Multi-modal data driven tourism demand causal inference method and system oriented to active policy making

Abstract

The invention discloses a multi-modal data-driven tourism demand causal inference method and system oriented to active policy making. The method comprises the following steps: multi-modal tourism data acquisition and characterization preprocessing; constructing a tourism demand causal map based on a PC algorithm and domain knowledge; policy causal effect estimation based on a double difference model; carrying out travel demand modeling and active policy simulation based on fusion of causal effects; the system comprises a data acquisition and preprocessing module, a causal atlas construction module, a causal effect estimation module, a demand modeling and policy simulation module, a data storage module and a visualization and interaction module. According to the method, the problems of insufficient causal relationship mining, policy evaluation lagging, insufficient multi-modal data utilization, difficulty in consideration of model interpretation and prediction performance and the like in an existing tourism demand analysis and policy making method are solved, the policy effect can be actively evaluated, scientific decision making is supported, and tourism demands are driven.

Description

Technical Field

[0001] The present invention relates to the field of data processing and artificial intelligence technology, and in particular to a multimodal data-driven tourism demand causal inference method and system for proactive policy formulation. Background Art

[0002] As an integral part of the economy, tourism is affected by many complex factors, including economic conditions, seasonal climate, marketing promotion, festivals, emergencies, government policies, etc. Correctly grasping the relationship between tourism demand and various related factors and quantitatively analyzing the effects of policy implementation can help tourism management departments formulate scientific and reasonable policies and measures, rationally allocate tourism resources, and improve management efficiency. Currently, there are still many problems with the methods of tourism demand forecasting and policy evaluation. Traditional research methods mostly use statistical regression or time series analysis, which can only discover correlations between variables but cannot represent true causal relationships. This can easily lead to "confounding bias" or "spurious correlations" and mislead us into making incorrect decisions. Most policy effectiveness evaluations are carried out after the policy is implemented, without the ability to proactively simulate in advance and lack forward-looking analysis. Therefore, it is difficult to judge the extent of the policy's effects during the policy formulation stage. Many existing studies use the analysis results of structured data statistics, but lack the ability to deeply explore the value of increasingly rich and diverse multimodal unstructured data such as social media text, images, policy document text, and search engine data. The content contained in these data, such as public sentiment, hot topics, and policy orientation, has not been fully developed and utilized. Deep learning models have high prediction accuracy, but they become "black boxes", making it impossible to obtain causal paths. Traditional causal inference models have great explanatory advantages due to their sufficient theoretical understanding, but they have no good solution for complex high-dimensional nonlinear problems.

[0003] Therefore, what is needed now is to be able to bring together multimodal data at the same time, explore deeper causal relationships between factors affecting tourism demand, quantify policy interventions, and support technical and systematic means of active policy simulation, thereby enhancing the scientific nature and foresight of tourism policy making. Summary of the Invention

[0004] Purpose of the invention: The purpose of the present invention is to provide a multimodal data-driven tourism demand causal inference method and system for proactive policy making.

[0005] Technical Solution: The multimodal data-driven tourism demand causal inference method for proactive policy formulation described in the present invention includes the following steps:

[0006] (1) Multimodal tourism data collection and characterization preprocessing;

[0007] (2) Construction of tourism demand causal graph based on PC algorithm and domain knowledge;

[0008] (3) Estimation of policy causal effects based on the difference-in-differences model;

[0009] (4) Tourism demand modeling and proactive policy simulation integrating causal effects.

[0010] Furthermore, the data collected in step (1) includes multi-source heterogeneous tourism data such as policy text data, social media data, search engine data, economic and environmental data, and tourism business data.

[0011] Furthermore, the step (2) includes:

[0012] (2.1) Using all possible variables that influence tourism demand as all nodes, a completely undirected graph with edges connecting all nodes is constructed as the initial skeleton of the algorithm;

[0013] (2.2) After performing conditional independence tests in a loop, the corresponding edges are deleted from the initial skeleton. The conditional independence tests include zero-order independence tests and higher-order conditional independence tests, and the significance level α is set.

