A method for evaluating intervention effect of sports intangible cultural heritage protection policy

Abstract

The application provides a kind of sports non-material cultural heritage protection policy intervention effect evaluation method, comprising: collecting multi-source heterogeneous data, extracting structured policy-behavior-response elements, constructing dynamic policy knowledge graph and realizing the time series modeling of causal network;Adopting time series contrast learning encoder and twin network optimization policy variable embedding, realizing accurate extraction of policy effect after mixed variable elimination;Integrating do-calculus and Monte Carlo tree search for counterfactual causal reasoning, forming multi-path effect contribution and stability evaluation;Through Shapley value decomposition and visual report, the net causal effect of each policy tool is quantified and the effectiveness is determined, the application can realize multi-dimensional causal effect accurate identification for sports non-heritage policy intervention process, improve the scientificity and timeliness of attribution analysis.

Description

Technical Field

[0001] This invention relates to the fields of causal reasoning and knowledge graph modeling technology, and in particular to a method for evaluating the effectiveness of policy interventions in the protection of intangible cultural heritage in sports. Background Technology

[0002] Current academic and applied research on the evaluation of the effectiveness of policy interventions in the protection of intangible cultural heritage in sports primarily relies on methods such as correlation analysis, expert scoring, regression discontinuity, and propensity score matching. These mainstream approaches typically employ static data modeling, analyzing policy documents, records of heritage activities, and social survey results to explore the relationship between policy interventions and cultural heritage transmission behaviors. In recent years, knowledge graphs, network analysis, and causal inference have been gradually introduced into this field, driving the development from correlation-based statistical evaluation to causal explanation mechanisms. For example, some published methods utilize causal network structures, combined with counterfactual inference, to attempt attribution of effects in complex policy environments lacking randomized experiments. Existing technologies are mostly based on static or semi-dynamic data, lacking the ability to integrate and model the evolution of multi-source data such as policies, heritage behaviors, and social responses in real time. Representative solutions often use single-point-of-time samples or simplified causal chains for indicator attribution analysis. However, in real-world scenarios, the accuracy of identifying policy intervention variables is limited, and they are easily affected by external variables such as regional economic fluctuations and seasonal effects. For example, directly using the correlation test between policy implementation time and the frequency of heritage activities, or relying on expert groups to subjectively score changes in intangible cultural heritage projects, are difficult to effectively isolate the true causal path of policy intervention. The scientific validity and interpretability of the evaluation results are significantly limited. Although some causal inference models introduce causal structures and counterfactual inference concepts, insufficient identification of policy intervention variables or the presence of confounding biases leads to distorted causal path modeling, making it impossible to dynamically and accurately track the evolution of policy effects. Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, this invention provides a method for evaluating the effectiveness of policy interventions in the protection of intangible cultural heritage in sports.

[0004] The technical solution of this invention is implemented as follows: A method for evaluating the effectiveness of policy interventions in the protection of intangible cultural heritage in sports, comprising: S1: Collect multi-source heterogeneous data from government documents, media reports, social media, and interview records of intangible cultural heritage inheritors; extract structured elements including policy subjects, policy tools, implementation nodes, inheritance projects, participants, dissemination scope, and funding; and perform time-series alignment and standardized coding to form a policy-behavior-response related dataset. S2: Based on the associated dataset described in S1, construct a dynamic policy knowledge graph containing timestamps. Establish a multi-level causal network with policy intervention as the root node. The first layer connects the implementing agency and resource allocation nodes, the second layer connects the publicity and promotion and inheritance activity nodes, and the third layer connects the result nodes of public participation and the breadth of skill dissemination. The evolution trajectory of relationship weights is marked by timestamps. S3: Based on the dynamic policy knowledge graph described in S2, a time-series comparative learning encoder is constructed. A Siamese network structure is used to construct positive sample pairs for the implementation cases of the same policy in different regions / time periods. At the same time, a control region where the policy has not been implemented is introduced as a negative sample. The distance distribution of the feature space is optimized through metric learning to generate a policy variable embedding vector with time invariance. S4: Based on the policy variable embedding vector described in S3, perform confounding variable clarification operation, which specifically includes: calculating the correlation coefficient matrix between the feature vector and external factors such as seasonal fluctuations and economic indicators, using the orthogonal projection method to eliminate interference from non-policy factors, and outputting the corrected set of policy-driven variables as input for causal inference. S5: Deploy a counterfactual path reasoning engine in the dynamic policy knowledge graph described in S2, perform causal intervention operations based on do-calculus, use the Monte Carlo tree search algorithm to traverse all potential causal paths, combine the variable set corrected in S4 to calculate the effect contribution value of each path, and generate a path stability scoring function to evaluate cross-group consistency. S6: Based on the effect contribution value and path stability score described in S5, construct a causal effect attribution matrix, use the Shapley value decomposition method to quantify the net causal effect of each policy tool, generate a visual evaluation report containing the trend of changes in the intensity of the main effect path, and output the threshold judgment basis for the effectiveness of policy intervention.

[0005] The present invention provides a method for evaluating the effectiveness of policy interventions in the protection of intangible cultural heritage in sports, which has the following beneficial effects: (1) This invention significantly improves the accuracy of identifying policy intervention variables by constructing a dynamic knowledge graph that integrates multi-source heterogeneous data and introducing a variable clarification mechanism based on time-series comparative learning. This mechanism uses positive and negative sample pairs of policy implementation in different regions and time periods to effectively distinguish between change patterns driven by real policies and pseudo-correlation phenomena caused by non-target factors such as seasonality and socio-economic fluctuations in the latent semantic space, thereby overcoming the shortcomings of existing statistical models in terms of weak causal inference ability and large bias in the absence of randomized controlled experiments. At the same time, by combining edge relationship modeling with timestamps, the knowledge graph structure is dynamically updated with the policy cycle, which enhances the ability to characterize the long-term and phased policy impact process and makes the evaluation results more timely and context-adaptive. (2) This invention deploys a counterfactual path reasoning engine based on a dynamic knowledge graph. It adopts a combination of do-calculus operation and Monte Carlo tree search to simulate the counterfactual scenario of "if the policy had not been implemented". It comprehensively traverses the multi-level mediation path from policy intervention to final effect and accurately estimates the causal contribution of each path. A path stability scoring function is introduced to quantify the consistency of a causal chain in different populations, regions or subgroups. The main effect path with cross-group robustness is retained first, effectively filtering noise interference and local bias, and improving the reliability and generalization ability of the reasoning results. By comprehensively considering the difference between actual observations and counterfactual predictions, an evaluation report is generated that includes the net causal effect strength, the role weight of key mediation nodes and the visualization of the impact path. This not only realizes the leap from "relevant description" to "causal explanation", but also provides interpretable and traceable decision support basis for policy optimization. (3) This invention constructs a closed-loop, adaptive policy causal evaluation framework with good scalability and engineering implementation capabilities. The entire process covers the complete chain from multi-source data collection, semantic standardization, graph construction, variable clarification to counterfactual reasoning. It supports the decoupling and attribution analysis of the effects of various intervention strategies such as policy tool combinations, resource allocation efficiency, and publicity and promotion paths, and can flexibly adapt to the protection scenarios of different types of intangible cultural heritage projects. The lightweight encoder design and efficient search algorithm ensure the method's response speed and computational resource utilization, making it suitable for large-scale, high-frequency policy monitoring needs. This technology not only improves the scientific and refined level of sports-related intangible cultural heritage policy evaluation, but can also be extended to public cultural services, heritage education and dissemination, and other fields, promoting the formation of a data-driven, evidence-oriented new governance model. Attached Figure Description

[0006] Figure 1 This is a flowchart illustrating a method for evaluating the effectiveness of policy interventions in the protection of intangible cultural heritage in sports, as described in this invention. Figure 2 This is a sub-flowchart of a method for evaluating the effectiveness of policy interventions in the protection of intangible cultural heritage in sports, as proposed in this invention. Detailed Implementation

[0007] 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.

[0008] 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.

