Intelligent decision support system based on big data and artificial intelligence

Abstract

The application discloses an intelligent decision support system based on big data and artificial intelligence, and relates to the technical field of decision support, comprising: a decision data acquisition module, which is used for acquiring decision data, and performing feature extraction and anchoring effect feature marking; a propagation path identification module, which constructs a decision participant opinion influence bipartite graph based on a graph neural network, identifies a core anchor source and a strong cascading propagation path thereof, and outputs an anchoring effect strength value; a weight adjustment module, which is used for generating a weight adjustment factor based on the anchoring effect strength value, generating a weight through iterative adjustment and correlation correction, and generating a decision scheme according to weighted statistics; and a decision correction module, which is used for decomposing a historical decision implementation effect signal to identify a deviation component, generating an counterfactual decision scheme based on causal inference, and correcting a deviation of an actual decision scheme; and effectively solves the information cascade and collective irrationality problems caused by over-amplification of a few opinions or authority influence in traditional group decision-making.

Description

Technical Field

[0001] This invention relates to the field of decision support technology, specifically to an intelligent decision support system based on big data and artificial intelligence. Background Technology

[0002] Currently, in the field of group decision support, related technologies mainly revolve around the collection and preliminary processing of opinions from decision-makers, basic operational data of the scenario, and feedback data on the effects of historical decisions. For example, participants' opinions are obtained through methods such as transcribing meeting recordings, online questionnaires, and on-site discussions. Basic data cleaning methods are used to handle missing values ​​and normalize numerical data, or conventional semantic analysis tools are used to extract shallow features from the text. At the same time, structured data from urban operations and data on the implementation effects of historical decisions are collected, and opinions are integrated based on traditional weighted voting, arithmetic mean, and other methods to generate preliminary decision-making schemes, providing basic data support for group decision-making.

[0003] However, existing technologies have significant shortcomings. The system cannot effectively identify and quantify the "anchoring effect," the implicit influence of early speakers or authorities on subsequent opinions. This leads to the potential overemphasis on minority opinions and the systematic suppression of diverse viewpoints during the decision-making process, creating an "information cascade" phenomenon. Secondly, traditional weight adjustment methods lack dynamic adaptability, failing to adjust weights in real time based on opinion propagation paths and the strength of the anchoring effect, making it difficult to break the negative impact of "information cocoons" on decision quality. Simultaneously, existing systems are weak in correcting decision biases. Analysis of historical decision effects is typically limited to simple comparisons or regression models, failing to decompose the bias components caused by the anchoring effect and unable to generate correction schemes through counterfactual reasoning. For example, in scenarios such as emergency resource allocation during urban flooding, traditional methods struggle to distinguish whether decision bias stems from objective limitations or group cognitive biases, resulting in a lack of targeted optimization in subsequent decisions. These technological deficiencies make group decision-making prone to falling into the trap of "collective irrationality," hindering true scientific and intelligent decision-making. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent decision support system based on big data and artificial intelligence. By dynamically identifying and quantifying the opinion anchoring effect in the decision-making process, adjusting the weights of participants in real time, and generating counterfactual decision-making schemes by combining causal inference, it effectively solves the problems of information cascading and collective irrationality caused by the excessive amplification of minority opinions or the influence of authority in traditional group decision-making, thereby improving the scientificity and intelligence of decision-making.

[0006] (II) Technical Solution

[0007] To achieve the above objectives, the present invention provides the following technical solution: an intelligent decision support system based on big data and artificial intelligence, comprising:

[0008] The decision data collection module is used to collect opinion data from decision-making participants, basic data on urban operation, and historical feedback data on decision-making effects, and to perform feature extraction and anchoring effect feature labeling.

[0009] The propagation path identification module constructs a bipartite graph of the influence of decision participants' opinions based on a graph neural network, identifies the core anchor point source and its strong cascade propagation path, and outputs the anchoring effect strength value.

[0010] The weight adjustment module is used to generate weight adjustment factors based on the anchoring effect strength value, generate weights through iterative adjustment and correlation correction, and generate decision schemes based on weighted statistics.

[0011] The decision correction module is used to decompose the historical decision implementation effect signals to identify deviation components, generate counterfactual decision schemes based on causal inference, and correct deviations in the actual decision schemes.

[0012] Furthermore, the operations performed by the decision data acquisition module include:

[0013] When collecting opinion data from decision-making participants, the opinion text, speaking time sequence, and identity attributes are recorded. Identity attributes include authority level, professional matching degree, and historical contribution to decision-making. The BERT model is used to extract semantic feature vectors from the opinion text, and a rule engine is used to filter out repeated statements and extremely emotional statements, and to complete incomplete expressions.