[0014] (2.3) By identifying the V structure and determining the directional rules, directions are assigned to the edges in the graph, and the undirected graph is transformed into a directed graph. The graph generated by the algorithm is then corrected and confirmed based on the prior knowledge in the tourism field.

[0015] Furthermore, the zero-order independence test in step (2.2) includes: for each pair of variables (X, Y), checking X⊥Y, and if p-value>α, removing the edge (X, Y).

[0016] Furthermore, the high-order conditional independence test in step (2.2) includes: for each pair (X, Y) that still has an edge and its subset S of k adjacent points k , check X⊥Y|S k , if p-value>α, remove the edge (X, Y) and record the separation set S k >(X, Y), the test method can use partial correlation coefficient or conditional mutual information.

[0017] Furthermore, the double difference model in step (3) is expressed as:

[0018] Y it =β0+β1D i +β2P t +β3(D i ×P t )+ε it +X′ it γ+αi +δ t

[0019] Among them, Y it is the tourism demand index of unit i at time t, D i is the treatment group dummy variable, P t is a dummy variable for the policy implementation period, D i ×P t is the interaction term, ε it is a random disturbance term, X' is a covariate vector that changes with time and individuals, γ is the coefficient of the covariate, α i is the individual fixed effect, δ t is the time fixed effect, and β3 is the estimator of the net causal effect of the policy.

[0020] Furthermore, the construction of the tourism demand model in step (4) adopts a hybrid model that combines the long short-term memory network LSTM and the Transformer, including:

[0021] (4.1) LSTM layer: inputs the time series data of historical tourism demand and its related temporal features and obtains their long-term dependencies;

[0022] (4.2) The Transformer encoder layer takes the LSTM output, the social media feature vector, the policy feature vector, and the expected causal effect value of the specific policy calculated in step (3) as joint input;

[0023] (4.3) Transformer uses multi-head self-attention to learn complex correlations between different features in parallel;

[0024] (4.4) Output layer: Use the fully connected layer to output the predicted value of future tourism demand indicators.

[0025] Furthermore, the step (4.3) Transformer uses multi-head self-attention to include:

[0026] For each attention head i,

[0027] Attentioni(Q,K,V)=softmax(Q i K i T / dk)V i

[0028] Among them, Q i , K i , V i They represent the query matrix, key matrix, and value matrix obtained by linearly transforming the input, respectively. dk is the dimension of the key vector.

[0029] The attention heads are then combined using concatenation and linear mapping operations.

[0030] Furthermore, the active policy simulation includes:

[0031] (4.5) Define the hypothetical policy or policy combination to be evaluated;

[0032] (4.6) Use the utility estimation method in step (3) to calculate the value of the policy effects assumed above;

[0033] (4.7) Based on the input characteristics of the virtual policy to be tested and the estimated causal effect value corresponding to the virtual policy to be tested, use the trained LSTM-Transformer model to predict the tourism demand in the future under the intervention of these virtual policies;

[0034] (4.8) Analyze and compare the forecast results of different policy options to provide policymakers with data-based policy selection recommendations and reference for expected policy effects.

[0035] The multimodal data-driven tourism demand causal inference system for proactive policy formulation described in the present invention includes:

[0036] Data collection and preprocessing module; used for multimodal tourism data collection and characterization preprocessing;

[0037] Causal graph construction module; used to embed PC algorithm and domain knowledge base interface to realize the construction of tourism demand causal graph based on PC algorithm and domain knowledge;

[0038] Causal effect estimation module: used for DID-embedded DID, DID algorithm with covariates and fixed effects to complete policy causal effect estimation based on double difference model;

[0039] Demand modeling and policy simulation module: This module integrates the embedded LSTM-Transformer model and simulation engine to complete tourism demand modeling and proactive policy simulation that integrates causal effects during task allocation and scheduling.

[0040] Data storage module: used to save original data information, characteristic data information, causal graph, model parameters and simulation result information;

[0041] Visualization and interaction module; used to display causal maps, policy effects, forecast results, and allow users to customize simulated policy scenarios.