[0009] like Figure 1 As shown, this application provides a method for evaluating the effectiveness of policy interventions in the protection of intangible cultural heritage in sports, specifically including: S1: Collect multi-source heterogeneous data from government documents, media reports, social media, and interview records of intangible cultural heritage inheritors; extract structured elements including policy subjects, policy tools, implementation nodes, inheritance projects, participants, dissemination scope, and funding; and perform time-series alignment and standardized coding to form a policy-behavior-response related dataset. S2: Based on the associated dataset described in S1, construct a dynamic policy knowledge graph containing timestamps. Establish a multi-level causal network with policy intervention as the root node. The first layer connects the implementing agency and resource allocation nodes, the second layer connects the publicity and promotion and inheritance activity nodes, and the third layer connects the result nodes of public participation and the breadth of skill dissemination. The evolution trajectory of relationship weights is marked by timestamps. S3: Based on the dynamic policy knowledge graph described in S2, a time-series comparative learning encoder is constructed. A Siamese network structure is used to construct positive sample pairs for the implementation cases of the same policy in different regions / time periods. At the same time, a control region where the policy has not been implemented is introduced as a negative sample. The distance distribution of the feature space is optimized through metric learning to generate a policy variable embedding vector with time invariance. S4: Based on the policy variable embedding vector described in S3, perform confounding variable clarification operation, which specifically includes: calculating the correlation coefficient matrix between the feature vector and external factors such as seasonal fluctuations and economic indicators, using the orthogonal projection method to eliminate interference from non-policy factors, and outputting the corrected set of policy-driven variables as input for causal inference. S5: Deploy a counterfactual path reasoning engine in the dynamic policy knowledge graph described in S2, perform causal intervention operations based on do-calculus, use the Monte Carlo tree search algorithm to traverse all potential causal paths, combine the variable set corrected in S4 to calculate the effect contribution value of each path, and generate a path stability scoring function to evaluate cross-group consistency. S6: Based on the effect contribution value and path stability score described in S5, construct a causal effect attribution matrix, use the Shapley value decomposition method to quantify the net causal effect of each policy tool, generate a visual evaluation report containing the trend of changes in the intensity of the main effect path, and output the threshold judgment basis for the effectiveness of policy intervention.

[0010] Step S1 involves collecting multi-source heterogeneous data from government documents, media reports, social media, and interviews with intangible cultural heritage inheritors. Structured elements, including policy subjects, policy tools, implementation milestones, inheritance projects, participating populations, dissemination scope, and funding, are extracted and chronologically aligned and standardized to form a policy-behavior-response related dataset. Specifically, this includes: S1.1: Based on the government-issued policy texts on the protection of intangible cultural heritage related to sports, execute the named entity recognition algorithm in natural language processing to identify and extract structured policy elements such as policy subjects, policy tools, and implementation time points to obtain a standardized set of policy variables; Based on the government-released policy texts on the protection of intangible cultural heritage in the sports category, a bidirectional encoded named entity recognition algorithm (parameters: pre-trained model BERT, regularization coefficient 0.01, and label set covering policy subjects / policy tools / implementation time nodes) is used to determine the entity boundaries and categories of the input policy text. Furthermore, by using a Conditional Random Field (CRF) post-processing algorithm (parameters: the dimension of the transition matrix equals the number of entity labels, and the state constraint rules are based on a policy domain dictionary), the optimal sequence labeling of continuous multi-word entities is achieved, and a preliminary structured entity result set covering policy subjects, policy tools, and implementation time nodes is obtained. Furthermore, a context-related filtering algorithm based on dependency parsing (parameters: set of syntactic dependency relations, window span set to 5) is adopted to realize semantic dependency verification of the preliminary entity result set, eliminate erroneous entities caused by syntactic ambiguity, supplement missing policy tool detail attributes, and generate a policy element set with enhanced semantic consistency. Furthermore, for the implementation time node element, a time normalization algorithm is executed (parameters: base time format YYYY-MM-DD, fuzzy time parsing rules cover quarterly / annual expressions) to map the time description in the text into standardized timestamp data to support subsequent cross-source time series alignment; By using standardized coding mapping algorithms, policy subjects, policy tools, and implementation time nodes that have undergone semantic dependency verification and timestamp normalization are converted into a unified knowledge representation format, such as the triple <policy subject, policy tool used, implementation time node>, to realize the output of a set of policy variables that can be used for correlation analysis and knowledge graph construction. For example, given the policy text of the "Special Action Plan for the Protection of Intangible Cultural Heritage in Sports," the BERT named entity recognition model is applied with a maximum sequence length of 512, a batch size of 16, and a learning rate of 2e-5. The identified entities include "General Administration of Sport of China" as the policy subject, "Inheritance Activity Subsidy Policy" as the policy tool, and "First Quarter of 2022" as the implementation time point. After CRF post-processing, the entity boundaries of "General Administration of Sport of China" are adjusted to maintain a perfect match with the dictionary; dependency parsing is used to supplement the limited object "targeting youth participation groups" in the policy tool description; time normalization maps "First Quarter of 2022" to the timestamp 2022-03-31. The final standardized output triple format is <General Administration of Sport of China, Inheritance Activity Subsidy Policy (targeting youth participation groups), 2022-03-31>. This set significantly improves the retrieval accuracy of policy variables and the reliability of time-series alignment in subsequent steps. S1.2: Perform sentiment analysis and topic modeling algorithms on text information related to sports intangible cultural heritage in media reports and social media platforms to identify public response tendencies and dissemination popularity in order to generate social response feature vectors; For input data including text information related to sports-related intangible cultural heritage from online media reports and mainstream social media platforms, a convolutional neural network text feature extraction method (parameters: word vector dimension 300, convolution kernel size 3, 4, 5) is used to capture semantic patterns and perform preliminary feature processing of the original text within a local context window. Furthermore, a sentiment analysis method based on Bi-LSTM (parameters: 128 hidden layer units, Dropout ratio 0.5) is used to classify the sentiment polarity of key phrases extracted by convolution, and obtain a quantitative score of three dimensions for each text: positive response value, negative response value and neutral response value. Furthermore, the topic modeling algorithm of Latent Dirichlet Allocation (LDA) (parameters: number of topics K=20, number of iterations 500) is used to perform topic distribution analysis on the sentiment-annotated text set, generating a topic weight vector and topic tag set for each text. Furthermore, by using a weighted fusion method of sentiment polarity distribution and topic weight vector (parameters: sentiment weight 0.6, topic weight 0.4), a comprehensive measurement of public response tendency is achieved, and a social response feature vector representation of uniform length is generated, where each dimension corresponds to a response intensity value of a sentiment-topic combination; By using feature normalization and principal component analysis (PCA) (parameter: retain the first 50 principal components), the fusion vector from the previous step is transformed into a low-dimensional compressed social response feature vector, achieving the expected technical effect of reducing computational complexity and noise interference while retaining the main information. For example, in a social media data processing scenario targeting sports-related intangible cultural heritage protection policies, the input includes 5,000 news reports from 5 mainstream media platforms and 20,000 user comments from 3 social media platforms. A convolutional neural network (CNN) with a word vector dimension of 300, kernel sizes of 3, 4, and 5, and 128 kernels per kernel, performs convolution and max pooling on the text to obtain a preliminary feature matrix. A bidirectional LSTM layer with 128 hidden units and a Dropout ratio of 0.5 outputs sentiment analysis scores for each text, such as 0.72 for positive response, 0.15 for negative response, and 0.13 for neutral response. The LDA model sets the number of topics K=20, achieving a stable topic distribution after 500 iterations. Each text corresponds to a 20-dimensional topic weight vector, for example, topic 5 has a weight of 0.28, and topic 12 has a weight of 0.19. The sentiment distribution and topic weights were weighted and fused using 0.6 and 0.4 respectively, resulting in a fused response vector of length 60 (20 topics × 3 sentiment dimensions). After Z-score normalization of this fused vector set, PCA was used to retain the top 50 principal components, forming the final social response feature vector set. The mean variance retention rate was significantly improved, and when this feature vector was subsequently applied in the policy effect evaluation model, it could significantly enhance the stability and accuracy of causal path identification. S1.3: Perform speech recognition and semantic role labeling algorithms on the interview records of intangible cultural heritage inheritors to extract information on inheritance behavior descriptions, including the frequency of inheritance activities, characteristics of the participants, and the path of skill dissemination, in order to construct a cultural inheritance behavior feature matrix; For the digitized audio data of interviews with intangible cultural heritage inheritors, an adaptive noise suppression algorithm (parameters: noise estimation window length of 30 frames, smoothing factor of 0.9) is used to suppress background noise and environmental reverberation, so as to ensure that the signal-to-noise ratio of subsequent speech recognition meets the input condition of ≥20dB. Furthermore, an end-to-end deep speech recognition model (parameters: CTC decoder threshold set to 0.7, acoustic model of Conformer structure) is used to transcribe audio signals frame by frame and obtain word-by-word text sequences containing timestamps and confidence levels to support the temporal localization of subsequent transmission activities. Furthermore, a semantic role labeling algorithm (parameters: based on BiLSTM-CRF structure, word vector dimension of 300, window size of 5) is adopted to automatically parse the predicate-argument structure in the transcribed text and obtain semantic slot labeling results involving inheritance actions, locations, participating groups and frequencies; Furthermore, using regularized frequency analysis, periodic features are extracted from time-related statements in the annotation results, and frequency indices of heritage activities are calculated to quantify the periodic intensity of heritage activities. Furthermore, by combining place name recognition and mapping algorithms, the location entities in the interview records are matched to the geocoding database, the geographical distribution characteristics of the participants are extracted, and a statistical table containing the feature vectors of the participants (including age group, gender ratio, and geographical code) is generated to characterize the multidimensional features of the participants. Furthermore, by using social network analysis methods (parameters: node threshold ≥ 5, edge weight threshold ≥ 0.3), a relational network is constructed for the technology transmission path. Based on the interaction frequency and resource flow between entities, a topological structure matrix of the skill transmission path is generated to describe the transmission association strength between different transmission nodes. By using a multi-dimensional feature fusion algorithm (parameters: principal component number = 10, normalization method = Z-score), the frequency index of inheritance activities, the feature vector of participating groups, and the skill dissemination path matrix are fused into a cultural inheritance behavior feature matrix, thereby achieving a unified vectorized representation of inheritance behavior. For example, audio recordings of interviews with intangible cultural heritage inheritors collected in a certain city, after adaptive noise suppression to improve the signal-to-noise ratio to 22dB, were input into a speech recognition system with a Conformer acoustic model. The output text sequence was 1800 words long with an average confidence level of 0.85. A semantic role labeling model, using a window size of 5, parsed the interview text, extracting 20 categories of inheritance actions, 15 locations, and 9 characteristics of the participating groups. In frequency analysis, within the time window... Within a month, a total of [number] heritage activities were conducted. Frequency index =5. After location mapping, the participants were distributed across 8 districts and counties, with the age group concentrated between 30 and 50 years old and a gender ratio of 6:4. Analysis of the skill transmission path showed that the core node degree was 8, the average edge weight was 0.45, and 3 main transmission paths were formed. Through principal component analysis and Z-score normalization, a cultural heritage behavior feature matrix with a dimension of 10×500 was obtained, providing high-quality input for the next step of multimodal temporal alignment and significantly improving the accuracy and stability of subsequent causal path modeling. S1.4: Perform a multimodal temporal alignment algorithm on structured elements from different data sources, and align policy implementation nodes, media response times and inheritance behavior change cycles based on timestamps to establish a unified time dimension reference system; S1.5: Based on the policy variable set, social response feature vector, and cultural heritage behavior feature matrix extracted from S1.1 to S1.4, a standardized coding mapping algorithm is executed to map heterogeneous data into a triple representation in a unified format, so as to construct a policy-behavior-response related dataset.

