional causal potential of time correlation, transaction coupling, and subject crossover.

[0033] For example, in an e-commerce data product transaction market scenario, for a unified time-granularity feature vector set of 5 behavioral clusters, 5 different delay step values ​​are used. After averaging the mutual information values, a time adjacency matrix is ​​formed, with values ​​ranging from 0.12 to 0.65. The frequency statistics of cross-cluster order concurrency are completed within a natural week sliding window, and the calculated monetary correlation coefficient ranges from 0.20 to 0.72. Combined with the frequency, a direct assignment is made to form an order coupling strength matrix. The calculated intersection ratio of participating subject sets ranges from 0.10 to 0.50, forming a subject overlap matrix. The three matrices are concatenated to generate a 5×5×3 structural relationship matrix, which is applied to the subsequent S1.3 directed edge direction attribute determination and weight sparsity screening, showing a significant improvement in the accuracy of time causal inference and the synergistic optimization effect of multi-dimensional indicators.

[0034] S1.3: The directional attributes of directed edges are determined by the temporal adjacency measure in the relationship matrix between the behavior clusters. The edge weight threshold is set by combining the order coupling strength index and the main body overlap value to perform sparsification screening, so as to generate a preliminary connection topology structure composed of the initial set of behavior nodes and the weighted initial set of directed edges.

[0035] For the input relationship matrix data between behavioral clusters, read the temporal adjacency measurement elements in the matrix, determine the order relationship between two behavioral clusters based on the sign and magnitude of the time lag mutual information value, and assign a direction attribute from the former to the latter to behavioral clusters with significant positive time dependence.

[0036] For behavior clusters with defined directional attributes, the corresponding order coupling strength index value is read. A coupling strength value is constructed based on the combination of concurrency frequency and monetary correlation. Combined with the main body overlap value, a comprehensive edge weight candidate value is formed. The edge weight candidate value is calculated using the following formula: Where S represents the order coupling strength value, and O represents the main body overlap value.

[0037] The candidate edge weight values ​​are compared with a preset edge weight threshold. If the candidate value is lower than the edge weight threshold, the directed edge is determined to have a weak contribution to the causal graph and is removed from the candidate set to achieve sparsity processing. Directed edges with candidate edge weight values ​​higher than or equal to the threshold are retained and their weight parameters are recorded.

[0038] We perform weight normalization on the retained initial set of directed edges so that the sum of the weights of all directed edges under the same source node is equal to 1, thereby avoiding topological deviations caused by weights of different scales.

[0039] The initial set of directed edges, after direction determination and sparsification filtering, is paired and bound with the initial set of behavioral nodes in the behavioral cluster data to form a preliminary connection topology consisting of the initial set of behavioral nodes and the weighted initial set of directed edges.

[0040] By using the above processing method, the multidimensional correlation metric matrix of the previous step is transformed into a connection topology structure with directional attributes and sparse characteristics, realizing the clear node-edge mapping relationship required for coarse-grained causal graph construction.

[0041] For example, for the relationship matrix between behavioral clusters generated in a trading market, the temporal adjacency metric is calculated using time-lag mutual information, resulting in matrix element values ​​ranging from 0.05 to 0.87. When the value is greater than 0.3, a significant positive time dependency is determined, and the directional attribute is identified accordingly. The order coupling strength index is calculated by statistically analyzing the concurrency frequency and monetary correlation of cross-cluster orders for each pair of behavioral clusters, using a weighted average of 0.6 for the concurrency frequency score and 0.4 for the monetary correlation score, resulting in a strength value ranging from 0.1 to 0.9. The subject overlap is obtained by statistically analyzing the intersection ratio of identity identifiers, resulting in a value ranging from 0 to 1. In specific implementation, S and O are substituted into the formula to calculate the comprehensive edge weight candidate value. For example, if the S value for a pair of behavioral clusters is 0.8 and the O value is 0.6, then the w value is 0.7. A preset edge weight threshold of 0.5 is set. All directed edges with candidate values ​​below 0.5 are removed from the connection topology, while the remaining edges are retained and normalized according to the source node. For example, if the three edge weights of a source node are 0.8, 0.7, and 0.6, after normalization, they are 0.36, 0.32, and 0.27, respectively. In this embodiment, a preliminary connection topology structure is finally formed, containing 15 initial behavioral node sets and 27 weighted directed edges. The topology sparsity is improved, and the edge directions and weights maintain consistency in the causal inference space, providing high-quality basic data for subsequent connectivity verification and causal path identification.

[0042] S1.4: Perform connectivity verification and isolated node removal operations on the initial connection topology, and map the remaining valid nodes as behavioral entity objects in the causal inference space, and map the valid edges as potential causal transmission path objects, thereby generating a standardized coarse-grained causal graph of transaction behavior.

[0043] Step S2: Based on the coarse-grained causal graph of trading behavior and the pre-built causal stability evaluator, a sliding window is used to monitor the consistency of the predictive contribution of each directed edge to price trend changes in a continuous period, and when the contribution fluctuation exceeds a set threshold, weight decay is triggered to generate the identifier of the subgraph region to be reconstructed.

[0044] The workflow of the causal stability estimator specifically includes: S2.1: Obtain the initial set of directed edges and the corresponding historical transaction sequence data in the coarse-grained transaction behavior causal graph. Based on the time adjacency index, divide the continuous time segments into multiple equal-length sliding windows to generate a windowed transaction behavior data stream containing time-series slice information.

[0045] The process involves obtaining the initial set of directed edges and corresponding historical transaction sequence data from a coarse-grained causal graph of transaction behavior. A graph structure parsing module is then used to index and match the start and end nodes of each directed edge, establishing a bidirectional mapping table from behavior nodes to historical transaction records. Based on the time adjacency index, the timestamp field in the historical transaction records is read. A time difference calculation function is used to batch calculate the time difference values ​​of adjacent records, and the time difference values ​​are logically compared with a preset time adjacency threshold to filter out continuous time segments that meet the adjacency criteria. A window partitioning algorithm is then executed on the selected time segments, using an equal-length window size parameter W as the partitioning benchmark. The time segments are sequentially divided into multiple non-overlapping time windows, and the start and end time indices of the windows are recorded to form a window time boundary matrix. During the window partitioning process, a time-series slicing encapsulation component is called to associate and bind the window time boundary matrix with the transaction record index sequence, generating a slice metadata structure containing the transaction record index, time segment length, and window number. The slice metadata structure is used to batch-slice the historical transaction sequence, extracting the transaction record subsets corresponding to each window, and assembling them in chronological order into a windowed transaction behavior data stream conforming to the sliding window mechanism. By using the above windowing processing method, the edges of the coarse-grained causal graph of transaction behavior and historical transaction data in the previous step are transformed into a transaction behavior data stream organized by time window, which realizes the input data structure for the subsequent causal stability evaluator to monitor the consistency of prediction contribution over continuous periods.

[0046] For example, in a data product market transaction monitoring system, a coarse-grained transaction behavior causal graph contains 120 initial directed edges, each corresponding to a historical transaction record length of 180 days, with timestamp precision in seconds. The system's time adjacency threshold is set to 240 seconds, and the window length parameter W is set to 7 days. For edge E12, its starting and ending nodes correspond to the high-frequency bidding cluster and price response cluster, respectively. 180 days of transaction records are read from the mapping table, and the time difference between adjacent records is calculated in batches. The filtering condition is that the time difference is ≤240 seconds for record segments. After filtering, approximately 516 time segments are obtained. A window partitioning algorithm is called to divide the data at a 7-day time granularity. Each window contains approximately 300 transaction records, and the window time boundary matrix records the start and end indices of each window. The window time boundary matrix is ​​bound to the transaction record index sequence to generate a slice metadata structure containing window number, window length, and record index. The slice metadata is batch-cut to obtain approximately 25 window transaction record subsets, which are then arranged in chronological order to form a windowed transaction behavior data stream. When this data stream is input into the Granger causality test module in subsequent steps, it can significantly improve the time matching accuracy of causal strength calculation and ensure that the causal stability evaluator's judgment results on the continuous periodic prediction contribution of edge E12 and its adjacent structures are more reliable.

[0047] S2.2: For each initial directed edge in the windowed trading behavior data stream, the Granger causality test algorithm is used to calculate its predicted contribution value to the price trend change target within the current sliding window, so as to generate an edge contribution sequence representing the causal strength of a single period.

[0048] S2.3: Based on the edge contribution sequence, the rolling variance statistical method is used to calculate the fluctuation dispersion of the predicted contribution values ​​among multiple consecutive sliding windows, so as to generate an edge stability evaluation coefficient that reflects the stability of the causal mechanism.

[0049] Based on the edge contribution sequence output by the preceding sub-step S2.2, the predicted contribution values ​​of the same directed edge within multiple consecutive sliding windows are selected as input variables to construct a statistical analysis object for cross-period contribution differences.

[0050] For each directed edge, the contribution sequence is processed by window serialization, and arranged into a fixed-length numerical array in chronological order to ensure data alignment consistency during the rolling analysis.

[0051] A rolling variance calculation mechanism is applied to the windowed serialized contribution array, defining the rolling window length and sliding step size parameters to ensure that each calculation includes a preset number of periodic contribution values ​​and that the step size controls the analysis resolution.

[0052] The following formulas for mean and variance are used to calculate the dispersion of contribution fluctuations within the rolling window: in, This represents the single-cycle contribution value for the i-th cycle within the current scrolling window. The average contribution within this window. This represents the number of cycles within the window.

[0053] The rolling variance result sequence is mapped to stability evaluation coefficients. The higher the coefficient, the greater the fluctuation of the causal effect of the directed edge over multiple periods, and the lower the stability.

