Intelligent push method for securities information

Abstract

The present invention discloses an intelligent push method for securities information, including: obtaining real-time market data, user historical behavior data and investment portfolio data, performing normalization processing to generate standardized data, using segmented accumulation method and multi-level entropy weight method to generate context feature data and new information feature data; calculating market mutual information value to construct market correlation data, using market liquidity weighted information diffusion model to generate transmission intensity data; using LSTM network and adaptive spectrum analysis method to generate correlation score data; using graph neural network to construct information transmission graph to identify key path to generate path influence data; using multi-objective dynamic programming method to generate push sequence data. The present invention can accurately identify information value, accurately predict market reaction, and improve the accuracy and timeliness of information push.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent push, and in particular to an intelligent push method for securities information. Background Art

[0002] In the securities market, the timeliness and accuracy of information directly affect the effectiveness of investment decisions. With the globalization of financial markets and the development of information technology, the amount of securities-related information generated every day is growing exponentially, the speed of information dissemination is accelerating, and the linkage effect between markets is becoming increasingly significant. Investors are no longer facing the problem of information scarcity, but how to quickly identify and obtain valuable information from massive amounts of information. Especially in the current context of the prevalence of high-frequency trading and programmatic trading, the timeliness requirements for information processing have reached the millisecond level, and the transmission effect of information has become increasingly complex. Therefore, developing a set of technical solutions that can accurately understand the market structure, accurately identify the value of information, and realize intelligent push is of great significance to improving the decision-making efficiency of investors and the information efficiency of the market.

[0003] At present, securities information push mainly adopts rule-based filtering methods and simple machine learning models. Traditional rule filtering methods mainly rely on keyword matching and preset subject classification rules, and achieve information screening by establishing tag libraries such as securities codes, industry classifications, and subject concepts. Some systems have introduced collaborative filtering algorithms to perform similarity calculations and information recommendations based on users' historical behavior data. In terms of market correlation analysis, linear correlation indicators such as the Pearson correlation coefficient are mainly used to measure the degree of correlation between markets. The evaluation of information transmission effects mainly relies on simple statistical methods, such as calculating the lagged correlation of market indices or analyzing the relationship between markets through vector autoregression models. These methods perform well when dealing with structured data and static relationships, but they are no longer able to cope with the current rapidly changing market environment and complex information ecology.

[0004] The existing technical solutions have the following specific problems in practical applications: First, in the feature extraction link, the existing methods often simply superimpose the data features of different time scales without considering the differences in the importance of the features of each time level, resulting in the inability to reasonably distribute the weights of the long-term trend and short-term fluctuations. Second, in terms of market correlation analysis, the existing linear correlation indicators cannot effectively capture the nonlinear correlation and conditional correlation between markets, and do not consider the impact of market liquidity on the efficiency of information transmission, resulting in the identification bias of the transmission path. Third, in state transition analysis, traditional time series models are difficult to simultaneously handle the impact of long-term dependencies and emergencies, resulting in insufficient accuracy in state prediction. Fourth, the existing graph analysis methods mostly use static graph structures, which cannot reflect the dynamic changes of market structure in a timely manner, affecting the accuracy of transmission path identification. Finally, in terms of push sequence optimization, the existing methods often use a single-objective optimization strategy, which is difficult to balance multiple objectives such as timeliness, relevance and influence, resulting in unsatisfactory push effects. The existence of these technical problems has seriously affected the accuracy and practicality of securities information push, and new technical solutions need to be developed to solve them. Summary of the invention

[0005] The purpose of the invention is to provide a method for intelligently pushing securities information, in order to solve at least one technical problem existing in the prior art.

[0006] The technical solution, the intelligent push method of securities information, comprises the following steps:

[0007] S1. Obtain real-time market data, user historical behavior data and investment portfolio data, perform normalization processing, and generate standardized data; use the segmented accumulation method to process the standardized data to generate time series feature data; use the multi-level entropy weight method to process the time series feature data to obtain context feature data; obtain new information data and perform feature extraction to obtain new information feature data; real-time market data includes price, trading volume and volatility; user historical behavior data includes transaction records, browsing records and stay time; investment portfolio data includes positions, profits and losses, and turnover rate;

[0008] S2. Obtain historical market data, calculate the mutual information value between markets, and construct market correlation data; based on market correlation data and real-time market data, use the market liquidity weighted information diffusion model to generate transmission intensity data;

[0009] S3, based on the context feature data, new information feature data and pre-stored historical state data, using LSTM network processing to generate state transition data; based on the state transition data and new information feature data, using adaptive spectrum analysis method to calculate and generate correlation score data;

[0010] S4. Based on the transmission intensity data and the new information feature data, the graph neural network is used to construct the information transmission graph data; based on the information transmission graph data, the key transmission path is identified and the key path data is generated; the influence of the key path data is calculated to obtain the path influence data;

[0011] S5. Based on the relevance score data and the path influence data, the comprehensive score of the information is calculated to generate importance score data; based on the importance score data and the context feature data, a multi-objective dynamic programming method is used to obtain push sequence data.

[0012] Beneficial effects: The present invention establishes a comprehensive feature representation system, accurately grasps the law of information dissemination, realizes dynamic evaluation of information relevance and intelligent identification of transmission paths; it not only improves the accuracy and timeliness of information push, but also realizes accurate prediction of market reactions, reduces information lag and information overload problems, provides users with more valuable information services, and at the same time improves the scalability and adaptability of the system, realizing intelligent and accurate push of securities information. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is a flow chart of the present invention.

[0014] Figure 2 This is a flow chart of step S1 of the present invention.

[0015] Figure 3 This is a flow chart of step S2 of the present invention.

[0016] Figure 4 This is a flow chart of step S3 of the present invention.

[0017] Figure 5 This is a flow chart of step S4 of the present invention.

[0018] Figure 6 This is a flow chart of step S5 of the present invention. DETAILED DESCRIPTION

[0019] like Figure 1 As shown, the present application proposes a method for intelligently pushing securities information, comprising the following steps:

[0020] S1. Obtain real-time market data, user historical behavior data and investment portfolio data, perform normalization processing, and generate standardized data; use the segmented accumulation method to process the standardized data to generate time series feature data; use the multi-level entropy weight method to process the time series feature data to obtain context feature data; obtain new information data and perform feature extraction to obtain new information feature data; real-time market data includes price, trading volume and volatility; user historical behavior data includes transaction records, browsing records and stay time; investment portfolio data includes positions, profits and losses, and turnover rate;

[0021] S2. Obtain historical market data, calculate the mutual information value between markets, and construct market correlation data; based on market correlation data and real-time market data, use the market liquidity weighted information diffusion model to generate transmission intensity data;

[0022] S3, based on the context feature data, new information feature data and pre-stored historical state data, using LSTM network processing to generate state transition data; based on the state transition data and new information feature data, using adaptive spectrum analysis method to calculate and generate correlation score data;

[0023] S4. Based on the transmission intensity data and the new information feature data, the graph neural network is used to construct the information transmission graph data; based on the information transmission graph data, the key transmission path is identified and the key path data is generated; the influence of the key path data is calculated to obtain the path influence data;

[0024] S5. Based on the relevance score data and the path influence data, the comprehensive score of the information is calculated to generate importance score data; based on the importance score data and the context feature data, a multi-objective dynamic programming method is used to obtain push sequence data.

[0025] like Figure 2 As shown, according to one aspect of the present application, step S1 is further:

[0026] S11, obtaining real-time market data, user historical behavior data and investment portfolio data, and using the maximum and minimum value normalization method to process the real-time market data, user historical behavior data and investment portfolio data respectively to generate standardized data;

[0027] S12, calculating the volatility value of the standardized data and determining the length of the data segment; based on the length of the data segment, performing cumulative calculation on each segment of data to generate time series feature data;

[0028] S13, processing the time series feature data in layers according to hourly, daily, and weekly levels, calculating the information entropy of each layer; using the information entropy to generate layer weights; performing weighted fusion on the features of each layer according to the layer weights, and outputting context feature data;

[0029] S14. Acquire new information data containing intelligence source, content and time, perform word segmentation and entity recognition processing on the new information data to generate feature item data; perform semantic vectorization processing on the feature item data to generate semantic vector data; perform feature extraction based on the semantic vector data in combination with a preset feature extraction template to generate new information feature data.

