Big data driven market prediction system and system

Abstract

The invention discloses a market prediction system driven by big data, and relates to the technical field of big data processing and market prediction, and the system comprises the following components: a data collection and preprocessing module, a market ecological modeling module of a graph neural network, a model evolution module driven by an evolutionary algorithm, and a model evaluation and feedback module. By integrating data acquisition and preprocessing, market ecological modeling of a graph neural network and a model evolution module driven by an evolutionary algorithm, market data, including enterprise internal operation data, data published by an industry association, survey data of a market research institution and related transaction data of an internet platform, are collected from multiple source channels; the method guarantees the universality and accuracy of data, carries out the market ecological modeling through a graph neural network, can deeply mine the relation between market participants and a market dynamic change rule, provides a more accurate and reliable basis for market prediction, carries out the continuous optimization of a model through an evolutionary algorithm, can adapt to the market change, and improves the market prediction efficiency. And the prediction accuracy and stability are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of big data processing and market forecasting, and in particular to a market forecasting system and system driven by big data. Background Art

[0002] With the rapid development of information technology, big data has become an important basis for corporate decision-making and market forecasting. The sources of market data are becoming increasingly diversified, including but not limited to internal corporate operating data, data released by industry associations, survey data from market research institutions, and related transaction data from Internet platforms. These data contain rich market information and potential commercial value. How to efficiently collect, process and utilize these data to achieve accurate predictions of market dynamics has become an important topic in the current field of big data processing and market forecasting technology.

[0003] In traditional market forecasting systems, traditional methods often only focus on factors that directly affect market prices and supply and demand relationships, but ignore the profound impact of market innovation atmosphere, strategic alliances between enterprises, and external environmental factors (such as policy changes) on the market ecology. This results in the model being unable to fully capture market dynamics when predicting market changes, thereby affecting the accuracy and reliability of the forecast. At the same time, in the model evolution driven by evolutionary algorithms, traditional methods cannot adapt to market changes, thus affecting the accuracy and stability of the forecast.

[0004] In view of the shortcomings of the above-mentioned traditional technologies, it is particularly important to propose a big data driven market forecasting system. Summary of the invention

[0005] The purpose of the present invention is to make up for the shortcomings of the prior art and provide a big data-driven market forecasting system and system. The system uses graph neural networks to model the market ecology, which can deeply explore the relationships between market participants and the laws of market dynamics, providing a more accurate and reliable basis for market forecasting. At the same time, the model is continuously optimized through evolutionary algorithms, which can adapt to market changes and improve the accuracy and stability of forecasts.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: a big data driven market forecasting system, which includes a data collection and preprocessing module, a market ecology modeling module of a graph neural network, a model evolution module driven by an evolutionary algorithm, and a model evaluation and feedback module;

[0007] Data collection and preprocessing module: collects market data from multiple sources, including but not limited to internal operation data of enterprises, data released by industry associations, survey data of market research institutions, and relevant transaction data of Internet platforms. For internal operation data of enterprises, it is obtained by interfacing with the information system of enterprises. For data of industry associations, it regularly crawls the data published on their official websites through web crawler technology. For data of market research institutions, it is obtained through purchase and cooperation. For transaction data of Internet platforms, it is extracted according to the API interface provided by the platform.

[0008] Clean, convert and integrate the collected data, remove duplicate records, erroneous data and missing values ​​in the data, and use the mean filling method to handle missing values ​​if the data is numerical and the missing ratio is small. If the missing ratio is large, regression filling is used according to the data distribution characteristics. Unify the format and standardize the data from different sources;

[0009] The market ecology modeling module of graph neural network: Let the set of participants in the market be V = {v 1 ,v 2 ,…,v n}, taking the participants as nodes of the graph neural network, the node feature matrix X∈R n×d Determined by multi-dimensional comprehensive quantitative method, for each node v i , whose eigenvector x i It consists of three dimensions: basic attribute characteristics, behavioral characteristics, and market influence characteristics. The basic attribute characteristics include the inherent attributes of the participants. The behavioral characteristics are obtained by analyzing the interactive behavior data of the participants in the market. The market influence characteristics are determined by analyzing the degree of influence of the participants on the market price and supply and demand relationship. In the specific calculation, the basic attribute characteristics are directly quantified, the behavioral characteristics are obtained by statistically analyzing the behavioral data and normalizing them to the interval [0, 1], and the market influence characteristics are obtained by calculating the volatility coefficient of the relevant market indicators of the participant. For the enterprise node, if the change in its product price causes the average price change of the industry to be Δp, and the average price fluctuation of the industry is Δp, then the value of its market influence characteristic in this dimension is