[0014] The time series data were standardized with a unified timestamp format. Missing values ​​were filled by superimposing the average of adjacent time periods with the historical trend of the same period. Numerical data were normalized to the [0, 1] interval. A multi-source data association table was constructed with decision topic + time as the association key. The anchoring effect characteristics were specially marked, including the speaking time sequence characteristic mark, the opinion convergence characteristic mark, and the authority association characteristic mark.

[0015] Furthermore, the speech sequence feature labeling assigns a time sequence weight to the opinion with speech sequence number T, with the time sequence weight being 1 / T; the opinion convergence feature labeling calculates the cosine similarity between subsequent opinions and earlier opinions, and labels the convergence level according to the convergence similarity threshold; the authority association feature labeling maps the authority level to the authority influence coefficient and binds it to the participant's identity.

[0016] Furthermore, the operations performed by the propagation path identification module include:

[0017] Construct a bipartite graph of participant opinions: Participant node attributes include ID, authority influence coefficient, historical decision contribution, and professional matching degree; opinion node attributes include semantic vector, convergence feature value, and temporal weight; if the semantic similarity between participant B's opinion and participant A's opinion is greater than or equal to the semantic similarity threshold, then construct a directed edge from A to B, and the weight of the edge is the product of A's authority influence coefficient and B's convergence feature value.

[0018] Furthermore, the PageRank algorithm is used to calculate the importance score of nodes, and the top three participant nodes that meet the importance score and have been converged on ≥M times are selected as core anchor sources. The propagation path of the core anchor sources is extracted, the convergent node density is calculated, and the path with density ≥D is marked as a strong cascade path. A random forest model is constructed, and the authoritative influence coefficient of the core anchor source, the proportion of strong cascade paths, the mean of subsequent participants' opinion convergence, and the speaking time sequence weight are input, and the anchoring effect strength value is output.

[0019] Furthermore, the operations performed by the weight adjustment module include:

[0020] If the convergence characteristic value of participants' opinions is ≥0.8 and the corresponding anchoring effect strength value is ≥0.6, then the weight adjustment factor = 0.5; if the convergence characteristic value of participants' opinions is <0.5 and urban operation basic data is cited, then the weight adjustment factor = 1.8; iterative prediction weight: based on the weight of the previous round and the proportion of strong anchoring convergence opinions, predict the expected weight of the current round.

[0021] Furthermore, the Pearson correlation coefficient is used to calculate the correlation between participants' opinions and urban operation data. When the correlation is ≥0.7, the weights are positively adjusted by +0.1 to +0.2; when the correlation is <0.3, the weights are negatively adjusted by -0.1 to -0.15. When the weight change is <5% for 3 consecutive rounds or the number of iterations is >10 rounds, the iteration is stopped and the final weights are output. For discrete opinions, a weighted voting method is used to generate decision schemes, and for continuous opinions, a weighted mean method is used to generate decision schemes.

[0022] Furthermore, the operations performed by the decision correction module include:

[0023] The historical decision effect signal is decomposed into IMF components using the empirical mode decomposition algorithm; the Pearson correlation coefficient between each IMF component and the historical anchoring effect intensity sequence is calculated, and IMFs with correlation coefficients ≥ the correlation threshold are identified as bias components; a causal graph is constructed based on Do-calculus: intervention variable X = anchoring effect intensity, decision parameter Y, effect Z, and confounding variable C = urban operation data;

[0024] By blocking the causal path with do(X=0), a regression model of Y and C is established based on the reasonable contribution components, generating counterfactual solution parameters. When the deviation between the actual solution and the expected effect of the counterfactual solution is >0.5h, the counterfactual solution parameters are adopted. When the deviation is ≤0.5h, the actual solution parameters are fine-tuned, with the fine-tuning range being 50% of the difference between the counterfactual solution parameters and the actual solution parameters.

[0025] An electronic device includes a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement an intelligent decision support system based on big data and artificial intelligence.

[0026] A computer-readable storage medium storing a computer program that, when executed, implements an intelligent decision support system based on big data and artificial intelligence.

[0027] (III) Beneficial Effects

[0028] This invention provides an intelligent decision support system based on big data and artificial intelligence, which has the following beneficial effects:

[0029] (1) By extracting semantic features through the BERT model, filtering invalid information by combining the rule engine, and using methods such as time-series weighting and convergence feature labeling, the anchoring effect features in the opinions of decision participants are accurately identified and quantified, realizing the effective integration and cleaning of multi-source data, providing a high-quality data foundation for subsequent propagation path identification and weight adjustment, significantly improving the scientificity and accuracy of the decision support system, and avoiding decision errors caused by data bias in traditional methods.