[0030] In one embodiment of the present application, real-time market data M (including price, trading volume, volatility), user historical behavior data U (transaction records, browsing records, dwell time), and investment portfolio data P (positions, profits and losses, turnover rate) are obtained, and each type of data is normalized using the maximum and minimum normalization method to generate a standardized vector, and a standardized market vector Mn, a standardized user behavior vector Un, and a standardized combination vector Pn are output. Based on the standardized market vector Mn, the standardized user behavior vector Un, and the standardized combination vector Pn, an adaptive segmented accumulation method is used to calculate the data volatility V = Σ|ΔX| / n, where X is the input sequence and n is the sequence length; the segment length L = f(V) is dynamically determined according to the volatility V, and f is a decreasing function; the cumulative eigenvalue C = Σ(wi×xi) is calculated for each segment of data, where wi is the time decay weight, and the time series feature matrix Ts is output. Based on the time series feature matrix Ts, the multi-level entropy weight method is used to stratify the feature matrix according to different time scales t: hourly, daily, and weekly levels; the information entropy Hi = -Σ(pij×ln(pij)) of each level is calculated, where pij is the normalized eigenvalue, and the level weight Wi = (1-Hi) / Σ(1-Hi) is generated; features of different levels are fused: F = Σ(Wi×Fi), where Fi is the feature of each level, and the context feature vector Cv is output.

[0031] In another embodiment of the present application, the correspondence between the volatility value and the data segment length is: volatility piecewise function: L(σ) = L_base·(1 + α·σ_norm)·exp(-β·|σ_diff|); where L(σ) is the segment length; L_base is the basic segment length (such as 30 minutes); σ_norm is the normalized volatility; σ_diff is the volatility change rate; α is the volatility sensitivity coefficient; β is the change rate influence coefficient. The specific corresponding relationship is as follows: when σ_norm≤0.2, L=L_base, and the basic segment length remains unchanged; when 0.2<σ_norm≤0.5, L = L_base·(1 + α·σ_norm), and the segment length increases linearly; when 0.5<σ_norm≤0.8, L = L_base·(1 + α·σ_norm)·exp(-β·|σ_diff|), considering the impact of the rate of change; when σ_norm>0.8, L=L_base / 2, and the segment is shortened when volatility is high. Adaptive adjustment of volatility threshold: σ_threshold(t) =μ·σ_historical + (1-μ)·σ_realtime; where σ_threshold(t) is the volatility threshold at time t; σ_historical is the historical volatility mean; σ_realtime is the real-time volatility; μ is the smoothing coefficient. Adaptive volatility segmentation ensures optimal feature extraction under different market conditions.

[0032] Collaborative processing method for data at different time scales: Calculation of collaborative weights between levels: W_sync(i, j) = exp(-|H(i) - H(j)| / τ)·ρ(i, j); where W_sync(i, j) is the collaborative weight of levels i and j; H(i) and H(j) are the information entropy of each level; ρ(i, j) is the correlation coefficient between levels; τ is the temperature parameter. Hierarchical feature fusion function: F(t) = Σ(W_k·F_k(t) + Σ(W_sync(k, m)·Cross(F_k(t), F_m(t)))); where F(t) is the fused feature; F_k(t) is the feature of the kth level; W_k is the hierarchical weight; Cross(·,·) is the feature cross term; k, m are different time levels. Inter-level consistency constraint: C(t) =||F_hour(t)-DownSample(F_day(t))|| 2 +||F_day(t)-DownSample(F_week(t))|| 2; C(t) is the consistency loss; F_hour(t), F_day(t), and F_week(t) are hour, day, and week level features respectively; DownSample is the downsampling function. The collaborative processing of multi-level features improves the fusion effect of information at different time scales.

[0033] This embodiment achieves a comprehensive analysis of market data, user behavior data, and investment portfolio data through hierarchical data processing and feature extraction methods. First, the maximum and minimum normalization method is used to standardize different types of data, so that data of different dimensions and scales can be compared and analyzed under the same framework; then, the standardized data is processed by the segmented accumulation method, which can capture the changing characteristics of the data at different time scales, not only retaining the temporal characteristics of the data, but also reflecting the cumulative effect of the data; finally, the feature fusion of the multi-level entropy weight method is realized, specifically, the weights are determined by calculating the information entropy of features at different levels, and adaptive weighting of features is realized. This embodiment ensures the comprehensiveness and accuracy of feature extraction, so that the system can simultaneously consider market fluctuations, user preferences, and investment portfolio status, providing a reliable feature basis for subsequent information push, while reducing the complexity of data processing and improving the real-time response capability of the system.

[0034] According to one aspect of the present application, step S11 is further:

[0035] S111, collecting real-time market data, user historical behavior data and investment portfolio data, and determining a minimum time granularity; based on the minimum time granularity, aligning the real-time market data, user historical behavior data and investment portfolio data according to timestamps to generate time-aligned data;

[0036] S112, calculating the quantile statistics of the time-aligned data; based on the quantile statistics, identifying outliers and performing smoothing to generate smoothed data; performing integrity check on the smoothed data and completing missing values ​​to obtain completed data; verifying the completed data according to preset data quality rules to generate verification data;

[0037] S113, classifying the verification data into numerical type and proportional type to obtain numerical data and proportional data; applying the maximum and minimum value normalization method to the numerical data to generate numerical normalized data; performing logarithmic transformation on the proportional data and then normalizing it to obtain proportional normalized data;

[0038] S114, calculating the distribution characteristics of the numerical normalized data and the proportional normalized data; determining the fusion weight according to the distribution characteristics; and performing weighted fusion processing based on the fusion weight to generate standardized data.

[0039] In one embodiment of the present application, the normalization method is: X_norm(i) = α·(x(i) - min(x)) / (max(x) - min(x)) + β·log(1 + x(i)) / log(1 + max(x)); wherein X_norm(i) is the normalized eigenvalue; x(i) is the original eigenvalue; min(x) is the minimum eigenvalue; max(x) is the maximum eigenvalue; α is the linear normalization weight; β is the logarithmic normalization weight; α+β=1; i is the feature index.

[0040] This embodiment achieves standardized processing of heterogeneous data through multi-level data preprocessing and normalization methods. First, the system collects different types of market data, user behavior data, and portfolio data, and ensures the temporal consistency of the data through timestamp alignment; then, the quantile statistics and outlier detection methods are used to smooth the data and fill in missing values, thereby improving the data quality; finally, according to the different data types, the maximum and minimum value normalization and logarithmic transformation methods are used for standardization. This embodiment not only ensures the comparability of data from different sources, but also maintains the original distribution characteristics of the data and avoids data distortion. At the same time, through the calculation of distribution characteristics and weight fusion, the effective integration of different types of data is achieved, providing a high-quality data foundation for subsequent feature extraction, and improving the data processing efficiency and accuracy of the system.

[0041] According to one aspect of the present application, step S12 is further:

[0042] S121. Read standardized data, use the exponentially weighted moving average method to calculate short-term volatility to generate short-term volatility data, use the Parkinson method to calculate medium-term volatility to generate medium-term volatility data, calculate long-term volatility based on the Yang-Zhang method to generate long-term volatility data, perform weighted combination processing on the three types of volatility data, and output volatility data.

[0043] S122. Obtain volatility data, calculate the local maximum and minimum values ​​of volatility to generate extreme value data, identify the trend change points of volatility based on the extreme value data to obtain trend data, determine the initial segmentation points based on the trend data to generate initial segmentation data, calculate the volatility difference between adjacent segments to optimize and adjust the initial segmentation data, and output the segmentation data.

[0044] S123. Read the segmented data, calculate the linear cumulative sum of the data in each segment to obtain the sum data, calculate the exponential cumulative sum of the data in the segment to generate the exponential sum data, calculate the logarithmic cumulative sum of the data in the segment to obtain the logarithmic sum data, calculate the weighted cumulative product of the data in the segment to obtain the product data, combine the sum data, exponential sum data, logarithmic sum data and product data in chronological order to generate time series feature data.