[0010] The relationship between entities is used as an edge to form an edge set E, and the feature matrix of the edge A∈R n×n×e Determined by the method of dynamic evolution of relations, for the edge (v i ,v j ), its l-th dimension edge feature a ijl By analyzing the historical evolution data of the relationship between the two, it is obtained that the relationship between the two is in the time series t = {t 1 ,t 2 ,…,t m There are multiple states S = {s1 ,s 2 ,…,s k}, the duration of each state is τ pq , where p represents the state number, q represents the time segment number in this state, and the importance weight of the relationship w pq Determined by the degree of impact on key market indicators under this state, if the state s 1 The average change rate of market trading volume is α 1 , the sum of the average change rates of market trading volume in all states is but Then the l-th edge feature a ijl for: Among them, h l It is a nonlinear mapping function for the l-th dimension edge feature, customized according to the relationship characteristics of different market fields;

[0011] Defining the message passing function of graph neural network For the kth layer to pass messages from node j to node i, the formula is: Among them, ξ is an adaptive activation function, which is in the form of Among them, μ and v are parameters dynamically determined according to the distribution pattern of market data, w 1kq ∈R 1×m 、w 2kr ∈R m×1 is a learnable weight matrix, based on the market supply and demand balance indicator b t With the preset balance threshold b 0 In contrast, if b t >b 0 , then the weight update direction tends to reduce the intensity of message transmission between nodes. If b t 0 , it tends to increase the intensity of message transmission, and the weight update step size is adaptively adjusted according to the degree of market imbalance;

[0012] Node update function The formula used to update the features of node i in layer k is: Among them, ∈ is a small positive number to prevent the denominator from being zero, N(i) is the set of neighbor nodes of node i, and w 3k ∈R d×m is the weight matrix, and the competitive network density ρ of the market where the node is located is used as a reference. If ρ is high, it indicates that the market competition is fierce, and the initial value of the weight matrix tends to strengthen the weight of the node characteristics and the messages of the nodes with strong competitive relationships. If ρ is low, the initial value of the weight matrix tends to balance the weight of the node characteristics and the messages of the neighboring nodes. 3k ∈R m ​The bias vector is divided into four phases: recovery, prosperity, depression and recession according to the market cycle theory. The current market cycle phase is determined by analyzing the market macroeconomic indicators and industry-specific indicators, and then the adjustment range of the bias vector is determined.

[0013] After propagation through K layers of graph neural network, the final node feature H = {h 1 ,h 2 ,…,h n}, where h i Used to predict market dynamics related indicators y, the prediction function P is: Among them, δ is a very small positive number to prevent the denominator from being zero, w 4i ∈R m×1 is a weight vector. For different market segments, the contribution rate of the segment to the overall market value growth is β i To determine the weight w 4i ,Right now b 4 ∈R is the bias, which is determined by quantile regression analysis of historical market stability period data, and the weight w 1kq、 w 2kr、 w 3k、 w 4i and bias b 3k、 b 4 By using the back propagation algorithm to optimize on a large data set, consisting of historical market data, including data on the attributes of participants and the relationships between them, the predicted index y is minimized with respect to the real market dynamics index y true The mean square error is used to update the weights and biases, that is: Where N is the number of samples in the data set, and t is the sample index;

[0014] Model evolution module driven by evolutionary algorithm: After the market ecology modeling module of the graph neural network obtains the initial prediction model, the evolutionary algorithm evolves the model, assuming that the structural parameter vector of the model is θ = θ 1 ,θ 2 ,…,θ p ), where p is the number of parameters. The mutation operation in the evolutionary algorithm is performed in the following way. For each parameter θ i , with probability p m Mutation, the mutated parameter θ ′ i for: Among them, α is the asynchronous length of the change. If the growth rate of the number of adopters of innovative products or technologies in the market is g t , the total number of potential adopters in the market is G, then r is a random number θ uniformly distributed in the interval [-1, 1] median is the median of the current model parameters, θ max and θ min are the maximum and minimum values ​​of the current model parameters respectively;

[0015] The selection operation is based on the fitness function F(θ), which is composed of the prediction performance evaluation index on the validation data set, the accuracy Acc, and the formula is: Among them, TP is the number of true positives, TN is the number of true negatives, FP is the number of false positives, and FN is the number of false negatives. According to the fitness function, the excellent structural parameters of the mutated model are selected to replace the structural parameters in the original model to achieve model evolution;

[0016] Model evaluation and feedback module: Use an independent test data set to evaluate the evolved model. In addition to the accuracy mentioned above, the evaluation indicators also include recall, F1-score, mean absolute error, and root mean square error. For classification prediction problems, the recall and F1-score are calculated by constructing a confusion matrix. For numerical prediction problems, the mean absolute error and root mean square error are calculated.

[0017] The evaluation results are fed back to the model evolution module driven by the evolutionary algorithm and the market ecology modeling module of the graph neural network. If the model evaluation results do not meet the preset performance standards, the model evolution module driven by the evolutionary algorithm adjusts the mutation probability and step length parameters according to the feedback information and re-evolves the model. The market ecology modeling module of the graph neural network adjusts the method of determining node features and edge features according to the feedback information and re-models the market ecology to optimize the performance of the market forecasting system.

