A financial information aggregation recommendation method and system based on unsupervised cross-modal learning

Abstract

This invention belongs to the field of computer technology and discloses a financial information aggregation and recommendation method and system based on unsupervised cross-modal learning. It constructs a multi-channel data collection crawler engine to dynamically acquire heterogeneous data from cross-platform data sources; deploys a heterogeneous dual-engine architecture of semantic and numerical engines to perform semantic parsing and extract transaction features from the acquired heterogeneous data; constructs a cross-modal dynamic alignment matrix to generate a joint representation vector; acquires real-time search keywords and click behaviors to form short-term intent, analyzes historical cross-platform behavior trajectories to form long-term preferences, and constructs a two-layer intent recognition network based on short-term intent and long-term preferences; and outputs dynamic recommendation results integrating business objectives and user satisfaction indicators. This invention, based on unsupervised cross-modal learning technology, achieves cross-platform multi-channel financial information aggregation, deep user intent analysis, and personalized recommendation, thereby solving the problems existing in the prior art.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, specifically relating to a financial information aggregation and recommendation method and system based on unsupervised cross-modal learning. Background Technology

[0002] In financial information recommendation technology, existing financial platforms (such as banking systems, social media, and news websites) employ independent data structures, making it difficult to aggregate cross-platform information and form a comprehensive user profile. For example, the integration efficiency of transaction data (structured) and user comments (unstructured) is low, affecting real-time recommendation performance. Traditional recommendation systems often rely on single-modal data (such as text or numerical values) and lack the ability to jointly model multimodal data (text, images, transaction records). This results in the incomplete exploration of the correlation between textual and graphical information and user risk preferences in financial product recommendations, making it difficult to capture real-time shifts in user preferences and changes in channel weights.

[0003] To address the issues of data silos and difficulties in heterogeneous integration in existing financial information recommendation methods, which result in insufficient cross-modal intent parsing capabilities and poor dynamic adaptability, there is an urgent need for a financial information aggregation and recommendation method and system that integrates cross-platform heterogeneous data and cross-modal feature fusion to meet the application requirements of financial information recommendation. Summary of the Invention

[0004] In view of this, the purpose of this invention is to propose a financial information aggregation and recommendation method and system based on unsupervised cross-modal learning. Based on unsupervised cross-modal learning technology, it realizes cross-platform and multi-channel financial information aggregation, deep analysis of user intent, and personalized recommendation, so as to solve the above-mentioned problems existing in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] To achieve the above objectives, in a first aspect, the present invention provides a financial information aggregation and recommendation method based on unsupervised cross-modal learning, comprising the following steps:

[0007] Step S10: Build a crawler engine for multi-channel data collection and use dynamic crawlers to obtain heterogeneous data from cross-platform data sources;

[0008] Step S20: Deploy a heterogeneous dual-engine architecture of semantic engine and numerical engine to perform semantic parsing on the acquired heterogeneous data and extract transaction features;

[0009] Step S30: Construct a cross-modal dynamic alignment matrix, use the contrastive learning loss function to eliminate semantic bias between text descriptions, chart data and numerical sequences, map them to a unified semantic space and generate a joint representation vector;

[0010] Step S40: Obtain real-time search keywords and click behavior to form short-term intent, analyze historical cross-platform behavior trajectories to form long-term preferences, and construct a two-layer intent recognition network based on short-term intent and long-term preferences;

[0011] Step S50: Input the semantic parsing and transaction features of the heterogeneous data into a two-layer intent recognition network to capture the implicit relationships between financial entities, drive the generation of cross-cycle recommendation strategies, and output dynamic recommendation results that integrate business objectives and user satisfaction indicators.

[0012] As a further aspect of the present invention, step S10, when the dynamic crawler acquires heterogeneous data from a cross-platform data source, includes the following steps:

[0013] By using a web crawler engine to identify the structural features of the target website or platform in the data source, dynamic web crawlers are used to simulate user behavior, process dynamically loaded web pages, and extract heterogeneous data.

[0014] Remove HTML tags and duplicate content from the extracted heterogeneous data and convert the encoding format to structure the data into a standard format;

[0015] The processed heterogeneous data is stored, and an incremental crawling mechanism is set up to periodically update the heterogeneous data in the database based on the data update timestamp.

[0016] As a further aspect of the present invention, in step S20, the semantic engine uses a dynamically pruned semantic parsing model to process text and image data, and the numerical engine extracts transaction features through a temporal autoencoder.

[0017] As a further aspect of the present invention, the semantic engine uses a pre-trained BERT model from the financial domain as a semantic parsing model to perform text semantic parsing, and combines it with a ResNet-50 network to extract image features.