[0054] Through the above-mentioned rolling variance statistics and stability quantification, the predicted contribution values ​​between multiple consecutive sliding windows are transformed into edge stability evaluation coefficients that reflect the stability of the causal mechanism, thereby realizing the monitoring of causal path quality based on the degree of fluctuation and the basis for triggering subsequent weight decay.

[0055] For example, for an initial directed edge in the causal graph of transaction behavior, its contribution sequence over the past 10 periods is 0.42, 0.47, 0.45, 0.50, 0.40, 0.38, 0.55, 0.53, 0.49, and 0.46. With a rolling window length of 5 and a sliding step of 1, the mean contribution of the first window is calculated to be 0.448, and the square root of the variance yields a rolling standard deviation of 0.037. The mean of the second window is 0.44, and the standard deviation is 0.046, and so on, generating a rolling standard deviation sequence. The standard deviation sequence is normalized to the [0,1] interval as a stability evaluation coefficient. When this coefficient exceeds a preset threshold of 0.04, the directed edge is deemed unstable, and exponential smoothing weight decay processing is triggered in subsequent step S2.4. The final output stability evaluation coefficient can significantly improve the accuracy of identifying the subgraph region to be reconstructed, thereby enhancing the robustness of the evolution of the causal graph of transaction behavior.

[0056] S2.4: Compare the edge stability evaluation coefficient with the preset contribution fluctuation threshold, and perform exponential smoothing weight decay processing on the initial directed edges that exceed the threshold to generate an adjusted edge weight set with dynamic suppression characteristics.

[0057] Each edge stability evaluation coefficient is compared with a preset contribution fluctuation threshold, and a matching test result table is established using the edge identifier as the index key. Initial directed edges that exceed the threshold in the comparison results are extracted as a set of high-fluctuation edges, and their historical weight decay curve time series data are retained. For each edge in the high-fluctuation edge set, the exponential smoothing function is called to calculate and update the weight value, specifically using the formula: in, For smoothing coefficients, This represents the weight value for the current period. The previous period's weight values ​​are used. Difference statistics are performed between the updated and original weight values ​​to construct an edge weight adjustment dataset containing the magnitude of weight changes. For edges whose weight changes exceed a set suppression threshold, additional decay compensation operations are performed to prevent instability of the causal graph structure in the event of a sudden drop in edge weights. All processed edge weights are remapped to the connection topology of the causal graph according to their edge identifiers, forming an adjusted edge weight set with dynamic suppression characteristics. Through exponential smoothing and difference statistics, the stability evaluation coefficients from the previous step are transformed into quantifiable and updatable edge weight data, achieving adaptive suppression of causal contribution fluctuations.

[0058] For example, in a market forecasting task for a data product, the high-volatility edge set contains 12 initial directed edges. The stability evaluation coefficient of each edge is between 0.45 and 0.78, the preset contribution volatility threshold is 0.40, and the smoothing coefficient α is set to 0.6. Exponential smoothing is performed on the current period weight value of edge E07 (0.52) and the previous period weight value (0.69), calculated as follows: The updated weight value was 0.582. A difference analysis was then performed between this value and the original value; the difference was 0.108, exceeding the suppression threshold of 0.05, triggering compensation decay. A smoothing offset was added to the weight update result at a ratio of 0.02, reducing the weight value to 0.562. Similar calculations were applied to all high-volatility edges and mapped back to the global graph structure. In the final output set of adjusted edge weights, the number of edges with significantly reduced weights decreased from 12 to 5, effectively weakening the transmission effect of short-term abnormal fluctuations in the causal chain and significantly improving the stability of trend prediction in policy intervention scenarios.

[0059] S2.5: Based on the initial directed edges and their adjacent node topological relationships in the adjusted edge weight set after the exponential smoothing weight decay process, aggregate them to form a connected subgraph region and mark it as the subgraph region identifier to be reconstructed, so as to output a spatial range instruction to guide the re-identification of local causal edges.

[0060] like Figure 3 As shown, step S3: Perform causal edge identification based on conditional independence test on the identifier of the subgraph region to be reconstructed, update the local graph structure connection relationship, thereby generating a multi-level transaction behavior causal graph with incremental evolution characteristics. Specifically, this includes: S3.1: Obtain the identifier of the subgraph region to be reconstructed and the corresponding original transaction behavior clustering data. Use the time adjacency constraint and subject overlap screening mechanism to extract the node set of the original transaction behavior clustering data to generate a local candidate node set containing the behavior node to be updated and its associated order coupling strength data.

[0061] Based on the input of the subgraph region identifier and its corresponding original transaction behavior clustering data, the node data parsing module is invoked to perform a deconstruction operation on the original transaction sequence within the identifier range, extracting the transaction record set that makes up the behavior node and its basic attributes such as timestamp, order identifier, and subject identity. Temporal adjacency constraint filtering is performed on the extracted transaction record set, and the time lag mutual information value between nodes is calculated using a window partitioning mechanism. Node pairs with mutual information values ​​below a preset lower limit are removed from the candidate range, forming a preliminary node subset with time correlation. Subject overlap filtering is performed on the preliminary node subset, calculating the intersection ratio of transaction subject identities between any two nodes, and using the subject relationship matrix to filter node pairs with subject overlap below a threshold, retaining node combinations with subject correlation, thus obtaining a complete set of nodes to be updated with both time and subject constraints. The coupling strength of associated orders is calculated for the set of nodes to be updated, and the correlation coefficient between cross-node order concurrency frequency and amount is calculated and assigned to the node object attribute field, forming a local candidate node set containing order coupling strength indicators. By using temporal adjacency constraints and subject overlap filtering mechanisms, the original clustering data pointed to by the subgraph region identifiers to be reconstructed is transformed into a node set with characteristics of temporal correlation, subject association, and order coupling strength, thus realizing the construction of structured input before local causal edge identification.

[0062] For example, in a data product market transaction behavior analysis scenario, for a subgraph region to be reconstructed triggered by a market intervention event, the input raw transaction behavior clustering data includes 500 behavior nodes. Each node records attributes such as transaction time, order number, transaction subject ID, and transaction amount. The node data within this region is divided into execution windows with a time granularity of 7 days. The temporal adjacency value between nodes is calculated using time lag mutual information, with a mutual information threshold set to 0.25. When the mutual information value of a node pair is less than 0.25, that node pair is removed from the candidate set, resulting in approximately 200 nodes with time correlation. Subject overlap statistics are then performed on these 200 nodes, calculating the intersection ratio of subject IDs between nodes. The overlap threshold is set to 0.4. After removing node pairs with an overlap of less than 0.4, approximately 150 nodes are retained. Finally, the correlation coefficient between cross-node order concurrency frequency and amount is calculated for the retained nodes, using the following formula: Here, A represents the order concurrency frequency of the two nodes within the same time window, and B represents the correlation coefficient of the corresponding order amounts. This calculation result is assigned to the order coupling strength field of the node attributes for use as a reference in subsequent conditional independence tests. The execution result provides each node with three indices: temporal adjacency, subject overlap, and order coupling strength, forming a local candidate node set. This set can be directly used for causal edge identification, significantly improving the accuracy and stability of local structure updates.

[0063] S3.2: Based on the local candidate node set, perform a multivariate conditional independence test algorithm to calculate the partial correlation coefficient statistics of any two behavioral nodes under the condition of other nodes, so as to generate an initial undirected edge connection matrix that characterizes the potential direct causal dependency between nodes.

[0064] Using a local set of candidate nodes as input, the data preprocessing module required by the multivariate conditional independence test algorithm is called to center and standardize the numerical values ​​of order coupling strength and time adjacency constraint parameters associated with nodes in a unified numerical space, so as to ensure the comparability of features of different dimensions in statistical tests.

[0065] After numerical preprocessing, a condition set is constructed for any two nodes in the candidate node set. The condition set is derived from the feature vectors of all nodes except the pair of nodes, and is used as a control variable to input into the test model in order to calculate the partial correlation coefficient given these control variables.

[0066] The partial correlation coefficient calculation formula is used. The corresponding element values ​​are extracted by calculating the inverse of the covariance matrix. The partial correlation coefficients are then calculated in batches for each node pair, outputting a statistical table of partial correlation coefficients for all node pairs under control conditions, serving as the original causal dependency strength data. A significance test is performed on the partial correlation coefficient statistical table. Node pairs with a significance level less than a preset threshold are marked as having a potential direct causal dependency, and these results are mapped to a 1 value in the initial undirected edge connection matrix. Node pairs whose statistics are insufficient to support causal dependency are effectively eliminated and marked as 0 values ​​in the matrix.

[0067] Through the above processing method, the local candidate node set of the previous step is transformed into an initial undirected edge connection matrix that represents the potential direct causal dependencies between nodes, thereby realizing a quantifiable basic output for causal edge identification in the subgraph region to be reconstructed.

[0068] For example, in a data product market transaction behavior analysis scenario, the local candidate node set consists of 10 nodes, each containing order coupling strength and subject overlap features for 30 time slices. After centering and standardizing this set, a condition set of 8 control variables is constructed. The partial correlation coefficient formula is used to calculate the coefficient between node 1 and node 2 under the condition of controlling the other 8 nodes, which is 0.45. This value is calculated by extracting elements from the inverse of the covariance matrix. Using the t-statistic formula, with n=30 and k=8, the t-value is 4.12, which is considered to indicate a potential causal dependency at a significance level of 0.01. After calculating the entire set, the initial undirected edge connection matrix is ​​10×10, with diagonal elements of 0. Significant node pairs are represented by 1 in the matrix. This matrix serves as the main input source in the subsequent directed processing in S3.3, significantly improving the robustness and accuracy of local causal structure identification.