[0018] Furthermore, in the data collection and preprocessing module, for the acquisition of Internet platform transaction data, when extracting data through the API interface provided by the platform, it is necessary to consider the timeliness and integrity of the data. To ensure the timeliness of the data, a timed collection mechanism is set up. According to the data update frequency of different platforms, for professional service platforms with low transaction frequency, data is collected once a day. For data integrity, data verification is performed during the collection process. If data is found to be missing or abnormal, a data supplement request is sent to the platform to obtain complete data. If the platform cannot provide complete data, a data repair algorithm is used. If only transaction quantity data exists, the repair is estimated based on the average transaction amount of the same type of goods on the platform.

[0019] Furthermore, in the data collection and preprocessing module, when non-numerical data is converted into quantitative features, when natural language processing technology is used to process market strategy information of text description type, in addition to lexical and syntactic analysis, semantic analysis is also required. By constructing a semantic knowledge base specific to the market field, semantic annotation is performed on the keywords in the text, and corresponding quantitative weights are assigned according to the semantic knowledge base. The word vector model in deep learning is used to convert the words in the text into vector representations, and the similarity between the vectors is calculated to measure the similarity of the market strategies, thereby providing richer quantitative features for market forecasting.

[0020] Furthermore, in the market ecology modeling module of the graph neural network, the market influence characteristics are determined based on the impact on market prices and supply and demand relationships, as well as the impact on the market innovation atmosphere. For enterprise nodes, the number of new products and new technology applications launched during the period are counted and compared with the industry average. Suppose the number of new products launched by the enterprise is n new , the average number of new products in the industry is Its value in the innovation influence dimension is This value is weighted and combined with the price and supply-demand impact values ​​to obtain a more comprehensive market influence characteristic. The weight can be determined based on the characteristics of the industry in which the market is located.

[0021] Furthermore, in the market ecology modeling module of the graph neural network, the importance weight w of the relationship pq In determining the impact of external environmental factors, in addition to the impact on key market indicators, the impact of external environmental factors must also be considered. In a policy-sensitive market, if the sales growth rate of new energy vehicles under the cooperative relationship between enterprises is γ after the policy is released, policy , the sum of the sales growth rates of this partnership under all policies is S is the number of policies, and the weights of the policy subsidy-related dimensions are This weight is combined with the weight determined by the degree of influence of key market indicators to obtain a more accurate edge feature weight.

[0022] Furthermore, the message passing function in the market ecosystem modeling module of the graph neural network In the adaptive activation function ξ, the parameters μ and v are not only dynamically determined according to the distribution of market data, but also need to consider the impact of market emergencies. By monitoring relevant event indicators, when an event occurs, μ and v are adjusted urgently. If the event severity index is 1, μ and v are scaled in a certain proportion according to the size of 1, so that the activation function can quickly adapt to the changes in market data caused by emergencies and ensure the effectiveness of message transmission.

[0023] Furthermore, the node update function U of the market ecology modeling module in the graph neural network i k In the equation, the weight matrix w 3k In the initialization method based on market competition network analysis, in addition to considering the market competition network density ρ, the market strategic alliance situation of the enterprise must also be considered. For enterprise nodes with strategic alliance relationships, the weight of enterprise node messages within the alliance is increased when the weight matrix is ​​initialized. The alliance relationship is determined by analyzing the strategic alliance agreements and cooperation projects between enterprises. If enterprise i and enterprise i have an alliance relationship, a higher message transmission weight is given between the two when the weight matrix is ​​initialized to better reflect the impact of alliance relationships in the market on the update of enterprise node features.

[0024] Furthermore, in the prediction function P of the market ecology modeling module of the graph neural network, the weight vector w 4i When determining the method based on market segment value contribution analysis, in addition to considering the contribution rate β to the overall market value growth i In addition, the risk level of the market segment should be considered. In the high-risk emerging market segment, the value contribution rate should be adjusted for risk. The risk coefficient of this segment is set as r. risk , then the adjusted weight w 4 ′ i =w 4i ·(1-r risk ), the risk coefficient can be determined by analyzing the market fluctuations and technical uncertainties in this field, so that the prediction function can take into account both the value contribution and the market risk.

[0025] Furthermore, in the model evolution module driven by the evolutionary algorithm, the variable step length α is determined based on the method of evaluating the market innovation diffusion speed, in addition to the growth rate of the number of adopters of innovative products and technologies g t In addition to the total number of potential adopters G in the market, the substitution effect of innovative products and technologies must also be considered. If the substitution rate of innovative products for existing products is s alt , then the asynchronous length is adjusted to By considering the substitution effect, the model evolution can better adapt to the complex changes in the process of innovative products replacing existing products in the market, thereby improving the model's adaptability to market dynamics.