[0018] As a further aspect of the present invention, step S30 involves constructing a cross-modal dynamic alignment matrix and performing cross-modal dynamic alignment, including the following steps:

[0019] A multi-granularity attention mechanism is used for text modalities to identify the weight between each word or sentence in the text, obtain the text information most relevant to user needs, and classify them according to regulatory intensity;

[0020] A morphological parser is constructed for image modalities to extract the implicit patterns of technical indicators in candlestick charts. The features of the images are analyzed to extract technical indicators related to market trends.

[0021] We use a time-series causal discovery algorithm on numerical modes to explore the lagged relationship between user transaction behavior and macroeconomic indicators, as well as the time delay relationship between transaction behavior and economic variables.

[0022] By mapping features from different modalities in text, images, and numerical data to a unified semantic space, a cross-modal alignment matrix is ​​constructed to fuse the features of text, image, and numerical data.

[0023] As a further aspect of the present invention, step S40, which involves obtaining real-time search keywords and click behavior to constitute short-term intent and analyzing historical cross-platform behavior trajectories to constitute long-term preferences, includes the following steps:

[0024] The system obtains users' real-time search keywords through an integrated search engine, processes and cleans these keywords, and removes noisy information.

[0025] Information is extracted from real-time user click behavior data, recording the time, page, and clicked object of each click, and a click log is generated by tracking user click behavior.

[0026] Feature extraction is performed on real-time search keyword and click behavior data. Short-term intent features are extracted and intent classification is performed using a short-term intent recognition network.

[0027] Predict users' long-term preferences based on historical behavior trajectories, and model them using a long-term preference recognition network. Combine short-term intentions with long-term preferences to output intention recognition.

[0028] Secondly, this invention also provides a financial information aggregation and recommendation system based on unsupervised cross-modal learning, comprising the following components:

[0029] The data integration layer includes a multi-channel crawler engine and an incremental crawling mechanism, used to acquire heterogeneous data from cross-platform data sources and update the data sources regularly;

[0030] The data processing layer includes a semantic engine and a numerical engine. By deploying a heterogeneous dual-engine architecture of semantic engine and numerical engine, it performs semantic parsing on the acquired heterogeneous data and extracts transaction features.

[0031] The intelligent processing layer includes a heterogeneous computing module, a cross-modal dynamic alignment matrix, and a knowledge graph service layer. It is used to eliminate semantic biases between text descriptions, chart data, and numerical sequences by using a contrastive learning loss function, mapping them to a unified semantic space and generating joint representation vectors.

[0032] The intent recognition layer is used to obtain short-term intents by real-time search keywords and click behaviors, analyze historical cross-platform behavior trajectories to form long-term preferences, and construct a two-layer intent recognition network based on short-term intents and long-term preferences.

[0033] The service output layer is used to input the semantic parsing and transaction features of the heterogeneous data into a two-layer intent recognition network, capture the implicit relationships between financial entities, drive the generation of cross-cycle recommendation strategies, and output dynamic recommendation results that integrate business objectives and user satisfaction indicators.

[0034] As a further aspect of the present invention, the semantic engine is responsible for semantic parsing of text data, using the BERT model in the financial field for text semantic parsing, and combining it with the ResNet-50 network to extract features from image data; the numerical engine uses a temporal autoencoder to extract transaction features and analyze users' transaction behavior patterns.

[0035] As a further aspect of the present invention, the heterogeneous computing module includes an FPGA-accelerated semantic engine and a GPU cluster-supported numerical engine.

[0036] As a further aspect of the present invention, the cross-modal dynamic alignment matrix includes:

[0037] Matrix operation unit that supports real-time alignment of multimodal data;

[0038] Built-in configurable contrastive loss function library.

[0039] As a further aspect of the present invention, the knowledge graph service layer includes a graph database that dynamically updates the relationships between financial entities and an integrated GNN inference interface.

[0040] In another aspect, the present invention provides a computer device including a memory and a processor, the memory storing a computer program which, when executed by the processor, performs any of the above-described financial information aggregation and recommendation methods based on unsupervised cross-modal learning according to the present invention.

[0041] In another aspect, the present invention provides a computer-readable storage medium storing computer program instructions that, when executed, implement any of the above-described financial information aggregation and recommendation methods based on unsupervised cross-modal learning according to the present invention.

