Financial knowledge graph data fusion and retrieval middle platform system for investment advice

Abstract

The application discloses a financial knowledge graph data fusion and retrieval middle platform system for investment suggestions, and belongs to the technical field of financial data processing. The system comprises an investment analysis enhancement module and a knowledge graph construction module which interact through a distributed message queue. A multi-scale time series convolution network is adopted to dynamically calculate and add investment analysis coefficients, including a relationship transmission coefficient and an Alpha sentiment token, to entities and relationships in the knowledge graph. The module dynamically calculates the investment analysis coefficients of entities and relationships through the multi-scale time series convolution network, thereby accurately depicting market risk and yield characteristics. The relationship transmission coefficient can quantize the risk transmission strength between associated assets, thereby assisting dynamic hedging and risk control of investment portfolios. The sentiment token can capture market emotion fluctuations in real time, thereby generating differentiated signals for quantitative strategies. In combination with the efficient interaction of the distributed message queue, the module significantly improves the real-time performance and accuracy of investment decisions, and is particularly suitable for high-frequency trading and multi-factor strategy optimization.

Description

Technical Field

[0001] This invention relates to the field of financial data processing technology, and in particular to a financial knowledge graph data fusion and retrieval platform system for investment advice. Background Technology

[0002] In the investment research and decision-making process, financial professionals need to process massive amounts of multi-source, heterogeneous data from various channels. This data includes, but is not limited to: company fundamentals, real-time market data, macroeconomic indicators, news and public opinion, social media information, and analyst research reports.

[0003] Currently, mainstream financial data terminals primarily offer isolated, tabular data queries. These terminals suffer from several inherent limitations: they lack deep semantic connections between data points, failing to effectively reveal complex relationships such as "the potential impact of a company's technological breakthrough on the stock prices of its upstream and downstream supply chain companies." Information retrieval methods rely on keyword matching, making it difficult to understand user query intent and hindering multi-hop reasoning queries. Traditional systems cannot quantitatively integrate unstructured textual information (such as news sentiment) with structured market data (such as stock price fluctuations), thus failing to provide a unified, quantifiable, intelligent analytical framework for investment decisions.

[0004] Knowledge graph technology, as an emerging technology capable of representing complex relationships between entities, offers a potential solution to the aforementioned problems. However, most existing knowledge graph-based financial systems remain at the stage of simple association queries, lacking in-depth optimization for the investment field, and in particular, lacking mathematical models and mechanisms that can quantify relationship strength, incorporate market sentiment, and directly serve the investment decision-making process. Summary of the Invention

[0005] To address the aforementioned problems, this invention aims to solve the issues described above. One objective of this invention is to provide a financial knowledge graph data fusion and retrieval platform system for investment advice, which solves the problems described above.

[0006] The solution adopted in this embodiment is: a financial knowledge graph data fusion and retrieval platform system for investment advice, including: a multi-source data access and preprocessing module, configured to obtain raw data from multiple heterogeneous financial data sources, and to automatically clean, standardize, and extract entities from the raw data using deep learning-based multimodal data fusion technology, generating standardized data stored in a distributed columnar database; the heterogeneous financial data sources include macroeconomic indicator databases, corporate financial report databases, real-time market data streams, news and public opinion text data, and analyst research reports.

[0007] The knowledge graph construction module for the investment domain is connected to the data access and preprocessing module via a high-speed data bus. It is used to construct and store a knowledge graph containing entities, relationships, and attributes based on a predefined investment domain ontology model and through entity linking, relation extraction, and attribute alignment operations based on graph neural networks.

[0008] The investment analysis enhancement module interacts with the knowledge graph construction module through a distributed message queue. It uses a multi-scale temporal convolutional network to dynamically calculate and attach investment analysis coefficients to entities and relationships in the knowledge graph, including relationship transmission coefficients and Alpha sentiment tokens.

[0009] The intelligent investment research and retrieval module is connected to the investment field knowledge graph construction module and the investment analysis enhancement module. It uses a semantic parser based on the Transformer architecture to receive natural language queries in the investment field, generates graph query commands through a graph query optimization algorithm, and uses a multi-objective optimization algorithm based on reinforcement learning to reason, filter, and sort the query results.

[0010] The portfolio strategy API gateway is connected to the intelligent investment research and retrieval module through a microservice architecture, and uses a containerized deployment method to encapsulate the retrieval results into standardized investment advice output.

[0011] Furthermore, the relationship transmission coefficient calculation in the investment analysis enhancement module employs a multi-path aggregation graph attention network, which is invoked in real time when the intelligent investment research retrieval module executes a query. The relationship transmission coefficient is calculated in the following ways: using an attention mechanism to assign different importance weights to multiple relationship paths between entities; fusing prior weights of relationship types and semantic information of relationship embedding vectors; introducing graph structure information, and normalizing the associations generated by nodes with high heights through the in-degree and out-degree of entity nodes; applying an adaptive decay factor, dynamically adjusting according to the relationship path length and the decay coefficient learned by the Long Short-Term Memory network based on real-time market volatility, and outputting the result through a Sigmoid activation function.

[0012] Furthermore, the Alpha sentiment token generation in the investment analysis enhancement module employs a multi-scale temporal convolutional gating network. This token is periodically calculated and updated for key entities in the knowledge graph within the investment analysis enhancement module. The system continuously monitors multiple data sources, extracting news sentiment, research report opinions, and price momentum features for each entity in real time, and calculating the current AST value in the following ways: adjusting the dynamic weights of each data source in real time according to the current market state through a gating recurrent unit network, calculating the comprehensive score of multi-source sentiment features; introducing a temporal memory effect, fusing historical Alpha sentiment token values, and applying a time decay coefficient for weighted summation; performing layer normalization on the results to ensure the numerical stability of the output value, and appending the calculated AST value to the corresponding entity node in the knowledge graph.