[0026] Furthermore, in the model evaluation and feedback module, when a qualitative evaluation method based on market expert knowledge is adopted, in order to improve the objectivity and accuracy of the evaluation, an expert evaluation database is established to record the evaluation history of each expert, including the evaluated market forecasting model, the given evaluation index value, the comparison between the actual market development situation and the evaluation results. The weights of the experts are dynamically adjusted according to their evaluation accuracy, and higher weights are given to experts with high evaluation accuracy. When integrating the qualitative evaluation results of multiple experts, a weighted average is taken according to the weights, and experts are regularly trained to update their market knowledge to ensure that their evaluation capabilities are in line with market development trends.

[0027] Compared with the prior art, an intelligent design method, system and detection method for architectural decoration based on user portraits have the following beneficial effects:

[0028] 1. The present invention realizes comprehensive and efficient processing and analysis of market data by integrating multiple modules, including data collection and preprocessing, market ecology modeling of graph neural networks, model evolution driven by evolutionary algorithms, and model evaluation and feedback. It can automatically collect market data from multiple sources, including internal enterprise operation data, data released by industry associations, survey data from market research institutions, and related transaction data on Internet platforms, thereby ensuring the breadth and accuracy of the data. The use of graph neural networks for market ecology modeling can deeply explore the relationship between market participants and the laws of market dynamics, providing a more accurate and reliable basis for market forecasting. At the same time, the model can be continuously optimized through evolutionary algorithms to adapt to market changes and improve the accuracy and stability of forecasts.

[0029] 2. The present invention evaluates the model through a variety of evaluation indicators, including accuracy, recall rate, F1-score, mean absolute error and root mean square error, to ensure that the prediction performance of the model is optimal. It also supports qualitative evaluation methods based on market expert knowledge. By establishing an expert evaluation database and dynamically adjusting expert weights, the objectivity and accuracy of the evaluation are improved. These beneficial effects make the big data-driven market prediction system of the present invention more reliable and practical in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are only some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0031] Figure 1 Operational flow chart of the big data-driven market forecasting system. DETAILED DESCRIPTION

[0032] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0033] Embodiment 1

[0034] This embodiment describes a large e-commerce platform that plans to build a big data-driven market forecasting system in order to accurately grasp market dynamics and optimize operational strategies. The system aims to use the transaction data, user behavior data, and merchant operation data within the platform, combined with external market intelligence, to predict future market trends, changes in consumer preferences, and potential hot products, thereby guiding inventory management, product recommendation strategies, and marketing activity planning.

[0035] Regularly collect platform transaction data through API interfaces, covering key information such as product sales, transaction amounts, user reviews, and browsing records; integrate merchant operation data, including basic merchant information, promotion records, and new product releases; use web crawler technology to capture public data from competitor platforms, such as hot-selling product lists, user reviews, and promotion information; purchase or cooperate to obtain in-depth reports on the e-commerce market from market research institutions; understand industry trends and consumer behavior analysis; clean the collected data to remove duplicates, errors, and missing values; for missing values, use mean filling, regression filling, or data repair algorithms to process them according to the data type and missing ratio; quantify non-numeric data, such as using natural language processing technology to analyze the sentiment tendencies in user reviews and convert text information into quantifiable feature values.

[0036] Merchants, users and commodities are used as nodes of the graph neural network to construct a relationship diagram of market participants. Merchant node features include basic attributes, behavioral features and market influence features. User node features can cover user portraits, purchasing behaviors, and social relationships. Commodity node features include commodity attributes, prices, sales, and reviews. For each node, a feature matrix is ​​constructed based on its feature vector. The feature vector consists of three dimensions: basic attribute features, behavioral features, and market influence features, and is determined through a multi-dimensional comprehensive quantitative method.

[0037] Analyze the competitive relationship between merchants, the social relationship between users, and the complementary or substitutive relationship between commodities, construct edge sets, and determine the edge feature matrix according to the importance weight of the relationship. Suppose the relationship between the two is in the time series t=t 1 ,t 2 ,…,t m There are multiple states S = s1 ,s 2 ,…,s k , the duration of each state is τ pq , where p represents the state number, q represents the time segment number in this state, and the importance weight of the relationship w pq Determined by the degree of impact on key market indicators under this state, if the state s 1 The average change rate of market trading volume is α 1 , the sum of the average change rates of market trading volume in all states is but Then the l-th edge feature Among them, h l It is a nonlinear mapping function for the l-th dimension edge feature, customized according to the relationship characteristics of different market fields;

[0038] Defining the message passing function of graph neural network For the kth layer to pass messages from node j to node i, the formula is: Among them, ξ is an adaptive activation function, which is in the form of Among them, μ and v are parameters dynamically determined according to the distribution pattern of market data, w 1kq ∈R 1×m 、w 2kr ∈R m×1 It is a learnable weight matrix whose parameters are dynamically determined according to the distribution pattern of market data and the impact of emergencies. The initialization of the weight matrix takes into account factors such as the density of market competition networks and the situation of corporate market strategic alliances to optimize the efficiency of information transmission.