[0030] (2) By constructing a bipartite graph of opinion influence through a graph neural network, and combining the PageRank algorithm to identify the core anchor source and its strong cascading propagation path, and using the random forest model to quantify the anchoring effect strength value, it is possible to accurately track the dynamic path of opinion propagation, reveal the implicit influence of authoritative opinions on group decision-making, effectively solve the problem that traditional methods are difficult to capture information cascading, provide a scientific basis for the weight adjustment module, significantly improve the objectivity and accuracy of opinion integration in the decision-making process, and avoid the excessive dominance of minority opinions.

[0031] (3) By dynamically generating weight adjustment factors, and combining the anchoring effect strength value and opinion convergence characteristics, the weights of participants are iteratively optimized and the correlation is corrected, which effectively suppresses the excessive influence of authoritative opinions, while increasing the weight of data-supported diverse opinions, avoiding the information cocoon problem in traditional decision-making. By quantifying the matching degree between opinions and urban operation data through Pearson correlation coefficient, the weights are further corrected to ensure that the decision-making scheme is more scientific and balanced, and significantly improves the objectivity and adaptability of group decision-making.

[0032] (4) By separating the bias component in the historical decision effect signal through the empirical mode decomposition algorithm and generating counterfactual schemes by combining causal inference (Do-calculus), the negative impact of the anchoring effect on the decision results can be effectively identified and quantified. The actual decision scheme can be dynamically corrected. When the bias value is significant, the counterfactual scheme can be directly adopted. When the bias is slight, the parameters can be fine-tuned, thereby significantly improving the objectivity and scientific nature of the decision, avoiding systematic decision-making errors caused by group cognitive bias or authority dominance, and ensuring that the scheme is more in line with actual needs and urban operation data. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the intelligent decision support system based on big data and artificial intelligence according to the present invention;

[0034] Figure 2 This is a schematic diagram of the propagation path identification process of the intelligent decision support system of the present invention. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] Please see Figures 1 to 2 This invention provides an intelligent decision support system based on big data and artificial intelligence, comprising: a decision data acquisition module, a propagation path identification module, a weight adjustment module, and a decision correction module, wherein:

[0037] The decision data collection module is used to collect opinion data from decision-making participants, basic data on urban operation, and historical feedback data on decision-making effects, and to perform feature extraction and anchoring effect feature labeling.

[0038] Specifically, for specific decision-making topics, such as the allocation of resources for emergency response to urban flooding, the adjustment of parking fees in the core area, and the prioritization of renovation of old residential areas, three types of decision support data are collected, including data on the opinions of decision-makers, basic data on urban operation, and feedback data on the effects of historical decisions.

[0039] The data on the opinions of decision-making participants includes: the text of the opinions, obtained through transcription of meeting recordings, online questionnaires, and on-site discussion records, preserving the original expressions; the speaking sequence, recording the speaking time of each participant and marking the speaking order number; identity attributes, including authority level: provincial expert = 0.9, municipal expert = 0.7, department head = 0.6, community representative = 0.4, general public = 0.3; professional matching degree: directly related to the decision-making topic = 1, indirectly related to the decision-making topic = 0.6, unrelated to the decision-making topic = 0.3; historical contribution to decision-making: based on the implementation effect of similar decision-making schemes participated in by the participant in the past 3-5 years, with a maximum score of 1.

[0040] The city's basic operational data is collected in a differentiated manner according to the decision-making scenario. For example, in the case of urban flooding emergency, historical coordinates of flooding points and records of water depth are collected, as well as real-time rainfall intensity, emergency material reserves, traffic capacity, and distribution of public service facilities. Emergency material reserves include the number of water pumps and the tonnage of sandbags. Traffic capacity includes real-time road congestion index. Distribution of public service facilities includes the coordinates of hospitals, schools, and nursing homes. In the case of parking fee adjustment, data such as the number of parking spaces in the core area, the average daily parking turnover rate, and the passenger flow of surrounding public transportation are collected.

[0041] Historical decision-making effect feedback data is collected from complete data of similar decisions over the past 5 years, including: the original decision plan, such as the proportion of material allocation and the adjustment range of fees; the quantitative effects after implementation, such as the duration of flooding and changes in the parking supply and demand gap; and subjective feedback data, including the results of public satisfaction surveys and interview records on the impact on business operations. Among them, satisfaction is assigned values ​​of 1, 0.8, 0.5, 0.2, and 0 respectively, corresponding to very satisfied, satisfied, neutral, dissatisfied, and very dissatisfied.