[0013] The preferred technical solution is that the relation transmission coefficient calculation unit is specifically configured as follows: a relation weight optimization mechanism based on meta-learning is adopted, and the basic weight of relation type is dynamically calculated through a multilayer perceptron network. The input includes relation embedding vector, entity embedding vector and market state vector; a multi-path attention aggregation algorithm is used to comprehensively consider path length decay, node degree normalization and relation weight factor; an adaptive decay coefficient is introduced, and the decay parameter is dynamically adjusted by learning market fluctuation characteristics in real time through a long short-term memory network.

[0014] The preferred technical solution is that the Alpha sentiment token generation unit specifically includes: a multi-source sentiment feature extraction submodule, configured to extract sentiment features from news text, research report content, and market data; a dynamic weight adjustment submodule, which uses a reinforcement learning-based optimization algorithm to calculate the contribution weight of each data source through an attention mechanism; a temporal feature fusion submodule, which uses a gated recurrent unit network to fuse historical sentiment token values; and an output normalization submodule, which uses layer normalization technology to ensure the numerical stability of the output value.

[0015] The preferred technical solution is that the intelligent investment research and retrieval module is specifically configured as follows: it includes a query optimizer based on deep reinforcement learning, which uses a double-Q learning network to optimize the selection of query paths; it integrates a graph attention network for accurate localization and feature extraction of core entity nodes; it deploys a multi-hop inference engine to support complex multi-step reasoning queries in heterogeneous information networks; and it uses a path optimization algorithm based on policy gradients, combined with relationship transmission coefficients and sentiment token values ​​for result filtering.

[0016] The preferred technical solution is that the intelligent investment research and retrieval module further includes: a differentiable graph query engine that supports gradient-based query optimization and parameter learning; a query rewriting optimizer that automatically converts natural language queries into efficient graph query statements; a result diversity guarantee mechanism that avoids result centralization through the maximum boundary correlation algorithm; a query caching module that uses an LRU caching strategy to store frequently used query results; and a real-time query monitoring subsystem that provides query performance monitoring and optimization suggestions.

[0017] The preferred technical solution is that the portfolio strategy API gateway further includes: a risk correlation penetration analysis service, which uses graph convolutional networks to calculate the risk transmission matrix of the portfolio; a dynamic asset allocation engine, which is based on a mean-variance optimization model and introduces transaction cost constraints and liquidity constraints; a stress testing subsystem, which uses Monte Carlo simulation to generate risk assessments under extreme market scenarios; and a real-time risk monitoring module, which calculates portfolio risk indicators in real time through stream processing technology.

[0018] The preferred technical solution is that the risk correlation penetration analysis service specifically includes: a graph neural network computing unit configured to learn the complex risk transmission relationships between entities in the portfolio; a multi-factor risk model integrating market risk factors, credit risk factors, and liquidity risk factors; a risk transmission path analyzer to identify key risk transmission paths and systemically important nodes; a risk value calculation engine that uses historical simulation and parametric methods to calculate the risk value of the portfolio; and a sensitivity analysis module to assess the marginal impact of each risk factor on the portfolio value.

[0019] The preferred technical solution is that the multi-source data access and preprocessing module further includes: a BERT-based event extractor that uses a multi-head attention mechanism to identify investment event trigger words and arguments; an event coreference resolution module that uses graph convolutional networks to solve the event referencing problem; a time series data alignment engine that uses a dynamic time warping algorithm to process multi-source time series data; a data quality monitoring platform that monitors data quality indicators in real time and generates quality reports; and a metadata management system that maintains the metadata information and data lineage of the data sources.

[0020] Compared with existing technologies, the financial knowledge graph data fusion and retrieval platform system for investment advice of this invention has the following technical advantages:

[0021] 1. This application discloses a financial knowledge graph data fusion and retrieval platform system for investment advice, including an investment analysis enhancement module. This module interacts with the knowledge graph construction module via a distributed message queue. It uses a multi-scale temporal convolutional network to dynamically calculate and attach investment analysis coefficients, including relationship transmission coefficients and Alpha sentiment tokens, to entities and relationships in the knowledge graph. This module dynamically calculates the investment analysis coefficients of entities and relationships through a multi-scale temporal convolutional network, achieving a precise characterization of market risk and return characteristics. Relationship transmission coefficients quantify the risk transmission strength between related assets, aiding in dynamic hedging and risk control of investment portfolios; Alpha sentiment tokens can capture market sentiment fluctuations in real time, generating differentiated signals for quantitative strategies. Combined with the efficient interaction of the distributed message queue, this module significantly improves the real-time performance and accuracy of investment decisions, making it particularly suitable for high-frequency trading and multi-factor strategy optimization.

[0022] 2. An intelligent investment research retrieval module is employed, connected to the investment domain knowledge graph construction module and the investment analysis enhancement module. This module uses a Transformer-based semantic parser to receive natural language queries from the investment domain, generates graph query commands through a graph query optimization algorithm, and utilizes a reinforcement learning-based multi-objective optimization algorithm to reason, filter, and rank the query results. The Transformer-based semantic parser accurately converts natural language queries into graph query commands. Through graph optimization algorithms and reinforcement learning-driven multi-objective ranking, it achieves intelligent filtering and reasoning of cross-market, multi-dimensional data. This module not only supports structured queries but also dynamically integrates real-time market data, public opinion data, and fundamental data, providing investment researchers with decision support that combines depth and timeliness, significantly reducing information retrieval costs and improving strategy generation efficiency. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be arbitrarily combined with each other.