[0039] Use the node update function to update node features: in, is the node update function used to update the features of node i in the kth layer, ∈ is a small positive number to prevent the denominator from being zero, N(i) is the set of neighbor nodes of node i, w 3k ∈R d×m The weight matrix is ​​used to adjust the bias vector in combination with the market cycle theory. After propagation through the multi-layer graph neural network, the final node features are obtained to predict market dynamics-related indicators. The prediction function takes into account the value contribution rate and risk level of market segments, and improves the prediction accuracy by optimizing weights and biases.

[0040] The initial prediction model is evolved using an evolutionary algorithm, and the structural parameter vector of the model is set to θ = θ 1 ,θ 2 ,…,θ p ), where p is the number of parameters. The mutation operation in the evolutionary algorithm is performed in the following way. For each parameter θ i , with probability p mMutation, the mutated parameter θ ′ i for: Among them, α is the asynchronous length of the change. If the growth rate of the number of adopters of innovative products or technologies in the market is g t , the total number of potential adopters in the market is G, then r is a random number θ uniformly distributed in the interval [-1, 1] median is the median of the current model parameters, θ max and θ min are the maximum and minimum values ​​of the current model parameters respectively;

[0041] The selection operation is based on the fitness function F(θ), which is composed of the prediction performance evaluation index on the validation data set, the accuracy Acc, and the formula is: Among them, TP is the number of true positives, TN is the number of true negatives, FP is the number of false positives, and FN is the number of false negatives. According to the fitness function, the excellent structural parameters of the mutated model are selected to replace the structural parameters in the original model to achieve model evolution;

[0042] An independent test data set is used to evaluate the evolved model. Evaluation indicators include accuracy, recall, F1-score, mean absolute error, and root mean square error. The evaluation results are fed back to the model evolution module and the market ecology modeling module. The model parameters and feature determination methods are adjusted based on the feedback to optimize system performance. A qualitative evaluation method based on market expert knowledge is introduced, and an expert evaluation database is established to record and adjust the weights of expert evaluation results to improve the objectivity and accuracy of the evaluation.

[0043] Embodiment 2

[0044] This embodiment describes a plan by a well-known electronic product manufacturer to use a big data-driven market forecasting system to accurately grasp the dynamic trends of the smartphone market in the coming year. The system will focus on predicting the market demand for new products, the movements of competitors, and market price fluctuations, providing data support for the company's strategic planning.

[0045] Obtain key data from the company's internal operating system to collect historical sales data, covering sales volume, return rate, and user feedback of various smartphone models; obtain product line planning information, including specifications, functions, and expected launch dates of new products to be released; use web crawler technology to regularly capture industry reports published on the official website of industry associations, focus on market size, growth rate, and competitive landscape information; obtain smartphone market research reports published by market research institutions; analyze competitors' market strategies and changes in user preferences; establish cooperative relationships with mainstream e-commerce platforms; obtain transaction data in real time through API interfaces, including smartphone sales volume, price change trends, and user reviews; perform data cleaning on e-commerce platform transaction data to remove duplicate and erroneous data; and use mean filling and regression filling methods to effectively handle missing values ​​to ensure data integrity and accuracy.

[0046] Market participants, including but not limited to competitors, consumers, and supply chain partners, are taken as nodes of the graph neural network. A feature matrix is ​​constructed for each node, including basic attributes such as brand awareness, market share, financial status, behavioral characteristics such as marketing efforts, user purchasing behavior, product replacement speed, and market influence characteristics, such as the impact of price changes on the industry average price and the ability to regulate supply and demand. According to the relationship between the participants, an edge set is constructed, and the edge feature matrix is ​​determined using historical data. The importance weights of the relationships are comprehensively considered, such as the closeness of the cooperative relationship and the intensity of the competitive relationship. Message passing functions and node update functions are defined to achieve information exchange and feature update between nodes. According to the density of the market competition network and the market strategic alliance of the enterprise, the initial value of the weight matrix is ​​adjusted to more accurately reflect the market reality. The market cycle theory is introduced, and the bias vector is adjusted according to the current market cycle phase so that the model can capture changes in market trends.