[0069] S3.3: Apply a directed rule inference engine to the initial undirected edge connection matrix, and perform directed mapping processing based on time lag mutual information gain and Granger causal inference logic to generate a set of local directed edge sequences with clear causal flow.

[0070] When performing directional processing on the initial undirected edge connection matrix generated by the multivariate conditional independence test, the node pairs corresponding to each element of the matrix and their partial correlation coefficients are introduced into the directional rule inference engine as input.

[0071] For each pair of nodes to be targeted, the time lag mutual information gain value between the node pairs is calculated. The time delay pattern between node-triggered events and target trend responses is statistically analyzed using a sliding window to construct a lag mutual information gain index matrix.

[0072] Using Granger causal inference logic, bidirectional regression fitting is performed on each pair of nodes at a given lag order. The difference in prediction error between the two-way models is calculated, and a matrix of bidirectional causal significance determination results is formed.

[0073] The lagged mutual information gain index matrix is ​​matched with the two-way causal significance determination result matrix by rules, and the direction with large mutual information gain value and significant one-way Granger causality is selected as the directional mapping direction.

[0074] Perform directed edge construction processing on the selected direction, mapping node pairs to local directed edges with definite start and end points, and recording the edge weight values ​​as a weighted combination of mutual information gain and Granger statistical significance to generate a set of local directed edge sequences.

[0075] Through the above reasoning mapping process, the undirected edge connection matrix of the previous step is transformed into local directed edge data that reflects causal directionality, thereby realizing the direction recognition of the local causal graph structure.

[0076] S3.4: Based on the set of local directed edges and the stable edge data in the coarse-grained transaction behavior causal graph, perform graph fusion operation, and use a weight decay compensation strategy to perform topological consistency verification on conflicting edges to generate incremental evolution subgraph data that has completed local structure update.

[0077] Based on the set of locally directed edge sequences and the stable edge data in the coarse-grained causal graph of transaction behavior, the graph fusion operation module is used to perform structured merging of the two types of edge data. The input conditions for the fusion process are the direction attribute and weight parameter of the locally directed edges, and the weight consistency index of the stable edges. A conflict detection algorithm is executed on the merged edge set to screen edge records that have the same node pair in the global graph topology index but have significant differences in direction attribute or weight, and the conflicting edges are extracted as subsequent compensation targets. For conflicting edges, the weight decay compensation strategy module is called, and the stability weight ratio and local causal strength coefficient are introduced when calculating the decay coefficient. The compensation value is calculated using the following formula: Among them, W local For the local directed edge weights, W stable To stabilize edge weights, k is the compensation rate adjustment factor, and e is the natural constant. The compensation value is weighted and summed with the attenuated weights of the original conflicting edges to generate updated edge weight values. A topological consistency check is performed on the fused edge set, using adjacency matrix connectivity analysis to ensure that newly added edges do not disrupt the reachability and directionality constraints of existing causal paths. During the check, newly added edges with detected loop structures undergo direction adjustment or further weight attenuation to maintain the unidirectionality of causal propagation and the stability of the order definition. The edge set after weight compensation and topological consistency check is encapsulated as incremental evolutionary subgraph data output. This processing method fuses the local causal edge identification results from the previous step with the global stable edge information into structurally complete and weighted subgraph update data, achieving consistency between local structural updates and global evolution.

[0078] For example, in a data product market scenario, the set of locally directed edge sequences contains two edges, A→B and B→C, where the weight of A→B is 0.32 and the weight of B→C is 0.45. In the stable edge data, the corresponding weights of A→B and B→C are 0.5 and 0.46. Conflict detection identifies a weight difference of 0.18 between the A→B and B edges, and performs a compensation formula calculation: taking k=2, W... local =0.32, W stable Substituting 0.5 into the formula, we get the compensation value as 0.32 + (0.5 - 0.32) / (1 + e) 2The updated A→B edge weight is approximately 0.4786, close to the stable edge weight, and retains local causal strength information. Topological consistency verification confirms that the fused subgraph A→B→C path has an acyclic structure and maintains its order. The incrementally evolved subgraph data significantly improves the consistency of path activation probability in prediction tests.

[0079] S3.5: Embed the incremental evolution subgraph data into the global graph topology, perform connectivity analysis and transitive closure calculation operations to generate a multi-level transaction behavior causal graph with incremental evolution characteristics.

[0080] The incremental evolution subgraph data is input as local topology update information into the global graph topology embedding module. Indexing is performed according to the mapping relationship between node identifiers and edge weights to determine the storage location of the corresponding nodes and their edges in the global graph. Based on the indexing results, a topology fusion algorithm is invoked to merge the newly added edges in the incremental evolution subgraph data with the original edge set of the global graph. Weight smoothing is performed on edges with conflicting weights to ensure the continuity of the global edge weight distribution. The fused global edge set and node set are input into the connectivity analysis module. A combination of depth-first and breadth-first traversal strategies is used to calculate the node coverage and path redundancy of each connected branch in the graph, marking structural risk areas where broken chains or isolated nodes may occur. For these structural risk areas, the transitive closure calculation module is invoked to perform reachability matrix construction and Boolean product operations, adding indirectly reachable node pairs as virtual connecting edges in the global graph to improve the global causal transmission link structure. The results of transitive closure calculation and connectivity analysis are updated synchronously to output the final multi-level causal graph of transaction behavior with global node reachability and continuous optimization of edge weights. Through embedding fusion and closure inference processing, the local structure update results of the previous step are transformed into topologically consistent data covering the whole, so as to achieve the expected technical effect of causal path identification and enhanced transmission stability.

[0081] For example, in a data product market transaction behavior analysis scenario, the incremental evolution subgraph data contains node N. 45 N 46 and the directed edges (N) between them identified by the conditional independence test. 45 →N 46 The edge weight is 0.72. The global graph topology contains 1000 behavioral nodes and 3500 directed edges, with node N. 45 There is already a connection to N in the global graph. 12 N 33 Edge, node N 46 Previously, these were isolated nodes. The embedded computation module locates N in the incremental subgraph using the node identifier index. 45 With N 46The positions are mapped to global graph storage units, and edges (N) are merged. 45 →N 46 To the global edge set. Weight smoothing adjustment uses an exponential smoothing coefficient α = 0.3, applied to edges existing in the global graph and related to N. 45 The relevant edge weights are adjusted to adapt to the changed distribution of newly added edge weights. The connectivity analysis module detects newly added edges through combinatorial traversal, making N... 46 By incorporating the globally connected branch, redundancy is reduced to 0.15. The transitive closure calculation module constructs a reachability matrix R, R i,j =1 indicates that node i is reachable from node j, through Boolean product operation R new =R∨(R·R) adds multiple pairs of indirect reachability relationships. For example, node N 46 N can be reached via a two-hop path. 12 Closure computation adds virtual connection edges (N) to the global graph. 46 →N 12 The virtual edge weight was set to 0.45 to improve the integrity of the causal path. After synchronous updates, the global connectivity index of the final multi-order transaction behavior causal graph was significantly improved, the risk of edge disconnection was reduced, and the model showed higher stability and interpretability in subsequent path order identification.

[0082] like Figure 4 As shown, step S4: Define the causal path order based on the shortest number of hops from any behavior node to the price trend response node in the multi-order transaction behavior causal graph. Automatically identify three typical paths covering direct influence, intermediary transmission, and synergistic amplification modes, and assign differentiated semantic labels to generate a set of semantically labeled causal paths. The differentiated semantic labels include: high-frequency bidding driving short-term premium label, bulk purchasing triggering inventory expectation adjustment label, and cross-platform arbitrage superimposed with policy response label.

[0083] Step S4 specifically includes: S4.1: Obtain all behavior nodes and price trend response nodes in the multi-level trading behavior causal graph with incremental evolution characteristics, and use the breadth-first search algorithm to traverse and calculate the connection links from any behavior node to the price trend response node in the graph to generate an original reachable path set containing the sequence of starting nodes, intermediate nodes and ending nodes.

[0084] In this specific embodiment, the price trend response node is a special type of terminal node in the multi-level trading behavior causal graph. Its core feature is that it carries and represents quantifiable market price trend signals. Technically, the attribute vector of this type of node is defined as a time series of price-sensitive indicators, such as the return, volatility, or event labels of price breaking through thresholds over the next N periods. The price trend response node is the logical endpoint of all causal transmission paths. Directed edges emanating from various behavioral nodes in the graph ultimately point to this node, and the numerical changes embedded in the node are considered a quantitative response to the causal effects produced by the aforementioned behavioral sequences. The existence of this node gives the abstract causal analysis a clear and calculable economic meaning, namely, measuring the contribution and transmission efficiency of various trading behavior patterns to the formation of the final price trend.

[0085] Obtain all behavior nodes and price trend response nodes in the multi-level transaction behavior causal graph, and call the node retrieval module to read the unique identifier of the node and its adjacency list data in the global topology.

[0086] For each behavior node, perform adjacency list parsing, extract the initial set of directed edges directly connected to it, and register the matching task between the starting point and the target node to be traversed in the task queue.

[0087] For each matching task in the queue, run the breadth-first search algorithm initialization phase, push the starting node into the hierarchical traversal queue, and allocate an empty path sequence container for each level to store temporary links.

[0088] During the breadth-first search process, the direction and weight of all outgoing edges of the current level node are determined. Adjacent nodes that meet the criteria of causal flow pointing to the price trend response node and edge weight not lower than a preset threshold are selected and added to the next level traversal queue. At the same time, the node identifier is appended to the corresponding path sequence container.

[0089] When a node meets the price trend response node condition during the traversal, the starting node, all intermediate nodes, and the current ending node in the path sequence container are combined in sequence to generate an original reachable path object containing the node sequence, and the object is stored in the result set.