[0024] The following details the financial knowledge graph data fusion and retrieval platform system used for investment advice.

[0025] The financial knowledge graph data fusion and retrieval platform system for investment advice includes: a multi-source data access and preprocessing module, configured to acquire raw data from multiple heterogeneous financial data sources, and to automatically clean, standardize, and extract entities from the raw data using deep learning-based multimodal data fusion technology, generating standardized data stored in a distributed columnar database; the heterogeneous financial data sources include macroeconomic indicator databases, corporate financial report databases, real-time market data streams, news and public opinion text data, and analyst research reports.

[0026] The knowledge graph construction module for the investment domain is connected to the data access and preprocessing module via a high-speed data bus. It is used to construct and store a knowledge graph containing entities, relationships, and attributes based on a predefined investment domain ontology model and through entity linking, relation extraction, and attribute alignment operations based on graph neural networks.

[0027] The investment analysis enhancement module interacts with the knowledge graph construction module through a distributed message queue. It uses a multi-scale temporal convolutional network to dynamically calculate and attach investment analysis coefficients to entities and relationships in the knowledge graph, including relationship transmission coefficients and Alpha sentiment tokens.

[0028] The intelligent investment research and retrieval module is connected to the investment field knowledge graph construction module and the investment analysis enhancement module. It uses a semantic parser based on the Transformer architecture to receive natural language queries in the investment field, generates graph query commands through a graph query optimization algorithm, and uses a multi-objective optimization algorithm based on reinforcement learning to reason, filter, and sort the query results.

[0029] The portfolio strategy API gateway is connected to the intelligent investment research and retrieval module through a microservice architecture, and uses a containerized deployment method to encapsulate the retrieval results into standardized investment advice output.

[0030] The specific implementation of the multi-source data access and preprocessing module is as follows: In its deployment, this module uses Apache Kafka as a distributed message queue to build a unified data access platform. For different data sources, dedicated connectors are deployed: For structured data such as macroeconomic indicator databases and corporate financial report databases, a JDBC connector is used for periodic polling or listening to CDC logs to incrementally acquire data changes. For real-time market data streams, a high-performance TCP / UDP listening service is deployed to parse the FIX protocol or a custom binary protocol, and Apache Flink is used for preliminary streaming cleaning and window aggregation. For unstructured data such as news and public opinion text data and analyst research reports, a web crawler cluster (such as a Scrapy distributed cluster) and API calling clients are deployed, and APIs from major financial data providers are connected.

[0031] The data cleaning process employs deep learning-based anomaly detection algorithms. Specifically, for numerical data, such as financial indicators and market data, an autoencoder-based reconstruction error is used to identify outliers; for text data, a pre-trained language model, such as BERT, is used to detect semantic anomalies or fake news.

[0032] Entity extraction processing employs a hybrid neural network model based on BERT-BiLSTM-CRF. First, BERT is used to obtain dynamic word vectors from the text, fully capturing contextual semantic information. Then, a bidirectional LSTM is used to capture long-distance dependencies in the text. Finally, a CRF layer is used for sequence labeling to identify financial entities in the text with high accuracy, such as company names, person names, product names, and event names. This model is fine-tuned on labeled financial text corpora.

[0033] The generated standardized data is stored in distributed columnar databases such as Apache Cassandra or Apache HBase. These databases offer excellent scalability and can efficiently store and query massive amounts of time-series and wide-table data. During storage, data is partitioned according to themes such as market data, financial reports, and news, and secondary indexes are created for frequently queried fields such as stock codes and timestamps to optimize query performance.

[0034] The specific implementation of the investment domain knowledge graph construction module: This module communicates asynchronously with the preprocessing module via a high-speed data bus, such as Apache Pulsar or RabbitMQ, to receive standardized data. The predefined investment domain ontology model is constructed using the OWL language and managed using tools such as Protégé. This ontology explicitly defines core entity types and their object attributes, such as Company, Person, Industry, Financial Product, and MacroEconomicIndicator, for example, hasCEO (company and person), belongsTo (company and industry), isSupplierOf (company and company), and isAffectedBy (company and macroeconomic indicator). It also defines data attributes to describe entity characteristics, such as companyName, revenue, and P / E ratio.

[0035] The entity linking operation based on graph neural networks is as follows: First, using knowledge representation learning algorithms such as TransE, TransR, or the more advanced RotatE, existing entities and relations in the knowledge graph are mapped to a low-dimensional vector space. Then, for a newly extracted entity, the semantic similarity between its context embedding vector and candidate entity vectors in the graph is calculated. Finally, a ranking model, such as LightGBM or a neural network, is used to integrate features like semantic similarity, entity popularity, and context matching. The mention is then linked to the most correct entity in the knowledge graph. If a link to an existing entity cannot be established, a new entity creation process is triggered.

[0036] Relation extraction employs a joint learning framework based on pre-trained language models, such as ERNIE and CPM. Entity recognition and relation classification are performed simultaneously. The model input is a pair of entities and their context, and the output is the type of relation between the entities. Training corpora are automatically generated using structured databases through remote supervision methods, such as supply chain databases and equity relationship databases, significantly expanding the scale of the training data.

[0037] Attribute alignment is an operation that addresses attribute values ​​describing the same entity from different data sources. For example, a company's market capitalization might be 100 billion from supplier A and 105 billion from supplier B. This operation uses a structured similarity calculation model based on the Siamese network, comprehensively comparing factors such as numerical differences, the credibility of the data source, and the data update time, ultimately generating a unified attribute value that resolves conflicts.

[0038] The completed knowledge graph is stored in native graph databases such as Neo4j, Nebula Graph, or JanusGraph. During storage, indexes and materialized views are created for frequently traversed relation types and query patterns to accelerate subsequent graph queries and analysis.