[0047] Based on the initial prediction model, the model is evolved using the evolutionary algorithm. The structural parameter vector of the model is θ = (θ 1 ,θ 2 ,…,θ p ), where p is the number of parameters. The mutation operation in the evolutionary algorithm is performed in the following way. For each parameter θ i , with probability p m Mutation, the mutated parameter θ ′ i for: Among them, α is the asynchronous length of the change. If the growth rate of the number of adopters of innovative products or technologies in the market is g t , the total number of potential adopters in the market is G, then r is a random number θ uniformly distributed in the interval [-1, 1] medianis the median of the current model parameters, θ max and θ min are the maximum and minimum values ​​of the current model parameters respectively;

[0048] The selection operation is based on the fitness function F(θ), which is composed of the prediction performance evaluation index on the validation data set, the accuracy Acc, and the formula is: Among them, TP is the number of true positives, TN is the number of true negatives, FP is the number of false positives, and FN is the number of false negatives. According to the fitness function, the excellent structural parameters of the mutated model are selected to replace the structural parameters in the original model, so as to realize the evolution of the model and optimize the structural parameters of the model. According to the growth rate of the number of adopters of innovative products and technologies, the total number of potential adopters in the market, and the substitution effect of innovative products, the mutation step length is dynamically adjusted so that the model can better adapt to market changes. The evolved model is evaluated using an independent test data set. The evaluation indicators include accuracy, recall, F1-score, mean absolute error, and root mean square error. The evaluation results are fed back to the model evolution module and the market ecology modeling module. According to the feedback, the model parameters and the method of determining the node and edge features are adjusted to continuously optimize the performance of the market prediction system. If conditions permit, a qualitative evaluation method based on market expert knowledge can be adopted to establish an expert evaluation database, dynamically adjust the expert weights, and improve the objectivity and accuracy of the evaluation.

[0049] Based on the optimized market forecasting system, the overall scale, growth rate, and price trend of the smartphone market in the next year are predicted, the market demand for new products is analyzed, the sales volume and market share of various smartphone models are predicted, and the trends of competitors are monitored, including their product strategies, marketing efforts, and price adjustments. The forecast results are applied to the company's product line planning, marketing strategy formulation, and price strategy adjustments. According to market demand and competitor trends, the company's resource allocation and production plans are adjusted to better meet market demand and enhance competitiveness.