[0090] By traversing the termination condition controller to determine that all queue tasks have been completed, the breadth-first search stops, and the original reachable path set consisting of multiple original reachable path objects is output, thus realizing the structured extraction of reachable paths.

[0091] By combining breadth-first search with direction and weight selection logic, the incremental evolutionary causal graph of the previous step is transformed into a set of original reachable paths containing a sequence of start nodes, intermediate nodes, and end nodes, thereby enabling the computation of causal paths and preparing input for subsequent order analysis.

[0092] For example, in a certain e-commerce transaction data product, the multi-order transaction behavior causal graph contains 120 behavior nodes and 3 price trend response nodes. The adjacency list data is stored in a sparse matrix, and the edge weight threshold is set to 0.4. After the node retrieval module reads all node IDs and adjacency information, it establishes 120 × 3 = 360 matching tasks from the starting point to the target node. During the traversal, the edge direction is determined based on the condition that the time lag mutual information gain > 0, and the weight is determined based on the condition that the edge weight ≥ 0.4. The breadth-first search queue depth is limited to 5 levels. The path sequence container is appended with the current node ID each time it is expanded. The final generated original reachable path object is such as [node A, node B, node C, price trend node P1]. A total of 820 different path sequences are obtained in the result set. The application effect is that it can significantly improve the completeness of path classification when calculating the order, ensuring that the three types of paths—direct influence, mediating transmission, and synergistic amplification—are all identified, and supporting the improvement of the accuracy of trend pattern mapping.

[0093] S4.2: Based on the number of directed edges contained in each path in the original reachable path set, perform a shortest hop count operation to quantify the path length, define the path length as a causal path order index, and thus generate a candidate causal path list with order attribute labels.

[0094] S4.3: Based on the causal path order index in the candidate causal path list with order attribute markers, perform hierarchical filtering to identify specific patterns, classify paths of order one as direct influence patterns, paths of order two as mediating transmission patterns, and paths of order three as collaborative amplification patterns, thereby generating a classification path group covering three typical path patterns.

[0095] Based on a candidate causal path list with order attribute labels as input, the order index of each path in the list is read, and a path index sequence arranged in ascending order is established for subsequent pattern filtering. The order index is matched against a preset pattern order mapping table, and a logical judgment module filters out the path set with an order of one, generating a path index label set corresponding to the direct influence pattern in the filtering results. Records with an order of two from the remaining path set are extracted to construct a path index label set corresponding to the mediation transmission pattern, and the intermediate node attributes of each path in this set are recorded for subsequent semantic mapping enhancement. In the further processing stage, a logical judgment is performed on the remaining path set for an order of three, constructing a path index label set corresponding to the collaborative amplification pattern, and attaching multi-source trigger node feature vectors to the paths in this set for association effect analysis. The path index label sets corresponding to the three patterns are encapsulated and stored according to pattern category, forming a categorized path group data structure covering three typical path patterns: direct influence, mediation transmission, and collaborative amplification. By using hierarchical filtering and pattern classification processing, the order labeling results from the previous step are transformed into path classification indicators with clear pattern categories, thereby achieving structured pattern recognition of multi-order causal paths.

[0096] For example, in an e-commerce transaction data analysis scenario, the input candidate causal path list contains 50 path records, with the order index ranging from 1 to 3. According to a preset mapping table, the order value for the direct impact mode is set to 1, the order value for the intermediary transmission mode is set to 2, and the order value for the collaborative amplification mode is set to 3. After executing the hierarchical filtering logic, 17 paths with an order of 1 are selected, whose starting nodes are mostly high-frequency bidding behavior, and whose ending nodes all directly point to short-term price fluctuation responses; 21 paths with an order of 2 are selected, mostly containing intermediate nodes of bulk purchasing type; and 12 paths with an order of 3 are selected, whose starting node combinations include two trigger sources: cross-platform arbitrage behavior and policy announcement response. The filtering results are stored separately according to mode categories, forming categorized path groups.

[0097] S4.4: For each typical path pattern in the classified path group, call the preset semantic rule library to perform a label mapping operation, map the direct impact pattern to the high-frequency bidding-driven short-term premium label, map the intermediary transmission pattern to the bulk purchase-induced inventory expectation adjustment label, and map the collaborative amplification pattern to the cross-platform arbitrage superimposed policy response label, thereby generating a path object to be labeled with differentiated semantic descriptions.

[0098] To map labels for various typical path patterns within the categorized path groups, it's necessary to pre-read classification data for direct impact patterns, mediating transmission patterns, and collaborative amplification patterns from a multi-level transaction behavior causal graph, and use these as the semantic mapping execution objects. A pre-built semantic rule base is invoked to match the triggering conditions and transaction behavior semantic features for each path pattern, establishing a mapping index table from path pattern to label. For direct impact patterns, the triggering conditions for high-frequency bidding behavior are retrieved from the rule base and mapped to the label "High-frequency bidding drives short-term premium"; for mediating transmission patterns, the conditional constraints of bulk purchasing triggering inventory expectation adjustments are retrieved from the rule base and mapped to the label "Bulk purchasing triggers inventory expectation adjustments"; for collaborative amplification patterns, the collaborative conditions of cross-platform arbitrage behavior and policy announcement responses are retrieved from the rule base and mapped to the label "Cross-platform arbitrage superimposed with policy response". These mapped labels are then bound to corresponding path instances, generating unlabeled path objects containing differentiated semantic descriptions. This binding structure unifies the structural information and business semantic information of the path patterns into the same data record, enabling direct invocation by the subsequent path distillation encoder. By calling the matching operation and label mapping processing method of the pre-set semantic rule library, the classification path group of the previous step is transformed into a path object to be labeled with semantic interpretation capability, thereby achieving the semantic processing effect of causal path set.

[0099] For example, in a set of transaction data, the starting node directly affecting the pattern path is the high-frequency bidding behavior cluster, the intermediate node is empty, and the ending node is the price trend response cluster. The triggering condition for this path in the rule base is defined as a bidding frequency greater than 50 times / hour and a bidding amount exceeding 0.4% of the total daily transaction amount. After matching, the mapping label is determined as "high-frequency bidding drives short-term premium". The starting node of the intermediary transmission pattern path is the bulk purchasing behavior cluster, the intermediate node is the inventory expectation adjustment cluster, and the ending node is the price trend response cluster. The triggering condition for bulk purchasing in the rule base is a single order quantity exceeding 500 units and an inventory adjustment frequency within 5 days. After matching, the label is mapped as "bulk purchasing triggers inventory expectation adjustment". The starting node of the collaborative amplification pattern path is the cross-platform arbitrage behavior cluster, the intermediate node is the policy announcement response cluster, and the ending node is the price trend response cluster. The triggering condition for this pattern in the rule base is an arbitrage profit ratio exceeding 0.2% and a policy announcement being released within 24 hours of the arbitrage behavior occurring. After matching, the label is mapped as "cross-platform arbitrage superimposed with policy response". After binding the three types of path objects to the original topology data and label descriptions, a set of path objects to be labeled is generated for the path distillation encoder to call directly. During the verification process, the correspondence between path labels and transaction behavior patterns is accurate, which improves the model's causal explanation ability and the operability of trend prediction.

[0100] S4.5: Integrate the path objects to be labeled with differentiated semantic descriptions and their corresponding original topological structure data, perform structured encapsulation processing to bind path topological information and semantic description information, and finally generate a set of semantically labeled causal paths that can be directly called by the path distillation encoder.

[0101] The process involves taking the path objects with differentiated semantic descriptions and their corresponding original topology data as input, calling the topology relation parser to extract the start point, end point, and all intermediate node numbers of each path, forming a node sequence index set. The node sequence index set is then processed to extract associated edge weights, mapping the weights and directions of each directed edge on the path to an edge attribute list, thus constructing a complete path structure parameter set. Key-value binding is performed between the semantic description fields and the path structure parameter set. A structured encapsulation strategy combines the node sequence index set, edge attribute list, and semantic labels into a unified data object, marking the path order and pattern category within the object to ensure the encoder can directly identify the path type and causal transmission meaning. The encapsulated data object undergoes format validation, including field integrity verification, index validity verification, and consistency verification between semantic labels and pattern categories, eliminating non-compliant path objects to ensure data quality. Finally, the validated encapsulated data objects are sorted according to a preset index and stored in the path call interface buffer, serving as direct input for the path distillation encoder. Through the above-mentioned structured mapping and parameter binding processing, the differentiated semantic description and original topological information of the previous step are transformed into a set of semantically annotated causal paths with a unified calling interface and conforming to the same coding standard, thereby achieving efficient compatibility and enhanced semantic interpretability in the encoder input stage.

[0102] For example, in the causal graph of trading behavior, a direct influence path of the "high-frequency bidding driving short-term premium" pattern is selected, with its node sequence being [node ID12, node ID45]. The starting node represents the high-frequency bidding behavior cluster, and the ending node is the price trend response node, with a path order of 1. In the original topology data, the directed edge weight of this path is 0.82, the direction is from node ID12 to node ID45, and the pattern category is direct influence. The parser extracts the node sequence index set {12, 45}, the edge attribute list {weight: 0.82, direction: 12→45}, and the semantic label is "high-frequency bidding driving short-term premium". Structured encapsulation is performed, combining node information, edge attributes, order markers, and semantic labels into a unified object: Object_A. During format verification, it is detected that all node IDs exist in the global graph node index table, the weight range is between 0 and 1 and the direction is valid, and the semantic label matches the pattern category. Object_A is stored in the call interface buffer, and the encoder can directly read its node sequence and label fields to infer the path probability distribution. In this path set, the encoder can significantly improve the accuracy of interpreting short-term premiums in high-frequency bidding behavior during the training phase, and significantly improve the matching degree of trend prediction pattern output in the test scenario, achieving simultaneous optimization of interpretability and execution efficiency.