[0039] The relationship transmission coefficient calculation in the investment analysis enhancement module employs a multi-path aggregation graph attention network, which is invoked in real-time during query execution by the intelligent investment research retrieval module. The relationship transmission coefficient is calculated as follows: an attention mechanism is used to assign different importance weights to multiple relationship paths between entities; prior weights of relationship types and semantic information of relationship embedding vectors are fused; graph structure information is introduced, and the associations generated by nodes with high height are normalized using the in-degree and out-degree of entity nodes; an adaptive decay factor is applied, dynamically adjusted based on the relationship path length and the decay coefficient learned by the Long Short-Term Memory network from real-time market volatility, and the result is output through a Sigmoid activation function. The formula for calculating the relationship transmission coefficient is:

[0040] Where σ represents the Sigmoid activation function, k represents the k-th relation path, and α k The path attention weights are represented by an attention mechanism α, which comprehensively considers all possible relationship chains between entities. k Assign different importance weights to each path;

[0041] The numerator of the formula Incorporating prior weights of relation types Semantic information φ(r) of relation embedding vectors i ), denominator This introduces graph structure information to normalize the associations generated by the height-number nodes, d in (e i ) and d out (e j ) represent in-degree and out-degree, respectively;

[0042] As an adaptive attenuation factor, It will depend on the path length l k The adaptive decay coefficient λ is dynamically adjusted, and λ is learned by the LSTM network based on real-time market volatility. This formula achieves "quantitative calculation and dynamic optimization of association strength." It transforms the "knowledge graph" from a static relational database: the formula gives the "edges" in the graph dynamically changing, quantifiable weights (RTC values), rather than simple 0 or 1. This allows the graph to evolve into a weighted dynamic network. It realizes the core function of the "investment analysis enhancement" module: the module is "enhanced" precisely because this formula adds the key investment analysis coefficient RTC to entities and relationships. It provides a ranking basis for "intelligent retrieval": this gives the ranking function of the "intelligent investment research retrieval module" a scientific and quantitative basis, arranging data in descending order of RTC value, rather than simply ranking based on keyword matching degree.

[0043] This module interacts with the graph construction module via a distributed message queue, such as Redis Streams or Kafka, to listen for incremental update events in the graph. The Relationship Transmission Coefficient (RTC) is calculated using a multi-path aggregation graph attention network. The specific implementation steps are as follows: Path Sampling: For a given entity pair (e1, e2), a biased random walk algorithm is used to sample isocentrality indicators in the knowledge graph to correct for biases, preventing "large companies" from dominating all associations. Adaptive Decay: λ is a key parameter in the path decay factor. This system uses a lightweight LSTM network to dynamically learn the value of λ. The input to this LSTM is real-time market volatility indices such as the VIX index, industry sector volatility, and macroeconomic uncertainty indices. In this way, the RTC coefficient can adaptively adjust according to the overall market risk situation, focusing more on direct strong associations during periods of high volatility and appropriately considering longer-term indirect associations during periods of low volatility.

[0044] Attention aggregation and normalization: After the scores of all paths have been calculated, an attention mechanism is used to assign a weight α to each path. k Finally, the weighted sums are processed using the Sigmoid function to obtain the final RTC value between 0 and 1. The calculation of attention weights takes into account the semantic information of the path itself and the current query context.

[0045] The specific implementation of the relation transmission coefficient calculation unit: The core of the relation weight optimization mechanism based on meta-learning is a meta-learner, which adopts a model-independent meta-learning (MAML) framework. Its goal is to learn a weight initialization strategy, enabling the system to quickly adjust relation weights when facing new relation types or new market paradigms. .

[0046] Input: The input to the meta-learner includes: 1) relation embedding vector φ(r) i 1) Knowledge representation learning; 2) Connected entity embedding vectors; 3) Market state vectors (volatility, trend, etc.).

[0047] Process: Internally, a multilayer perceptron (MLP) is used to map the above inputs to relational weights. The parameters of this MLP are trained on various market scenarios (tasks) through meta-learning to obtain a well-initialized set of parameters. When the market environment changes, only a small number of samples (e.g., which relationships are more effective recently) are needed to fine-tune the MLP parameters, thereby quickly generating relationship weights that adapt to the new environment.

[0048] Output: Dynamically calculated basic weights of the relation type .

[0049] The multi-path attention aggregation algorithm uses a variant of the Graph Attention Network (GAT). Instead of simply considering one-hop neighbors, it computes attention across multiple sampled paths. For each path, an importance score is calculated using a neural network, and the final representation of a path is the result of aggregating the embeddings of all relations along that path using GAT. Finally, the representations of all paths are aggregated again to obtain a comprehensive association representation.

[0050] The dynamic adjustment of the adaptive decay coefficient is implemented using an LSTM network. This LSTM takes market volatility sequences and macroeconomic uncertainty indices as inputs, and its last hidden state is mapped through a fully connected layer, outputting a scalar value as λ. This LSTM network is trained on historical data, with the goal of maximizing the performance of the RTC coefficient in predicting future stock price correlations by adjusting λ.

[0051] Furthermore, the Alpha sentiment token generation in the investment analysis enhancement module employs a multi-scale temporal convolutional gating network. This token is periodically calculated and updated for key entities in the knowledge graph within the investment analysis enhancement module. The system continuously monitors multiple data sources, extracting news sentiment, research report opinions, and price momentum features for each entity in real time, and calculating the current AST value in the following ways: adjusting the dynamic weights of each data source in real time according to the current market state through a gating recurrent unit network, calculating the comprehensive score of multi-source sentiment features; introducing a temporal memory effect, fusing historical Alpha sentiment token values, and applying a time decay coefficient for weighted summation; performing layer normalization on the results to ensure the numerical stability of the output value, and appending the calculated AST value to the corresponding entity node in the knowledge graph.