[0011] Step S2: Based on the associated dataset described in S1, a dynamic policy knowledge graph containing timestamps is constructed. A multi-level causal network is established with policy intervention as the root node. The first layer connects implementing agencies and resource allocation nodes, the second layer connects publicity and promotion and inheritance activity nodes, and the third layer connects result nodes of public participation and the breadth of skill dissemination. The evolution trajectory of relationship weights is marked by timestamps. Specifically, this includes: S2.1: Perform entity recognition and relation extraction processing on the structured elements in the policy-behavior-response association dataset described in S1 to obtain the entity set of policy subject, policy tool, implementation node, inheritance project, participating population, dissemination scope and funding investment and the semantic relationships between each pair, forming an initial set of triples; For the structured elements in the policy-behavior-response association dataset described in S1, an entity recognition method that integrates rule templates and sequence labeling models is adopted (parameters: pre-trained BERT embedding dimension 768, initial value of CRF decoding layer transition matrix is ​​uniformly distributed) to achieve accurate classification and boundary determination of category labels such as policy subjects, policy tools, implementation nodes, inheritance projects, participants, dissemination scope and funding in text data; Furthermore, a semantic dependency analysis algorithm (parameter: dependency tree depth not exceeding 5 orders) is used to extract syntactic and semantic dependency relationships between entities, and weighted entity connection data is obtained to determine the subject-verb, verb-object, and causal relationship types and their strengths between entities. Furthermore, a relation extraction method based on pattern matching is adopted (parameter: the matching pattern library contains 56 policy context templates) to achieve entity relation recognition across sentences or paragraphs and generate a relation set that conforms to the definition of knowledge graph triples; Furthermore, a relation classification and normalization algorithm (parameter: the number of categories in the Softmax classification layer is set to 12) is applied to classify the extracted relations by type and perform unified encoding on similar relation labels to eliminate relation redundancy caused by synonyms; Furthermore, a duplicate entity fusion algorithm is adopted (parameters: entity similarity threshold is set to 0.85, fusion strategy is weighted average) to merge multiple representations of the same entity caused by differences in data sources into a single node representation, generating a redundant entity set and its semantic relationships; By using entity recognition and relation extraction, the structured elements from the previous step are transformed into an initial set of triples that conform to the knowledge graph construction specifications, thereby achieving a precise mapping of multi-source policy intervention information to a computable causal network. For example, in a collection of policy texts related to the protection of intangible cultural heritage in the sports category, each policy document, after being processed by the BERT-CRF model, outputs a tag sequence of 1200 tokens, with a significant improvement in the accuracy of policy subject identification. In the dependency analysis phase, 420 subject-verb relationships and 180 verb-object relationships were extracted, with an average dependency tree depth of 3. The pattern matching phase identified 112 cross-sentence relationships, covering the causal relationship between funding and implementing agencies, as well as the promotional relationship between publicity and public participation. In the relationship classification phase, funding relationships were uniformly encoded as "funding support," and policy implementation was categorized as "policy implementation behavior," ultimately resulting in 12 standard relationship tags. During the duplicate entity fusion process, "Municipal Cultural Bureau" and "XX Municipal Cultural Bureau" were merged into a unified node, reducing the size of the entity set from 320 to 280. The final output initial triple set contained 280 entity nodes and 920 relationship edges, laying a complete and accurate data foundation for the subsequent construction of a static policy knowledge graph. S2.2: Based on the set of triples obtained in S2.1, construct the basic topology of the static policy knowledge graph. With policy intervention as the root node, establish a hierarchical causal network architecture with the first layer connecting the implementing agency and resource allocation nodes, the second layer connecting the publicity and promotion and inheritance activity nodes, and the third layer connecting the result nodes of public participation and the breadth of skill dissemination. Using the initial set of triples output by S2.1 as input data, a graph structure construction algorithm (parameters: node type = policy intervention, implementing agency, resource allocation, publicity and promotion, inheritance activities, public participation, breadth of skill dissemination; relation type = causal relationship) is adopted to set the policy subject node as the root node of the causal network; Furthermore, by using a hierarchical layout method (parameters: number of layers = 3, inter-layer association rules = policy intervention links implementation and resources, implementation and resources link publicity and promotion and inheritance activities, publicity and promotion and inheritance activities link to result nodes), the direct association structure between the first-layer node set and the root node is realized, and the first-layer structure data containing the implementation agency node and the resource allocation node is obtained. Furthermore, through the causal relationship mapping algorithm (parameter: mapping basis = semantic relationship label in triple), the association matching between the second-level nodes—promotion and publicity nodes and heritage activity nodes—and the first-level nodes is realized, and the relationship edge set of this layer is generated; Furthermore, through the result indicator mapping algorithm (parameters: indicator type = public participation, skill dissemination breadth), directed edges are connected between the third-layer nodes and the second-layer nodes, and a final set of nodes reflecting the result variables is generated; By integrating node attributes and generating relationship matrices, the results of the previous step are transformed into the topological data structure of a static policy knowledge graph, thus achieving the initial modeling of a complete multi-level causal network. For example, in a sports intangible cultural heritage protection policy case, the input is a set of triplets containing the nodes "2020 Sports Heritage Support Policy," "Municipal Cultural Bureau" (the implementing agency), "5 million yuan in special funds" (resource allocation), "television promotion" and "offline lectures" (promotional activities), "twice-monthly skill demonstrations" (frequency of inheritance), "average number of participants of 200" (public participation rate), and "coverage of 10 provinces" (scope of skill dissemination). The algorithm is constructed with the root node "2020 Sports Heritage Support Policy," the first-level connection nodes being "Municipal Cultural Bureau" and "5 million yuan in special funds," the second-level connection nodes being "television promotion" and "offline lectures," which are connected to the implementing agency and resource allocation nodes of the first level, respectively; and the third-level connection nodes being "average number of participants of 200" and "coverage of 10 provinces," which are connected to the promotional and inheritance activity nodes of the second level, respectively. In the process of generating the relation matrix, the weight of the first-level relation is set as follows: The weight of the second-level relationship is The weight of the third-level relationship is The generated static policy knowledge graph exhibits complete node structure and clear causal relationships during the verification phase, providing an accurate initial topology for subsequent timestamp annotation and dynamic graph evolution modeling. S2.3: Extract and standardize the time information of the multi-source heterogeneous data collected in S1, and use timestamps to mark the time of each policy intervention and its subsequent response events in order to construct event sequence data in the time dimension as the time basis for the evolution of the dynamic map; S2.4: Based on the time series data obtained in S2.3, a time-aware graph neural network is used to dynamically evolve the static knowledge graph constructed in S2.2. A timestamp attribute is added to each node and edge in the graph, and a weight update mechanism based on a time window is constructed to realize the dynamic evolution capability of the graph structure. Based on the time series data obtained in S2.3, a time-aware graph neural network (parameters: time window size Δt, node embedding dimension d, edge weight initialization method) is used to perform dynamic graph modeling on the static knowledge graph constructed in S2.2, so as to realize the structural update of nodes and edges as they evolve over time. Furthermore, by using the temporal convolution method (parameters: kernel width k, stride s), convolution is performed on the node feature matrix of the time series data to extract local temporal patterns, mapping the state changes of the same node in different time slices into high-dimensional time-aware features; Furthermore, the sin-cos position embedding function (parameters: fundamental frequency ω, dimension mapping matrix W) is used to transform the timestamps of nodes and edges into continuous trainable time vectors, which are then fused with the original semantic embedding vectors element-wise to obtain the time-annotated node and edge representations. Furthermore, through the temporal attention mechanism (parameters: number of attention heads h, scaling factor α), the temporal correlation weights between nodes are calculated within the time window Δt, and the edge weights within the same time slice are updated according to the weights, ensuring that the node connection relationship is sensitive to changes in key time slices during the temporal evolution process; Furthermore, a weight update mechanism based on a sliding time window (parameters: sliding step size p, update threshold τ) is adopted to recalculate the average value and variance of the edge weights between nodes within each sliding window, and adjust the weight increase or decrease of the edges according to the preset threshold τ to realize the dynamic evolution of the graph over time. By using a time-aware graph neural network and a sliding time window algorithm, the results of the previous step are transformed into a dynamic graph structure containing timestamp attributes, enabling the dynamic evolution of nodes and edges in the time dimension, and providing a foundation for subsequent dynamic updates of relation weights and causal path analysis. For example, in a sports intangible cultural heritage policy knowledge graph containing 500 policy intervention events, the time window size Δt is set to 30 days, the node embedding dimension d is 128, and the edge weights are initialized to mutual information values ​​based on the initial topology of S2.2. A temporal indexed convolution algorithm is used, with a kernel width k of 5 and a stride s of 1, performing convolution on the node time-series feature matrix to extract the local temporal patterns of each node over the past 150 days. The temporal encoding function uses sin-cos positional embedding with a fundamental frequency ω of 0.01. The dimension mapping matrix W is randomly orthogonally initialized to convert timestamps into 128-dimensional time vectors, which are then fused with the original semantic embeddings through element-wise addition. The temporal attention mechanism uses 4 attention heads h and a scaling factor α of 0.5 to calculate the temporal correlation weights between nodes within the 30-day window. The resulting correlation score matrix shows a significant strengthening trend during the key policy release period. The sliding time window weight update mechanism has a sliding step size p of 15 days and an update threshold τ of 0.05. In each window period, the mean and variance of the edge weights are recalculated. Edges with a mean increase exceeding τ undergo a weight gain operation, while edges with a mean decrease exceeding τ undergo a weight decay operation. After implementing the above configuration, the output dynamic graph shows a significant reconstruction of the node association structure during the policy release period. The dynamic change curve of the edge weights highly matches the time-series trajectory of the policy impact, providing high-precision time-series input for the next step of dynamic update of relationship weights in S2.5. S2.5: Perform dynamic update operation on relation weights for the dynamic policy knowledge graph constructed in S2.4. Based on the actual impact intensity of policy intervention events on outcome nodes (such as public participation and the breadth of technology dissemination) at different time nodes, a sliding time window mechanism is used to periodically evaluate and adjust the weights of the edges between nodes in order to generate a causal path network with temporal evolution characteristics.

