A method and system for multivariate time series prediction explanation based on information theory and causal reasoning

Abstract

The present application relates to the technical field of artificial intelligence model explanation, in particular to a multivariate time series prediction explanation method and system based on information theory and causal reasoning. The method comprises the following steps: acquiring original multivariate time series data; performing data preprocessing on the acquired original multivariate time series data; performing causal information reasoning on the preprocessed multivariate time series data; calculating the dynamic feature weight of the multivariate time series data based on the causal information reasoning; performing local fitting and explanation generation based on the dynamic feature weight; and obtaining a time-feature two-dimensional heat map. Through the multidimensional feature relationship mining, the optimized causal relationship reasoning algorithm, the dynamic weight calculation and the intuitive time-feature two-dimensional heat map generation mechanism, the present application effectively considers the time continuity, presents the prediction explanation in an intuitive and easy-to-understand manner, and enables the user to quickly understand the artificial intelligence model decision-making process.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence model explanation, in particular to a multivariate time series prediction explanation method and system based on information theory and causal reasoning. BACKGROUND

[0002] Artificial intelligence is widely used in time series prediction, covering medical, physical, energy, sensor data and other fields. In these application scenarios, it is crucial to ensure that users can quickly understand the reasons for artificial intelligence decision-making, especially in time-sensitive decision assistance scenarios. Explanation methods help to explore the factors behind the decision, thereby increasing human trust in the model and promoting the widespread and in-depth application of artificial intelligence.

[0003] Currently, there are some saliency methods for explaining the importance of input features to prediction, mainly including: (1) perturbation-based methods, such as feature occlusion FO and RISE, which provide explanations by perturbing inputs and comparing results; (2) gradient-based methods, such as integrated gradients, DeepLIFT, and GradSHAP, which use the gradient of feature representation for explanation; (3) attention mechanism combined with saliency methods, the effectiveness of these methods in explaining black-box models is still controversial; (4) other methods, such as SHAP uses Shapley value, and LIME generates explanations by randomly combining records and their neighborhoods and weighting them according to the proximity of the neighborhood.

[0004] In addition, some studies have begun to consider the time-sensitive nature of time series applications. For example: FIT quantifies the importance of observations over time by evaluating the contribution of observations to the prediction output; Dynamask considers the time dependence of time series data in its design and perturbs the input by dynamically combining adjacent values in the input.

[0005] The existing methods have the following shortcomings when explaining multivariate time series prediction.

[0006] (1) Using saliency methods based on perturbation, gradient, etc. to directly explain time series prediction often produces results that are difficult for humans to understand. Although it is possible to score feature importance on a fine time scale to generate a saliency map, the saliency score of each feature has an exact numerical value within a time range, and the fragmented results presented are difficult for users to provide meaningful and easy-to-understand explanations.

[0007] (2) Although the FIT method quantifies the importance of features over time, it still scores in numerical form, and the output results are still not intuitive enough for users.

[0008] (3) The Dynamask method can identify changes in feature importance over time and strives to maintain this continuity in interpretation, making it more user-friendly than previous algorithms such as FIT. However, such methods usually evaluate feature importance through other internal states of the model (such as gradients), which limits their practical applications. Summary of the Invention

[0009] In response to the problems of existing artificial intelligence model prediction and interpretation methods, such as the lack of consideration of time continuity, non-intuitive interpretation results, and strong model dependence, this proposal proposes a multivariate time series prediction and interpretation method based on information theory and causal reasoning. This proposal helps users better understand the decision-making process of artificial intelligence models in time series prediction through a multivariate time series prediction and interpretation method based on information theory and causal reasoning that can effectively consider time continuity, has intuitive and easy-to-understand interpretation results, and is model-independent.

[0010] In a first aspect, the present invention provides a multivariate time series prediction and interpretation method based on information theory and causal reasoning, which adopts the following technical solutions:

[0011] Get the original multivariate time series data;

[0012] Perform data preprocessing on the acquired original multivariate time series data;

[0013] Perform causal information inference on preprocessed multivariate time series data;

[0014] Calculate dynamic feature weights of multivariate time series data based on causal information reasoning;

[0015] Perform local fitting and interpret generation based on dynamic feature weights;

[0016] Get the time-feature two-dimensional heat map.