[0052] The formula for calculating AST is:

[0053]

[0054] in, The numerical value represents the result of multi-source information fusion, expressed through dynamic weighting. Sentiment characteristics from different data sources The overall score; This represents the dynamic weight of the s-th data source at time t. The GRU network adjusts in real time based on the current market conditions;

[0055] This represents the numerical value of the temporal memory effect, where T represents the total length of the retracement period. The Alpha sentiment token value representing an entity at a historical moment (t-τ), The time decay coefficient is represented by LayerNorm, indicating layer normalization. This formula achieves "fusion and temporal modeling of multi-source heterogeneous sentiment information," deeply integrating the value of "multi-dimensional data." This formula dynamically weights and merges three distinct data sources—news (text), research reports (text), and market data (numerical data)—into a unified, comparable quantitative indicator, truly realizing multimodal data fusion. It also endows the system with "market memory" through a time-series memory effect term. The system possesses "experience" similar to that of human investors, understanding that emotions persist and diminish, thus avoiding drastic jumps in emotion scores caused by data fluctuations, resulting in smoother and more reasonable output. It provides an emotion dimension to "investment recommendations": the recommendations output by the portfolio strategy API gateway are not only based on fundamental correlation (RTC) but also overlaid with market sentiment (AST) dimensions, making the recommendations more comprehensive. For example, it recommends targets with "strong correlation and positive sentiment." This system quantifies and deepens analytical capabilities through two core formulas, and constructs an end-to-end intelligent closed loop from data access to investment decision-making through the organic linkage of five modules, ultimately achieving the core technological effect of improving the depth, accuracy, and intelligence level of investment analysis.

[0056] The Alpha Emotional Token (AST) is generated using a multi-scale temporal convolutional gating network. The specific process is as follows:

[0057] Multi-source feature extraction: News sentiment Using a BERT model pre-trained in the financial field, sentiment analysis is performed on news headlines and body text, outputting a sentiment score between [-1, 1].

[0058] Research report viewpoints We use a text classification model to analyze research report summaries and investment recommendations, such as "Buy," "Neutral," and "Sell," and map them to quantitative scores. We also extract the percentage difference between the target price and the current price as another feature.

[0059] Price Momentum Calculate a series of technical indicators as features, such as the past N-day return, the Relative Strength Index (RSI), and the proportion of abnormal trading volume.

[0060] Dynamic weight adjustment Dynamic weights It is generated by a gated recurrent unit (GRU) network. The input of the GRU is the current market state vector, such as market volatility, industry price fluctuations, and fund flows. The hidden states are used to calculate attention weights, thereby determining which data source's sentiment characteristics are more trustworthy at the current moment.

[0061] Temporal memory fusion: Fusion weights of historical AST values Using exponential decay, such as = γ^τ (0<γ<1). The value of γ can be adjusted according to the characteristics of the industry in which the entity is located. For example, for the rapidly changing technology industry, a smaller value of γ makes historical sentiment decay faster; for the stable public utility industry, a larger value of γ makes sentiment last longer.

[0062] Layer Normalization: Finally, layer normalization (LayerNorm) is applied to the weighted summation result to ensure that the output AST value remains stable even if the distribution of the input features changes, facilitating comparisons between different entities and at different time points. The calculated RTC and AST coefficients are updated as attributes to the corresponding nodes and edges in the knowledge graph in real time.

[0063] Specific implementation of the Alpha sentiment token generation unit: Multi-source sentiment feature extraction submodule: For news and research report texts, a RoBERTa-based sequence classification model is used for sentiment analysis. This model is fine-tuned on a large number of financial news and research report texts to understand specific expressions and implicit sentiments in the financial field.

[0064] For price momentum features, a series of standardized indicators, such as MACD, Bollinger Bands %B, and momentum values, are calculated using technical analysis libraries like ta-lib, and then concatenated into a feature vector. The dynamic weight adjustment submodule employs a reinforcement learning-based optimization algorithm. It treats the allocation of data source weights as a decision-making process.

[0065] State: Current market environment, recent AST value of the entity, and recent confidence levels of various data sources, measured by the correlation between predicted and actual returns.

[0066] Action: Assign a weight to each data source. .

[0067] Reward: The higher the correlation between the AST value calculated after weighting and future short-term stock price returns, the greater the reward. A neural network policy function is trained using a policy gradient method, such as the REINFORCE algorithm, taking the state as input and outputting weighted actions.

[0068] The temporal feature fusion submodule uses a gated recurrent unit (GRU) network. The input to the GRU is the multi-source fused features at the current time step (…). The hidden state is used to integrate historical information. The gating mechanism of GRU can learn which historical information is important and needs to be retained, and which is outdated and needs to be forgotten, thus effectively integrating the information of the historical sentiment token value AST(t-τ).

[0069] The output normalization submodule employs layer normalization. It normalizes all features of a single sample to a mean of 0 and a variance of 1. This effectively addresses the internal covariate shift problem, making deep network training more stable and the output AST values ​​more reliable.