[0012] like Figure 2 As shown, step S3 involves constructing a time-series comparative learning encoder based on the dynamic policy knowledge graph described in S2. A Siamese network structure is used to construct positive sample pairs for implementation cases of the same policy in different regions / time periods. Simultaneously, control regions where the policy was not implemented are introduced as negative samples. Metric learning is used to optimize the distance distribution in the feature space, generating a policy variable embedding vector with time-invariant properties. Specifically, this includes: S3.1: Based on the multi-level causal network structure in the dynamic policy knowledge graph described in S2, a temporal contrastive learning encoder with a Siamese network architecture is constructed. This encoder consists of two LSTM-Transformer hybrid models with shared weights, which are used to encode the temporal feature vectors of the target policy case and its contrast sample, respectively. The input of the multi-level causal network structure based on the dynamic policy knowledge graph described in S2 includes the time-series correlation information of policy intervention nodes, implementing agency nodes, resource allocation nodes, publicity and promotion nodes, inheritance activity nodes, public participation nodes, and skill dissemination breadth nodes; A twin network architecture design method is adopted (parameters: shared weight mechanism is set to True, and the input layer dimension is consistent with the feature dimension of the causal network node) to realize a parallel dual-branch temporal coding structure to support synchronous feature extraction of target policy cases and comparison samples; Furthermore, a long short-term memory unit is constructed using an LSTM module (parameters: the number of hidden layer units is set to 128, and the time step is the time window length of the dynamic knowledge graph) to realize the local temporal dependency modeling of node time series features and output a feature vector sequence that captures short-term causal changes. Furthermore, through the Transformer encoding module (parameters: the number of multi-head attention heads is set to 8, and the position encoding range covers the entire time window length), global feature association modeling across time slices is realized, and a high-dimensional global temporal feature vector containing time-dependent weight distribution is generated; Furthermore, a feature-level connection processing method is adopted to concatenate the local temporal vector output by LSTM and the global temporal vector output by Transformer in the feature dimension, forming a comprehensive temporal feature representation that integrates short-term and long-term dependencies; By using a Siamese network structure with shared weight parameters, the above-mentioned comprehensive temporal features are synchronously encoded in the target policy case branch and the comparison sample branch, and paired high-dimensional temporal embedding vectors are output to achieve policy variable encoding under temporal pattern alignment. For example, in a sports-related intangible cultural heritage protection policy case, the input time window of the dynamic policy knowledge graph is set to 12 months, the node feature dimension is 64, and the causal network contains a 3-layer structure with a total of 18 nodes. Both branches of the Siamese network are configured with a 128-unit LSTM module and an 8-head attention Transformer module, with a position encoding range of 12, aligned with the time window of the dynamic knowledge graph. The target policy case branch receives the feature sequences of policy intervention nodes and related result nodes in a certain region over a continuous 12-month period, while the comparison sample branch receives data from another region within the corresponding time window. The LSTM module sequentially performs local dependency modeling on the time series of each node, and the Transformer module establishes an attention matrix for the global temporal evolution pattern between nodes. After feature concatenation, a comprehensive temporal feature vector is obtained. Shared weights ensure consistency in the encoding patterns of the two branches, ultimately outputting two sets of 128×12-dimensional high-dimensional embedding vectors for subsequent sample pair construction and metric learning optimization. The test results of this configuration show that feature fusion can significantly improve the accuracy of capturing the temporal invariance of policy variables and support the efficient execution of subsequent confounding factor clarification steps. S3.2: Construct sample pairs for the standardized coded policy-behavior-response association dataset in S2. Select implementation records of the same policy in different regions or at different times as positive sample pairs, and introduce data from control regions that did not implement the policy as negative sample pairs to enhance the model's ability to distinguish policy variables. S3.3: Perform time-aware feature encoding on each sequence data in the constructed sample pair, use the LSTM module to extract local temporal dependencies, and model the global temporal evolution mode through the self-attention mechanism of Transformer to generate a high-dimensional temporal embedding vector as the encoder output; For the sample pair sequence data constructed by S3.2, a time-aware long short-term memory network (LSTM, parameters: number of hidden units h=128, time step T is determined by the effective period of the dynamic knowledge graph node) is used to model the local temporal dependencies and extract feature representations that reflect the change patterns of policy variables in adjacent time windows. Furthermore, by superimposing a Transformer encoder based on a multi-head self-attention mechanism (parameters: number of attention heads a=8, number of coding layers l=4, positional coding form is sine-cosine hybrid coding) on ​​the LSTM output sequence, the global temporal evolution pattern is captured, and a high-dimensional embedding representation that can retain long-term dependencies across time periods is obtained. Furthermore, the structural stability of the Transformer encoder output is enhanced through layer normalization and residual connection operations, maintaining the numerical stability of high-dimensional embeddings and providing a discriminable spatial structure for distance calculation in metric learning. Furthermore, a feature concatenation and weighted fusion method is adopted to combine the local temporal feature vector extracted by LSTM and the global evolutionary feature vector extracted by Transformer according to their weights. , The vectors are then fused to form the final high-dimensional temporal embedding vector. Through the above hybrid encoding structure, the sample pair sequence data output in the previous step is transformed into policy variable embedding vectors with significantly enhanced time invariance in the latent feature space, thereby greatly improving the similarity measurement effect under subsequent triplet loss optimization. For example, when encoding a sequence of sports intangible cultural heritage policy samples, the LSTM is set with 128 hidden units and a time step of 30, corresponding to a natural month cycle in the policy implementation phase. The input sequence is a standardized policy-behavior-response matrix. The LSTM output size is 128×30, which is processed by a Transformer encoder with parameters of 8 attention heads and 4 encoding layers. Position encoding uses a sine-cosine hybrid encoding, and the output is a 128×30 context-enhanced feature. Layer normalization and residual connections are performed on the Transformer output, using weighted... =0.6、 =0.4 The local features of LSTM and the global features of Transformer are weighted and fused to generate a high-dimensional embedding vector of length 128. In the metric space corresponding to this embedding, the Euclidean distance for positive sample pairs is reduced to For negative sample pairs, the Euclidean distance is increased to This significantly improves the feature separability in subsequent triplet loss function optimization and achieves robust representation of policy variables under cross-regional and cross-time period comparisons; S3.4: Based on the embedding vectors generated in S3.3, the triplet loss function is used for metric learning optimization, in which the Euclidean distance between positive sample pairs is compressed, while the distance between negative samples and anchor points is increased, thereby forming a time-invariant clustering structure in the latent feature space and generating policy variable embedding vectors. Based on the high-dimensional temporal embedding vector generated by S3.3, the triplet loss function (parameters: Margin threshold is 0.5, distance metric is Euclidean distance) is used to optimize the inter-class separation and intra-class compactness of the feature space. Furthermore, using the Euclidean distance calculation formula between embedding vectors, distance compression is performed between positive sample pairs and anchor samples, and distance stretching is performed between negative samples and anchor samples to form a clustering structure with time invariance and causal consistency. (Euclidean distance) The calculation formula is as follows;