[0017] Furthermore, the data preprocessing of the original multivariate time series data includes converting the original multivariate time series data into Standardization is performed, where D represents the number of features, T represents the number of time observations, Characterized by d In time t data points; for each feature dimension d , through the formula Normalize, where and Characteristics d The mean and standard deviation of the time series; at the same time, the time series is divided into subsequence of sThe subsequence is represented as .

[0018] Furthermore, the causal information reasoning of the pre-processed multivariate time series data includes calculating the mutual information between different features for each subsequence, and for the feature and features , calculate the joint distribution probability of the two features and marginal distribution probability 、 , and then calculate the mutual information value, and construct the mutual information matrix by calculating the mutual information value of each two features separately , record the mutual influence between each feature dimension; the mutual information value calculation is expressed as:

[0019] .

[0020] Furthermore, the causal information reasoning of the pre-processed multivariate time series data also includes mining the causal relationship between features using a constraint-based causal discovery algorithm based on the calculated mutual information matrix, wherein a completely undirected graph is constructed. , the nodes in the figure correspond to the various feature dimensions of the multivariate time series, and the nodes are connected by undirected edges, indicating that there may be associations between the features in the initial state. According to the time sequence of the time series, a time constraint mark is added to each edge to record the time sequence of the association between the two features; the size of the condition set currently used in the iteration process is recorded as , set the maximum condition set size ,from First, for each edge in the graph , given different condition sets composed of other nodes , a conditional independence test is performed.

[0021] Furthermore, the causal information reasoning of the pre-processed multivariate time series data also includes, in the orientation stage, if there is time constraint information indicating that the feature Changes always precede features ,and For undirected edges, when other directional rules do not conflict, the edge is preferentially directed to i→j, and the directional rules are repeatedly applied until all edges are directed, and finally a causal relationship graph is constructed.

[0022] Furthermore, the dynamic feature weights of multivariate time series data are calculated based on causal information reasoning, including respectively calculating the causal centrality index, time series change trend coefficient and information gain rate of the multivariate time series data features, and integrating them into the final weights ,in 、 and is the adjustment factor, and , where the feature is first measured The scope of influence on other characteristics as a cause ,in Representation characteristics The number of edges pointing to other features, is the total number of nodes, and then consider the characteristics To the prediction target The causal path length and strength of ,in Indicates that from the feature To the prediction target The set of all directed paths of For path The product of the weights, For path length, is the attenuation coefficient, combined with the above indicators, calculate the characteristics Causal centrality ,in is the weight coefficient, and .

[0023] Furthermore, the method of calculating the dynamic feature weights of multivariate time series data based on causal information reasoning also includes calculating the time series change trend coefficient, wherein the feature The average rate of change of adjacent time points within the subsequence ,in Characterized by In time The value of is the time interval, and then the time decay mechanism is introduced to give recent data a higher weight ,in is the attenuation coefficient, Characterized by The mean value within the subsequence. Comprehensive calculation of the change trend coefficient ,in weight coefficient, and .

[0024] Furthermore, the local fitting and interpretation generation based on dynamic feature weights includes, during the prediction process, for a given input subsequence , the weighted Euclidean distance is used to calculate the similarity between the input subsequence and each subsequence in the training data, which is expressed as:

[0025] ,

[0026] in, Characterized by By introducing dynamic weights, similarity calculation focuses on dimensions with key causal features and obvious temporal trends.

[0027] Furthermore, the local fitting and interpretation generation based on the dynamic feature weights also includes selecting a number of subsequences with the smallest distance as a similar subsequence set, and constructing a weighted linear regression model for the selected similar subsequence set: ,in Characterized by The regression number is solved by minimizing the weighted mean square error: .

[0028] Second, a multivariate time series prediction and interpretation system based on information theory and causal reasoning, including:

[0029] The data acquisition module is configured to acquire raw multivariate time series data;

[0030] The preprocessing module is configured to perform data preprocessing on the acquired original multivariate time series data;

[0031] The information reasoning module is configured to perform causal information reasoning on the preprocessed multivariate time series data;

[0032] The weight module is configured to calculate the dynamic feature weights of multivariate time series data based on causal information reasoning;

[0033] The fitting module is configured to perform local fitting and interpretation generation based on dynamic feature weights;

[0034] The output module is configured to obtain a time-feature two-dimensional heat map.