[0042] The BERT model is used to extract semantic features from opinion texts, outputting multi-dimensional semantic vectors. Meaningless expressions are removed through a rule engine, with rules including: filtering duplicate statements and continuous expressions with semantic similarity ≥ 0.95; removing emotional statements and expressions containing extreme words such as "absolutely unreasonable" or "completely useless" without substantive suggestions; and completing incomplete expressions, such as completing "prioritize allocation to the east of the city" to "prioritize allocation of emergency water pumps in the east of the city", with the completion based on the context.

[0043] For time-series data such as rainfall intensity and speaking time, a unified timestamp format is used; for missing values, the method of using the average of adjacent time periods plus historical trend correction is used to fill in the missing values. For example, if the rainfall intensity of a certain time period is missing, the average rainfall intensity of the hour before and after that time period is first calculated, and then corrected according to the fluctuation trend of the rainfall intensity of the same period in the past 3 years, such as an average increase of 10% for the same season and the same date, to avoid the deviation caused by single interpolation; for numerical data such as material reserves and satisfaction, normalization processing is performed and mapped to the interval [0, 1].

[0044] Using decision topic and time as the association key, a multi-source data association table is constructed. For example, the text of expert A's opinion, real-time rainfall intensity, and the 4-hour flood receding time in the same period of 2021 in the flood emergency scenario on July 10, 2023 are associated with the same decision event ID, providing a unified data entry point for subsequent analysis.

[0045] The anchoring effect-related features are specifically labeled, including the speaking time sequence feature label, the opinion convergence feature label, and the authority association feature label. The speaking time sequence feature label assigns a time sequence weight to the opinion of each participant, with the following rule: the speaking order number is T, T≥1, the time sequence weight = 1 / T, such as the first speaker T=1, the time sequence weight = 1, the tenth speaker T=10, the time sequence weight = 0.1, and the labeling results are stored in the time sequence weight field.

[0046] The opinion convergence feature labeling uses the cosine similarity algorithm to calculate the semantic similarity between the opinions of subsequent participants and the opinions of early potential anchors. For example, the opinions of the first 3 speakers are labeled according to the convergence similarity threshold: ≥0.8 is high convergence, labeled as 1; 0.5-0.8 is medium convergence, labeled as 0.5; <0.5 is low convergence, labeled as 0. The results are stored in the convergence feature value field.

[0047] The authoritative association feature is marked, and the authoritative influence coefficient is assigned based on the authority level. It is directly mapped to a numerical value, such as provincial expert = 0.9 and general public = 0.3. It is also bound to the participant's identity attribute and stored in the authoritative influence coefficient field.

[0048] By extracting semantic features using the BERT model, filtering invalid information using a rule engine, and employing methods such as time-series weighting and convergence feature labeling, the anchoring effect characteristics in the opinions of decision-makers are accurately identified and quantified. This achieves effective integration and cleaning of multi-source data, providing a high-quality data foundation for subsequent propagation path identification and weight adjustment. It significantly improves the scientificity and accuracy of the decision support system and avoids decision-making errors caused by data bias in traditional methods.

[0049] The propagation path identification module constructs a bipartite graph of the influence of decision participants' opinions based on a graph neural network, identifies the core anchor point source and its strong cascade propagation path, and outputs the anchoring effect strength value.

[0050] Specifically, a graph neural network (GNN) is used to construct a bipartite graph of the influence of decision-making participants' opinions. The bipartite graph contains two types of nodes: participant nodes and opinion nodes. The attributes of participant nodes include participant ID, authority influence coefficient, historical decision contribution, and professional matching degree; the attributes of opinion nodes include opinion semantic vector, convergence feature value, and time series weight. Each opinion node is uniquely associated with the corresponding participant node.

[0051] If the semantic similarity between participant B's opinion and participant A's opinion is greater than or equal to the semantic similarity threshold, such as ≥0.7, meaning that B's opinion is similar to or references A, then a directed edge is constructed from participant node A to participant node B. The weight of the edge is equal to (A's authority influence coefficient × B's convergence characteristic value). For example, if A is a provincial water conservancy expert with an authority influence coefficient of 0.9, and B's opinion has a convergence characteristic value of 1.0, which is highly similar, then the edge weight is 0.9 × 1.0 = 0.9.