[0042] Compared with existing technologies, the financial information aggregation and recommendation method and system based on unsupervised cross-modal learning proposed in this invention have the following beneficial effects:

[0043] 1) This invention effectively maps heterogeneous data such as text, images, and numerical data to a unified semantic space by constructing a cross-modal dynamic alignment matrix. Cross-modal data fusion of information carried by different modal data (such as transaction behavior, news articles, financial charts, etc.) improves the expressive power and accuracy of the data, avoids the limitations of traditional single-modal methods, and enhances the multi-dimensional expressive power and accuracy of information.

[0044] 2) This invention uses unsupervised learning to improve the system's generalization ability and adaptability. It can automatically discover potential relationships from a large amount of heterogeneous data without relying on manually labeled data. This enables the system to adapt to the ever-changing market environment and shows strong robustness on diverse financial data. It effectively avoids the time and labor costs of traditional supervised learning in acquiring labeled data.

[0045] 3) This invention also combines short-term intent recognition with long-term preference recognition to form a two-layer intent recognition network, enabling the system to more accurately capture users' immediate needs and potential preferences. Short-term intents, such as real-time search keywords and click behavior, reflect users' current needs, while long-term preferences, based on the analysis of historical behavioral trajectories, can reveal users' potential interests and long-term investment strategies. Through the application of the two-layer intent recognition network, the system can combine the features of both to output a comprehensive prediction of users' intent, providing users with more accurate personalized and dynamic recommendation services;

[0046] 4) By analyzing users' short-term intentions and long-term preferences, and combining factors such as seasonal changes and cyclical fluctuations in the market, this invention can generate cross-cycle recommendation strategies. This makes the recommendation results not only limited to the user's needs at the current moment, but also able to provide users with long-term and stable investment advice based on market trends, meeting the needs of users in different time dimensions and improving the applicability and timeliness of the recommendations.

[0047] In summary, the financial information aggregation and recommendation method and system based on unsupervised cross-modal learning of this invention enhances the expressive power of financial information through data fusion and semantic alignment, enabling multi-faceted analysis of user needs. Unsupervised learning strengthens the system's adaptability and generalization capabilities, reducing the need for manual annotation. Two-layer intent recognition improves the accuracy of identifying short-term and long-term needs, precisely matching users' investment goals. Cross-period recommendation strategies meet changing user needs across different time dimensions. A dynamic recommendation mechanism continuously optimizes user experience and supports the achievement of business objectives. Efficient data collection and processing mechanisms ensure real-time performance and data quality, enhancing the flexibility and applicability of the recommendation system. Supporting multiple types of financial data, the system can quickly adapt to changes in user behavior and automatically provide optimal recommendations. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention or related technologies, the accompanying drawings used in the description of the exemplary embodiments or related technologies will be briefly introduced below. The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation thereof. In the drawings:

[0049] Figure 1 This is a flowchart illustrating a financial information aggregation and recommendation method based on unsupervised cross-modal learning, according to an embodiment of the present invention.

[0050] Figure 2 This is a flowchart illustrating the process of obtaining heterogeneous data from cross-platform data sources in a financial information aggregation and recommendation method based on unsupervised cross-modal learning, according to an embodiment of the present invention.

[0051] Figure 3 This is a flowchart illustrating cross-modal dynamic alignment in a financial information aggregation and recommendation method based on unsupervised cross-modal learning, according to an embodiment of the present invention.

[0052] Figure 4 This is a flowchart illustrating the analysis of historical cross-platform behavioral trajectories to form long-term preferences in a financial information aggregation and recommendation method based on unsupervised cross-modal learning, according to an embodiment of the present invention. Detailed Implementation

[0053] The present application will now be further described in conjunction with the accompanying drawings and specific embodiments. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0054] To make the objectives, technical solutions, and advantages of this invention clearer, the embodiments of this invention will be further described in detail below with reference to specific examples and the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit this application.

[0055] It should be noted that all uses of "first" and "second" in the embodiments of the present invention are for the purpose of distinguishing two different entities or different parameters with the same name. Therefore, "first" and "second" are merely for convenience of expression and should not be construed as limiting the embodiments of the present invention. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, such as other steps or units inherent in a process, method, system, product, or device that includes a series of steps or units.

[0056] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0057] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content and operations / steps, nor does it necessarily have to be performed in the order described. For example, some operations / steps can be broken down, combined, or partially merged, so the actual execution order may change depending on the actual situation.

[0058] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0059] To address the problems of data silos and difficulties in heterogeneous integration in existing financial information recommendation methods, which result in insufficient cross-modal intent parsing capabilities and poor dynamic adaptability, this invention proposes a financial information aggregation and recommendation method and system based on unsupervised cross-modal learning. Based on unsupervised cross-modal learning technology, it realizes cross-platform and multi-channel financial information aggregation, deep user intent parsing, and personalized recommendation, thereby solving the aforementioned problems in existing technologies.