[0035] In a third aspect, the present invention provides a computer-readable storage medium storing a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device for a multivariate time series prediction and interpretation method based on information theory and causal reasoning.

[0036] In a fourth aspect, the present invention provides a terminal device comprising a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, and the instructions are suitable for being loaded and executed by the processor to implement the multivariate time series prediction and interpretation method based on information theory and causal reasoning.

[0037] In summary, the present invention has the following beneficial technical effects:

[0038] This invention optimizes the traditional PC algorithm for time series prediction and interpretation. By introducing time constraint information during the initialization phase and closely integrating it with the temporal order of time series data, the mined causal relationships are more consistent with actual temporal logic. During the conditional independence test phase, a dynamic threshold adjustment mechanism is introduced to adaptively adjust the conditional independence judgment criteria based on data fluctuations, effectively avoiding the accidental or missed deletion of edges due to fixed thresholds. During the orientation phase, time constraint information is combined with traditional orientation rules to further improve the accuracy of edge orientation. This allows the algorithm to closely integrate the temporal characteristics of time series with the dynamic changes in data, more accurately mining true and reliable causal relationships.

[0039] This invention effectively considers time continuity through multi-dimensional feature relationship mining, optimized causal reasoning algorithm, dynamic weight calculation and intuitive time-feature two-dimensional heat map generation mechanism, presenting prediction explanations in an intuitive and easy-to-understand manner, allowing users to quickly understand the decision-making process of the artificial intelligence model.

[0040] Through time constraint labeling, dynamic threshold adjustment and improved directional rules, the present invention can more accurately mine the causal relationship in multivariate time series data, providing a more reliable basis for subsequent prediction and interpretation. Compared with existing methods, it is more accurate and reliable in causal relationship analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 Schematic diagram of a multivariate time series prediction and interpretation method based on information theory and causal reasoning according to Example 1 of the present invention. DETAILED DESCRIPTION

[0042] The present invention will be further described in detail below with reference to the accompanying drawings.

[0043] Example 1

[0044] Reference Figure 1 In this embodiment, a multivariate time series prediction and interpretation method based on information theory and causal reasoning includes:

[0045] Get the original multivariate time series data;

[0046] Perform data preprocessing on the acquired original multivariate time series data;

[0047] Perform causal information inference on preprocessed multivariate time series data;

[0048] Calculate dynamic feature weights of multivariate time series data based on causal information reasoning;

[0049] Perform local fitting and interpret generation based on dynamic feature weights;

[0050] Get the time-feature two-dimensional heat map.

[0051] Specifically:

[0052] 1. Data preprocessing,

[0053] The original multivariate time series data Perform standardization, for each feature dimension d , through the formula Normalize, where and Characteristics d The mean and standard deviation of the time series. At the same time, the time series is divided into Each subsequence is represented as .

[0054] 2. Calculate the mutual information between features,

[0055] Multivariate time series data features can be obtained in the following ways.

[0056] (1) In the medical record scenario, patient indicators (such as red blood cell count, blood oxygen saturation, heart rate, and body temperature) are extracted as features based on the hospital information system and diagnostic equipment, and sorted by timestamp to form a time series. Similarly, in financial transaction scenarios, key indicators (such as opening price, transaction amount, and position size) are extracted from the business system log as features, and sorted by timestamp to form a time series.

[0057] (2) In scenarios such as industrial monitoring and environmental perception, multi-dimensional time series data are collected in real time by deploying multiple types of sensors (such as temperature, pressure, and flow sensors). Each sensor corresponds to a feature dimension. d ( d =1,2,..., D ), the time series length is T .

[0058] The calculation process of joint distribution probability and marginal distribution probability includes the following steps: X (s) , calculate the joint distribution probability and marginal distribution probability between different features. and features As an example, the data discretization method is used to calculate the joint distribution probability and marginal distribution probability 、 , the specific steps are as follows.