[0103] Step S5: Construct a path distillation encoder based on the semantically labeled causal path set, mapping the original transaction sequence to a path activation probability distribution. Then, through training with expert rule combinations in a teacher-student architecture and KL divergence constraints on a lightweight graph neural network structure, generate a path activation probability distribution vector that implicitly learns causal transmission logic. Specifically, this includes: S5.1: Obtain the set of causal paths with semantic annotations and the original transaction sequence, and define the logical formalization of various typical path patterns based on the combination of expert rules to generate an expert knowledge constraint set containing deterministic transmission rules in order to establish the behavioral benchmark of the teacher model.

[0104] In this specific embodiment, the raw transaction sequence refers to a set of unprocessed, original transaction records directly extracted from a specified transaction data source, such as an exchange market data database or order book log, within a preset target observation time window. The acquisition method is as follows: based on the start and end times set in the process, all transaction events occurring within that window are queried from the data source and output. Each record contains core fields such as timestamp, transaction type identifier, transaction price, and quantity, and is strictly arranged in ascending order of timestamp.

[0105] The semantically labeled causal path set and the corresponding original transaction sequence output from step S4 are used as input conditions. A path pattern index structure is established by calling the node topology information and semantic label data from the causal path set to support rapid retrieval and matching of path patterns. For each typical path pattern, key elements such as node roles, edge direction attributes, and propagation order within the path are extracted based on a pre-built trend-driven logic library to construct a combination of logical rules, including path triggering conditions, propagation conditions, and response conditions. Using formal definition methods, the condition combination is transformed into a set of Boolean logic expressions, condition operators, and threshold parameters. An executable expert rule instance is generated through a rule compiler, and this instance set constitutes a preliminary expert knowledge constraint set.

[0106] A consistency check is performed on the initial expert knowledge constraint set. A rule conflict detection algorithm is used to eliminate mutually exclusive and duplicate conditions, and threshold standardization is applied to maintain the consistency of parameter ranges, resulting in a corrected expert knowledge constraint set. The corrected expert knowledge constraint set is then loaded into the path matching engine of the teacher model to establish the behavioral benchmark of the teacher model in the path distillation encoder. This benchmark defines the causal transmission expectation of the teacher model, ensuring that subsequent student model training can be aligned with clear and verifiable standards.

[0107] By formalizing expert rules and performing consistency checks, the semantically annotated causal path set from the previous step is transformed into an executable set of expert knowledge constraints, thereby enabling the teacher model to set standardized causal transmission logic in the path distillation encoder.

[0108] For example, in a data product trading market, the input set of semantically labeled causal paths includes three typical patterns: the direct impact pattern path is "high-frequency bidding → short-term premium," the intermediary transmission pattern path is "bulk purchasing → inventory expectation adjustment → medium-term price correction," and the collaborative amplification pattern path is "cross-platform arbitrage behavior + policy announcement response → market consensus reconstruction → long-term trend inflection point." The original transaction sequence includes 90 consecutive days of timestamp records, transaction amounts, order types, participant identities, and price level changes. For the direct impact pattern, the extracted node roles are bidding behavior nodes and price response nodes, the edge direction attribute is positive time, the transmission order is 1, and the constructed logical rule expression is: Where Δp is a preset price change threshold. This expression is compiled into a teacher model rule instance, setting the bidding frequency threshold to ≥20 times per day and the price change rate threshold to ≥0.5 yuan / day. For the intermediary transmission model, the node roles are extracted in the order of purchasing behavior node, inventory expectation node, and price adjustment node, and the logical rule expression is constructed as follows: Where Q is the purchase quantity threshold, Δk is the inventory reduction threshold, and Δm is the price adjustment threshold. The example parameters are set as Q≥1000 units, Δk≥5%, and Δm≥2 yuan.

[0109] For the collaborative amplification mode, node roles are extracted as arbitrage behavior nodes, policy response nodes, market consensus reconstruction nodes, and long-term trend nodes. The logical rule is that the arbitrage trigger condition and the policy announcement response condition must be met simultaneously, and the long-term trend change rate must exceed Δl, with Δl ≥ 1.5 yuan / month. Conflict detection is performed on all rule instances, revealing a shared node conflict between the market consensus reconstruction condition in the collaborative amplification mode and the inventory expectation decline condition in the intermediary transmission mode. This conflict is resolved by adjusting the threshold range and splitting the condition operators. The expert knowledge constraint set, after consistency verification, is loaded into the teacher model path matching engine, forming executable rule standards for each typical mode. This achieves the setting of the causal transmission logic benchmark for the teacher model and ensures that the subsequent student model path activation probability distribution training remains consistent with this benchmark. The output expert knowledge constraint set significantly improves the accuracy and stability of path trigger judgment in market fluctuation test scenarios.

[0110] S5.2: Use the expert knowledge constraint set to perform rule matching operation on the original transaction sequence, calculate the theoretical triggering intensity of each causal path within a specific time window, and generate a teacher path distribution vector that reflects the ideal causal transmission mechanism, which serves as the target truth value for student model learning.

[0111] Specifically, for each causal path and its corresponding logical formal definition, transaction sequence data is scanned within a given specific time window. When the combination of events in the sequence satisfies the triggering condition of that path, a valid match is recorded. By counting the number of all valid matches within the window and combining this with the weighting coefficients defined in the rules, the theoretical triggering strength of that path within this window is calculated. After normalization, the theoretical triggering strengths of all paths constitute the teacher path distribution vector. This vector reflects the strength distribution of the ideal causal transmission mechanism defined by prior expert knowledge and will serve as a supervisory signal to guide the parameter optimization of the student model.

[0112] S5.3: Construct a lightweight graph neural network structure based on the original transaction sequence, map the behavior nodes to the initial node embedding representation, and aggregate neighborhood information through a multi-layer message passing mechanism to generate a student path latent feature vector containing local topological features.

[0113] Based on the original transaction sequence and the corresponding behavior node data in the semantically labeled causal path set, a lightweight graph neural network structure is constructed as the input layer. A node embedding initialization mechanism is used to map each behavior node to an initial embedding vector of fixed dimension. This mapping process can be achieved by looking up the embedding matrix. In the initialization stage, the embedding value is determined by using historical transaction frequency and behavior category label together.

[0114] Based on the initial embedding vector, a graph neural network structure with a multi-layer message passing mechanism is configured. Each layer contains a neighborhood information aggregation operator and a node feature update operator. The neighborhood information aggregation operator performs weighted summation aggregation based on the directed edge weights in the causal path set. The node feature update operator maps the aggregation result and the current node embedding vector to a new node feature representation through a non-linear activation function.

[0115] In the message passing process, a path order-aware masking matrix is ​​introduced to perform a masking operation on edges that do not match the target path order in the neighborhood aggregation operation, retaining only the neighborhood features that conform to the current order pattern, thereby obtaining feature propagation results with path pattern selectivity.

[0116] After multiple layers of message passing are stacked, a global pooling operation is performed on the node features to aggregate the node features belonging to the same path into a path-level representation vector. The pooling method can be average pooling or attention-weighted pooling, and the attention weight is calculated by the matching degree between the path semantic label and the node features.

[0117] The generated path-level representation vector is processed by feature concatenation and linear transformation. The concatenated high-dimensional features are mapped to low-dimensional student path latent feature vectors through weight matrix multiplication and bias vector addition, which serve as direct inputs for subsequent probability normalization and divergence calculation.

[0118] Through the above-mentioned multi-layer message passing and path-level pooling processing, the original transaction sequence node data of the previous step is transformed into a student path latent feature vector containing local topological features of causal paths, thereby realizing the causal structure encoding of the input end of the path activation probability distribution.

[0119] For example, in a data product trading market scenario, the original transaction sequence contains 500 behavioral nodes. The historical transaction frequency of each node is quantized as an integer between 0 and 50, and the behavioral category label is encoded by three patterns as 0, 1, and 2. The initial node embedding dimension is set to 32, and the embedding matrix is ​​randomly initialized and then adjusted during training. A lightweight graph neural network structure is configured with a 3-layer message passing mechanism. Each layer's neighborhood aggregation operator performs operations according to the directed edge weights, with edge weights ranging from 0.1 to 1.0. The path order-aware mask matrix is ​​generated by three order patterns, and the masking ratio is dynamically adjusted according to the order matching condition. Global pooling adopts an attention-weighted approach, and the attention weight calculation formula is: in, This represents the attention weight corresponding to path p. For trainable weight vectors, The node feature vector is represented by a 64-dimensional path-level representation vector obtained after global pooling. This vector is then linearly transformed and mapped to a 16-dimensional student path latent feature vector. In the validation set, this latent feature significantly improves the causal consistency of subsequent probability mappings, demonstrating that the trend prediction model maintains a stable distribution of activated paths under policy change scenarios.

[0120] S5.4: Perform probability normalization on the student path latent feature vector to map it to a student path activation probability distribution that conforms to the probability axiom, and calculate the difference metric between the student path activation probability distribution and the teacher path distribution vector to generate a divergence loss scalar for measuring causal logical consistency.

[0121] The student path latent feature vector generated by the lightweight graph neural network structure is input into the normalization processing module. The Softmax normalization operator is used to perform exponential scaling on the activity of each dimension in the vector to ensure that all components satisfy the probability axiom and the sum is one.

[0122] The normalized student path activation probability distribution vector and teacher path distribution vector are input into the difference calculation module. The difference is defined using the KL divergence formula, and the difference components are extracted for each path category and summed cumulatively.