[0013] in, Let i be the i-th dimension embedding value of the anchor sample. To compare the i-th dimension of the embedding value of the sample, where n is the dimension of the embedding vector; Furthermore, using the triplet loss function optimization strategy, the following objective function is constructed. :

[0014] in, This represents the distance between anchor point a and positive sample p. This represents the distance between anchor point a and negative sample n. This is a preset interval threshold parameter; Furthermore, by updating the encoder weights based on the gradient of the objective function through backpropagation algorithm, the geometric optimization of temporal features in the latent space is achieved, so that implementation cases affected by the policy are clustered in the same cluster in the feature space, while implementation cases unaffected by the policy are separated into different clusters. Furthermore, by iteratively updating the triplet set in the training batches, the most "difficult" positive and negative sample combinations are selected using a dynamic sample mining strategy, thereby improving the model's ability to distinguish heterogeneous policy intervention paths. By optimizing through metric learning, the embedding vectors maintain similar distance distribution characteristics across different time periods, thereby forming a stable time-invariant policy variable representation in the latent feature space and generating policy variable embedding vectors. For example, in the scenario of evaluating policy interventions related to intangible cultural heritage in sports, the embedding vectors of the same policy implemented in the eastern region in 2020 and the embedding vectors of the same policy implemented in the western region in 2021 are used as positive sample pairs, and the embedding vectors of the southern region, where the policy was not implemented in 2020, are selected as negative samples. Assuming the embedding vector dimension n is 128, after normalization, the mean distance between the anchor sample and the positive sample is calculated using Euclidean distance. The mean distance between the anchor sample and the negative sample is Margin threshold The result of the triplet loss function calculation is as follows: = Therefore, this batch of samples will not generate gradient updates. Triplet combinations with distance differences close to the Margin threshold are selected through dynamic sample mining; for example, the distance between the anchor sample and the positive sample is... The negative sample distance is At that time, the loss value is This triggers the gradient update process. After optimization, the intra-class variance of the embedding vectors of the same policy in different regions in the feature space is significantly reduced, the similarity structure of cross-regional policy embeddings is stable, and it exhibits a higher interference removal capability in the subsequent confounding variable clarification step; S3.5: In the optimized feature space, perform nearest neighbor retrieval operation on all policy cases to identify and filter out pseudo-related samples that have similar implementation paths to the target policy but are affected by non-policy factors, and use them as the input candidate set for the subsequent confounding variable clarification module.

[0015] Step S4: Based on the policy variable embedding vector described in S3, perform a confounding variable clarification operation, specifically including: calculating the correlation coefficient matrix between the feature vector and external factors such as seasonal fluctuations and economic indicators; using orthogonal projection to eliminate interference from non-policy factors; and outputting the corrected set of policy-driven variables as input for causal inference. Specifically, this includes: S4.1: Based on the policy variable embedding vector output by S3, construct a multi-dimensional feature space matrix, where each row represents a high-dimensional feature vector of a policy implementation case, and each column represents a standardized policy-related variable dimension, to form the initial input data structure for identifying confounding factors; Based on the policy variable embedding vector output by S3, a matrix data structure mapping method is used (parameters: embedding vector dimension = H, number of policy cases = N) to encode each policy implementation case as a row vector. Furthermore, by using a dimensionality standardization algorithm (parameters: mean μ and standard deviation σ are derived from training set statistics), the numerical range of each embedding dimension is unified, and a standardized embedding matrix is ​​obtained. Furthermore, a variable dimension labeling method (parameter: dimension index and policy element mapping table) is adopted to realize the explicit correspondence between the columns of the embedded matrix and the policy-related variable dimensions, and to generate a feature matrix with semantic labels; Furthermore, a sparsification algorithm (parameter: sparsity threshold τ) is used to remove redundant dimensions that are close to zero in all cases and to compress the matrix storage space. By constructing an index using matrix retrieval, the results of the previous step are transformed into an initial input data structure that supports the identification of confounding factors, thereby achieving the expected technical effects of subsequent covariate expansion and correlation analysis. For example, in the scenario of evaluating the effectiveness of sports intangible cultural heritage protection policies, let N=500 policy implementation cases, with each case having an embedding vector dimension H=128. A matrix mapping method is used to stack all vectors into a 500×128 two-dimensional matrix M. Z-score standardization is then performed on matrix M, as shown in the following formula:

[0016] in, For embedding vector elements, This represents the mean of the corresponding dimension. The standard deviation is the corresponding dimension. After standardization, the first column is labeled as "policy implementation time span" and the second column as "resource input intensity" through a mapping table, until all dimension labels are bound by the 128th column. A sparsity threshold τ = 0.001 is set, and L1 norm calculation is performed on each column of matrix M. Columns with norm results lower than τ are filtered out. After compression, the matrix dimensions are reduced to 500×120, retaining dimensions that significantly contribute to the differences in policy variables. Finally, an initial input data structure with a retrieval index is constructed. Under hardware conditions, this structure can query all feature dimensions of any policy case within milliseconds, significantly improving the efficiency and stability of subsequent confounding variable identification and causal analysis. S4.2: The feature space matrix constructed in S4.1 is expanded by covariate expansion, introducing non-policy external variables such as seasonal fluctuation indicators, regional economic indices, and population density changes, in order to construct an expanded feature matrix containing potential confounding factors for subsequent correlation analysis and variable separation. Based on the multidimensional feature space matrix formed in S4.1, an external data association method (parameters: seasonal fluctuation index data source, regional economic operation index database, population statistics system) is used to introduce non-policy external covariates into the current feature space for subsequent identification and separation of confounding factors. Furthermore, by using time series interpolation (parameter: timestamp alignment window length is set to 1 month), the seasonal fluctuation indicator is accurately mapped to the time scale of each policy case, and a seasonal covariate array consistent with the embedded vector dimension is obtained; Furthermore, a factor scaling normalization algorithm is adopted (parameter: the maximum-minimum scaling interval is set to 0 to 1) to achieve the scaling of regional economic index data and generate an economic covariate vector to keep it consistent with the embedded spatial scale; Furthermore, by using a geographic weighted regression method (parameter: spatial weight matrix based on geographic distance decay function), the population density change data is mapped to the policy implementation area, and the population change covariates corresponding to each sample instance are obtained; By performing matrix column concatenation operations, the aforementioned seasonal covariate array, economic covariate vector, and population change covariate are added column by column to the original multidimensional feature space matrix, thereby constructing the extended feature matrix. For example, in an evaluation scenario of the effectiveness of a sports intangible cultural heritage protection policy, the initial feature space matrix has a dimension of 100×50, where rows represent 100 policy implementation cases and columns represent 50 standardized policy-related variables. When introducing non-policy covariates, the seasonal fluctuation index data comes from the National Meteorological Center and is updated monthly. Time series interpolation is used to align the implementation time of each policy to the corresponding month, resulting in a 100×1 column of seasonal covariates. The regional economic index data comes from provincial statistical bureaus, using quarterly year-on-year GDP growth rate data, and is normalized to a 100×1 column of economic covariates. The population density change data comes from the National Population Census and Annual Population Sampling Survey. Geographically weighted regression is used to map the implementation areas of each policy to population change values, resulting in a 100×1 column of population change covariates. The extended feature matrix after concatenating the matrix columns has a dimension of 100×53, with the three newly added columns being seasonal, economic, and population change covariates. In this embodiment, the mean value of the seasonal fluctuation covariate is 0.45, the mean value of the economic covariate is 0.62, and the mean value of the population change covariate is 0.08. The expanded feature matrix can significantly improve the ability of subsequent correlation analysis to identify confounding factors and provide complete input for the calculation of Pearson correlation coefficient in step S4.3. S4.3: Based on the extended feature matrix output by S4.2, calculate the Pearson correlation coefficient matrix between the policy variable embedding vector and each external confounding factor, identify the set of confounding variables that are highly correlated with the policy variables, and quantify the influence of non-policy interference factors on the policy variable embedding space. S4.4: The orthogonal projection method is used to perform linear algebra processing on the set of confounding variables identified in S4.3, and the original policy variable embedding vector space is projected into a subspace orthogonal to the confounding factors to eliminate the linear influence of the confounding factors in the feature space and generate a biased policy-driven embedding vector. S4.5: Normalizes and aligns the debiased policy embedding vector output from S4.4 to generate a standardized set of policy-driven variables, which serves as input to the subsequent counterfactual path inference engine to support accurate modeling and evaluation of causal effect paths.

[0017] Step S5: Deploy a counterfactual path reasoning engine in the dynamic policy knowledge graph described in S2, perform causal intervention operations based on do-calculus, traverse all potential causal paths using the Monte Carlo tree search algorithm, calculate the effect contribution value of each path using the variable set corrected in S4, and generate a path stability scoring function to assess cross-group consistency. Specifically, this includes: S5.1: Based on the topological structure of the dynamic policy knowledge graph described in S2, construct a causal reasoning graph model, take the policy intervention node as the causal intervention variable, perform do-calculus operation to simulate the counterfactual scenario of 'if the policy had not been implemented', and generate a causal graph structure containing the state before and after the intervention, as the basic input for path reasoning; Based on the topological structure of the dynamic policy knowledge graph described in S2, a directed acyclic graph modeling method is adopted (parameters: node type = policy intervention, implementing agency, resource allocation, publicity and promotion, inheritance activities, public participation, breadth of skill dissemination; edge type = causal relationship, directed attribute) to achieve the initial construction of a causal reasoning graph model; Furthermore, a graph structure regularization algorithm (parameters: structure sparsity threshold = 0.15, node degree limit = maximum 5) is used to remove redundant or low-contribution edges in the model and obtain a basic causal graph dataset with simplified topology. Furthermore, an intervention variable specification algorithm (parameter: intervention node identifier = set of policy intervention nodes) is adopted to set policy intervention nodes as exogenous variables in the causal graph and generate a call list of intervention operations as the entry point for causal inference; Furthermore, by executing causal intervention operations through do-calculus rules (rule base: do-operator three-step calculus rules), the intervention variables in the causal graph structure are removed from their upstream dependencies, and a post-intervention graph structure that meets the requirements of counterfactual modeling is obtained; Furthermore, a state slice generation algorithm (parameters: time dimension resolution = month, state variable category = all result nodes) is adopted to realize the structured slice extraction of the state before and after intervention, and generate a mapping relationship matrix to represent the value comparison of each node before and after intervention; Through the above processing method of the causal reasoning graph generation module, the set of policy-driven variables in the previous step is transformed into a counterfactual causal graph structure that includes the pre-intervention and post-intervention states, so as to achieve the expected technical effect of subsequent path reasoning. For example, in evaluating the effectiveness of a sports intangible cultural heritage policy, the input consists of a set of bias-reduced policy-driven variables based on the output of S4 and the topology of the dynamic knowledge graph of S2. The policy intervention node includes "Policy Number P123," which connects to intermediate nodes such as "Implementing Agency A," "Resource Allocation B," and "Publicity and Promotion C," ultimately pointing to the result nodes "Public Participation R1" and "Breadth of Skill Dissemination R2." A directed acyclic graph (DAG) modeling method is used to configure node types and directed causal edges, forming an initial causal graph with 7 nodes and 12 edges. The execution graph structure is regularized, removing two low-contribution edges under a sparsity threshold of 0.15, while retaining key causal links. An intervention variable specification algorithm is used to set the policy intervention node as an exogenous variable, and a do-calculus rule is invoked to remove the upstream associations of this node, allowing it to act independently on the downstream result nodes. A state slicing algorithm was used to extract node states at monthly resolution before intervention (e.g., public participation = 210, skill dissemination breadth = 38) and after intervention (e.g., public participation = 260, skill dissemination breadth = 47), generating a mapping matrix containing 2 rows (before / after intervention) × 2 columns (R1 / R2 outcome indicators). Referring to this matrix provides a clear structured input for effect estimation in S5.2, enabling preparation for counterfactual path reasoning across time and conditions. S5.2: Based on the policy-driven variable set output by S4, construct a causal effect estimation function, use the potential outcome framework to model the expected response value on each causal path, estimate the causal effect of the variable set corresponding to each path in the graph model, calculate the expected difference in the counterfactual scenario, and form a preliminary assessment result of the path effect contribution. S5.3: Based on the preliminary evaluation results of the path effect contribution generated in S5.2, the Monte Carlo tree search algorithm is introduced to perform a depth-first extended search of causal paths in the dynamic knowledge graph. The search depth threshold and path pruning rules are set to avoid excessive consumption of computing resources by redundant paths, and a set of potential high-contribution causal paths is output. S5.4: Based on the set of high-contribution causal paths output by S5.3, a path stability scoring function is constructed. The cross-group consistency test method is used to perform variance analysis on the effect contribution of each path in multiple subgroups, calculate its inter-group consistency score, and screen out the main effect paths that are stable in multiple groups. S5.5: Based on the main effect path and its stability score output by S5.4, the path effect contribution value and stability score are integrated to construct a comprehensive evaluation index for causal paths. The output is a causal path evaluation table containing path number, causal effect intensity, path stability score and time evolution characteristics, which serves as the input basis for subsequent attribution analysis and visualization. Using the main effect paths and their cross-group stability scores obtained from S5.4 screening as the input dataset, a weighted fusion algorithm was employed (parameters set to the effect contribution values). Weighting of stability score This enables the construction of a comprehensive evaluation index for multidimensional causal paths; Furthermore, by using a normalization algorithm (with parameters set to the minimum-maximum normalization interval [0,1]), the numerical ranges of the effect contribution values ​​and stability scores under different scales are unified, and a normalized dual-index matrix is ​​obtained. Furthermore, a linear weighted summation method is employed (with weight parameters taken from the correlation coefficients between the stability score and the effect strength output in S5.4) to calculate the comprehensive score for each main effect pathway and generate a comprehensive score sequence. ,in Contribution value to the effect Assess stability score; Furthermore, a time series decomposition algorithm (parameters: sliding window length L=5, trend extraction method is moving average) is used to extract the temporal evolution features of the main effect path during the evaluation period and form a time-rating two-dimensional matrix; Furthermore, by using a path numbering encoding algorithm (encoding rule: root node identifier + hierarchical sequence number + result node identifier), the aforementioned comprehensive score and time evolution characteristics are associated with path numbers to generate a structured causal path evaluation table, and output a complete data structure containing path number, causal effect strength, stability score and time evolution trend indicators. Through the above-mentioned weighted fusion and time-series feature extraction processing methods, the main effect path screening results of the previous step are transformed into quantifiable and dynamic causal path comprehensive evaluation indicators, realizing the multi-dimensional interpretability and visualization usability of causal effect analysis. For example, in the scenario of evaluating the effectiveness of sports intangible cultural heritage policies, the 10 main effect paths output by S5.4 are processed. The numerical range of the effect contribution value matrix E is set to 0.25 to 0.85, and the numerical range of the stability score matrix R is set to 0.40 to 0.95. A minimum-maximum normalization method is used to map them to the [0,1] interval. The weighting parameters are set as follows: =0.6、 =0.4, the comprehensive scoring formula is: The comprehensive score sequence for each path is calculated. For example, the score for path P1 is... =0.84. A moving average trend extraction with a sliding window length of L=5 was performed on the score sequence to obtain the score change trend curves for each path during the evaluation period. The path number (e.g., "POL-01-ACT05-RES02"), corresponding score, stability, and time trend were correlated according to coding rules to generate a causal path evaluation table. During the validation phase, this evaluation table was used in the downstream visualization module to generate trend heatmaps and path intensity flow maps, significantly improving the interpretability of policy effect evaluation and the relevance of analytical decision-making.