[0059] (1) Data discretization: Use the equal interval binning method to discretize the continuous time series data and map the eigenvalues ​​to a finite number of intervals (bins). , let the discrete interval set be ,in K is the number of intervals, and the interval width is , No. k The interval is ,feature Similarly, let the discrete interval set be ,in M is the number of intervals, and the interval width is , No. m The interval is For example, if the feature The value of is [10,30], and it is divided into 5 intervals using equidistant binning, so the intervals are [10,14), [14,18), ..., [26,30].

[0060] (2) Calculation of marginal distribution probability,

[0061] Statistical features and features The number of samples in each interval of the subsequence data 、 ; The marginal distribution probability is the ratio of the number of interval samples to the total number of samples, .

[0062] (3) Calculation of joint distribution probability,

[0063] Statistical feature pairs Falling in the joint interval The number of samples within , the joint distribution probability is the ratio of the number of joint interval samples to the total number of samples .

[0064] Calculation of mutual information value,

[0065] Based on the discretized probability distribution, the features are calculated using the following formula and features The mutual information value of

[0066] .

[0067] Mutual information reflects the dependency between two features. The higher the mutual information value, the closer the correlation between the two features. Calculate the mutual information value of each two features separately and construct the mutual information matrix , thereby recording the mutual influence between each feature dimension.

[0068] Unlike traditional methods that analyze feature importance from a single perspective, this proposal explores the inherent dependencies between features in multivariate time series data from an information-theoretic perspective. Existing technologies struggle to fully capture the complex relationships between features. Mutual information calculations, however, quantify the degree of dependency between different feature dimensions, uncovering underlying data patterns and hidden relationships. This allows for a deeper understanding of multivariate time series data, providing rich and accurate foundational data for subsequent causal reasoning, and ultimately enabling more comprehensive and accurate forecasting and interpretation.

[0069] 3.Causal reasoning,

[0070] Based on the calculated mutual information matrix, a constraint-based causal discovery algorithm (PC algorithm) is used to mine causal relationships between features. Traditional PC algorithms, when applied to time series causal relationship mining, fail to fully consider the temporal characteristics of time series and the dynamic nature of data changes, leading to problems such as misjudgment of causal relationships and inaccurate edge orientation. This patent proposal optimizes the traditional PC algorithm for time series prediction scenarios. The specific steps are as follows.

[0071] (1) Initialization phase: building a completely undirected graph The nodes in the graph correspond to the feature dimensions of the multivariate time series. The nodes are connected by undirected edges, indicating that there may be associations between the features in the initial state. At the same time, according to the chronological order of the time series, a time constraint tag is added to each edge to record the chronological order of the association between the two features.

[0072] Constructing a completely undirected graph . in is a node set corresponding to the multivariate time series Feature dimensions ; It is a set of undirected edges, indicating that there may be a correlation between any two features in the initial state. The initial weight of each undirected edge is the mutual information value, which reflects the dependence strength between features.

[0073] The time sequence relationship is determined by Granger causality test. First, the feature and features , establish hysteresis q The regression model of the period Secondly, test the null hypothesis (Right now no Granger reasons), if we reject And the average lag time ,but ,express Before On the contrary, if yes Granger cause, then If the time sequence cannot be determined, such as simultaneous changes, it will be marked as The final initialization stage output: feature dimension V , with initial weights Undirected edges, and time constraint labels .

[0074] (2) Conditional independence test phase: The size of the condition set currently used in the iteration process is , set the maximum condition set size ,from First, for each edge in the graph , given different condition sets composed of other nodes In the case of , a conditional independence test is performed. The conditional independence test introduces a dynamic threshold adjustment mechanism, the threshold Dynamic calculation based on the data fluctuation level of the current subsequence. The formula is as follows:

[0075] ,

[0076] in, is the adjustment coefficient, and Characteristics and features At time step The value of and Characteristics and features The mean value in the current subsequence, is the subsequence length.

[0077] Conditional independence is determined by calculating conditional mutual information. and features , given a set of conditions The conditional mutual information formula is:

[0078] ,

[0079] If in a certain condition set Down, Less than the dynamic threshold , then it is considered that under this condition the characteristic and features Conditionally independent, delete the edge After each round of deletion, the number of edges is dynamically adjusted based on the connection status of the remaining edges and data characteristics. If the remaining edges are sparse, increase The above process is repeated until no more edges can be deleted.