[0050] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. Big data driven market forecasting system, characterized by: The system includes a data collection and preprocessing module, a market ecology modeling module based on graph neural networks, a model evolution module driven by evolutionary algorithms, and a model evaluation and feedback module: Data collection and preprocessing module: collects market data from multiple sources, including but not limited to internal operation data of enterprises, data released by industry associations, survey data of market research institutions, and relevant transaction data of Internet platforms. For internal operation data of enterprises, it is obtained by interfacing with the information system of enterprises. For data of industry associations, it regularly crawls the data published on their official websites through web crawler technology. For data of market research institutions, it is obtained through purchase and cooperation. For transaction data of Internet platforms, it is extracted according to the API interface provided by the platform. Clean, convert and integrate the collected data, remove duplicate records, erroneous data and missing values ​​in the data, and use the mean filling method to handle missing values ​​if the data is numerical and the missing ratio is small. If the missing ratio is large, regression filling is used according to the data distribution characteristics. Unify the format and standardize the data from different sources; The market ecology modeling module of graph neural network: Let the set of participants in the market be V = {v1,v2,…,v n }, taking the participants as nodes of the graph neural network, the node feature matrix X∈R n×d Determined by multi-dimensional comprehensive quantitative method, for each node v i , whose eigenvector x i It consists of three dimensions: basic attribute characteristics, behavioral characteristics, and market influence characteristics. The basic attribute characteristics include the inherent attributes of the participants. The behavioral characteristics are obtained by analyzing the interactive behavior data of the participants in the market. The market influence characteristics are determined by analyzing the degree of influence of the participants on the market price and supply and demand relationship. In the specific calculation, the basic attribute characteristics are directly quantified, the behavioral characteristics are obtained by statistically analyzing the behavioral data and normalizing them to the interval [0, 1], and the market influence characteristics are obtained by calculating the volatility coefficient of the relevant market indicators of the participant. For the enterprise node, if the change in its product price causes the average price change of the industry to be Δp, the average price fluctuation of the industry is The value of its market influence characteristic in this dimension is The relationship between entities is used as an edge to form an edge set E, and the feature matrix of the edge A∈R n×n×e Determined by the method of dynamic evolution of market relations, for edge (v i ,v j ), its l-th dimension edge feature a ijl By analyzing the historical evolution data of the relationship between the two, it is obtained that the relationship between the two is in the time series t = {t1, t2, …, t m There are multiple states S = {s1, s2, ..., s} k }, the duration of each state is τ pq , where p represents the state number, q represents the time segment number in this state, and the importance weight of the relationship w pq Determined by the degree of influence on key market indicators under this state, if the average change rate of market trading volume under state s1 is α1, the sum of the average change rates of market trading volume under all states is but Then the l-th edge feature a ijl for: Among them, h l It is a nonlinear mapping function for the l-th dimension edge feature, customized according to the relationship characteristics of different market fields; Define the message passing function of the graph neural network For the k-th layer, it is used to pass messages from node j to node i, and the formula is: where ξ is an adaptive activation function, and its form is where μ and v are parameters dynamically determined according to the distribution form of market data, and w 1kq ∈R 1×m 、w 2kr ∈R m ×1 are learnable weight matrices. According to the comparison between the market supply and demand balance index b t and the preset balance threshold b0, if b t >b0, the weight update direction tends to reduce the message passing intensity between nodes. If b t <b0, it tends to increase the message passing intensity, and the weight update step size is adaptively adjusted according to the market imbalance degree; Node update function The formula used to update the features of node i in layer k is: Among them, ∈ is a small positive number to prevent the denominator from being zero, N(i) is the set of neighbor nodes of node i, and w 3k ∈R d×m is the weight matrix, and the competitive network density ρ of the market where the node is located is used as a reference. If ρ is high, it indicates that the market competition is fierce, and the initial value of the weight matrix tends to strengthen the weight of the node characteristics and the messages of the nodes with strong competitive relationships. If ρ is low, the initial value of the weight matrix tends to balance the weight of the node characteristics and the messages of the neighboring nodes. 3k ∈R m The bias vector is divided into four phases: recovery, prosperity, depression and recession according to the market cycle theory. The current market cycle phase is determined by analyzing the market macroeconomic indicators and industry-specific indicators, and then the adjustment range of the bias vector is determined. After propagation through K layers of graph neural network, the final node feature H = {h1,h2,…,h n }, where h i Used to predict market dynamics related indicators y, the prediction function P is: Among them, δ is a very small positive number to prevent the denominator from being zero, w 4i ∈R m×1 is a weight vector. For different market segments, the contribution rate of the segment to the overall market value growth is β i To determine the weight w 4i ,Right now b4∈R is the bias, which is determined by quantile regression analysis of historical market stability period data, and the weight w 1kq、 w 2kr、 w 3k、 w 4i and bias b 3k、 b4 is optimized by using the back propagation algorithm on a large data set, which consists of historical market data, including data on the attributes of participants and the data on their relationships, by minimizing the difference between the predicted indicator y and the real market dynamics indicator y true The mean square error is used to update the weights and biases, that is: Where N is the number of samples in the data set, and t is the sample index; Model evolution module driven by evolutionary algorithm: After the market ecology modeling module of the graph neural network obtains the initial prediction model, the evolutionary algorithm evolves the model, assuming that the structural parameter vector of the model is θ = θ1, θ2, …, θ p ), where p is the number of parameters. The mutation operation in the evolutionary algorithm is performed in the following way. For each parameter θ i , with probability p m Mutation, the mutated parameter θ ′ i for: Among them, α is the asynchronous length of the change. If the growth rate of the number of adopters of innovative products or technologies in the market is g t , the total number of potential adopters in the market is G, then r is a random number θ uniformly distributed in the interval [-1, 1] median is the median of the current model parameters, θ max and θ min are the maximum and minimum values ​​of the current model parameters respectively; The selection operation is based on the fitness function F(θ), which is composed of the prediction performance evaluation index on the validation data set, the accuracy Acc, and the formula is: Among them, TP is the number of true positives, TN is the number of true negatives, FP is the number of false positives, and FN is the number of false negatives. According to the fitness function, the excellent structural parameters of the mutated model are selected to replace the structural parameters in the original model to achieve model evolution; Model evaluation and feedback module: Use an independent test data set to evaluate the evolved model. In addition to the accuracy mentioned above, the evaluation indicators also include recall, F1-score, mean absolute error, and root mean square error. For classification prediction problems, the recall and F1-score are calculated by constructing a confusion matrix. For numerical prediction problems, the mean absolute error and root mean square error are calculated. The evaluation results are fed back to the model evolution module driven by the evolutionary algorithm and the market ecology modeling module of the graph neural network. If the model evaluation results do not meet the preset performance standards, the model evolution module driven by the evolutionary algorithm adjusts the mutation probability and step length parameters according to the feedback information and re-evolves the model. The market ecology modeling module of the graph neural network adjusts the method of determining node features and edge features according to the feedback information and re-models the market ecology to optimize the performance of the market forecasting system.

2. The big data driven market forecasting system according to claim 1, characterized in that: In the data collection and preprocessing module, for the acquisition of Internet platform transaction data, when extracting data through the API interface provided by the platform, it is necessary to consider the timeliness and integrity of the data. In order to ensure the timeliness of the data, a timed collection mechanism is set up. According to the data update frequency of different platforms, for professional service platforms with low transaction frequency, data is collected once a day. For data integrity, data verification is performed during the collection process. If data is found to be missing or abnormal, a data supplement request is sent to the platform to obtain complete data. If the platform cannot provide complete data, a data repair algorithm is used. If only transaction quantity data exists, the repair is estimated based on the average transaction amount of the same type of goods on the platform.