[0045] In one embodiment of the present application, the volatility calculation method is: VOL(t) = λ1·σ_ewma(t) + λ2·σ_park(t) + λ3·σ_yz(t); wherein VOL(t) is the comprehensive volatility; σ_ewma(t) is the exponentially weighted moving average volatility; σ_park(t) is the Parkinson volatility; σ_yz(t) is the Yang-Zhang volatility; λ1, λ2, λ3 are adaptive weight coefficients; and t is the time index.

[0046] In another embodiment of the present application, the time series data is further standardized, specifically: set a fixed time window size T (such as 30 minutes), segment the context feature data according to the time window T, and generate time window sequence data; calculate the feature mean and standard deviation in each time window, use the interpolation method to process the missing time points, and generate a feature sequence of equal interval sampling; set the sequence length N (such as 48, indicating that the sampling is performed at 30 minutes within 24 hours), truncate the super long sequence, fill the short sequence with zero or repeat, and output the feature sequence of fixed length N. The feature dimension alignment is specifically as follows: construct context features: [price, trading volume, volatility, ...], with a dimension of d1; construct new information features: [entity type, sentiment score, importance, ...], with a dimension of d2; construct historical state: [previous state vector], with a dimension of d3; set a unified feature dimension D = max(d1, d2, d3), use principal component analysis (PCA) to reduce the dimension or use an autoencoder to transform the features, and output a feature vector of a unified dimension D.

[0047] This embodiment achieves accurate feature extraction of market data through multi-dimensional volatility calculation and adaptive segmentation methods. First, the system combines volatility calculation methods of three different time scales: short-term, medium-term and long-term, to comprehensively capture market volatility characteristics; then, it identifies trend change points based on volatility data and realizes adaptive segmentation of data; finally, through different types of cumulative calculations, it obtains multi-dimensional features of the data. This embodiment can not only accurately identify the market's volatility cycle, but also adaptively adjust the segment length, avoiding the problem of feature loss that may be caused by fixed windows. Through a variety of cumulative calculation methods such as linear, exponential, logarithmic and product, the system can comprehensively capture the changing characteristics of the data, provide rich feature information for subsequent analysis, and improve the accuracy and comprehensiveness of feature extraction.

[0048] According to one aspect of the present application, step S13 is further:

[0049] S131, grouping the time series feature data according to hourly intervals to obtain hourly data; grouping the time series feature data according to daily intervals to obtain daily data; grouping the time series feature data according to weekly intervals to obtain weekly data; calculating the mean, variance, skewness and kurtosis characteristics of the hourly data, daily data and weekly data to generate stratified data;

[0050] S132. Based on the stratified data, a fixed-width binning method is used to calculate the information entropy at the hour level to obtain the hourly information entropy; based on the stratified data, an adaptive binning method is used to calculate the information entropy at the day level to obtain the daily information entropy; based on the stratified data, a kernel density estimation method is used to calculate the information entropy at the week level to obtain the weekly information entropy; the hourly information entropy, the daily information entropy and the weekly information entropy are combined to generate information entropy data;

[0051] S133, based on the information entropy data, calculating the normalized value of the information entropy of each level to obtain the normalized entropy value; based on the normalized entropy value, calculating the initial weight of each level; performing variance constraint optimization on the initial weight, and outputting the level weight data;

[0052] S134. Multiply the hourly hierarchical features in the hierarchical data by the corresponding weights in the hierarchical weight data to obtain the hourly feature weights; multiply the daily hierarchical features in the hierarchical data by the corresponding weights in the hierarchical weight data to obtain the daily feature weights; multiply the weekly hierarchical features in the hierarchical data by the corresponding weights in the hierarchical weight data to obtain the weekly feature weights; perform fusion based on the hourly feature weights, daily feature weights and weekly feature weights to generate contextual feature data.

[0053] In one embodiment of the present application, the entropy weight calculation method is: W(k) = -H(k)·ln(p(k)) / Σ(-H(j)·ln(p(j))); wherein W(k) is the weight of the kth level; H(k) is the information entropy of the kth level; p(k) is the probability distribution of the kth level; j is the level index; k is the current level.

[0054] This embodiment realizes the adaptive fusion of features through multi-level entropy weight analysis. First, the system processes the data in layers according to the three time scales of hours, days, and weeks, and calculates the statistical features of each level; then, different binning methods are used to calculate the information entropy of each level to ensure the accuracy of the entropy value calculation; finally, the level weights are calculated based on the normalized entropy value to realize the adaptive fusion of features. This embodiment can automatically identify the importance of features at different time scales, avoiding the subjective problems that may be caused by artificially setting weights. Through variance constraint optimization, the system can also ensure the rationality of weight allocation and improve the stability of feature fusion. This embodiment not only improves the accuracy of feature representation, but also enhances the system's ability to adapt to market changes, providing a reliable feature basis for subsequent analysis.

[0055] According to one aspect of the present application, step S14 is further:

[0056] S141. Acquire new information data, use a rule filtering method to remove noise information to obtain filtered data, use a word segmentation algorithm to segment the filtered data to obtain word segmentation data, identify entities in the word segmentation data based on a pre-trained model to obtain entity data, and integrate the word segmentation data and entity data to generate feature item data.

[0057] S142, read feature item data, calculate the word frequency-inverse document frequency value to obtain TF-IDF data, use the word vector model to calculate the semantic similarity between words to obtain similarity data, weight the TF-IDF data and similarity data based on the attention mechanism to obtain weighted data, perform dimensionality reduction processing on the weighted data, and output semantic vector data.

[0058] S143, reading a preset feature extraction template to obtain template data, matching the semantic vector data with the template data to obtain matching data, calculating a correlation score of the matching data to generate score data, filtering the score data according to a preset threshold, and outputting new information feature data.

[0059] This embodiment realizes intelligent feature extraction of new information through deep learning and semantic analysis methods. First, the system uses rule filtering and word segmentation algorithms to pre-process the information, remove noise and extract keywords; then, entity recognition is performed through a pre-trained model to identify key entities in the information; finally, the semantic features of the information are calculated by combining TF-IDF and word vector models. This embodiment can not only accurately identify the key content in the information, but also capture the deep semantic features of the information. Through the weighting and dimensionality reduction processing of the attention mechanism, the system can extract the most representative features, improving the efficiency and accuracy of feature extraction. At the same time, the feature matching mechanism based on the preset template ensures the practicality and interpretability of the extracted features, and provides reliable feature support for the correlation analysis of information.

[0060] like Figure 3 As shown, according to one aspect of the present application, step S2 is further:

[0061] S21. Obtain historical market data including stocks, bonds and foreign exchange, and calculate mutual information values ​​between different markets based on the historical market data; and construct directed weighted market association data according to the mutual information values;

[0062] S22. Based on market correlation data and real-time market data, calculate the liquidity index of each market; calculate the inter-market transmission coefficient based on the liquidity index and mutual information value; generate transmission intensity data based on the inter-market transmission coefficient.

[0063] In one embodiment of the present application, each market historical data set Md (historical data of stocks, bonds, foreign exchange and other markets) is obtained, the mutual information MI(X, Y) = Σp(x, y)log(p(x, y) / (p(x)p(y))) between markets is calculated, and a directed weighted graph G(V, E) is constructed, where V is a market node and E is an associated edge weight, to obtain a market association graph Mg; based on the market association graph Mg and real-time market data M, a market liquidity weighted information diffusion model is used to calculate the market liquidity index Li = Vi / (Pi×σi), where Vi is the trading volume, Pi is the price, and σi is the volatility; and a transmission coefficient Tij = α×MI(i, j)×(Li×Lj) is constructed. 0.5 , α is the adjustment parameter; construct the conduction matrix T = [Tij] and output the conduction intensity matrix Tm.

[0064] In another embodiment of the present application, the mutual information calculation method is: MI(X, Y) = Σ(p(x, y)·log(p(x, y) / (p(x)·p(y)))) + γ·(1 - exp(-|ρ(X, Y)|)); wherein MI(X, Y) is the mutual information value; p(x, y) is the joint probability; p(x) and p(y) are the marginal probabilities; ρ(X, Y) is the correlation coefficient; and γ is the adjustment coefficient.