[0080] (3) Orientation stage: Based on the traditional orientation rules, the edge orientation is combined with the time constraint information. In addition to following the original orientation rules such as collision structure, if there is time constraint information indicating the feature Changes always precede features ,and For undirected edges, if other directional rules do not conflict, the edge is preferably directed from i to j. These directional rules are applied repeatedly until all edges are directed, and finally a causal relationship graph is constructed.

[0081] 1) Traditional directional rules,

[0082] Collision structure orientation: If there are three nodes , satisfying ①i and j are connected, j and k are connected, but i and k are not connected; ②under a certain condition set S, i and k are conditionally independent, and .Then the forced orientation is , forming a collision structure.

[0083] Non-collision structure orientation: If the edge With oriented edges and Composition path , then the edge Targeted , to avoid ring formation.

[0084] 2) Time constraint priority directional rules,

[0085] Time constraint marking based on initialization phase , undirected edges are defined according to the following rules.

[0086] Direction when time sequence is clear: if the side Time constraint tag , and no directed edge If not targeted by traditional rules, the priority is On the contrary, if , then the orientation is .

[0087] Conflict handling between time constraints and traditional rules: If the time constraint orientation conflicts with the traditional collision structure orientation, the traditional rules will prevail. But traditional rules require , then the traditional rules shall prevail and the time constraint mark shall be adjusted to , to avoid violating causal logic.

[0088] 3) The mathematical iterative process of edge orientation,

[0089] Define the directional state matrix ,in Indicates that the edge is oriented , Indicates that the edge is oriented , Represents an undirected edge or undirected.

[0090] 4) Directed iteration step,

[0091] First, initialize the orientation matrix (All edges are undirected); secondly, apply the traditional rule to traverse all triples , detect the collision structure and update if the conditions are met or ; In addition, the time constraint orientation is applied, and for the unoriented edges ,like , then update: ; When the directed matrix is ​​no longer updated or all edges are directed, the iteration is terminated and a directed causal graph is finally generated, including: feature dimensions V , directed edges , where the edge weight (Normalized mutual information value, reflecting causal strength).

[0092] This patented optimized PC algorithm, by introducing temporal constraints and tightly integrating the temporal order of time series data, ensures that the mined causal relationships are more consistent with actual temporal logic. A dynamic threshold adjustment mechanism adaptively adjusts the conditional independence judgment criteria based on data fluctuations, effectively avoiding the accidental or missed deletion of edges due to fixed thresholds. Improved orientation rules further enhance the accuracy of edge orientation. In summary, the optimized algorithm can more accurately mine true and reliable causal relationships from multivariate time series data.

[0093] 4. Dynamic weight calculation,

[0094] For each feature in a subsequence, the weight of the feature is dynamically calculated based on its causal role in the causal relationship graph, the time series change trend, and the information gain rate. The specific calculation is as follows.

[0095] (1) Calculation of causal centrality index,

[0096] First measure the features The scope of influence on other characteristics as a cause ,in Representation characteristics The number of edges pointing to other features, is the total number of nodes. Next, consider the feature To the prediction target The causal path length and strength of ,in Indicates that from the feature To the prediction target The set of all directed paths of For path The product of the weights, For path length, is the attenuation coefficient. Combined with the above indicators, calculate the characteristics Causal centrality ,in is the weight coefficient, and .

[0097] (2) Calculation of time series trend coefficient,

[0098] First calculate the features The average rate of change of adjacent time points within the subsequence ,in Features In time The value of is the time interval. Secondly, the time decay mechanism is introduced to give recent data a higher weight. ,in is the attenuation coefficient, Features The mean value within the subsequence. Comprehensive calculation of the change trend coefficient ,in weight coefficient, and .

[0099] (3) Information gain rate calculation,

[0100] Measuring characteristics For the prediction target Information contribution ,in Features and prediction target The mutual information of Features The entropy of .

[0101] (4) Dynamic weight integration,

[0102] Integrate the above indicators into features The final weight ,in 、 and is the adjustment factor, and , which can be optimized according to the actual data characteristics.