[0018] Step S6: Based on the effect contribution value and path stability score described in S5, construct a causal effect attribution matrix, quantify the net causal effect of each policy tool using the Shapley value decomposition method, generate a visual assessment report containing the trend of main effect path strength changes, and output the threshold for judging the effectiveness of policy intervention. Specifically, this includes: S6.1: Based on the counterfactual path effect contribution value and path stability score output by S5, construct a causal effect attribution matrix, where rows represent policy tool nodes, columns represent outcome indicator nodes, and matrix elements represent the causal effect strength of the policy tool on the corresponding outcome indicator after controlling for confounding variables, in order to support the synergistic effect analysis of multidimensional policy tools; Based on the counterfactual path effect contribution value and path stability score output by S5, a matrix construction method (parameters: policy tool node set, outcome indicator node set, effect contribution numerical matrix) is used to initialize and define the dimensions of the causal effect attribution matrix. Furthermore, through a node mapping algorithm (parameters: unique identifiers for policy tool nodes and unique identifiers for outcome indicator nodes), precise indexing of matrix row and column coordinates is achieved, resulting in a structured row (policy tool) and column (outcome indicator) mapping table. Furthermore, a weighted allocation calculation method (parameters: effect contribution value, stability score) is adopted to realize the comprehensive calculation of matrix element values, where the matrix element values ​​are defined as the causal effect strength of the corresponding policy tool on the outcome indicator after controlling for confounding variables, and the full matrix numerical distribution is generated. Furthermore, by using a data normalization processing algorithm (parameters: row-based and column-based normalization modes), the effect intensity is standardized and adjusted at different scales to eliminate the differences in the units of different policy tools and outcome indicators, thus obtaining a normalized causal effect attribution matrix. Furthermore, a collaborative analysis model (parameters: principal components of the matrix, correlation determination threshold) is adopted to achieve a quantitative assessment of the synergistic effect of multiple policy tools and generate a collaborative mode index, forming a matrix structure that supports the collaborative analysis of multi-dimensional policy tools. By using matrix construction and normalization, the results of the previous step are transformed into a highly consistent and comparable causal effect attribution matrix, enabling quantitative characterization and synergistic effect analysis of the causal relationship between multidimensional policy tools and outcome indicators. For example, in the scenario of evaluating the effectiveness of sports-related intangible cultural heritage policies, the input includes four policy tool nodes (such as special fund allocation, public awareness campaigns, inheritor training, and sports events) and three outcome indicator nodes (public participation, breadth of skill dissemination, and improvement of cultural awareness), as well as the corresponding counterfactual effect contribution value matrix and stability score. For instance, the contribution of special fund allocation to public participation is 0.75, with a stability score of 0.80; the contribution of inheritor training to the breadth of skill dissemination is 0.68, with a stability score of 0.85. A weighted formula is used:

[0019] in, For matrix element values, Contribution value to the effect To score stability, the matrix elements for calculating the impact of special fund allocation on public participation are: By performing row-by-row and column-by-column normalization on the matrix, the maximum value of policy tools in the same row is adjusted to 1, and the remaining elements are scaled proportionally to form a normalized matrix [[1.00,0.72,0.55],[0.68,1.00,0.64],[0.59,0.78,1.00],[0.63,0.70,0.66]]. Principal component analysis of this matrix shows a synergistic effect between special fund allocation and public awareness in terms of public participation, and a high collaborative model index between competitions and inheritor training in terms of the breadth of skill dissemination. The final causal effect attribution matrix provides a quantitative basis for subsequent Shapley value decomposition and policy optimization. S6.2: Perform Shapley value decomposition on the causal effect attribution matrix described in S6.1, and allocate the total causal effect to each policy tool according to the marginal contribution ratio based on cooperative game theory, so as to quantify the net causal contribution of each policy tool to the overall intervention effect. S6.3: Based on the Shapley value decomposition results output in S6.2, a ranking table of the net effects of policy tools is constructed. The comprehensive influence index of each policy tool is calculated using the information entropy weighting method to identify core intervention factors and assist in policy optimization direction decision-making. S6.4: Based on the net causal effect described in S6.2 and the timestamp information described in S5, generate a trend diagram of the main effect path intensity change. Use a combination of dynamic heat map and flow map to visualize the evolution of the impact of each policy tool on key outcome indicators. S6.5: Based on the net causal effect distribution and preset threshold judgment rules output by S6.2, perform the policy intervention effectiveness judgment operation, use the two-sided hypothesis testing method to evaluate whether the effect value is significantly higher than the random level, and output the binary judgment result and confidence interval of whether the policy intervention is effective; Based on the net causal effect distribution data output by S6.2 and the preset threshold judgment rules, the effect value significance test method (parameters: two-sided hypothesis test, significance level α) is used to make a statistical judgment on whether the policy intervention effect exceeds the range of random fluctuations. Furthermore, by constructing the null hypothesis H0: the mean effect of policy intervention equals the random level, the test statistic is calculated using the corresponding formula:

[0020] in, The sample mean. The average value of random water. The standard deviation of the sample is 1. The number of samples; Furthermore, the corresponding p-value is calculated using the cumulative distribution function of the two-tailed t-distribution, and the significance of the effect value is judged in combination with the preset significance level α, and the binary determination result of the effect effectiveness is obtained. Furthermore, based on the sample mean and standard deviation, the (1) effect of the policy intervention is calculated. α) Confidence interval, using the formula:

[0021] in, For t, the distribution is in degrees of freedom n The critical value below 1. Furthermore, the judgment results and confidence interval information are summarized to form a data structure containing significance judgment, confidence interval range, and judgment basis, which serves as the evaluation output for decision-making reference; By using the statistical values ​​and confidence intervals calculated through two-sided hypothesis testing, the net causal effect distribution results of the previous step are transformed into quantifiable effectiveness indicators, thereby achieving the expected technical effects of scientific evaluation and threshold control of policy intervention effects. For example, in an evaluation of the effectiveness of a certain sports-related intangible cultural heritage protection policy, the net causal effect distribution sample size is 50, the sample mean is 0.82, the standard deviation is 0.15, and the presupposition α = 0.05. Calculate the test statistic:

[0022] The t-value was approximately 38.6. Further calculation of the p-value showed a result significantly less than 0.05, indicating that the policy intervention effect was significantly higher than the random level. Based on the t-distribution critical value of 2.009, the confidence interval was calculated as follows:

[0023] The obtained interval is approximately [0.77, 0.87]. The test showed that the policy effect was statistically significant and stable within the confidence interval, leading to the conclusion that the intervention was effective.