[0042] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0043] (1) Revealing deep causal relationships: Compared with correlation analysis alone, the causal graph obtained based on PC algorithms and domain knowledge can better identify the factors affecting tourism demand and their mechanism of action;

[0044] (2) Quantifying policy effects: Using causal inference models such as DID can find the net effect of policy intervention from a large number of interfering factors, and provide data support for policy evaluation as a method of measuring policy effects;

[0045] (3) Active policy simulation: Combining causal effects with deep learning prediction models can simulate proposed policies in advance and predict their impact, thereby achieving a shift from "ex post evaluation" to "ex ante prediction" and improving the initiative and foresight of policy making;

[0046] (4) Making full use of heterogeneous data from multiple sources, such as policy texts, social media, and search engines, can better capture market information and the attitudes and opinions of the masses, making the analysis more powerful;

[0047] (5) Improve the scientific nature of decision-making: Help tourism management departments create decision-making support tools that are data-driven and result-oriented, achieve optimal policy combinations, more efficient resource allocation, and more guaranteed policy effects. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 is a flow chart of the method of the present invention;

[0049] Figure 2 This is a schematic diagram of the multimodal data characterization processing flow;

[0050] Figure 3 An example of a simplified causal relationship diagram for factors affecting tourism demand in City X;

[0051] Figure 4 Schematic diagram of the double difference model (DID) principle;

[0052] Figure 5 Schematic diagram of the LSTM-Transformer demand forecasting model that integrates causal effects;

[0053] Figure 6 This is a functional module architecture diagram of the system of the present invention. DETAILED DESCRIPTION

[0054] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0055] like Figure 1As shown, the multimodal data-driven causal inference method for tourism demand for proactive policy formulation described in the present invention includes the following steps: multimodal tourism data collection and characterization preprocessing; construction of a tourism demand causal graph based on a PC algorithm and domain knowledge; estimation of policy causal effects based on a double difference model; and tourism demand modeling and proactive policy simulation that integrates causal effects. The following is an example of proactive spring tourism policy formulation in City X:

[0056] Step 1: Multimodal tourism data collection and characterization preprocessing. The specific process is as follows: Figure 2 As shown:

[0057] Data Source:

[0058] Social media data: Over the past five years, we have collected approximately 5 million public posts (including text and associated images / short videos) related to keywords such as "X City Tourism," "X City Spring Outing," and "X City Cultural Festival" from various online platforms. We also collected data on likes, comments, and reposts corresponding to these posts. We can use APIs or crawler technology to set keywords and time periods, and periodically (for example, once a day) capture this data.

[0059] Search engine data: Over the past five years, the corresponding indexes of keywords such as "X City Attractions," "X City Food," "X City Accommodation," and "X City Cultural Experience" on the daily search engine index data, as well as the search popularity of related terms for each of these keywords on Baidu, and the regional and age distribution of these search users.

[0060] Official statistics: Combined with data released by the X City Statistics Bureau and Tourism Bureau over the past five years on monthly tourist arrivals (domestic and foreign tourists), total tourism revenue, number of visitors to A-level scenic spots, average hotel occupancy rates and room rates, monthly tourist arrivals to major source areas (by province), average monthly temperature, precipitation, and AQI in X City, as well as data on the national and major source provinces' CPI and per capita disposable income for the same period.

[0061] Historical policy documents: Over the past five years, we collected over 50 tourism-related policy documents issued by City X and higher-level governments, recording their release dates, main contents, subsidy amounts involved, and preferential policies.

[0062] OTA platform data: Summary statistics on hotel bookings, air ticket bookings, and scenic spot ticket bookings in City X over the past five years, including relevant prices and user ratings.

[0063] Feature extraction:

[0064] Policy text: LDA extracts topics, conducts sentiment analysis (positive, neutral, negative) and counts the number of policy mentions and coverage; policy measures are extracted for each policy document; policy strength is reflected in the subsidy amount range and preferential margin given in the policy; policy targets are the people who benefit from the policy or the type of visitors; policy implementation time is the period of policy implementation specified in the document, which mostly refers to the time period for the specific implementation of the current people-benefiting policy in the context of tourism benefit policies.