[0103] The dynamic weight calculation method proposed in the patent introduces the causal centrality index to quantify the core position of the feature in the causal relationship network from the structural level; combines the time series trend coefficient to capture the influence of the dynamic change of the feature in the time dimension on the prediction; and introduces the information gain rate to evaluate the information contribution of the feature to the prediction target. The combination of the three breaks through the limitation of considering only a single dimension in traditional methods and can more accurately identify key causal features and their importance changes at different time points.

[0104] 5. Local fitting and explanation generation,

[0105] (1) Similar subsequence retrieval,

[0106] In the prediction process, for a given input subsequence , the weighted Euclidean distance is used to calculate the similarity between the input subsequence and each subsequence in the training data. The formula is as follows:

[0107] ,

[0108] where is the dynamic weight of the feature in step 4. By introducing the dynamic weight, the similarity calculation pays more attention to the dimensions with obvious causal key features and time series trends.

[0109] (2) Weighted linear regression fitting,

[0110] Select a number of subsequence with the smallest distance as the similar subsequence set, and construct a weighted linear regression model for the selected similar subsequence set: where is the regression number of the feature , and the solution is obtained by minimizing the weighted mean square error: .

[0111] Combined with the causal relationship graph constructed in step 3, if the feature is the direct cause of the prediction target (exists directed edge ), then the regression coefficient is forced to be positive (positive causal), avoiding the model learning parameters that violate causal logic.

[0112] (3) Dynamic explanation generation mechanism,

[0113] Based on the regression coefficient and the dynamic weight , a time-feature two-dimensional heat map is generated. The horizontal axis is the time step , the vertical axis is the feature , and the pixel value is , highlighting the feature-time combination that has the greatest impact on the prediction results.

[0114] Example 2

[0115] This embodiment provides a multivariate time series prediction and interpretation system based on information theory and causal reasoning, including:

[0116] The data acquisition module is configured to acquire raw multivariate time series data;

[0117] The preprocessing module is configured to perform data preprocessing on the acquired original multivariate time series data;

[0118] The information reasoning module is configured to perform causal information reasoning on the preprocessed multivariate time series data;

[0119] The weight module is configured to calculate the dynamic feature weights of multivariate time series data based on causal information reasoning;

[0120] The fitting module is configured to perform local fitting and interpretation generation based on dynamic feature weights;

[0121] The output module is configured to obtain a time-feature two-dimensional heat map.

[0122] A computer-readable storage medium stores a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device, a multivariate time series prediction and interpretation method based on information theory and causal reasoning.

[0123] A terminal device includes a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, wherein the instructions are suitable for being loaded and executed by the processor to implement a multivariate time series prediction and interpretation method based on information theory and causal reasoning.

[0124] The above are all preferred embodiments of the present invention, and are not intended to limit the scope of protection of the present invention. Therefore, any equivalent changes made based on the structure, shape, and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. A multivariate time series prediction and interpretation method based on information theory and causal reasoning, characterized by: include: Get the original multivariate time series data; In industrial monitoring scenarios, multiple types of sensors, including temperature, pressure, and flow sensors, are deployed to collect multi-dimensional time series data in real time. Each sensor corresponds to a feature dimension. d , d =1,2,..., D , D represents the number of features, and the length of the time series is t ; Perform data preprocessing on the acquired original multivariate time series data; Perform causal information inference on preprocessed multivariate time series data; Calculate dynamic feature weights of multivariate time series data based on causal information reasoning; Perform local fitting and interpret generation based on dynamic feature weights; Get the time-feature two-dimensional heat map; where the horizontal axis is the time step , the vertical axis is the characteristic , the pixel value is ; The dynamic feature weights of multivariate time series data are calculated based on causal information reasoning, including respectively calculating the causal centrality index, time series change trend coefficient and information gain rate of the multivariate time series data characteristics, and integrating them into the final weights ,in 、 and is the adjustment factor, and , where the feature is first measured The scope of influence on other characteristics as a cause ,in Representation characteristics The number of edges pointing to other features, is the total number of nodes, and then consider the characteristics To the prediction target The causal path length and strength of ,in Indicates that from the feature To the prediction target The set of all directed paths of For path The product of the weights, For path length, is the attenuation coefficient, combined with the above indicators, calculate the characteristics Causal centrality ,in is the weight coefficient, and ; Representative measurement characteristics For the prediction target Information contribution; ,in Features and prediction target The mutual information of Features The entropy of ; The method of calculating the dynamic feature weight of multivariate time series data based on causal information reasoning also includes calculating the time series change trend coefficient, wherein the feature The average rate of change of adjacent time points within the subsequence ,in Features In time The value of is the time interval, and then the time decay mechanism is introduced to give recent data a higher weight ,in is the attenuation coefficient, Features The mean value in the subsequence is used to comprehensively calculate the change trend coefficient ,in weight coefficient, and ; Represents the length of the subsequence; The local fitting and explanation generation based on dynamic feature weights includes, during the prediction process, for a given input subsequence , the weighted Euclidean distance is used to calculate the similarity between the input subsequence and each subsequence in the training data, which is expressed as: ， in, Features By introducing dynamic weights, similarity calculations focus on key causal features and dimensions with obvious temporal trends. The local fitting and interpretation generation based on the dynamic feature weights also includes selecting several subsequences with the smallest distance as similar subsequence sets, and constructing a weighted linear regression model for the selected similar subsequence sets: ,in Features The regression number is solved by minimizing the weighted mean square error: .