[0065] Transmission coefficient calculation method: TC(i, j) = MI(i, j)·(1 + β·LIQ(i))·(1 + β·LIQ(j))·exp(-α·d(i, j)); where TC(i, j) is the transmission coefficient from market i to j; MI(i, j) is the mutual information value; LIQ(i) and LIQ(j) are market liquidity indicators; d(i, j) is the market distance; α is the attenuation coefficient; β is the liquidity impact coefficient.

[0066] This embodiment achieves an accurate characterization of the market transmission effect by constructing a correlation network between markets and calculating an information diffusion model weighted by market liquidity. First, the mutual information value is calculated based on historical market data, which can capture the nonlinear correlation between markets and better reflect the real correlation between markets than the traditional correlation coefficient method; then, the market liquidity index is introduced to weight the transmission coefficient. This weighting mechanism takes into account the actual trading activity of the market, so that the calculation of the transmission intensity is more in line with the actual market situation. This embodiment can not only identify the direct influence relationship between markets, but also discover the implicit indirect transmission path, thereby providing a scientific quantitative basis for the cross-market transmission of information; it improves the accuracy of market correlation analysis, so that the system can better predict the transmission effect and influence range of information between different markets.

[0067] According to one aspect of the present application, step S21 is further:

[0068] S211, obtaining historical market data, using wavelet transform to remove high-frequency noise, and obtaining noise-reduced data; using exponential smoothing method to process the noise-reduced data, and obtaining trend data; performing outlier detection and correction on the trend data, and outputting preprocessed data;

[0069] S212. Based on the preprocessed data, calculate the marginal entropy of each market to obtain marginal entropy data; based on the preprocessed data, calculate the joint entropy between market pairs to obtain joint entropy data; calculate the conditional entropy based on the marginal entropy data and the joint entropy data to generate conditional entropy data; calculate the mutual information value between different markets based on the conditional entropy data;

[0070] S213. Based on the mutual information value and the preset mutual information threshold, filter the filtered mutual information value; based on the filtered mutual information value, construct the connection relationship between markets to obtain connection data; assign the mutual information value as the weight of the edge to the connection data to construct directed weighted market association data.

[0071] This embodiment achieves the accurate construction of the market correlation network through multi-level data processing and mutual information analysis. First, the system uses wavelet transform to perform noise reduction on historical market data, removing high-frequency noise while retaining important market information; then, by calculating edge entropy, joint entropy and conditional entropy, the information dependency between markets is accurately evaluated; finally, a market correlation network is established based on the mutual information value. This embodiment can capture the nonlinear correlation between markets and avoid the complex correlation relationships that may be ignored by the traditional correlation coefficient method. Through the screening of mutual information thresholds and the construction of directed weighted networks, the system can not only identify significant market correlations, but also reflect the direction and intensity of influence between markets, providing a reliable network structure foundation for subsequent conduction analysis and improving the accuracy and reliability of market correlation analysis.

[0072] According to one aspect of the present application, step S22 is further:

[0073] S221. Read real-time market data, calculate the bid-ask spread to obtain a spread index, calculate the trading volume change rate to obtain a trading volume index, calculate the turnover rate to obtain a turnover rate index, and output a liquidity index based on a comprehensive calculation of the three indicators.

[0074] S222. Obtain the mutual information value in the liquidity indicator and the market-related data, calculate the liquidity weighting factor to obtain the weighting factor, multiply the weighting factor by the mutual information value, and generate a transmission coefficient.

[0075] S223, read the conduction coefficient, construct a conduction matrix to obtain a conduction matrix, analyze the conduction matrix to identify the main conduction path to obtain path data, calculate the conduction strength according to the path data, and generate conduction strength data.

[0076] This embodiment achieves an accurate assessment of market transmission intensity through liquidity weighting and transmission coefficient calculation. First, the system comprehensively considers factors such as bid-ask spread, volume change rate, and turnover rate to calculate the market's liquidity index; then, the liquidity index is combined with the mutual information value to generate a weighted transmission coefficient; finally, the final transmission intensity data is obtained by constructing a transmission matrix and analyzing the main transmission paths. This embodiment not only reflects the information correlation between markets, but also takes into account the actual trading activity of the market, making the assessment of transmission intensity more in line with actual market conditions. Through the analysis of the transmission matrix and the identification of the main paths, the system can accurately grasp the law of information dissemination in the market network, provide a scientific basis for the priority judgment of information push, and improve the accuracy and practicality of market impact prediction.

[0077] like Figure 4 As shown, according to one aspect of the present application, step S3 is further:

[0078] S31, based on the context feature data, the new information feature data and the pre-stored historical state data, the LSTM network is used for processing, including input gate processing, forget gate processing, output gate processing and cell state processing, to generate state transition data;

[0079] S32, performing wavelet transform on the state transition data and the new information feature data to obtain time-frequency features; based on the time-frequency features, calculating the correlation scores of different frequency bands; performing weighted synthesis on the correlation scores, and outputting correlation score data.

[0080] In one embodiment of the present application, based on the context feature vector Cv and the historical state sequence Hs, the context state transition probability is modeled using an LSTM network, including an input gate: it = σ(Wi×[h(t-1),xt] + bi); a forget gate: ft= σ(Wf×[h(t-1),xt] + bf); an output gate: ot = σ(Wo×[h(t-1),xt] + bo); a cell state: ct =ft×c(t-1) + it×tanh(Wc×[h(t-1),xt] + bc); and a state transition matrix St is obtained. Based on the state transition matrix St and the new information feature vector Nv, an adaptive spectrum analysis method is used to perform a wavelet transform on the information feature and the state sequence to obtain a time-frequency feature; and the correlation scores of different frequency bands are calculated: R(f) = Σ(X(f)×Y*(f)) / sqrt(Σ|X(f)| 2 ×Σ|Y(f)| 2); Weighted synthesis final score: S = Σ(w(f)×R(f)), w(f) is the frequency weight, and the output relevance score vector Rv. Among them, σ is the activation function, Wi is the weight matrix of the input gate, xt is the input vector of the current time step, bi is the bias vector of the input gate, Wf is the weight matrix of the forget gate, h(t-1) is the hidden state vector of the previous time step, bf is the bias vector of the forget gate, Wo is the weight matrix of the output gate, bo is the bias vector of the output gate, c(t-1) is the cell state vector of the previous time step, Wc is the weight matrix of the cell state update, bc is the bias vector of the cell state update, X(f) is the information feature vector at frequency f, Y*(f) is the conjugate complex of the state sequence feature vector at frequency f, and Y(f) is the state sequence feature vector at frequency f.

[0081] In another embodiment of the present application, the LSTM state update method is: h(t) = o(t) Θtanh(c(t)); c(t) = f(t) Θc(t-1) + i(t) Θg(t); wherein h(t) is the hidden state; c(t) is the cell state; o(t) is the output gate; f(t) is the forget gate; i(t) is the input gate; g(t) is the candidate memory unit; Θ is the Hadamard product.

[0082] Spectral correlation calculation method: Corr(f) = η·ρ_p(f) + (1-η)·ρ_s(f)·exp(-Δ·|f-f_0|); where Corr(f) is the correlation score of frequency band f; ρ_p(f) is the Pearson correlation coefficient; ρ_s(f) is the Spearman correlation coefficient; f_0 is the reference frequency; η is the correlation coefficient weight; Δ is the frequency attenuation coefficient.

[0083] This embodiment realizes the dynamic evaluation of information relevance through the combined application of LSTM network and adaptive spectrum analysis. First, the three-gate structure (input gate, forget gate, output gate) of LSTM network can effectively handle long-term dependency problems. Through the cell state update mechanism, the system can retain important historical information features while filtering out irrelevant information; then, the correlation of different frequency bands is analyzed using adaptive spectrum analysis method, which can capture the relevant features of information at different time scales, so as to more comprehensively evaluate the relevance of information. By combining time domain analysis (LSTM) and frequency domain analysis (spectrum analysis), this embodiment can not only accurately identify the immediate correlation between information, but also discover potential long-term correlations, improve the accuracy and stability of correlation evaluation, and provide reliable guarantee for the accuracy of information push.