[0070] The intelligent investment research and retrieval module also includes: a differentiable graph query engine, supporting gradient-based query optimization and parameter learning; a query rewriting optimizer, automatically converting natural language queries into efficient graph query statements; a result diversity guarantee mechanism, avoiding result centralization through the maximum boundary correlation algorithm; a query caching module, using an LRU caching strategy to store frequently used query results; and a real-time query monitoring subsystem, providing query performance monitoring and optimization suggestions. The specific implementation of the intelligent investment research and retrieval module: The query optimizer based on deep reinforcement learning models the process of converting natural language queries into graph queries as a sequential decision problem. The agent needs to select traversal paths step by step on the knowledge graph. State: The current entity node, the traversed paths, and the semantic representation of the query. Action: Starting from the current node, select the next relation direction or entity to traverse. Reward: The degree of matching between the final retrieved entity list and the user's actual expected results, and the query time. A double-Q learning algorithm is used to train this agent to avoid overestimating the Q-value. The trained agent can quickly plan efficient retrieval paths for new queries.

[0071] Graph Attention Networks (GATs) are used for precise localization and feature extraction of core entity nodes. When parsing natural language queries, the initial entities mentioned in the query are first encoded using GATs. By calculating their attention weights with other entities in the graph, it is initially determined which adjacent entities and relationships are most relevant to the current query, providing guidance for subsequent path searching.

[0072] The multi-hop inference engine is based on subgraph matching and rule-based reasoning. It supports predefined investment logic rules, such as, "If a company is a supplier of another company, and the latter releases a new product, the former may benefit." The system matches queries against these rules, triggering corresponding multi-hop inference processes to discover deeper connections.

[0073] The policy gradient-based path optimization algorithm is similar to the reinforcement learning query optimizer, but it focuses more on using RTC and AST coefficients to filter and sort intermediate results in real time on a determined retrieval path. It learns a policy function that decides whether to continue deeper traversal or backtrack based on the current path's RTC strength, the entity's AST value, and the distance to the target, ultimately outputting a list of entities sorted by estimated relevance.

[0074] The intelligent investment research and retrieval module's additional component implementation: the differentiable graph query engine, is a cutting-edge exploratory component. It attempts to represent the entire knowledge graph query process as a differentiable computational graph. This allows for the use of gradient descent to optimize query parameters and even learns from user feedback how to better interpret query intent. The specific implementation may be based on frameworks such as NeuSL.

[0075] The query rewrite optimizer is based on rules and machine learning. It includes a rule base for common query equivalence transformations, such as early filtering and selection pushdown. Simultaneously, it uses a machine learning model to predict the cost of different query execution plans. This model is trained by collecting historical query performance data to select the execution plan with the lowest cost.

[0076] The result diversity guarantee mechanism employs the maximum boundary relevance algorithm. This algorithm maximizes the diversity between result sets while ensuring result relevance. When ranking candidate results, it considers not only their relevance score to the query but also their similarity to the already selected result set, prioritizing entities with high relevance that are different from existing results.

[0077] The query caching module uses an in-memory database such as Redis to implement an LRU cache. The cache key is the semantic hash value of the query statement, and the cache value is the serialized data of the retrieval result. Simultaneously, a cache invalidation mechanism is established so that when the underlying data of the knowledge graph is updated, the relevant cache entries are automatically invalidated.

[0078] The real-time query monitoring subsystem is built on monitoring libraries such as Micrometer. It collects metrics for each query, such as response time, number of results, cache hit rate, and the time consumed by each component. These metrics are then visualized using tools like Grafana, and alert rules are set to provide warnings and logs for abnormal queries.

[0079] The portfolio strategy API gateway also includes: a risk correlation penetration analysis service, which uses graph convolutional networks to calculate the risk transmission matrix of the portfolio; a dynamic asset allocation engine, based on a mean-variance optimization model, incorporating transaction cost constraints and liquidity constraints; a stress testing subsystem, which uses Monte Carlo simulation to generate risk assessments under extreme market scenarios; and a real-time risk monitoring module, which calculates portfolio risk indicators in real time through stream processing technology. The specific implementation of the portfolio strategy API gateway is as follows: The risk correlation penetration analysis service uses graph convolutional networks to calculate risk transmission. The assets held in the portfolio are considered as seed nodes in the graph. Through message passing on the knowledge graph using GCN, risks such as credit risk and liquidity risk are simulated, showing how they propagate and amplify through these relationships within the network. Ultimately, the infection risk value of each asset and the systemic risk exposure of the entire portfolio are calculated.

[0080] The dynamic asset allocation engine is an extension of the Black-Litterman model. This engine not only considers historical returns and covariance matrices derived from the mean-variance model, but also incorporates insights from knowledge graph analysis as input. For example, based on RTC and AST, it identifies industries with strong correlations or positive sentiment. Using Bayesian methods, it combines market equilibrium returns with these subjective opinions to generate new expected returns, which are then used to optimize asset allocation. The optimization process strictly incorporates transaction cost models and liquidity constraints, such as maximum position sizes and trading volume limits.

[0081] The stress testing subsystem employs Monte Carlo simulation. It first simulates thousands of possible market scenarios based on historical data and models such as GARCH, including sharp interest rate increases and stock market crashes. Then, under these extreme scenarios, it re-evaluates the value of all assets in the portfolio and considers the risk transmission effects revealed by the knowledge graph, calculating the portfolio's potential loss distribution to obtain the value at risk and expected deficit under these extreme scenarios.

[0082] The real-time risk monitoring module is built on the Apache Flink stream processing engine. It continuously monitors real-time market data, news feeds, and knowledge graph update events. For the monitored portfolio, it calculates its risk indicators in real time, such as Greeks like Delta, Gamma, and Vega, and Value at Risk (VaR). It also sets risk thresholds through a rules engine, triggering alerts or automatic position liquidation if these thresholds are exceeded.