2. A multivariate time series prediction and interpretation method based on information theory and causal reasoning according to claim 1, characterized in that: The data preprocessing of the original multivariate time series data includes converting the original multivariate time series data into Standardization is performed, in which D represents the number of features, T represents the number of time observations, Features d In time t data points; for each feature dimension d , through the formula Normalize, where and Characteristics d The mean and standard deviation of the time series; at the same time, the time series is divided into subsequence of s The subsequence is represented as .

3. A multivariate time series prediction and interpretation method based on information theory and causal reasoning according to claim 2, characterized in that: The causal information reasoning of the pre-processed multivariate time series data includes calculating the mutual information between different features for each subsequence. and features , calculate the joint distribution probability of the two features and marginal distribution probability 、 , and then calculate the mutual information value, and construct the mutual information matrix by calculating the mutual information value of each two features separately , record the mutual influence between each feature dimension; the mutual information value calculation is expressed as: 。 4. A multivariate time series prediction and interpretation method based on information theory and causal reasoning according to claim 3, characterized in that: The causal information reasoning of the pre-processed multivariate time series data also includes mining the causal relationship between features using a constraint-based causal discovery algorithm based on the calculated mutual information matrix, wherein a completely undirected graph is constructed. , the nodes in the figure correspond to the various feature dimensions of the multivariate time series, and the nodes are connected by undirected edges, indicating that there may be associations between the features in the initial state. According to the time sequence of the time series, a time constraint mark is added to each edge to record the time sequence of the association between the two features; the size of the condition set currently used in the iteration process is recorded as , set the maximum condition set size ,from First, for each edge in the graph , given different condition sets composed of other nodes , a conditional independence test is performed.

5. The multivariate time series prediction and interpretation method based on information theory and causal reasoning according to claim 4 is characterized in that: The causal information reasoning of the pre-processed multivariate time series data also includes, in the orientation stage, if there is time constraint information indicating that the feature Changes always precede features ,and For undirected edges, when other directional rules do not conflict, the edge is preferentially directed to i→j, and the directional rules are repeatedly applied until all edges are directed, and finally a causal relationship graph is constructed.

6. A multivariate time series prediction and interpretation system based on information theory and causal reasoning, which executes the multivariate time series prediction and interpretation method based on information theory and causal reasoning as claimed in claim 1, characterized in that: include: The data acquisition module is configured to acquire raw multivariate time series data; The preprocessing module is configured to perform data preprocessing on the acquired original multivariate time series data; The information reasoning module is configured to perform causal information reasoning on the preprocessed multivariate time series data; The weight module is configured to calculate the dynamic feature weights of multivariate time series data based on causal information reasoning; The fitting module is configured to perform local fitting and interpretation generation based on dynamic feature weights; The output module is configured to obtain a time-feature two-dimensional heat map.

Images