[0123] The formula for calculating KL divergence is: in This represents the theoretical trigger probability value of the i-th path in the teacher path distribution vector. This represents the model prediction probability of the i-th path in the student path activation probability distribution.

[0124] The cumulative value of the output difference is used as a divergence loss scalar, which measures the fit between the student model's causal logic expression and the teacher model, and serves as the input for subsequent gradient backpropagation.

[0125] By using a combined approach of normalization and difference measurement, the latent feature vector generated in the previous step is transformed into a student path activation probability distribution that conforms to probability axioms and can be further evaluated for causal logic consistency, thereby achieving a quantitative assessment of the effectiveness of causal transmission explanatory learning.

[0126] For example, in a data product market transaction behavior prediction scenario, the student path latent feature vector has a length of 20, and the value of each component ranges from... The normalization process uses the Softmax operator to calculate the exponent value of each component and sum them up to 173.4. Dividing the exponents of each component by the sum yields a 20-dimensional probability distribution, with the largest probability component being 0.145 and the smallest being 0.021. The teacher path distribution vector is obtained through rule matching, with the largest probability component being 0.152 and the smallest being 0.019. Substituting the normalized student and teacher distributions into the KL divergence formula, the calculated divergence loss scalar is 0.0037, indicating that the student model has a very small deviation from the causal logic of the teacher model. This loss value can be used for gradient backpropagation optimization in substep S5.5, significantly improving the ability to preserve the causal mechanism in trend prediction generation in this scenario.

[0127] S5.5: Backpropagate the gradient signal based on the divergence loss scalar to update the connection weight parameters of the lightweight graph neural network structure, iteratively optimize until the student path activation probability distribution converges to the neighborhood space of the teacher path distribution vector, and finally output the path activation probability distribution vector of implicitly learned causal transmission logic.

[0128] Based on the divergence loss scalar output from the previous sub-step S5.4 as the optimization driving signal, a gradient backpropagation path is constructed to transmit error information to the connection weight parameters of each layer in the lightweight graph neural network structure. Partial derivatives are calculated on the divergence loss scalar to extract the gradient tensor for each layer's network weight matrix. This gradient tensor is then matched with the node embedding representation of the current layer to form a precise mapping of the weight update increment. The gradient tensor is scaled using a preset learning rate coefficient to generate a weight update vector with amplitude adjustment characteristics. Weight regularization operators are used to suppress update oscillations caused by excessive gradients, improving network stability. This weight update vector is added to the weight parameter matrix of the current iteration to form the updated network weight configuration. Immediately after the parameter update, the next round of path activation probability distribution inference is performed. The convergence of the KL divergence values ​​between the student path activation probability distribution and the teacher path distribution vector is monitored during continuous iterations. When the convergence speed falls below a preset threshold, an adaptive adjustment process for fine-tuning the learning rate and regularization coefficients is triggered to further shorten the iteration convergence cycle. By freezing the weight parameters when the iteration termination condition is met, the final student path activation probability distribution vector is fixed as the encoded output of the implicitly learned causal transmission logic, thus achieving a stable mapping from the original transaction sequence to the causal path activation probability space.

[0129] By using gradient backpropagation and weight iteration optimization, the divergence loss result from the previous step is transformed into accurate weight update data for a lightweight graph neural network structure, achieving the expected technical effect of converging the student model path activation probability distribution vector to the neighborhood space of the teacher path distribution vector.

[0130] For example, in a training scenario containing 120 action nodes and three typical path patterns, the initial value of the divergence loss scalar is 0.85. The current teacher path distribution P is calculated using the KL divergence formula. t With student path distribution P s The difference value was determined by setting a learning rate of 0.01 and an L2 regularization coefficient of 0.001, and reducing the divergence value to 0.12 through 5 iterations. During the gradient update process, the gradient of the first-layer weight matrix W1 is denoted as G1, and after scaling, the update vector ΔW1 is obtained. After cumulative iterations, the weight magnitude of W1 remains within the range of [-0.5, 0.5], avoiding gradient explosion. In the final output student path activation probability distribution vector, the activation probability of the first-order path is concentrated between 0.72 and 0.78, the activation probability of the second-order path is concentrated between 0.63 and 0.69, and the activation probability of the third-order path is concentrated between 0.54 and 0.58. Moreover, the path pattern is stably triggered in the prediction stage, which accelerates the accurate generation of subsequent trend pattern labels.

[0131] Step S6: Input the path activation probability distribution vector into the trend prediction generation process, calculate the activation intensity sequence of each path, and combine it with a preset path effect mapping table to map different path combinations into trend pattern label sequences containing accelerated upward movement, plateau oscillation, or structural reversal patterns. Specifically, this includes: S6.1: Obtain the path activation probability distribution vector as input data, and perform weighted aggregation operation on the multidimensional probability components in the vector based on the time sliding window mechanism to generate a path activation intensity sequence that represents the degree of causal transmission activity at different time granularities.

[0132] S6.2: Based on the path activation intensity sequence, extract three types of key feature parameters: first-order direct influence intensity value, second-order mediation transmission intensity value, and third-order collaborative amplification intensity value. Use a multi-dimensional feature splicing algorithm to construct a composite path feature vector containing causal contribution information of all orders.

[0133] Based on the input data of the path activation intensity sequence, the predefined order partitioning retrieval module is called to divide the sequence into three subsequences according to the causal path order label: first-order direct influence, second-order mediation transmission, and third-order synergistic amplification. The time index synchronization relationship is maintained to ensure the comparability of data of different orders.

[0134] For each first-order subsequence, perform statistical operations on the mean and maximum value. Use the mean as the benchmark quantity that directly affects the intensity value, and the maximum value as a supplementary reference indicator for short-term strong response, forming a dual-indicator system for the direct influence intensity parameter.

[0135] The sliding window integral operation is performed on the second-order subsequence. The integral result is used to characterize the cumulative contribution of mediation transmission. At the same time, the variance within the window is calculated to reflect the stability of the transmission path. The two are combined into the mediation transmission strength parameter.

[0136] Cross-correlation coefficients are calculated for the third-order subsequences to extract the synchronicity features of the coordinated amplification. The peak detection algorithm is then used to identify multipath resonance points, and the resonance amplitude value is used as the coordinated amplification intensity parameter.

[0137] The intensity parameters of the three orders are input into the multidimensional feature concatenation algorithm in a fixed order. The matrix is ​​concatenated column by column to form a composite path feature vector containing causal contribution information of all orders, ensuring that the vector structure is consistent with the index of the path effect mapping table.

[0138] By using a multi-dimensional feature concatenation process, the path activation intensity sequence from the previous step is transformed into a composite path feature vector, achieving a unified quantitative expression of causal contributions across all orders, and providing stable input features for subsequent trend pattern rule matching.

[0139] S6.3: Read the rule base defined in the pre-stored path effect mapping table, and perform logical comparison processing between the composite path feature vector and the threshold conditions in each rule base to filter out the set of candidate trend pattern rules triggered under the current market conditions.

[0140] The input includes the composite path feature vector output by step S6.2 and the rule base index information of the pre-stored path effect mapping table. The rule base records the path combination threshold conditions corresponding to different trend pattern labels.

[0141] The key feature parameters of each order in the composite path feature vector are loaded into the logical comparison module, and the threshold condition set of each rule in the mapping table is read in sequence. The threshold conditions include numerical boundaries such as the first-order direct influence strength, the second-order mediation transmission strength, and the third-order cooperative amplification strength.

[0142] For each rule's threshold condition set, a condition matching operation is performed. The corresponding intensity value is judged to meet the interval conditions set in the rule base by using a feature-by-feature comparison method. The matching result is marked as triggered or not triggered in the internal state matrix.

[0143] The rule entries marked as triggered in the state matrix are summarized by rule number to construct a set of candidate trend pattern rules under the current market state, and the trigger time slice index of each rule is recorded for subsequent priority sorting.

[0144] By using a threshold satisfaction rate statistical method, rules with insufficient trigger counts are filtered out, while highly stable rules that are triggered in most time slices are retained, forming the final candidate trend pattern rule set.

[0145] This comparison process transforms the composite path feature vector from the previous step into a set of candidate trend pattern rules under the current market conditions, achieving the technical effect of initial screening of trend pattern labels.

[0146] For example, in a data product market analysis scenario, the numerical configuration of the composite path feature vector is as follows: first-order direct influence strength value is 0.78, second-order mediation transmission strength value is 0.65, and third-order collaborative amplification strength value is 0.49. The threshold conditions for candidate rule R1 in the pre-stored path effect mapping table are: first-order ≥ 0.75, second-order ≥ 0.60, and third-order ≥ 0.45; the threshold conditions for rule R2 are: first-order ≥ 0.70, second-order ≥ 0.65, and third-order ≥ 0.55. During the comparison process, all three conditions for rule R1 are met, and the state matrix is ​​marked as triggered; the third-order condition for rule R2 is not met, and the state matrix is ​​marked as not triggered. Statistical results show that R1 is triggered in all five consecutive time slices, exhibiting high triggering stability, and is included in the candidate trend pattern rule set. Inputting this set into the subsequent conflict resolution step can significantly improve the accuracy of trend label selection.

[0147] S6.4: Perform conflict resolution and priority sorting operations on the candidate trend pattern rule set, select the optimal matching rule based on the confidence score of each rule, and thus determine the unique target trend pattern label category.

[0148] S6.5: Generate result identifiers containing semantic information of accelerated upward movement, plateau oscillation, or structural reversal based on the target trend pattern label category, and assemble the result identifiers into a complete time-seriesd trend pattern label sequence as the final output.