[0083] The risk correlation penetration analysis service specifically includes: a graph neural network (GNN) computing unit configured to learn complex risk transmission relationships between entities in the portfolio; a multi-factor risk model integrating market risk factors, credit risk factors, and liquidity risk factors; a risk transmission path analyzer to identify key risk transmission paths and systemically important nodes; a risk value calculation engine that uses historical simulation and parametric methods to calculate the portfolio's risk value; and a sensitivity analysis module to assess the marginal impact of each risk factor on the portfolio value. The specific implementation of the risk correlation penetration analysis service: The GNN computing unit typically employs inductive learning algorithms such as GraphSAGE or GAT. These algorithms can learn the embedding representation of each node, which aggregates information from its multi-hop neighbors. In the risk transmission task, the node's label can be its historical risk event occurrence rate, and the GNN can predict the risk score of each node under the current network structure through supervised learning.

[0084] Multi-factor risk models combine traditional risk factors, such as the Fama-French three-factor model (market, market capitalization, and valuation), with knowledge graph factors. For example, a "supply chain concentration" factor is defined to calculate the RTC-weighted sum of a company's customer concentration; a "market sentiment beta" factor is defined to measure the sensitivity of a company's stock price to changes in the sentiment of a group of entities associated with the knowledge graph. Panel data regression analysis is used to assess the explanatory power of these new factors on returns and incorporate them into the risk model.

[0085] The risk transmission path analyzer is based on a modified Dijkstra's algorithm or the K-shortest path algorithm. It searches the knowledge graph for all important paths connecting two risk events or entities and sorts the risk transmission strength of the paths according to the RTC value of the relationships along the paths. This identifies the most critical risk transmission paths and nodes with high systemic importance in the network, whose failure would have a significant impact.

[0086] The Value at Risk (VaR) calculation engine offers several methods: Historical Simulation: Based on historical data, it calculates the daily profit and loss distribution of the portfolio over a past period, then takes the quantile at a given confidence level as VaR. Parametric Method: Assuming returns follow a normal distribution, it calculates VaR using the portfolio's variance and covariance matrices. The innovation of this system lies in the fact that the estimation of the covariance matrix is ​​not only based on historical price data but also incorporates the strength of inter-entity associations calculated from the knowledge graph as prior information, adjusting the covariance matrix to more quickly capture structural risk changes. The sensitivity analysis module calculates the gradient of VaR with respect to various risk factors. For example, it calculates the sensitivity of VaR to changes in a certain industry index or to changes in the strength of a certain association in the knowledge graph. This helps risk managers understand the marginal contribution of different sources of risk.

[0087] The multi-source data access and preprocessing module also includes: a BERT-based event extractor that uses a multi-head attention mechanism to identify investment event trigger words and arguments; an event coreference resolution module that uses graph convolutional networks to solve the event referencing problem; a time-series data alignment engine that uses a dynamic time warping algorithm to process multi-source time-series data; a data quality monitoring platform that monitors data quality indicators in real time and generates quality reports; and a metadata management system that maintains metadata information and data lineage of data sources. The additional components of the multi-source data access and preprocessing module are implemented as follows: The BERT-based event extractor uses a pipeline approach. First, the BERT-CRF model is used to identify event trigger words and label argument roles. Then, a classifier is used to determine the event type composed of trigger words and arguments, such as "merger and acquisition" or "profit warning." To address the complexity of financial events, the model needs to identify nested events and multi-argument events, such as the acquirer, acquired party, acquisition amount, and time in a merger and acquisition event.

[0088] The event coreference resolution module uses a graph convolutional network. It treats all event mentions in a document as nodes in a graph. The edges between nodes are constructed from various features, such as semantic similarity, temporal relationships, and syntactic relationships. GCN aggregates neighbor information through message passing, learns a rich representation for each node, and finally determines which mentions point to the same event in the real world based on the similarity of the node representations.

[0089] The time-series data alignment engine processes data with different frequencies and timestamps. For daily financial report data and second-level market data, it employs dynamic time warping algorithms or higher-order neural network time alignment models to find the optimal alignment path between different time series, thus ensuring that data at the correct time points can be matched during analysis. The data quality monitoring platform defines a set of data quality dimensional indicators, including completeness (whether there are missing data), accuracy (whether the values ​​are correct), consistency (whether multi-source data is contradictory), timeliness (whether the data is timely), and validity (whether the format is correct). The platform runs quality check jobs regularly, generates quality reports and dashboards, and triggers alarms for events that deteriorate in quality.

[0090] Metadata management systems utilize open-source frameworks such as Apache Atlas or LinkedIn DataHub. They crawl metadata from all data sources, including data structure, schema, lineage, and owner information. Data lineage records the cleaning, transformation, and fusion operations performed on the raw data to ultimately generate a specific entity or attribute in the knowledge graph. This is crucial for data traceability, impact analysis, and compliance auditing.

[0091] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that an article or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such an article or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of other identical elements in the article or apparatus that includes that element.