[0052] The PageRank algorithm is used to calculate the importance score of nodes, and the participant nodes that meet the following two conditions are selected as core anchor sources: the node importance score ranks in the top three; the number of times it is converged by subsequent participants is ≥M, such as 5 times, that is, the number of edges pointing to the node is ≥5. For example, a provincial water conservancy expert ranks first in importance score and is converged by 8 subsequent participants, so it is determined to be a core anchor source.

[0053] For the core anchor source, the propagation path is extracted by backtracking its outgoing edges, i.e., other nodes affected by the anchor source; the density of converging nodes on each path is calculated as (number of nodes with a converging feature value ≥ 0.8 on the path / total number of nodes on the path), and paths with a converging node density ≥ D are marked as strong cascaded paths, such as ≥ 0.6;

[0054] A random forest model is constructed, with input features including: the authority influence coefficient of the core anchor source, the proportion of strong cascade paths, the mean of the consensus of subsequent participants, and the speaking time weight of the anchor source. The proportion of strong cascade paths = the number of strong cascade paths / the total number of propagation paths. The mean of the consensus of subsequent participants is the average of the convergence feature values ​​of all participants influenced by the anchor source. The output index is the anchoring effect strength value, which is obtained by averaging the prediction values ​​of all decision trees using the arithmetic mean method. The value range is [0, 1], where 0 represents no anchoring effect and 1 represents a complete anchoring effect.

[0055] The decision bias rate in historical decision data is used as a label to train the model. The decision bias rate = (actual implementation effect - expected implementation effect) / expected implementation effect. For example, in the scenario of waterlogging, the expected duration of water receding is 4 hours, and the actual duration of water receding is 5 hours. Then the bias rate = (5-4) / 4 = 0.25. The anchoring effect strength value <0.3 is weak anchoring, 0.3-0.6 is medium anchoring, and >0.6 is strong anchoring.

[0056] By constructing a bipartite graph of opinion influence using a graph neural network, combining the PageRank algorithm to identify core anchor sources and their strong cascading propagation paths, and using a random forest model to quantify the anchoring effect strength, we can accurately track the dynamic path of opinion propagation, reveal the implicit influence of authoritative opinions on group decision-making, effectively solve the problem that traditional methods are difficult to capture information cascading, provide a scientific basis for the weight adjustment module, significantly improve the objectivity and accuracy of opinion integration in the decision-making process, and avoid the excessive dominance of minority opinions.

[0057] The weight adjustment module is used to generate weight adjustment factors based on the anchoring effect strength value, generate weights through iterative adjustment and correlation correction, and generate decision schemes based on weighted statistics.

[0058] Specifically, the basic weight is calculated as follows: Basic weight = (Participant's historical decision contribution × Professional matching degree) / (1 + Authority influence coefficient × 0.3). For example, Expert B: Historical decision contribution 0.8, Professional matching degree 1.0, Authority influence coefficient 0.9, Basic weight = (0.8 × 1.0) / (1 + 0.9 × 0.3) ≈ 0.63; Ordinary resident E: Historical decision contribution 0.2, Professional matching degree 0.3, Authority influence coefficient 0.3, Basic weight = (0.2 × 0.3) / (1 + 0.3 × 0.3) ≈ 0.06.

[0059] If a participant's opinion is strongly anchored and convergent, meaning it simultaneously satisfies a convergence characteristic value ≥ 0.8 and a corresponding core anchor source anchoring effect strength value ≥ 0.6, then the weight adjustment factor = 0.5, reducing the weight and suppressing the excessive propagation of convergent opinions. For example, if participant F's opinion has a convergence characteristic value of 0.9 with strong anchor source A, the adjustment factor = 0.5.

[0060] If a participant's opinion is novel, meaning it simultaneously meets the criteria of a convergence characteristic value < 0.5 and cites basic urban operation data (e.g., historical flooding depth at the Chengdong subway station reached 1.2 meters), then the weight adjustment factor = 1.8, increasing the weight to encourage diverse and evidence-based opinions. For example, if resident E proposes that water pumps should be prioritized at the Chengdong subway station, with a convergence characteristic value of 0.4 and citing historical flooding depth data, the adjustment factor = 1.8. Opinions that do not meet the above two conditions will have their weights unchanged.

[0061] Based on the weights and opinion dissemination trend indicators from the previous round, such as the proportion of strongly anchored consensus opinions, the expected weights for the current round are predicted as follows: ,in, The expected weight for the current round. For the previous round of weighting, To strongly anchor the proportion of opinions with a shared trend;

[0062] The correlation between opinions and urban operation data is calculated using the Pearson correlation coefficient to measure the degree of matching between opinions and real-time urban operation data. For example, the correlation between prioritizing the allocation of materials in the eastern part of the city and the real-time water depth in the eastern part of the city is calculated, and weight adjustments are made based on the correlation.