[0149] The unique target trend pattern label category determined in the preceding step S6.4 is used as the initial input. The pattern label encoding module is invoked to map this category into a trend pattern label data object containing explicit semantic fields and structural identifiers. The trend pattern label data object undergoes label semantic parsing processing, splitting the label category into three core semantic parameters: trend direction field, fluctuation characteristic field, and structural inflection point field. Each field is assigned a preset semantic value; for example, the "accelerating upward" category is assigned a positive value for the direction field, a high amplitude value for the fluctuation characteristic field, and an none value for the structural inflection point field. The parsed trend pattern label data objects are arranged in ascending order of timestamp information using a serialization sorting engine, forming a chronological sequence of label units. The sequence synthesis processing logic is invoked to format and encapsulate the label unit sequence, attaching a unique time index identifier to each label unit to ensure accurate reproduction of the label-time slice correspondence during subsequent decoder calls. Through this multi-stage processing method, the target trend pattern label category from the previous step is transformed into a structured, serialized time-series trend pattern label sequence, achieving a complete and callable output of the trend prediction semantic results.

[0150] For example, in a prediction scenario based on a data product trading market, the target trend pattern label category output by the preceding steps is "structural reversal". The label encoding module maps this category to a trend direction field = negative, a volatility characteristic field = high amplitude, and a structural inflection point field = present. The parsing module sets a value of -1 in the direction field, an amplitude value of 1.8 in the volatility characteristic field, and a state value of 1 in the structural inflection point field. These values ​​are all semantic numerical codes preset for this scenario. The sequence sorting engine generates 30 label units according to the trading data range from the timestamp 2024-05-01 to 2024-06-30. Each unit contains a direction value, an amplitude value, and an inflection point state value. The synthesis processing logic sorts these units by time and encapsulates them into a matrix form T = t × p, where t is the total number of time slices (30), and p is the parameter dimension (3) of each label unit, ultimately forming a 30 × 3 trend pattern label sequence matrix. This matrix, as input to the trend pattern decoder, can significantly improve the consistency of trend prediction in terms of time dimension and semantic interpretability, and has achieved accurate prediction of market price turning points in actual verification.

[0151] Step S7: Based on the trend pattern label sequence and historical price level data, perform interval regression calculation through the trend pattern decoder to generate a price trend interval prediction result with confidence interval. Specifically, this includes: S7.1: Obtain the trend pattern label sequence output from the previous steps and the preprocessed historical price level data. Perform sliding window segmentation on the historical price level data to generate multi-scale time slice price tensors, and map the trend pattern label sequence to the corresponding pattern constraint vector, thereby obtaining a fused input dataset containing time series features and semantic constraints.

[0152] S7.2: Based on the fused input dataset, construct a dynamic weight allocation network inside the trend pattern decoder, use the attention mechanism to calculate the influence coefficient of each trend pattern label in the pattern constraint vector on the current time slice price tensor, and generate a dynamic weight matrix that represents the price fluctuation sensitivity under different patterns, so as to quantify the moderating effect of semantic labels on numerical prediction.

[0153] The fused input dataset is input into the dynamic weight allocation network inside the trend pattern decoder to establish a dual-input structure of pattern constraint vector and multi-scale time slice price tensor.

[0154] A multi-head attention mechanism is configured within the network structure, using different trend pattern labels in the pattern constraint vector as query vectors, and multi-scale features of the time slice price tensor as key and value vectors.

[0155] For each attention head, a similarity calculation is performed, and the matching score between each morphological label and the time slice feature is calculated using the dot product form and the following formula: Where Q represents the query matrix in the morphological constraint vector, K represents the key matrix of the price tensor features, and d represents the dimension of the key vector.

[0156] The matching scores are normalized using softmax to form the attention weight distribution of each trend pattern label under different time slices, ensuring that the sum of the weights is 1.

[0157] By using the normalized attention weights and the price tensor value vector V for weighted summation, the weighted response value of each trend pattern label to the price characteristics of different time slices is obtained.

[0158] The weighted response values ​​of all trend pattern labels are assembled into a dynamic weight matrix in label order, where each row corresponds to a trend pattern label and each column corresponds to a time slice feature dimension.

[0159] By using a dynamic weight matrix, the fused input dataset from the previous step is transformed into a numerical matrix that can quantify the price volatility sensitivity under different trend patterns, thus enabling explicit modeling of the moderating effect of semantic labels on numerical prediction.

[0160] S7.3: The dynamic weight matrix is ​​used to perform weighted feature transformation on the multi-scale time slice price tensor, and the optimization iterative operation of the quantile regression loss function is performed to calculate the conditional quantile parameters corresponding to different confidence levels, thereby obtaining the initial quantile estimate sequence representing the upper and lower boundaries of the price distribution.

[0161] Based on the correspondence between the dynamic weight matrix and the multi-scale time slice price tensor, a weighting mapping is established. Each element in the dynamic weight matrix is ​​multiplied element-wise with the corresponding time slice feature vector in the price tensor to form a modulated multi-scale weighted feature tensor.

[0162] The multi-scale weighted feature tensor is subjected to scale normalization to adjust the weighting results at different time scales to a uniform range for subsequent regression calculations. The normalization factor can be determined based on the standard deviation or mean absolute deviation of the price series at each scale.

[0163] Based on the normalized weighted feature tensor, the input matrix of the quantile regression model is constructed. The time series features at each scale are concatenated into a high-dimensional vector structure, which serves as the set of independent variables for quantile regression, and the actual historical price values ​​are used as the set of dependent variables.

[0164] The quantile regression loss function is set to minimize the conditional quantile error. Iterative optimization is performed on the high-dimensional input matrix using this loss function. The regression coefficients are continuously adjusted by gradient descent or quasi-Newton method. After the optimization process is completed, each regression coefficient is applied to the fused input matrix to calculate the corresponding conditional quantile prediction value, forming an initial quantile estimate sequence containing the upper and lower boundaries of price prediction at multiple confidence levels.

[0165] Through the above-mentioned optimization iteration and calculation methods, the multi-scale time-slice price features modulated by the dynamic weight matrix are transformed into statistically significant conditional quantile parameters, thereby realizing the construction of the initial price distribution boundary sequence.

[0166] For example, in a market forecasting scenario for a data product, the historical price time series is set to 240 trading days, with a sliding window length of 20 days and a step size of 5 days, forming 12 time-slice scale price tensors. The dynamic weight matrix is ​​12×3 in size, where 3 represents the trend pattern label category. In the normalization process, the weight effect of each scale is divided by the standard deviation of the corresponding time-slice price tensor, with the standard deviation ranging from 15 to 45 at different scales. The dimension of the concatenated input matrix is ​​36. The quantile levels τ are set to 0.1, 0.5, and 0.9. A quantile regression loss function is used, and 200 gradient descent iterations are performed with a learning rate of 0.01. After convergence, the regression coefficient matrix is ​​36×3 in size, with each column corresponding to a τ value. Multiplying the regression coefficients by the input matrix yields three conditional quantile estimation sequences: the 0.1 quantile sequence represents the lower limit of price prediction, the 0.5 quantile sequence represents the central trend, and the 0.9 quantile sequence represents the upper limit of prediction. This process significantly improves the boundary accuracy of the prediction interval under multiple trend combinations.

[0167] S7.4: Based on the initial quantile estimate sequence, perform monotonicity verification and smoothing correction processing, use the shape-preserving prediction algorithm combined with historical residual distribution statistics to calculate the coverage probability correction factor, and generate calibrated price fluctuation boundary correction parameters to ensure the statistically effective coverage of the prediction interval.

[0168] S7.5: Perform final offset superposition operation on the initial quantile estimate sequence according to the price fluctuation boundary correction parameter, synthesize a triplet data structure containing the lower limit threshold, the center prediction value and the upper limit threshold, and output the price trend interval prediction result with a clear confidence interval, thus completing the complete mapping from pattern recognition to quantitative interval generation.

[0169] Based on the calibrated price fluctuation boundary correction parameters and the initial quantile estimate sequence, the upper and lower limit quantiles and center prediction values ​​for corresponding time slices are selected as calculation objects to construct an offset superposition calculation model to generate the final prediction interval. For each time slice, the initial upper limit quantile estimate is weighted by correction parameters, and the corrected offset is synthesized with the original value using vector element-level addition to obtain the updated price upper limit threshold sequence. The same correction parameter superposition operation is performed on the initial lower limit quantile estimate for each time slice to obtain the updated price lower limit threshold sequence. Offset neutralization is performed on the sequence corresponding to the historical price center prediction values, i.e., the correction parameters are split according to positive and negative polarities and evenly distributed to the center prediction values ​​to avoid the overall bias introduced by the upper and lower limit boundary corrections, thus obtaining the updated center prediction value sequence. A triple combiner is constructed to encapsulate the updated lower limit threshold, center prediction value, and upper limit threshold into structured triple data units in chronological order, ensuring that the internal boundary relationships of each triple satisfy the strict inequality of lower limit < center < upper limit. The triplet data units are serialized and output as price trend range prediction results, along with corresponding confidence level labels for each range, to achieve a precise one-to-one mapping between trend pattern labels and quantitative ranges.

[0170] By using the above offset superposition and triplet construction processing method, the shape constraint and price boundary correction results of the previous step are transformed into interval prediction data that can be directly used for quantitative analysis and risk assessment, thus achieving the complete technical effect of price trend prediction from qualitative shape recognition to quantitative interval generation.