[0065] Social media text: Use the BERT model for sentiment classification and the LDA algorithm to extract hot topics from comment data.

[0066] Social media images: In addition to ResNet and other CNN models being able to extract scene categories, they can also estimate crowd density.

[0067] All characteristics are aggregated by month to generate characteristic panel data.

[0068] Step 2: Construct a causal relationship graph of tourism demand in City X based on PC algorithm and domain knowledge; an example of a simplified causal relationship graph of factors affecting tourism demand in City X is as follows: Figure 3 As shown:

[0069] The monthly characteristic variables extracted in step 1 are used as the input nodes of the PC algorithm.

[0070] The significance level was set to α = 0.05.

[0071] Skeleton construction:

[0072] Zero-order independence test: For example, if it is found that "per capita income of source cities" and "average temperature of City X" are statistically independent, remove the edge between them.

[0073] Higher-order conditional independence test: For example, test whether "festival activities held" and "hotel bookings" are independent while controlling "seasonal factors" (using monthly dummy variables) and "promotion budget". If the p-value is greater than 0.05, it is determined that the two are conditionally independent and the edge can be deleted.

[0074] Edge Orientation:

[0075] Identify the V-structure: If "Spring Flower Show Promotion", "Flower Viewing Theme Discussion Volume" and "Increase in the Number of Sunny Days" appear, it is determined that "Spring Flower Show Promotion" is not directly related to "Increase in the Number of Sunny Days".

[0076] Domain knowledge correction:

[0077] Tourism expert review: Find the specific direction of "marketing and promotion investment", "tourist attention" and "tourist volume".

[0078] Step 3: Estimation of the effect of the "Spring Cultural Tourism Festival in City X" based on the DID model. The schematic diagram of the double difference model (DID) is as follows: Figure 4 As shown:

[0079] Research Subject: City X launched the “Spring Cultural Tourism Festival” for the first time in March this year.

[0080] data:

[0081] Treatment group: City X (numbered D-1).

[0082] Control group: Find a city C1 (D = 0) that is geographically close to City X before 2022, has a similar level of economic development, and has the same types of tourism resources as City X, but has never held a large-scale spring cultural festival.

[0083] Time: The following is an abbreviated version of the paragraph: Data from January 2019 to December 2023; the policy period (P=1) refers to data from March 2022 onwards.

[0084] Outcome variable Y it : Monthly number of domestic tourists.

[0085] Covariate X' it : Monthly average temperature, whether there are national holidays in that month, and the proportion of tourism-related positive emotion posts on social media in the corresponding city in that month.

[0086] Model: Y it =β0+β2P t +β3(D X市 ×P 政策后 )+ε it +X' it γ 1气温 +γ 2节假日 +δ t +γ 3社媒情感 +α1

[0087] Parallel trend test: Comparing the logarithmic change trends of the monthly number of tourists in City X and City C1 during the period from January 2019 to February 2022, it was found that they were basically parallel.

[0088] Results: The results of regression analysis showed that the estimated value of β3 was 0.078, p<0.01.

[0089] The Spring Cultural Tourism Festival led to an approximately 7.8% increase in the number of domestic tourists in City X each month. "Increase by 8 ± 2 percentage points" refers to the effect size or the average effect or range of effect sizes obtained by models such as X-Learner. "Increase by 8%" refers to the average effect. This article uses "increase by 8%" to represent the effect size.

[0090] Step 4: After introducing the causal effect, we take the tourism demand of City X as an example to establish a model and implement the active policy simulation of City X. The structural diagram of the LSTM-Transformer demand forecasting model integrating the causal effect is shown in the figure below: Figure 5 shown.

[0091] 1. Demand forecasting model construction:

[0092] LSTM layer: This layer converts data related to the monthly number of tourists in City X, the average temperature in each month, and the dummy variables of holidays into a time series.