[0063] If the correlation is ≥ 0.7: the expected weights are positively adjusted by +0.1 to +0.2, with a larger adjustment range for higher correlation. If the correlation is < 0.3: the expected weights are negatively adjusted by -0.1 to -0.15, with a larger adjustment range for lower correlation. If 0.3 ≤ correlation < 0.7: no adjustment is made, and the expected weights are maintained. Linear adjustment is used when making adjustments.

[0064] When the weight change is less than 5% for three consecutive rounds, the iteration stops and the final weight is output. If the condition is not met after more than 10 iterations, the iteration is forcibly terminated and the weight of the 10th round is taken as the final result to avoid efficiency loss caused by excessive iteration. For example, if resident E's initial weight is 0.05, adjustment factor = 1.8, previous round weight = 0.05, and the proportion of strongly anchored consensus opinions in this round = 0.6, then the expected weight = 0.05 × (1 + 0.2 × (0.5 − 0.6)) = 0.049, the correlation between their opinion and city operation data = 0.8, the positive correction magnitude = +0.15, and the final weight = 0.049 × 1.8 + 0.15 ≈ 0.24.

[0065] For the core decision-making dimensions of the decision-making topics, such as the area and quantity of material allocation in flood emergency scenarios, and the time period division and hourly rate in parking fee adjustment scenarios, the opinions of all participants were weighted and statistically analyzed.

[0066] For discrete opinions, such as prioritizing the allocation of materials in the east / west of the city: a weighted voting method is used to calculate the total weight of each opinion option. The option with the highest total weight is taken as the decision plan for that dimension. For example, if 6 out of 10 participants suggest the east of the city, the final total weight is 3.2; and 4 suggest the west of the city, the total weight is 1.8. In this case, prioritizing the allocation of materials in the east of the city is the candidate option.

[0067] For continuous opinions, such as the number of water pumps to be allocated being 50 / 60 and the charging standard being 10 yuan / hour / 12 yuan / hour: the weighted average method is used, and the calculation formula is: parameter value = Σ (participant's opinion parameter value × the final dynamic weight of the participant) / Σ (final dynamic weight of all participants); for example, if 3 experts' opinions on the number of water pumps are 50 units with a weight of 0.6, 60 units with a weight of 0.8, and 55 units with a weight of 0.7, then the parameter value = (50 × 0.6 + 60 × 0.8 + 55 × 0.7) / (0.6 + 0.8 + 0.7) = 55.48, which is rounded up to 55 units.

[0068] By dynamically generating weight adjustment factors and combining the anchoring effect strength value and opinion convergence characteristics, the weights of participants are iteratively optimized and the correlation is corrected. This effectively suppresses the excessive influence of authoritative opinions, while increasing the weight of data-supported diverse opinions. This avoids the information cocoon problem in traditional decision-making. By quantifying the matching degree between opinions and urban operation data through the Pearson correlation coefficient, the weights are further corrected to ensure that the decision-making scheme is more scientific and balanced, and significantly improves the objectivity and adaptability of group decision-making.

[0069] The decision correction module is used to decompose the historical decision implementation effect signals to identify deviation components, generate counterfactual decision schemes based on causal inference, and correct deviations in the actual decision schemes.

[0070] Specifically, core quantitative performance indicators after the implementation of decisions are selected as performance signal sequences. For example, in the case of urban flooding, the sequence of flood receding time is selected, such as the receding time of the last 5 urban flooding decisions: 4h, 6h, 3h, 7h, and 5h. In the case of parking fee adjustment, the sequence of parking turnover rate in the core area is selected.

[0071] The empirical mode decomposition algorithm is used to decompose the implementation effect signal of historical decisions into reasonable contribution components and deviation components. The specific decomposition steps are as follows: S1: Identify the local maxima and minima of the effect signal sequence, fit the upper envelope and lower envelope respectively by cubic spline interpolation, and calculate the mean of the envelope.

[0072] S2: Subtract the mean of the envelope from the original signal to obtain the first intrinsic mode function IMF1. If IMF1 satisfies that the mean of the upper and lower envelopes is 0 at any time and the number of extreme points is equal to or differs from the number of zero crossings by 1, then retain IMF1; otherwise, use IMF1 as a new original signal and repeat the above steps S1~S2 until an IMF1 that meets the conditions is obtained.