3. The big data driven market forecasting system according to claim 1, characterized in that: In the data collection and preprocessing module, when non-numerical data is converted into quantitative features, when natural language processing technology is used to process market strategy information of text description type, in addition to lexical and syntactic analysis, semantic analysis is also required. By constructing a semantic knowledge base specific to the market field, semantic annotation is performed on the keywords in the text, and corresponding quantitative weights are assigned according to the semantic knowledge base. The word vector model in deep learning is used to convert the words in the text into vector representations, and the similarity between the vectors is calculated to measure the similarity of the market strategies, thereby providing richer quantitative features for market forecasting.

4. The big data driven market forecasting system according to claim 1, characterized in that: In the market ecology modeling module of the graph neural network, the market influence characteristics are determined based on the impact on market prices and supply and demand relationships, as well as the impact on the market innovation atmosphere. For enterprise nodes, the number of new products and new technology applications launched during the period are counted and compared with the industry average. The number of new products launched by the enterprise is set as n. new , the average number of new products in the industry is Its value in the innovation influence dimension is This value is weighted and combined with the price and supply-demand impact values ​​to obtain a more comprehensive market influence characteristic. The weight can be determined based on the characteristics of the industry in which the market is located.

5. The big data driven market forecasting system according to claim 1, characterized in that: In the market ecology modeling module of the graph neural network, the importance weight w of the relationship pq In addition to the impact on key market indicators, the impact of external environmental factors must also be considered in determining the degree of impact on key market indicators. In a policy-sensitive market, if the sales growth rate of new energy vehicles under the cooperative relationship between enterprises is γ after the policy is released, policy , the sum of the sales growth rates of this partnership under all policies is S is the number of policies, and the weights of the policy subsidy-related dimensions are This weight is combined with the weight determined by the degree of influence of key market indicators to obtain a more accurate edge feature weight.

6. The big data driven market forecasting system according to claim 1, characterized in that: The message passing function of the market ecosystem modeling module in the graph neural network In the adaptive activation function ξ, the parameters μ and v are not only dynamically determined according to the distribution pattern of market data, but also need to consider the impact of sudden market events. By monitoring relevant event indicators, when an event occurs, μ and v are adjusted urgently. If the event severity index is 1, μ and v are scaled in a certain proportion according to the size of 1, so that the activation function can quickly adapt to the changes in market data caused by sudden events.

7. The big data driven market forecasting system according to claim 1, characterized in that: The node update function of the market ecology modeling module in the graph neural network In the equation, the weight matrix w 3k In the initialization method based on market competition network analysis, in addition to considering the market competition network density ρ, the market strategic alliance situation of the enterprise must also be considered. For enterprise nodes with strategic alliance relationships, the weight of enterprise node messages within the alliance is increased when the weight matrix is ​​initialized. The alliance relationship is determined by analyzing the strategic alliance agreements and cooperation projects between enterprises. If enterprise i and enterprise i have an alliance relationship, a higher message transmission weight is given between the two when the weight matrix is ​​initialized.

8. The big data driven market forecasting system according to claim 1, characterized in that: In the prediction function P of the market ecology modeling module of the graph neural network, the weight vector w 4i When determining the method based on market segment value contribution analysis, in addition to considering the contribution rate β to the overall market value growth i In addition, the risk level of the market segment should be considered. In the high-risk emerging market segment, the value contribution rate should be adjusted for risk. The risk coefficient of this segment is set as r. risk , then the adjusted weight w4 ′ i =w 4i ·(1-r risk ), the risk coefficient can be determined by analyzing the market fluctuations and technical uncertainties in this field, so that the prediction function can take into account both the value contribution and the market risk.

9. The big data driven market forecasting system according to claim 1, characterized in that: In the model evolution module driven by the evolutionary algorithm, the variable asynchronous length α is determined based on the method of evaluating the market innovation diffusion speed, in addition to the growth rate of the number of adopters of innovative products and technologies g t In addition to the total number of potential adopters G in the market, the substitution effect of innovative products and technologies must also be considered. If the substitution rate of innovative products for existing products is s alt , then the asynchronous length is adjusted to By considering the substitution effect, the model evolution can better adapt to the complex changes in the process of innovative products replacing existing products in the market.

10. The big data driven market forecasting system according to claim 1, characterized in that: In the model evaluation and feedback module, when a qualitative evaluation method based on market expert knowledge is adopted, in order to improve the objectivity and accuracy of the evaluation, an expert evaluation database is established to record the evaluation history of each expert, including the evaluated market forecast model, the given evaluation index value, the comparison between the actual market development situation and the evaluation results. The weight of the expert is dynamically adjusted according to the accuracy of the expert's evaluation, and a higher weight is given to experts with high evaluation accuracy. When the qualitative evaluation results of multiple experts are integrated, a weighted average is performed according to the weights, and experts are regularly trained to update their market knowledge.

Images