[0084] In another embodiment of the present application, the specific application method of the LSTM network is as follows: network structure design: the input layer is X(t) = [Context(t); Info(t); State(t-1)]; dimension = d_context + d_info + d_state; the hidden layer is a 3-layer LSTM structure, the first layer: 128 neurons, dropout rate 0.2; the second layer: 64 neurons, dropout rate 0.3; the third layer: 32 neurons, dropout rate 0.4; the output layer is a fully connected layer, and the output dimension = d_output. The detailed design of the LSTM unit is as follows: forget gate: f(t) = σ(W_f·[h(t-1), x(t)] + b_f); input gate: i(t) = σ(W_i·[h(t-1), x(t)] + b_i); candidate memory: g(t) = tanh(W_g·[h(t-1), x(t)] + b_g); output gate: o(t) = σ(W_o·[h(t-1), x(t)] + b_o); state update: c(t) = f(t)Θc(t-1) + i(t)Θg(t); hidden state update: h(t) = o(t)Θtanh(c(t)). The initial value of the learning rate is 0.001, and the Adam optimizer is used; the batch size is 64; the sequence length is 48 (indicates 24 hours of data, with a sampling interval of 30 minutes); the weights are initialized using Xavier; the activation function is GeLU, and the L2 regularization coefficient is 0.0001. The specific training strategy is as follows: In the pre-training stage: use historical data for pre-training; adopt the curriculum learning strategy, gradually increase from short sequences to the target sequence length; the number of pre-training rounds is 100 rounds. In the online fine-tuning stage: use sliding windows for incremental learning; the window size is 7 days; the update frequency is every hour; the number of fine-tuning rounds is 10 rounds. Attention mechanism: Attention(Q, K, V) = softmax(Q·K T / sqrt(d_k))·V; where Q is the query matrix; K is the key matrix; V is the value matrix; d_k is the attention dimension, T is transposed. Residual connection: h'_l = LayerNorm(h_l + Attention(h_l)); where h_l is the hidden state of the lth layer; LayerNorm is layer normalization. The deep LSTM structure and attention mechanism enhance the modeling ability of long-term dependencies.

[0085] According to one aspect of the present application, step S31 is further:

[0086] S311, read context feature data, new information feature data and historical status data, calculate the input gate control factor to obtain input gate data, perform activation function processing on the input gate data to obtain input control data, use the input control data to adjust the current input information, and generate input status data.

[0087] S312, obtaining historical state data, calculating the forget gate control factor to obtain forget gate data, performing activation function processing on the forget gate data to obtain forget control data, and using the forget control data to adjust the historical state information to generate forget state data.

[0088] S313, read the input state data, calculate the output gate control factor to obtain the output gate data, perform activation function processing on the output gate data to obtain the output control data, adjust the cell state in combination with the forgetting state data, and generate state transition data.

[0089] This embodiment realizes the intelligent fusion of historical states and current inputs through the gating mechanism of the LSTM network. First, the system controls the influence of the current input information through the input gate; then, the retention degree of historical state information is adjusted through the forget gate; finally, the final state transition data is generated through the output gate and the update of the cell state. This embodiment can automatically learn long-term dependencies and effectively solve the gradient vanishing problem in traditional RNN networks. Through the synergy of the three gating units, important historical information can be intelligently retained, while responding to new market changes in a timely manner, improving the accuracy and stability of state transitions. This embodiment can not only capture the long-term trends of the market, but also quickly adapt to sudden changes in the market, providing reliable state support for the evaluation of information relevance.

[0090] According to one aspect of the present application, step S32 is further:

[0091] S321, based on the state transition data and the new information feature data, using the wavelet basis function to perform multi-scale decomposition to obtain wavelet coefficients; performing time-frequency domain mapping on the wavelet coefficients to generate time-frequency features;

[0092] S322, dividing the time-frequency features into a high frequency band, a medium frequency band and a low frequency band according to a preset frequency band division standard to obtain segmented features; calculating the Pearson correlation coefficient and the Spearman correlation coefficient of each frequency band in the segmented features; performing weighted average of the Pearson correlation coefficient and the Spearman correlation coefficient to generate a correlation score for each frequency band;

[0093] S323. Based on the correlation score, calculate the importance weight of each frequency band to obtain the frequency band weight; use the frequency band weight to perform weighted summation processing on the correlation score of each frequency band to obtain correlation score data.

[0094] This embodiment realizes a multi-dimensional evaluation of information relevance through an adaptive spectrum analysis method. First, the system uses wavelet transform to perform multi-scale decomposition of the data to obtain the time-frequency characteristics of different frequency bands; then, the correlation scores of the high-frequency, medium-frequency and low-frequency bands are calculated respectively; finally, the final correlation score is obtained through adaptive calculation of the frequency band weights. This embodiment can simultaneously capture the relevant characteristics of information at different time scales, avoiding the one-sidedness that may be caused by a single time scale analysis. Through the weighted average of the two correlation coefficients and the adaptive adjustment of the frequency band weights, the degree of correlation between information can be more accurately evaluated, which improves the comprehensiveness and reliability of the correlation evaluation.

[0095] like Figure 5 As shown, according to one aspect of the present application, step S4 is further:

[0096] S41. Based on the conduction intensity data and the new information feature data, the graph neural network is used to perform node update and edge weight update processing to generate information conduction graph data;

[0097] S42, based on the information transmission diagram data, calculating the weight value of the transmission path; based on the weight value, using the Dijkstra algorithm, generating critical path data;

[0098] S43. Based on the critical path data and the conduction strength data, the path influence value is calculated and normalized to generate the path influence data.

[0099] In one embodiment of the present application, based on the conduction intensity matrix Tm and the new information feature vector Nv, a graph neural network is used to construct an information conduction path, and node updates are performed: hi' = σ(W×AGG({hj|j∈N(i)})); edge weight updates: eij' = f(hi, hj, eij); and the information conduction graph Ig is output. Based on the information conduction graph Ig, the path weight W(P) = Π(1 / rij) is constructed, where rij is the edge resistance; the Dijkstra algorithm is used to find the optimal conduction path, and the critical path set Kp is output. Based on the critical path set Kp and the conduction intensity matrix Tm, the path influence is calculated: I(P) = W(P)×L(P), where L(P) is the path length; normalization processing is performed: I'(P) = I(P) / ΣI(P), and the path influence vector Pv is output. Where W is the weight matrix, AGG is the aggregation function, hj is the feature vector of node j, N(i) is the set of neighbor nodes of node i, eij is the edge weight between node i and node j, f() is the edge weight update function, and Π is the product operator.

[0100] In one embodiment of the present application, a graph neural network updating method is: h'_v = σ(W·AGGREGATE({h_u:u∈N(v)}) + B·h_v); wherein h'_v is the updated feature of node v; h_u is the neighbor node feature; N(v) is the neighbor set of node v; W and B are weight matrices; σ is the activation function; AGGREGATE is the aggregation function.

[0101] Path weight calculation method: PW(p) = Π(TC(i, j))·exp(-λ·L(p))·(1 + μ·D(p)); where PW(p) is the weight of path p; TC(i, j) is the conduction coefficient between adjacent nodes on the path; L(p) is the path length; D(p) is the path divergence; λ is the length penalty coefficient; μ is the divergence influence coefficient.

[0102] This embodiment realizes the intelligent identification and influence assessment of information transmission paths through the combination of graph neural network and Dijkstra algorithm. First, the information transmission graph is constructed using graph neural network, which can automatically learn node features and edge weights to capture complex market correlation relationships; then, the key transmission path is identified through Dijkstra algorithm, which not only considers the path length, but also the node importance and edge transmission strength, making the identified key path more practical. This embodiment can effectively discover the law of information dissemination in the market network, accurately evaluate the scope and intensity of information influence, improve the accuracy of the system's prediction of market response, and provide a scientific basis for the priority sorting of information push. At the same time, through the influence assessment of key paths, the system can better grasp the timing and scope of information dissemination and improve the push effect.

[0103] According to one aspect of the present application, step S41 is further:

[0104] S411. Read the conduction strength data and the new information feature data, construct an initial graph structure according to the conduction strength to obtain the initial graph data, assign a feature vector to each node to obtain node feature data, assign a weight to each edge to obtain edge weight data, integrate the node feature data and the edge weight data to generate graph structure data.