[0093] Transformer encoder layer input:

[0094] Hidden state of the LSTM layer;

[0095] Average social media sentiment and the proportion of culture-related topics in each month in City X;

[0096] The monthly search index of City X is used as a representative of “City X tourism”.

[0097] Policy effect characteristics: Whether the dummy variable of the "Spring Cultural Tourism Festival" exists, the effect value corresponding to the "Spring Cultural Tourism Festival" (such as +0.08 indicates that the expected number of tourists this quarter will increase by 8%, etc.) are all considered as characteristics or included in the baseline forecast items for consideration.

[0098] Model training: Use data from 2018 to 2023 to build a model to predict the number of tourists in the next 1 to 3 months.

[0099] 2. Active policy simulation:

[0100] Scenario 1 (baseline): If the "Spring Cultural Tourism Festival" is not held in the spring of 2024, then based on the trend of other situations, the model can be used to predict the number of visitors X0.

[0101] Scenario 2 (maintaining the existing policy): Assuming that the 2024 "Spring Cultural Tourism Festival" will be held during next year's spring break, and setting the policy impact intensity factor to 8%, the number of tourists that the festival can bring into the model is X1.

[0102] Scenario 3 (enhanced policy): If, in addition to holding a cultural festival next spring, a certain type of activity, such as the "Hanfu Experience", is specifically experimented with or its effectiveness is evaluated by professionals, it may attract another 2% of traffic (the total number of people attracted is 10%). Then, by putting this new driving effect into the model, the actual number of tourists is X2.

[0103] Decision support: By comparing the X0, X1, and X2 indicators with the corresponding increased costs, the X City Tourism Bureau can better determine the 2024 spring tourism plan. If (X2-X1) / X1>increased cost-effectiveness, the measures in scenario three can be adopted.

[0104] like Figure 6 As shown, the multimodal data-driven tourism demand causal inference system for proactive policy formulation of the present invention includes:

[0105] Data collection and preprocessing module; used for multimodal tourism data collection and characterization preprocessing;

[0106] Causal graph construction module; used to embed PC algorithm and domain knowledge base interface to realize the construction of tourism demand causal graph based on PC algorithm and domain knowledge;

[0107] Causal effect estimation module: used for DID-embedded DID, DID algorithm with covariates and fixed effects to complete policy causal effect estimation based on double difference model;

[0108] Demand modeling and policy simulation module: This module integrates the embedded LSTM-Transformer model and simulation engine to complete tourism demand modeling and proactive policy simulation that integrates causal effects during task allocation and scheduling.

[0109] Data storage module: used to save original data information, characteristic data information, causal graph, model parameters and simulation result information;

[0110] Visualization and interaction module; used to display causal maps, policy effects, forecast results, and allow users to customize simulated policy scenarios.

Claims

1. A multimodal data-driven causal inference method for tourism demand for proactive policy making, characterized by: The steps include: (1) Multimodal tourism data collection and characterization preprocessing; (2) Construction of tourism demand causal graph based on PC algorithm and domain knowledge; (3) Estimation of policy causal effects based on the difference-in-differences model; (4) Tourism demand modeling and proactive policy simulation integrating causal effects.

2. The multimodal data-driven tourism demand causal inference method for proactive policy formulation according to claim 1 is characterized by: The data collected in step (1) includes tourism data from multiple sources including policy text data, social media data, search engine data, economic and environmental data, and tourism business data.

3. The multimodal data-driven tourism demand causal inference method for proactive policy formulation according to claim 1 is characterized by: The step (2) comprises: (2.1) Using all possible variables that influence tourism demand as all nodes, a completely undirected graph with edges connecting all nodes is constructed as the initial skeleton of the algorithm; (2.2) After performing conditional independence tests in a loop, the corresponding edges are deleted from the initial skeleton. The conditional independence tests include zero-order independence tests and higher-order conditional independence tests, and the significance level α is set. (2.3) By identifying the V structure and determining the directional rules, directions are assigned to the edges in the graph, and the undirected graph is transformed into a directed graph. The graph generated by the algorithm is then corrected and confirmed based on the prior knowledge in the tourism field.