[0073] S3: Subtract IMF1 from the original signal to obtain the residual signal. Repeat steps S1~S2 with the residual signal as the new original signal, and extract IMF2, IMF3... in sequence until the residual signal is a monotonic function, and stop the decomposition.

[0074] Calculate the Pearson correlation coefficient between each IMF component and the historical anchoring effect strength sequence, where the historical anchoring effect strength sequence corresponds to the anchoring effect strength value of historical decisions, such as 0.6, 0.8, 0.4, 0.9, and 0.5. If the correlation coefficient is greater than or equal to the correlation threshold, such as 0.7, the IMF component is determined to be a biased component; otherwise, it is determined to be a reasonable contribution component. For example, after decomposing the flood receding time sequence, the correlation coefficient between IMF2 and the anchoring intensity sequence is 0.82, which is determined to be a biased component; the correlation coefficient between IMF1 and the sequence is 0.35, which is determined to be a reasonable contribution component.

[0075] The Do-calculus method based on causal inference sets causal variables as follows: intervention variable X, anchoring effect strength, with a value range of [0, 1]; decision scheme parameters Y, such as the proportion of material allocation in the western part of the city in the flood scenario, and the hourly parking fee standard in the core area in the parking scenario; implementation effect Z, such as the duration of flood receding and parking turnover rate; and confounding variables C, such as basic urban operation data: rainfall intensity and number of parking spaces in the core area. A causal graph structure is constructed as follows: C→X, C→Y, X→Y, Y→Z, C→Z.

[0076] By blocking the causal path X→Y through the Do operator, i.e., setting the intervention do (X=0), assuming the anchoring effect strength is 0 and there is no anchoring influence; based on the reasonable contribution components in historical data, linear regression is used, with the quantified value of C as the input and the predicted value of Y as the output, to establish a regression model of Y and C, with the formula as follows: ,in, In a counterfactual scenario Y The value of , α For the intercept term, β For regression coefficients, ε For the error term, the model training uses Y and C data corresponding to the reasonable contribution components of historical decisions;

[0077] A random forest model is used, with inputs being decision scheme parameters Y and city operation data C, and output being the predicted value of implementation effect Z. It predicts the expected effects of the actual scheme and the counterfactual scheme respectively, and calculates the deviation value. If the deviation value > 0.5h, it is considered a strong deviation, and the counterfactual scheme parameters are directly adopted to adjust the actual scheme, such as adjusting the material allocation ratio in the west of the city from 60% to 45% and in the east of the city from 40% to 55%. If the deviation value ≤ 0.5h, it is considered a weak deviation, and the actual scheme parameters are fine-tuned. The fine-tuning range is 50% of the difference between the counterfactual scheme parameters and the actual scheme parameters, i.e., fine-tuning range = (counterfactual scheme parameters - actual scheme parameters) × 0.5.

[0078] By separating the bias components in historical decision-making effect signals using the empirical mode decomposition algorithm and combining them with causal inference (Do-calculus) to generate counterfactual solutions, the negative impact of anchoring effects on decision-making outcomes can be effectively identified and quantified. This allows for dynamic correction of actual decision-making solutions. When the bias value is significant, the counterfactual solution is directly adopted, while parameters are fine-tuned when the bias is slight. This significantly improves the objectivity and scientific nature of decision-making, avoids systemic decision-making errors caused by group cognitive bias or authority dominance, and ensures that the solutions are more in line with actual needs and urban operation data.

[0079] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement an intelligent decision support system based on big data and artificial intelligence.

[0080] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed, implements an intelligent decision support system based on big data and artificial intelligence.

[0081] In the application, the various formulas mentioned are all calculated by removing dimensions and taking their numerical values. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The coefficients in the formulas are set by those skilled in the art according to the actual situation.