[0092] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. The present invention has been described in detail with reference to preferred embodiments. Those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications and substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A financial knowledge graph data fusion and retrieval platform system for investment advice, characterized in that: include: The multi-source data access and preprocessing module is configured to acquire raw data from multiple heterogeneous financial data sources, and use deep learning-based multimodal data fusion technology to automatically clean, standardize, and extract entities from the raw data, generating standardized data stored in a distributed columnar database; the heterogeneous financial data sources include macroeconomic indicator databases, corporate financial report databases, real-time market data streams, news and public opinion text data, and analyst research reports. The knowledge graph construction module for the investment field is connected to the data access and preprocessing module via a high-speed data bus. It is used to construct and store a knowledge graph containing entities, relationships and attributes based on a predefined investment field ontology model and through entity linking, relation extraction and attribute alignment operations based on graph neural networks. The investment analysis enhancement module interacts with the knowledge graph construction module through a distributed message queue. It uses a multi-scale temporal convolutional network to dynamically calculate and append investment analysis coefficients to entities and relationships in the knowledge graph, including relationship transmission coefficients and Alpha sentiment tokens. The Alpha sentiment token calculates the current AST value as follows: a gated recurrent unit network dynamically adjusts the weights of each data source in real time according to the current market state, calculating the comprehensive score of multi-source sentiment features; a temporal memory effect is introduced, fusing historical Alpha sentiment token values ​​and applying a time decay coefficient for weighted summation; layer normalization is performed on the results to ensure the numerical stability of the output value, and the calculated AST value is appended to the corresponding entity node in the knowledge graph. The intelligent investment research and retrieval module is connected to the investment field knowledge graph construction module and the investment analysis enhancement module. It uses a semantic parser based on the Transformer architecture to receive natural language queries in the investment field, generates graph query commands through a graph query optimization algorithm, and uses a multi-objective optimization algorithm based on reinforcement learning to reason, filter and sort the query results. The portfolio strategy API gateway is connected to the intelligent investment research and retrieval module through a microservice architecture, and uses a containerized deployment method to encapsulate the retrieval results into standardized investment advice output. The relationship transmission coefficient calculation in the investment analysis enhancement module employs a multi-path aggregation graph attention network, which is invoked in real-time during query execution by the intelligent investment research retrieval module. The relationship transmission coefficient is calculated as follows: an attention mechanism is used to assign different importance weights to multiple relationship paths between entities; prior weights of relationship types and semantic information of relationship embedding vectors are fused; graph structure information is introduced, and the associations generated by nodes with high height are normalized using the in-degree and out-degree of entity nodes; an adaptive decay factor is applied, dynamically adjusted based on the relationship path length and the decay coefficient learned by the Long Short-Term Memory network from real-time market volatility, and the result is output through a Sigmoid activation function.

2. The system according to claim 1, characterized in that, The Alpha sentiment token generation in the investment analysis enhancement module employs a multi-scale temporal convolutional gating network. This token is periodically calculated and updated for key entities in the knowledge graph within the investment analysis enhancement module. The system continuously monitors diverse data sources and extracts news sentiment, research report opinions, and price momentum features for each entity in real time.

3. The system according to claim 2, characterized in that, The specific configuration of the relation transmission coefficient calculation unit is as follows: a relation weight optimization mechanism based on meta-learning is adopted, and the basic weight of relation type is dynamically calculated through a multilayer perceptron network. The input includes relation embedding vector, entity embedding vector and market state vector; a multi-path attention aggregation algorithm is used to comprehensively consider path length decay, node degree normalization and relation weight factor; an adaptive decay coefficient is introduced, and the decay parameter is dynamically adjusted by learning market fluctuation characteristics in real time through a long short-term memory network.

4. The system according to claim 3, characterized in that, The Alpha sentiment token generation unit specifically includes: a multi-source sentiment feature extraction submodule, configured to extract sentiment features from news text, research report content and market data; The dynamic weight adjustment submodule uses a reinforcement learning-based optimization algorithm to calculate the contribution weight of each data source through an attention mechanism. The temporal feature fusion submodule uses a gated recurrent unit network to fuse historical sentiment token values; the output normalization submodule uses layer normalization technology to ensure the numerical stability of the output values.

5. The system according to claim 1, characterized in that, The intelligent investment research and retrieval module is specifically configured as follows: it includes a query optimizer based on deep reinforcement learning, which uses a double-Q learning network to optimize the selection of query paths; it integrates a graph attention network for accurate localization and feature extraction of core entity nodes; it deploys a multi-hop inference engine to support complex multi-step reasoning queries in heterogeneous information networks; and it uses a path optimization algorithm based on policy gradients, combined with relationship transmission coefficients and sentiment token values ​​for result filtering.

6. The system according to claim 5, characterized in that, The intelligent investment research and retrieval module also includes: a differentiable graph query engine that supports gradient-based query optimization and parameter learning; a query rewriting optimizer that automatically converts natural language queries into efficient graph query statements; a result diversity guarantee mechanism that avoids result centralization through the maximum boundary correlation algorithm; a query caching module that uses an LRU caching strategy to store frequently used query results; and a real-time query monitoring subsystem that provides query performance monitoring and optimization suggestions.

7. The system according to claim 1, characterized in that, The portfolio strategy API gateway also includes: a risk correlation penetration analysis service that uses graph convolutional networks to calculate the risk transmission matrix of the portfolio; a dynamic asset allocation engine that uses a mean-variance optimization model and introduces transaction cost constraints and liquidity constraints; a stress testing subsystem that uses Monte Carlo simulation to generate risk assessments under extreme market scenarios; and a real-time risk monitoring module that uses stream processing technology to calculate portfolio risk indicators in real time.

8. The system according to claim 7, characterized in that, The risk correlation penetration analysis service specifically includes: a graph neural network computing unit configured to learn the complex risk transmission relationships between entities in the portfolio; a multi-factor risk model integrating market risk factors, credit risk factors, and liquidity risk factors; a risk transmission path analyzer to identify key risk transmission paths and systemically important nodes; a risk value calculation engine that uses historical simulation and parametric methods to calculate the risk value of the portfolio; and a sensitivity analysis module to assess the marginal impact of each risk factor on the portfolio value.

9. The system according to claim 1, characterized in that, The multi-source data access and preprocessing module also includes: a BERT-based event extractor that uses a multi-head attention mechanism to identify investment event trigger words and arguments; an event coreference resolution module that uses graph convolutional networks to solve the event referencing problem; a time-series data alignment engine that uses a dynamic time warping algorithm to process multi-source time-series data; a data quality monitoring platform that monitors data quality indicators in real time and generates quality reports; and a metadata management system that maintains the metadata information and data lineage of the data sources.

Images