[0105] S412. Obtain graph structure data, use the message passing mechanism of the graph neural network to update node features to obtain updated node data, calculate the importance of edges based on the attention mechanism to obtain updated edge weight data, combine the updated node data and the updated edge weight data, and output information conduction graph data.

[0106] This embodiment realizes the intelligent construction of the information transmission graph through the dynamic update mechanism of the graph neural network. First, the system constructs the initial graph structure based on the transmission strength and assigns initial features to the nodes and edges; then, the node features are updated through the message passing mechanism of the graph neural network; finally, the importance of the edges is dynamically adjusted using the attention mechanism. This embodiment can automatically learn the complex relationships between nodes, while highlighting important transmission paths through the attention mechanism; it can not only accurately represent the direct relationship between markets, but also discover potential indirect transmission relationships, thereby improving the accuracy and completeness of information transmission analysis. Through the dynamic update mechanism, the system can adapt to changes in the market structure in a timely manner, providing reliable graph structure support for subsequent path analysis.

[0107] According to one aspect of the present application, step S42 is further:

[0108] S421. Read the information conduction graph data, calculate the direct conduction strength between nodes to obtain direct conduction data, calculate the indirect conduction strength between nodes to obtain indirect conduction data, integrate the direct conduction data and the indirect conduction data, and generate a conduction path weight value.

[0109] S422, obtaining the conduction path weight value, using the Dijkstra algorithm to calculate multiple candidate paths to obtain candidate path data, weighting and screening the candidate path data, and outputting the key path data.

[0110] This embodiment uses the Dijkstra algorithm to achieve intelligent identification of key conduction paths. First, the system calculates the direct and indirect conduction strengths between nodes and generates a complete conduction weight value; then, the Dijkstra algorithm is used to calculate multiple candidate paths; finally, the final key path is obtained by weight sorting and screening. This embodiment not only considers the conduction strength of the path, but also the length and complexity of the path, and can more accurately identify the key conduction path in a practical sense. Through the calculation and screening mechanism of multiple paths, the system can discover different conduction channels, provide a variety of options for the formulation of information dissemination strategies, and improve the accuracy and practicality of conduction path identification.

[0111] According to one aspect of the present application, step S43 is further:

[0112] S431. Read the key path data and conduction strength data, calculate the influence of each node on the path to obtain the node influence data, calculate the conduction strength of each edge on the path to obtain the edge strength data, combine the node influence data and the edge strength data, and output the path influence value.

[0113] S432, obtaining the path influence value, calculating the maximum and minimum influence values ​​to obtain extreme value data, normalizing the path influence value according to the extreme value data, and generating the path influence data.

[0114] This embodiment achieves accurate quantification of path influence through a comprehensive evaluation method. First, the system calculates the influence of each node on the path and the conduction strength of the edge; then, the overall influence of the path is calculated by combining the node influence and the edge strength; finally, the standardized influence value is obtained through maximum and minimum normalization processing. This embodiment not only takes into account the importance of each node on the path, but also takes into account the intensity attenuation during the conduction process, which can more accurately reflect the actual influence range of information dissemination. Through normalization processing, the system can provide a unified influence measurement standard, provide a reliable quantitative basis for the priority sorting of information push, and improve the accuracy and comparability of influence evaluation.

[0115] According to one aspect of the present application, step S4 also includes: constructing node features. Constructing market node feature vectors: basic features: [market type, trading volume, volatility]; time series features: [price momentum, trading volume trend, turnover rate]; association features: [correlation coefficient matrix with other markets]. Constructing information node feature vectors: content features: [topic classification, sentiment score, importance]; timeliness features: [release time, update frequency, validity period]; dissemination features: [forward volume, comment volume, citation volume].

[0116] The specific edge construction rules are as follows: define direct correlation edges between markets based on mutual information values; define indirect correlation edges between markets based on common influencing factors; and define information-market correlation edges based on content relevance. Calculate edge weights: W(i, j) = α * MI(i, j) + β * Corr(i, j) + γ * Flow(i, j); where: MI(i, j) is the mutual information value, Corr(i, j) is the correlation coefficient, Flow(i, j) is the capital flow intensity, and α, β, γ are adaptive weights.

[0117] like Figure 6 As shown, according to one aspect of the present application, step S5 is further:

[0118] S51, calculating a weight coefficient based on the correlation score data and the path influence data; performing weighted processing on the correlation score data and the path influence data according to the weight coefficient to generate importance score data;

[0119] S52. Based on importance score data and context feature data, threshold constraints of timeliness, relevance and influence are constructed; based on the threshold constraints, a multi-objective dynamic programming method is used to generate push sequence data.

[0120] In one embodiment of the present application, based on the relevance score vector Rv and the path influence vector Pv, a comprehensive score is constructed: S = β×Rv + (1-β)×Pv, β is the weight coefficient; β is dynamically adjusted: β = g(t, v), t is the time factor, and v is the market volatility; the information importance score Is is output. Based on the information importance score Is and the context feature vector Cv, a multi-objective dynamic programming method is used to construct the objective function: F = w1×f1 (timeliness) + w2×f2 (relevance) + w3×f3 (influence); the constraints include the timeliness threshold T, the relevance threshold R, and the influence threshold I; dynamic programming is used to solve the optimal push sequence, and the final push sequence Fs is output.

[0121] In another embodiment of the present application, the importance score calculation method is: Score(i) = ω_r·R(i) +ω_p·P(i) + ω_t·exp(-θ·ΔT(i)); wherein Score(i) is the importance score of information i; R(i) is the relevance score; P(i) is the path influence; ΔT(i) is the time decay; ω_r, ω_p, ω_t are dynamic weights; θ is the time decay coefficient.

[0122] This embodiment realizes the intelligent optimization of the information push sequence through a multi-objective dynamic programming method. First, the system comprehensively considers the relevance score and path influence, and obtains the comprehensive importance score of the information through adaptive weight calculation; then, on this basis, the constraints of the three dimensions of timeliness, relevance and influence are considered, and the multi-objective dynamic programming algorithm is used to solve the optimal push sequence. This embodiment can not only balance the weights between multiple goals, but also dynamically adjust the push strategy according to the real-time market situation to ensure that the push sequence meets the timeliness requirements and maximizes the influence of the information. This embodiment can maximize the value of information while ensuring the timeliness of push, improve the push effect and user satisfaction, and avoid the problems of information overload and repeated push.

[0123] According to one aspect of the present application, step S51 is further:

[0124] S511. Read the correlation score data and the path influence data, calculate the statistical distribution characteristics of the data to obtain distribution characteristic data, calculate the information gain based on the distribution characteristic data to obtain gain data, generate an initial weight value according to the gain data, and output weight coefficient data.

[0125] S512: Obtain correlation score data, path influence data and weight coefficient data, perform weighted product operation to obtain weighted data, normalize the weighted data, and generate importance score data.

[0126] This embodiment achieves accurate assessment of information importance through an adaptive weight calculation method. First, the system analyzes the statistical distribution characteristics of relevance scores and path influence; then, the initial weight value is calculated based on information gain; finally, the final importance score is obtained through weighted product and normalization. This embodiment can automatically adapt to the characteristics of different types of information and avoid the subjectivity that may be caused by manually setting weights. Through the weighted fusion mechanism, the system can balance the importance of the two dimensions of relevance and influence, provide a scientific basis for the priority judgment of information push, and improve the accuracy and objectivity of importance assessment.

[0127] According to one aspect of the present application, step S52 is further:

[0128] S521. Read importance score data and context feature data, calculate the timeliness index of the information to obtain timeliness data, calculate the relevance index of the information to obtain relevance data, calculate the influence index of the information to obtain influence data, set thresholds according to the three indicators, and generate constraint condition data.

[0129] S522, obtaining constraint condition data, constructing a multi-objective optimization function to obtain optimization function data, setting a state transfer equation for dynamic programming to obtain state equation data, combining the optimization function data and the state equation data to generate planning data.

[0130] S523, read the planning data, use the dynamic programming algorithm to solve to obtain the solution sequence data, sort the solution sequence data in time series, and output the push sequence data.