[0082] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, and combinations thereof. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0083] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0084] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. An intelligent decision support system based on big data and artificial intelligence, characterized in that: include: The decision data collection module is used to collect opinion data from decision-making participants, basic data on urban operation, and historical feedback data on decision-making effects, and to perform feature extraction and anchoring effect feature labeling. When collecting opinion data from decision-making participants, the opinion text, speaking time sequence, and identity attributes are recorded. Identity attributes include authority level, professional matching degree, and historical contribution to decision-making. The BERT model is used to extract semantic feature vectors from the opinion text, and a rule engine is used to filter out repeated statements and extremely emotional statements, and to complete incomplete expressions. For time series data, a unified timestamp format is used; for missing values, the method of superimposing the average of adjacent time periods with the historical trend of the same period is used to fill in the missing values; and for numerical data, the data is normalized to the [0, 1] interval. A multi-source data association table is constructed using decision-making topic and time as the association key; anchoring effect characteristics are specifically labeled, including speaking time sequence characteristics, opinion convergence characteristics, and authority association characteristics. The propagation path identification module constructs a bipartite graph of the influence of decision participants' opinions based on a graph neural network, identifies the core anchor point source and its strong cascade propagation path, and outputs the anchoring effect strength value. The PageRank algorithm is used to calculate the importance score of nodes, and the participant nodes that meet the requirements of the top three importance scores and have been converged ≥ M times are selected as the core anchor source. Extract the propagation path of the core anchor source, calculate the density of convergent nodes, and mark the path with density ≥ D as a strong cascade path; construct a random forest model, input the authority influence coefficient of the core anchor source, the proportion of strong cascade paths, the mean of the convergence of opinions of subsequent participants, and the speaking time sequence weight, and output the anchoring effect strength value. The weight adjustment module is used to generate weight adjustment factors based on the anchoring effect strength value, generate weights through iterative adjustment and correlation correction, and generate decision schemes based on weighted statistics. The decision correction module is used to decompose the historical decision implementation effect signals to identify deviation components, generate counterfactual decision schemes based on causal inference, and correct deviations in the actual decision schemes. The historical decision effect signal is decomposed into IMF components using the empirical mode decomposition algorithm; the Pearson correlation coefficient between each IMF component and the historical anchoring effect intensity sequence is calculated, and IMFs with correlation coefficients ≥ the correlation threshold are identified as bias components; a causal graph is constructed based on Do-calculus: intervention variable X = anchoring effect intensity, decision parameter Y, effect Z, and confounding variable C = urban operation data; By blocking the causal path with do(X=0), a regression model of Y and C is established based on the reasonable contribution components, generating counterfactual solution parameters. When the deviation between the actual solution and the expected effect of the counterfactual solution is >0.5h, the counterfactual solution parameters are adopted. When the deviation is ≤0.5h, the actual solution parameters are fine-tuned, with the fine-tuning range being 50% of the difference between the counterfactual solution parameters and the actual solution parameters.

2. The intelligent decision support system based on big data and artificial intelligence according to claim 1, characterized in that: Speaking sequence feature labeling: Assign a time sequence weight to the opinion with speaking sequence number T, where the time sequence weight is 1 / T; Opinion convergence feature labeling: Calculate the cosine similarity between subsequent opinions and earlier opinions, and label the convergence level according to the convergence similarity threshold; Authority-related feature labeling: Mapping authority level to authority influence coefficient and binding it to participant identity.

3. The intelligent decision support system based on big data and artificial intelligence according to claim 1, characterized in that: The operations performed by the propagation path identification module include: Construct a bipartite graph of participant opinions: Participant node attributes include ID, authority influence coefficient, historical decision contribution, and professional matching degree; opinion node attributes include semantic vector, convergence feature value, and temporal weight; if the semantic similarity between participant B's opinion and participant A's opinion is greater than or equal to the semantic similarity threshold, then construct a directed edge from A to B, and the weight of the edge is the product of A's authority influence coefficient and B's convergence feature value.

4. The intelligent decision support system based on big data and artificial intelligence according to claim 1, characterized in that: The operations performed by the weight adjustment module include: If the convergence characteristic value of participants' opinions is ≥0.8 and the corresponding anchoring effect strength value is ≥0.6, then the weight adjustment factor = 0.5; if the convergence characteristic value of participants' opinions is <0.5 and urban operation basic data is cited, then the weight adjustment factor = 1.8; iterative prediction weight: based on the weight of the previous round and the proportion of strong anchoring convergence opinions, predict the expected weight of the current round.

5. The intelligent decision support system based on big data and artificial intelligence according to claim 4, characterized in that: The correlation between participants' opinions and urban operation data is calculated using the Pearson correlation coefficient. When the correlation is ≥0.7, the weights are positively adjusted by +0.1 to +0.2; when the correlation is <0.3, the weights are negatively adjusted by -0.1 to -0.

15. If the weight change is <5% for 3 consecutive rounds or the number of iterations is >10 rounds, the iteration is stopped and the final weights are output. For discrete opinions, a weighted voting method is used to generate decision schemes; for continuous opinions, a weighted average method is used to generate decision schemes.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the intelligent decision support system based on big data and artificial intelligence as described in any one of claims 1-5.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed, performs the steps of the intelligent decision support system based on big data and artificial intelligence as described in any one of claims 1-5.

Images