4. The multimodal data-driven tourism demand causal inference method for proactive policy formulation according to claim 3 is characterized by: The zero-order independence test in step (2.2) includes: for each pair of variables (X, Y), checking X⊥Y, and if p-value>α, removing the edge (X, Y).

5. The multimodal data-driven tourism demand causal inference method for proactive policy formulation according to claim 3 is characterized by: The high-order conditional independence test in step (2.2) includes: for each pair (X, Y) that still has an edge and its subset S of k adjacent points k , check X⊥Y|S k , if p-value>α, remove the edge (X, Y) and record the separation set S k >(X, Y), the test method can use partial correlation coefficient or conditional mutual information.

6. The multimodal data-driven tourism demand causal inference method for proactive policy formulation according to claim 1 is characterized by: The double difference model in step (3) is expressed as: Y it =β0+β1D i +β2P t +β3(D i ×P t )+e it +X' it c+a i +d t Among them, Y it is the tourism demand index of unit i at time t, D i is the treatment group dummy variable, P t is a dummy variable for the policy implementation period, D i ×P t is the interaction term, ε it is a random disturbance term, X' is a covariate vector that changes with time and individuals, γ is the coefficient of the covariate, α i is the individual fixed effect, δ t is the time fixed effect, and β3 is the estimator of the net causal effect of the policy.

7. The multimodal data-driven tourism demand causal inference method for proactive policy formulation according to claim 1 is characterized by: The construction of the tourism demand model in step (4) adopts a hybrid model that combines the long short-term memory network LSTM and the Transformer, including: (4.1) LSTM layer: inputs the time series data of historical tourism demand and its related temporal features and obtains their long-term dependencies; (4.2) The Transformer encoder layer takes the LSTM output, the social media feature vector, the policy feature vector, and the expected causal effect value of the specific policy calculated in step (3) as joint input; (4.3) Transformer uses multi-head self-attention to learn complex correlations between different features in parallel; (4.4) Output layer: Use the fully connected layer to output the predicted value of future tourism demand indicators.

8. The multimodal data-driven tourism demand causal inference method for proactive policy formulation according to claim 7 is characterized by: The step (4.3) Transformer uses multi-head self-attention including: For each attention head i, Attentioni(Q,K,V)=softmax(Q i K i T / dk)V i Among them, Q i , K i , V i They represent the query matrix, key matrix, and value matrix obtained by linearly transforming the input, respectively. dk is the dimension of the key vector. The attention heads are then combined using concatenation and linear mapping operations.

9. The multimodal data-driven tourism demand causal inference method for proactive policy formulation according to claim 1 is characterized by: The active policy simulation includes: (4.5) Define the hypothetical policy or policy combination to be evaluated; (4.6) Use the utility estimation method in step (3) to calculate the value of the policy effects assumed above; (4.7) Based on the input characteristics of the virtual policy to be tested and the estimated causal effect value corresponding to the virtual policy to be tested, use the trained LSTM-Transformer model to predict the tourism demand in the future under the intervention of these virtual policies; (4.8) Analyze and compare the forecast results of different policy options to provide policymakers with data-based policy selection recommendations and reference for expected policy effects.

10. A multimodal data-driven tourism demand causal inference system for proactive policy making, characterized by: include: Data acquisition and preprocessing module; Used for multimodal tourism data collection and characterization preprocessing; Causal graph building blocks; Used to embed PC algorithm and domain knowledge base interface to realize the construction of tourism demand causal graph based on PC algorithm and domain knowledge; Causal effect estimation module; The DID algorithm used to embed DID into DID, including covariates and fixed effects, completes the estimation of policy causal effects based on the double difference model; Demand modeling and policy simulation module: This module integrates the embedded LSTM-Transformer model and simulation engine to complete tourism demand modeling and proactive policy simulation that integrates causal effects during task allocation and scheduling. Data storage module; Used to save original data information, characteristic data information, causal diagrams, model parameters and simulation result information; Visualization and interaction module; Used to display causal maps, policy effects, forecast results, and allow users to customize simulated policy scenarios.

Images