[0131] In one embodiment of the present application, the multi-objective optimization function is F = w1*f_time + w2*f_relevance+ w3*f_impact; where f_time is the timeliness target, f_relevance is the relevance target, f_impact is the influence target, and w1, w2, and w3 are adaptive weights. Based on the historical push effect evaluation, Bayesian optimization is used to determine the initial weight, and the dynamic weight adjustment is: w_i(t+1) = w_i(t) + η * ▽Performance(t); where w_i(t) is the weight at time t, η is the learning rate, and ▽Performance(t) is the performance gradient. The weight constraint is: Σw_i = 1; 0≤w_i≤1.

[0132] In another embodiment of the present application, a feedback adjustment mechanism is also included, specifically: constructing performance evaluation indicators: user click rate, dwell time and transaction conversion rate; feedback update rules are: constructing short-term feedback based on real-time user behavior; constructing medium-term feedback based on daily level statistics; constructing long-term feedback based on periodic evaluation; weight optimization goal is: maximization: E[Performance] = Σ(w_i * metric_i); constraint is: Var[Performance] ≤ threshold. Where E[Performance] is the expected value of performance, metric_i is the i-th evaluation indicator, Var[Performance] is the variance of performance, and threshold is the performance variance threshold.

[0133] This embodiment realizes the intelligent optimization of the push sequence through a multi-objective dynamic programming method. First, the system calculates the timeliness, relevance, and influence indicators of the information, and sets the corresponding constraints; then, it constructs a multi-objective optimization function and a state transition equation; finally, it solves the optimal push sequence through a dynamic programming algorithm. This multi-objective optimization push sequence generation method can consider multiple push targets at the same time, and maximize the overall benefits while meeting various constraints. Through the dynamic programming algorithm, the system can efficiently solve large-scale sequence optimization problems and provide rapid decision support for real-time push. This embodiment not only ensures the timeliness and relevance of the push sequence, but also maximizes the influence of information, improving the accuracy and effect of information push.

[0134] In another embodiment of the present application, step S12 may also be:

[0135] S12a, reading standardized data, calculating the moving standard deviation of the data to obtain standard deviation data, calculating volatility statistics based on the standard deviation data to generate statistical value data, performing exponential smoothing on the statistical value data, and outputting volatility data.

[0136] S12b, obtain volatility data, set a threshold according to the volatility to generate threshold segmentation data, calculate the volatility difference between adjacent segments to obtain difference data, fine-tune the segmentation according to the difference data, and output the segmentation data.

[0137] S12c, read the segmented data, calculate the cumulative sum of the data in each segment to obtain the sum data, calculate the cumulative product of the data in the segment to obtain the product data, and integrate the sum data and the product data to generate time series feature data.

[0138] This embodiment can more accurately reflect the volatility of market data and reduce the impact of noise on volatility calculation by calculating the moving standard deviation and exponential smoothing. By setting thresholds and calculating the volatility differences between adjacent segments and fine-tuning the segments, it is possible to more accurately capture changes in market data and improve the accuracy of time series feature data. By calculating the cumulative sum and cumulative product of the data within the segment and integrating them to generate time series feature data, it is possible to more comprehensively reflect the dynamic changes of market data and provide richer feature information for subsequent analysis and prediction. This embodiment improves the accuracy and expressiveness of volatility calculation and time series feature data generation, and provides a solid data foundation for the intelligent push and analysis of securities information.

[0139] According to one aspect of the present application, step S13 may also be:

[0140] S13a, read the time series feature data, group the data according to hourly intervals to obtain hourly level data, group the data according to daily intervals to obtain daily level data, group the data according to weekly intervals to obtain weekly level data, calculate the statistical features of the three level data to generate statistical level data, perform standardization processing on the statistical level data, and output the level data.

[0141] S13b, obtain hierarchical data, use the fixed-width binning method to calculate the hourly hierarchical entropy value to obtain hourly entropy value data, use the adaptive binning method to calculate the daily hierarchical entropy value to obtain daily entropy value data, use the kernel density estimation method to calculate the weekly hierarchical entropy value to obtain weekly entropy value data, combine and normalize the three types of entropy value data to generate entropy value data.

[0142] S13c, read the hierarchical data, calculate the autocorrelation coefficient of each hierarchical data to obtain autocorrelation data, calculate the mutual correlation coefficient between the hierarchies to obtain mutual correlation data, construct a correlation matrix based on the autocorrelation data and the mutual correlation data, and generate correlation data.

[0143] S13d. Obtain entropy data and correlation data, use the maximum entropy principle to calculate the initial weights of each level to obtain initial weight data, use correlation constraints to optimize the initial weights to obtain optimized weight data, perform weighted fusion on the level data based on the optimized weight data, and generate context feature data.

[0144] This embodiment can more comprehensively capture data features on different time scales and improve the accuracy and depth of data analysis by dividing the time series feature data into three levels: hour, day and week, and calculating their statistical features and entropy values ​​respectively. The entropy values ​​of different levels are calculated respectively by fixed width binning, adaptive binning and kernel density estimation methods, and combined normalization processing is performed, which can effectively measure the uncertainty and information content of the data and provide a reliable basis for subsequent feature extraction and analysis. By calculating the autocorrelation coefficients of data at each level and the mutual correlation coefficients between levels, and constructing a correlation matrix, the intrinsic connection and mutual influence between data can be revealed, providing an important basis for feature fusion and model optimization. The maximum entropy principle is used to calculate the initial weights of each level, and the initial weights are optimized using correlation constraints, which can ensure the rationality and effectiveness of weight distribution, improve the quality of feature fusion and the accuracy of context feature data. This embodiment improves the accuracy and reliability of data feature extraction and analysis, and provides a solid data foundation for subsequent intelligent push and decision support.

[0145] According to one aspect of the present application, it also includes: S0, obtaining and processing new information feature data

[0146] S01. Obtain new information from multiple data sources in real time: obtain stock exchange announcements; obtain listed company announcements; obtain financial news; obtain research reports; obtain relevant discussions on social media.

[0147] S02. Extract features from the new information obtained: use natural language processing technology to extract key information from the text; identify information types (such as announcement types, news types, etc.); extract the securities codes and topics involved; extract timestamp information; calculate the initial importance score of the information,

[0148] Step S03, generating new information feature data: organizing the extracted features into a standardized data structure, establishing associations between information, and generating feature vectors of new information.

[0149] This embodiment ensures the comprehensiveness and timeliness of information by acquiring new information in real time from multiple data sources, including stock exchange announcements, listed company announcements, financial news, research reports, and social media related discussions. Natural language processing technology is used to extract key text information, identify information types, extract the securities codes and topics involved, and timestamp information, and can efficiently extract useful information from large amounts of text data. By calculating the initial importance score of the information, a preliminary assessment of the acquired new information can be made to help identify the most valuable information for users. The extracted features are organized into a standardized data structure, and the association relationship between the information is established to generate feature vectors of the new information, ensuring the consistency and availability of the data.

[0150] The present invention realizes the intelligent and accurate push of securities information by constructing a complete technical chain of data processing, feature extraction, market correlation analysis, path identification and sequence optimization. First, a comprehensive feature representation system is established through multi-level data processing and feature extraction; then, the law of information dissemination is accurately grasped through market correlation analysis and conduction effect evaluation; then, deep learning and graph analysis methods are used to realize dynamic evaluation of information relevance and intelligent identification of conduction paths; finally, the optimization of push sequence is realized through multi-objective optimization methods. The present invention not only improves the accuracy and timeliness of information push, but also realizes accurate prediction of market response, reduces information lag and information overload problems, provides users with more valuable information services, and improves the scalability and adaptability of the system.