[0024] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0025] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for evaluating the effectiveness of policy interventions in the protection of intangible cultural heritage in sports, characterized in that, Includes the following steps: S1: Collect multi-source heterogeneous data including government documents, media reports, social media, and interview records of intangible cultural heritage inheritors, extract structured elements, and perform time-series alignment and standardized coding to form a policy-behavior-response related dataset; S2: Based on the policy-behavior-response association dataset, construct a dynamic policy knowledge graph, establish a multi-level causal network with policy intervention as the root node, and mark the evolution trajectory of relationship weights through timestamps; S3: Based on the dynamic policy knowledge graph, a time-series comparative learning encoder is constructed. A Siamese network structure is used to construct positive sample pairs for the implementation cases of the same policy in different regions or at different times. At the same time, a control region where the policy has not been implemented is introduced as a negative sample. The distance distribution of the feature space is optimized through metric learning to generate policy variable embedding vectors. S4: Based on the policy variable embedding vector, perform confounding variable clarification operation, calculate the correlation coefficient matrix between the policy variable embedding vector and external factors, and output the corrected set of policy-driven variables; S5: Deploy a counterfactual path reasoning engine in the dynamic policy knowledge graph to perform causal intervention operations. Use the Monte Carlo tree search algorithm to traverse all potential causal paths, combine the corrected set of policy-driven variables to calculate the effect contribution value of each path, and generate a path stability scoring function to evaluate cross-group consistency.

2. The method for evaluating the effectiveness of policy interventions in the protection of intangible cultural heritage in sports, as described in claim 1, is characterized in that... Following step S5, the following is also included: S6: Based on the effect contribution value and the path stability score, construct a causal effect attribution matrix, quantify the net causal effect of each policy tool, generate a visual evaluation report containing the trend of changes in the intensity of the main effect path, and output the threshold judgment basis for the effectiveness of policy intervention.

3. The method for evaluating the effectiveness of policy interventions in the protection of intangible cultural heritage in sports, as described in claim 1, is characterized in that... Step S1 specifically includes: Based on the government's policy texts on the protection of intangible cultural heritage in the sports category, the named entity recognition algorithm in natural language processing is executed to identify and extract structured policy elements and obtain a standardized set of policy variables. We perform sentiment analysis and topic modeling algorithms on text information related to sports intangible cultural heritage in media reports and social media platforms to identify public response tendencies and dissemination popularity in response to policy implementation and generate social response feature vectors. Speech recognition and semantic role labeling algorithms were applied to interview records of intangible cultural heritage inheritors to extract description information of inheritance behavior and construct a feature matrix of cultural inheritance behavior. A multimodal temporal alignment algorithm is executed on the standardized policy variable set, the social response feature vector, and the cultural heritage behavior feature matrix to align them based on timestamps and establish a unified time dimension reference system. Based on the time-aligned standardized set of policy variables, the social response feature vector, and the cultural heritage behavior feature matrix, a standardized coding mapping algorithm is executed to map heterogeneous data into a unified format of triple representations, thereby constructing a policy-behavior-response related dataset.

4. The method for evaluating the effectiveness of policy interventions in the protection of intangible cultural heritage in sports, as described in claim 3, is characterized in that... The acquisition of the standardized policy variable set includes a BERT-based named entity recognition algorithm, combined with conditional random field post-processing and semantic verification based on dependency parsing. A time normalization algorithm is then executed to map the time descriptions in the text into standardized timestamp data. The time normalization algorithm adopts the baseline time format YYYY-MM-DD mapping. Through a standardized encoding mapping algorithm, the elements that have undergone semantic dependency verification and timestamp normalization are converted into a unified knowledge representation format, and the standardized policy variable set is output.

5. The method for evaluating the effectiveness of policy interventions in the protection of intangible cultural heritage in sports, as described in claim 1, is characterized in that... Step S2 specifically includes: Entity identification and relation extraction are performed on the structured elements in the policy-behavior-response association dataset to obtain the entity set of policy subjects, policy tools, implementation nodes, inheritance projects, participating groups, dissemination scope and funding, as well as the semantic relationships between each pair, forming an initial set of triples; Based on the initial set of triples, the basic topology of the static policy knowledge graph is constructed. The collected multi-source heterogeneous data is subjected to time information extraction and standardization processing. Timestamps are used to time-label each policy intervention and its subsequent response events to construct event sequence data in the time dimension. Based on the event sequence data in the time dimension, a time-aware graph neural network is used to dynamically evolve and model the static knowledge graph. A timestamp attribute is added to each node and edge in the graph, and a weight update mechanism based on a time window is constructed. The dynamic policy knowledge graph is dynamically updated by performing relation weight updates. Based on the actual impact of policy intervention events on the outcome nodes at different time points, a sliding time window mechanism is used to periodically evaluate and adjust the weights of the edges between nodes, generating a causal path network with temporal evolution characteristics.

6. The method for evaluating the effectiveness of policy interventions in the protection of intangible cultural heritage in sports, as described in claim 1, is characterized in that... Step S3 specifically includes: Based on the multi-level causal network structure in the dynamic policy knowledge graph, a temporal contrastive learning encoder with a Siamese network architecture is constructed. The temporal contrastive learning encoder consists of two LSTM-Transformer hybrid models with shared weights, which are used to encode the temporal feature vectors of the target policy case and its contrast sample, respectively. Sample pairs were constructed from the standardized coded policy-behavior-response association dataset. Implementation records of the same policy in different regions or at different times were selected as positive sample pairs, and data from control regions where the policy was not implemented were introduced as negative sample pairs. Time-aware feature encoding is performed on each sequence data in the constructed sample pair. Local temporal dependencies are extracted using the LSTM module, and the global temporal evolution mode is modeled through the self-attention mechanism of Transformer to generate a high-dimensional temporal embedding vector. Based on the high-dimensional temporal embedding vector, a triplet loss function is used for metric learning optimization to generate a policy variable embedding vector. In the optimized feature space, a nearest neighbor search operation is performed on all policy cases to identify and filter out pseudo-related samples that have similar implementation paths to the target policy but are affected by non-policy factors, and these samples are used as the input candidate set.

7. The method for evaluating the effectiveness of policy interventions in the protection of intangible cultural heritage in sports, as described in claim 6, is characterized in that... Step S3 further includes using twin network parallel encoding, synchronously inputting target policy cases and regional or time period comparison samples, setting the number of LSTM hidden layer units to 128, the time step to the time window length of the dynamic knowledge graph, setting the Transformer multi-head attention to 8 heads, and the position encoding range to cover the entire time window length.

8. The method for evaluating the effectiveness of policy interventions in the protection of intangible cultural heritage in sports, as described in claim 1, is characterized in that... Step S4 specifically includes: Based on the policy variable embedding vectors of the output, a multidimensional feature space matrix is ​​constructed, where each row represents a high-dimensional feature vector of a policy implementation case and each column represents a standardized policy-related variable dimension, forming the initial input data structure for identifying confounding factors. The multidimensional feature space matrix is ​​subjected to covariate expansion processing to introduce non-policy external variables and construct an extended feature matrix containing potential confounding factors. Based on the extended feature matrix, the Pearson correlation coefficient matrix between the policy variable embedding vector and each external confounding factor is calculated to identify the set of confounding variables that are highly correlated with the policy variables; The set of confounding variables is processed by linear algebra using the orthogonal projection method. The original policy variable embedding vector space is projected onto a subspace orthogonal to the confounding factors to generate a biased policy-driven embedding vector. The policy-driven embedding vectors are normalized and aligned with dimensions to generate a standardized set of policy-driven variables.

9. The method for evaluating the effectiveness of policy interventions in the protection of intangible cultural heritage in sports, as described in claim 1, is characterized in that... Step S5 specifically includes: Based on the topological structure of dynamic policy knowledge graph, a causal reasoning graph model is constructed. Policy intervention nodes are used as causal intervention variables. Do-calculus operations are performed to simulate counterfactual scenarios and generate a causal graph structure containing the states before and after intervention. Based on a policy-driven set of variables, a causal effect estimation function is constructed. The expected response value on each causal path is modeled using a potential outcome framework. Causal effect estimation is performed on the set of variables corresponding to each path in the graph model. The expected difference in the counterfactual scenario is calculated to form a preliminary assessment result of the path effect contribution. Based on the preliminary evaluation results of the path effect contribution, the Monte Carlo tree search algorithm is introduced to perform a depth-first extended search of causal paths in the dynamic knowledge graph. The search depth threshold and path pruning rules are set to output a set of potential high-contribution causal paths. Based on the set of potential high-contribution causal paths, a path stability scoring function is constructed. A cross-group consistency test method is used to perform variance analysis on the effect contribution of each path in multiple subgroups, calculate its inter-group consistency score, and screen out the main effect paths that are stable in multiple groups. Based on the main effect pathway and its stability score, the path effect contribution value and stability score are integrated to construct a comprehensive evaluation index for causal pathways and output a causal pathway evaluation table.

10. The method for evaluating the effectiveness of policy interventions in the protection of intangible cultural heritage in sports, as described in claim 9, is characterized in that... The causal path evaluation table includes path number, causal effect strength, path stability score, and time evolution characteristics.

Images