[0171] For example, in a data product price prediction scenario, the 90% confidence upper limit sequence of the initial quantile estimates is 120.5, 121.0, 121.8, 122.3, and 122.9 for the last five consecutive trading days, with corresponding shape-preserving prediction correction parameters of 0.8, 0.6, 0.7, 0.5, and 0.4; the initial lower limit sequence is 118.2, 118.6, 119.0, 119.4, and 119.8, with corresponding correction parameters of -0.5, -0.4, -0.6, -0.3, and -0.2; and the center prediction sequence is 119.3, 119.8, 120.4, 120.8, and 121.3. When performing offset overlay calculations, the upper limit is calculated using the following formula: in, This is the initial upper limit value. To correct for parameter offset, the lower limit is set using the following formula: in, This is the initial lower limit value. To correct for parameter offsets, the center prediction value is corrected using the following formula: in, The initial center prediction value is used. Substituting the above parameters into the formula sequentially yields the updated upper limit sequence of 121.3, 121.6, 122.5, 122.8, and 123.3; the lower limit sequence of 117.7, 118.2, 118.4, 119.1, and 119.6; and the center prediction value sequence of 119.15, 119.55, 120.25, 120.75, and 121.25. The triples for each time slice are output in the format of [lower limit, center, upper limit]. For example, the triples for day one are [117.7, 119.15, 121.3], and the triples for day five are [119.6, 121.25, 123.3]. This prediction result significantly improves the coverage and stability of the price prediction range in practical verification, enabling the prediction to maintain a high degree of matching between the payment price range and the actual price range even under multiple market intervention scenarios.

[0172] Step S8: Using the deviation between the price trend range prediction result and the actual market feedback data, the contribution fluctuation threshold of the causal stability evaluator and the mapping weights of the path effect mapping table are corrected in reverse, completing the dynamic closed-loop optimization of the model parameters. Specifically, this includes: S8.1: Obtain the price trend range prediction results and actual market feedback data, perform difference calculation on the two to generate a prediction deviation sequence, and calculate the root mean square error index based on the prediction deviation sequence as the basis for model performance quantification.

[0173] S8.2: Based on the model performance quantification, perform gradient backpropagation to map the prediction deviation sequence to the parameter adjustment amount of the causal stability evaluator, so as to generate an initial correction vector for the contribution fluctuation threshold.

[0174] S8.3: The initial correction vector is used to dynamically update the sliding window monitoring logic inside the causal stability evaluator, and the contribution fluctuation judgment benchmark of each directed edge is reset to output an updated contribution fluctuation threshold with adaptive characteristics.

[0175] S8.4: Based on the distribution error of different trend pattern labels in the predicted deviation sequence, perform weighted least squares regression analysis to calculate the weight offset of each path combination to the trend pattern label in the path effect mapping table, thereby generating the mapping weight matrix to be updated.

[0176] S8.5: Based on the mapping weight matrix to be updated, perform a full replacement operation on the path effect mapping table, and synchronize the conversion rules of mapping different path combinations to trend pattern labels to the trend prediction generation process, so as to complete the dynamic closed-loop optimization of model parameters including the updated contribution fluctuation threshold and the updated mapping weight matrix.

[0177] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0178] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0179] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A data product pricing trend prediction method based on transaction behavior clustering analysis, characterized in that, Specifically, it includes: S1: Obtain the data of each behavior cluster output by the transaction behavior clustering module, and generate a coarse-grained transaction behavior causal graph based on unsupervised indicators; S2: Based on the coarse-grained causal graph of trading behavior and the pre-built causal stability evaluator, a sliding window is used to monitor the consistency of the predictive contribution of each directed edge to price trend changes in a continuous period, and when the contribution fluctuation exceeds a set threshold, weight decay is triggered to generate the identifier of the subgraph region to be reconstructed. S3: Perform causal edge identification operation on the identifier of the subgraph region to be reconstructed, update the local graph structure connection relationship, and generate a multi-level transaction behavior causal graph. S4: Define the causal path order based on the shortest number of hops from any behavior node to the price trend response node in the multi-order transaction behavior causal graph, identify typical paths and assign differentiated semantic labels to generate a causal path set; S5: Construct a path distillation encoder based on the causal path set, map the original transaction sequence to a path activation probability distribution, and generate a path activation probability distribution vector through expert rule combination in the teacher-student architecture and KL divergence constraint training of the graph neural network structure. S6: Based on the path activation probability distribution vector, calculate the activation intensity sequence of each order path, and combine it with the preset path effect mapping table to map different path combinations into trend pattern label sequences. S7: Based on the trend pattern label sequence and historical price level data, perform interval regression calculation to generate price trend interval prediction results.

2. The data product pricing trend prediction method based on transaction behavior clustering analysis according to claim 1, characterized in that, After step S7, the following is included: S8: Using the deviation between the price trend range prediction results and the actual market feedback data, the contribution fluctuation threshold of the causal stability evaluator and the mapping weight of the path effect mapping table are reversed.

3. The data product pricing trend prediction method based on transaction behavior clustering analysis according to claim 1, characterized in that, The coarse-grained transaction behavior causal graph includes an initial set of behavior nodes and an initial set of directed edges.

4. The data product pricing trend prediction method based on transaction behavior clustering analysis according to claim 3, characterized in that, Step S2 specifically includes: Obtain the initial set of directed edges and the corresponding historical transaction sequence data, and divide the continuous time segments into multiple equal-length sliding windows based on the time adjacency index to generate a windowed transaction behavior data stream; For each initial directed edge in the windowed trading behavior data stream, the Granger causality test algorithm is used to calculate its predictive contribution to the price trend change target within the current sliding window, generating an edge contribution sequence. Based on the edge contribution sequence, the rolling variance statistical method is used to calculate the fluctuation dispersion of the predicted contribution values ​​among multiple consecutive sliding windows, and an edge stability evaluation coefficient is generated. If the edge stability evaluation coefficient exceeds the preset contribution fluctuation threshold, the initial directed edge is subjected to exponential smoothing weight decay processing to generate an adjusted edge weight set. Based on the initial directed edges and their adjacent node topological relationships in the adjusted edge weight set after the exponential smoothing weight decay process, a connected subgraph region is aggregated and marked as the identifier of the subgraph region to be reconstructed.

5. The data product pricing trend prediction method based on transaction behavior clustering analysis according to claim 1, characterized in that, Step S4 specifically includes: Obtain all behavioral nodes and price trend response nodes in the multi-level trading behavior causal graph, and use the breadth-first search algorithm to traverse and calculate the connection links from any behavioral node to the price trend response node in the graph to generate the original set of reachable paths. Based on the number of directed edges contained in each path in the original reachable path set, a shortest hop count operation is performed to quantify the path length, and the path length is defined as the causal path order index to generate a candidate causal path list. Based on the causal path order index in the candidate causal path list, hierarchical filtering is performed to identify specific patterns and generate categorized path groups. For the aforementioned path classification group, a preset semantic rule library is invoked to perform a label mapping operation, generating path objects to be labeled; The path objects to be labeled with differentiated semantic descriptions and their corresponding original topological structure data are integrated, and structured encapsulation processing is performed to bind path topological information and semantic description information, ultimately generating a set of semantically labeled causal paths that can be directly called by the path distillation encoder.

6. The data product pricing trend prediction method based on transaction behavior clustering analysis according to claim 5, characterized in that, The original reachable path set includes a sequence of starting nodes, intermediate nodes, and ending nodes.

7. The data product pricing trend prediction method based on transaction behavior clustering analysis according to claim 5, characterized in that, The process of performing hierarchical filtering to identify specific patterns includes: Paths of order one are classified as direct influence patterns, paths of order two are classified as mediating transmission patterns, and paths of order three are classified as synergistic amplification patterns, generating a classification path group that covers the direct influence patterns, mediating transmission patterns, and synergistic amplification patterns.

8. The data product pricing trend prediction method based on transaction behavior clustering analysis according to claim 7, characterized in that, The step of calling a preset semantic rule library to perform label mapping operation for the classified path group includes: The direct impact mode is mapped as a high-frequency bidding-driven short-term premium label, the intermediary transmission mode is mapped as a bulk purchase-induced inventory expectation adjustment label, and the collaborative amplification mode is mapped as a cross-platform arbitrage superimposed with a policy response label.

9. The data product pricing trend prediction method based on transaction behavior clustering analysis according to claim 7, characterized in that, Step S5 specifically includes: Obtain the original transaction sequence and the set of causal paths, and logically formalize the direct influence mode, the mediator transmission mode and the synergistic amplification mode based on the combination of expert rules to generate an expert knowledge constraint set; The original transaction sequence is subjected to rule matching operation using the expert knowledge constraint set. The theoretical triggering intensity of each causal path within a specific time window is calculated to generate the teacher path distribution vector, which serves as the target ground value for student model learning. A graph neural network structure is constructed based on the original transaction sequence. The behavior nodes are mapped to the initial node embedding representation. Neighborhood information is aggregated through a multi-layer message passing mechanism to generate student path latent feature vectors. The student path latent feature vector is subjected to probability normalization processing to map it to a student path activation probability distribution that conforms to the probability axiom. The difference metric between the student path activation probability distribution and the teacher path distribution vector is calculated to generate a divergence loss scalar. Backpropagation of gradient signals based on the divergence loss scalar is used to update the connection weight parameters of the graph neural network structure. Iterative optimization is performed until the student path activation probability distribution converges to the neighborhood space of the teacher path distribution vector, and finally the path activation probability distribution vector is output.

10. The data product pricing trend prediction method based on transaction behavior clustering analysis according to claim 1, characterized in that, The unsupervised indicators include: temporal adjacency, order coupling strength, and subject overlap; the typical paths cover direct influence, intermediary transmission, and synergistic amplification modes; the trend pattern label sequence includes accelerated upward movement, platform oscillation, or structural reversal patterns.

Images