[0151] The preferred embodiments of the present invention are described in detail above; however, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. An intelligent push method for securities information, characterized in that: The steps include: S1. Obtain real-time market data, user historical behavior data and investment portfolio data, perform normalization processing, and generate standardized data; use the segmented accumulation method to process the standardized data and generate time series feature data; The time series feature data is processed by using a multi-level entropy weight method to obtain context feature data; new information data is acquired and feature extraction is performed to obtain new information feature data; Real-time market data includes price, trading volume and volatility; historical user behavior data includes transaction records, browsing records and dwell time; portfolio data includes positions, profit and loss and turnover rate; S2. Obtain historical market data, calculate the mutual information value between markets, and construct market correlation data; based on market correlation data and real-time market data, use the market liquidity weighted information diffusion model to generate transmission intensity data; S3, based on the context feature data, new information feature data and pre-stored historical state data, using LSTM network processing to generate state transition data; based on the state transition data and new information feature data, using adaptive spectrum analysis method to calculate and generate correlation score data; S4. Based on the transmission intensity data and new information feature data, the information transmission graph data is constructed using graph neural network; Based on the information transmission diagram data, identify the key transmission path and generate the key path data; calculate the influence of the key path data and obtain the path influence data; S5. Calculate the comprehensive score of the information based on the relevance score data and the path influence data to generate importance score data; Based on the importance score data and context feature data, a multi-objective dynamic programming method is used to obtain the push sequence data.

2. The intelligent push method for securities information according to claim 1, characterized in that: Step S1 is further as follows: S11, obtaining real-time market data, user historical behavior data and investment portfolio data, and using the maximum and minimum value normalization method to process the real-time market data, user historical behavior data and investment portfolio data respectively to generate standardized data; S12, calculating the volatility value of the standardized data and determining the length of the data segment; based on the length of the data segment, performing cumulative calculation on each segment of data to generate time series feature data; S13, processing the time series feature data in layers according to hourly, daily, and weekly levels, calculating the information entropy of each level; and generating the level weight using the information entropy; According to the hierarchical weights, the features of each level are weighted and fused to output contextual feature data; S14, obtaining new information data including intelligence source, content and time, performing word segmentation and entity recognition processing on the new information data, and generating feature item data; The feature item data is semantically vectorized to generate semantic vector data; based on the semantic vector data, in combination with a preset feature extraction template, feature extraction is performed to generate new information feature data.

3. The intelligent push method for securities information according to claim 2, characterized in that: Step S2 is further as follows: S21. Obtain historical market data including stocks, bonds and foreign exchange, and calculate mutual information values ​​between different markets based on the historical market data; and construct directed weighted market association data according to the mutual information values; S22. Based on market correlation data and real-time market data, calculate the liquidity index of each market; calculate the inter-market transmission coefficient based on the liquidity index and mutual information value; generate transmission intensity data based on the inter-market transmission coefficient.

4. The intelligent push method for securities information according to claim 3, characterized in that: Step S3 is further as follows: S31, based on the context feature data, the new information feature data and the pre-stored historical state data, the LSTM network is used for processing, including input gate processing, forget gate processing, output gate processing and cell state processing, to generate state transition data; S32, performing wavelet transform on the state transition data and the new information feature data to obtain time-frequency features; based on the time-frequency features, calculating the correlation scores of different frequency bands; performing weighted synthesis on the correlation scores, and outputting correlation score data.

5. The intelligent push method for securities information according to claim 4, characterized in that: Step S4 is further as follows: S41. Based on the conduction intensity data and the new information feature data, the graph neural network is used to perform node update and edge weight update processing to generate information conduction graph data; S42, based on the information transmission diagram data, calculating the weight value of the transmission path; based on the weight value, using the Dijkstra algorithm, generating the critical path data; S43. Based on the critical path data and the conduction strength data, the path influence value is calculated and normalized to generate the path influence data.

6. The intelligent push method for securities information according to claim 5, characterized in that: Step S5 is further as follows: S51, calculating a weight coefficient based on the correlation score data and the path influence data; performing weighted processing on the correlation score data and the path influence data according to the weight coefficient to generate importance score data; S52. Based on importance score data and context feature data, threshold constraints of timeliness, relevance and influence are constructed; Based on threshold constraints, a multi-objective dynamic programming method is used to generate push sequence data.

7. The intelligent push method for securities information according to claim 6, characterized in that: Step S11 is further as follows: S111, collecting real-time market data, user historical behavior data and investment portfolio data, and determining a minimum time granularity; based on the minimum time granularity, aligning the real-time market data, user historical behavior data and investment portfolio data according to timestamps to generate time-aligned data; S112, calculating quantile statistics of the time-aligned data; Based on the quantile statistics, outliers are identified and smoothed to generate smoothed data; integrity checks are performed on the smoothed data and missing values ​​are filled to obtain filled data; based on the preset data quality rules, the filled data is verified to generate verification data; S113, classifying the verification data into numerical type and proportional type to obtain numerical data and proportional data; applying the maximum and minimum value normalization method to the numerical data to generate numerical normalized data; The proportional data is logarithmically transformed and then normalized to obtain proportional normalized data; S114, calculating the distribution characteristics of the numerical normalized data and the proportional normalized data; and determining the fusion weight according to the distribution characteristics; Based on the fusion weights, a weighted fusion process is performed to generate standardized data.

8. The intelligent push method for securities information according to claim 6, characterized in that: Step S13 is further as follows: S131, grouping the time series feature data according to hourly intervals to obtain hourly data; grouping the time series feature data according to daily intervals to obtain daily data; grouping the time series feature data according to weekly intervals to obtain weekly data; calculating the mean, variance, skewness and kurtosis characteristics of the hourly data, daily data and weekly data to generate stratified data; S132. Based on the stratified data, a fixed-width binning method is used to calculate the information entropy at the hour level to obtain the hourly information entropy; based on the stratified data, an adaptive binning method is used to calculate the information entropy at the day level to obtain the daily information entropy; based on the stratified data, a kernel density estimation method is used to calculate the information entropy at the week level to obtain the weekly information entropy; the hourly information entropy, the daily information entropy and the weekly information entropy are combined to generate information entropy data; S133. Based on the information entropy data, a normalized value of the information entropy of each level is calculated to obtain a normalized entropy value; based on the normalized entropy value, an initial weight of each level is calculated; Perform variance constraint optimization on the initial weights and output the hierarchical weight data; S134. Multiply the hourly hierarchical features in the hierarchical data by the corresponding weights in the hierarchical weight data to obtain the hourly feature weights; multiply the daily hierarchical features in the hierarchical data by the corresponding weights in the hierarchical weight data to obtain the daily feature weights; multiply the weekly hierarchical features in the hierarchical data by the corresponding weights in the hierarchical weight data to obtain the weekly feature weights; perform fusion based on the hourly feature weights, daily feature weights and weekly feature weights to generate contextual feature data.

9. The intelligent push method for securities information according to claim 6, characterized in that: Step S21 is further as follows: S211, obtaining historical market data, using wavelet transform to remove high-frequency noise, and obtaining noise-reduced data; using exponential smoothing method to process the noise-reduced data, and obtaining trend data; Detect and correct outliers on trend data and output preprocessed data; S212. Calculate the marginal entropy of each market based on the preprocessed data to obtain marginal entropy data; Based on the preprocessed data, the joint entropy between the market pairs is calculated to obtain the joint entropy data; According to the edge entropy data and the joint entropy data, the conditional entropy is calculated to generate the conditional entropy data; Based on conditional entropy data, calculate the mutual information value between different markets; S213, based on the mutual information value and the preset mutual information threshold, screening to obtain a filtered mutual information value; Based on filtering mutual information values, the connection relationship between markets is constructed to obtain connection data; the mutual information value is assigned to the connection data as the weight of the edge to construct directed weighted market association data.

10. The intelligent push method for securities information according to claim 6, characterized in that: Step S32 is further as follows: S321, based on the state transition data and the new information feature data, using the wavelet basis function to perform multi-scale decomposition to obtain wavelet coefficients; performing time-frequency domain mapping on the wavelet coefficients to generate time-frequency features; S322, dividing the time-frequency features into a high frequency band, a medium frequency band and a low frequency band according to a preset frequency band division standard to obtain segmented features; Calculate the Pearson correlation coefficient and Spearman correlation coefficient of each frequency band in the segmented features; perform weighted average of the Pearson correlation coefficient and the Spearman correlation coefficient to generate the correlation score of each frequency band; S323, calculating the importance weight of each frequency band based on the correlation score to obtain a frequency band weight; The frequency band weights are used to perform weighted summation on the correlation scores of each frequency band to obtain the correlation score data.

Images