[0060] See Figure 1 As shown, embodiments of the present invention provide a financial information aggregation and recommendation method based on unsupervised cross-modal learning, the method comprising the following steps:

[0061] Step S10: Build a crawler engine for multi-channel data collection and use dynamic crawlers to obtain heterogeneous data from cross-platform data sources.

[0062] In this step, see Figure 2 As shown, when using dynamic web crawlers to obtain heterogeneous data from cross-platform data sources, the specific steps include:

[0063] Step S101: Use a crawler engine to identify the structural features of the target website or platform in the data source, use a dynamic crawler to simulate user behavior, process dynamically loaded web pages, and extract heterogeneous data.

[0064] Step S102: Remove HTML tags and duplicate content from the extracted heterogeneous data and convert the encoding format to structure the data into a standard format;

[0065] Step S103: Store the processed heterogeneous data and set up an incremental crawling mechanism to periodically update the heterogeneous data in the database based on the data update timestamp.

[0066] Furthermore, the dynamic crawler engine integrates a task scheduling module to allocate crawling tasks based on data source priority; it adopts a failure retry mechanism, automatically retrying up to 3 times when crawling fails; and it sets a traffic control strategy to limit the number of requests per second.

[0067] In this embodiment, when allocating crawling tasks according to data source priority, for each data source... Assign a dynamic priority score The score is determined by the update frequency. User access popularity and data value weight The decision is made jointly. The calculation formula is: ;in, , , These are weighting coefficients, and Update frequency Calculated based on the reciprocal of the historical update interval of the data source; user access popularity. Data value weight is derived from statistics of recent user query logs. The task allocation is based on the importance of the data source. During task allocation, the scheduling module uses a priority-based round-robin algorithm, assuming the total crawling resources are... Each data source Allocated resources The proportion of its priority score to the total score: Where n is the total number of data sources.

[0068] For example, if the system needs to monitor three data sources: data source A, data source B, and data source C, the priority quantification model can be used to calculate:

[0069] Among them, update frequency ( ): A=0.8 (every hour interval), B=0.6 (every two hours interval), C=0.4 (every four hours interval);

[0070] User access popularity ( ): A=0.9 (high frequency access), B=0.7 (medium frequency), C=0.5 (low frequency);

[0071] Data value weight ( A = 0.8 (high value), B = 0.9 (high value), C = 0.6 (medium value);

[0072] Weighting coefficients ;

[0073] The priority score calculation result is:

[0074] ;

[0075] ;

[0076] .

[0077] If the total resource R = 100, then the resource allocation is as follows:

[0078] ;

[0079] ;

[0080] .

[0081] Step S20: Deploy a heterogeneous dual-engine architecture of semantic engine and numerical engine to perform semantic parsing on the acquired heterogeneous data and extract transaction features.

[0082] In this step, the semantic engine employs a dynamically pruned semantic parsing model to process text and image data, while the numerical engine extracts transaction features through a temporal autoencoder. Specifically, the semantic engine uses a pre-trained BERT model from the financial domain as the semantic parsing model for text semantic parsing, and combines it with a ResNet-50 network to extract image features. The dynamically pruned semantic parsing model refers to dynamically pruning and optimizing the BERT and ResNet-50 models from the financial domain; that is, automatically pruning unimportant parameters based on the input data during model inference, thereby improving processing speed.

[0083] In this embodiment, the semantic engine uses the BERT model from the financial domain to process the input text sequence. Context vector representation is generated through a multi-layer Transformer encoder. . No. The formula for calculating the self-attention mechanism of a layered Transformer is as follows:

[0084] ;

[0085] in, Each by input Linear projection yields, is the dimension of the key vector.

[0086] For example, if the input text is: "The central bank cuts the reserve requirement ratio by 0.5 percentage points, which is expected to release about 1 trillion yuan of long-term funds", the output after processing by the BERT model is: word embedding vector dimension:

[768] ; semantic parsing results: "central bank" → financial institution entity; "reduction of reserve requirement ratio" → monetary policy action; "0.5 percentage points" → numerical parameter; "release of long-term funds" → market impact; attention weights are shown as: "reduction of reserve requirement ratio" and "1 trillion yuan" get the highest attention scores (0.85+), indicating that these words are the most critical to financial semantic parsing.

[0087] Step S30: Construct a cross-modal dynamic alignment matrix, use the contrastive learning loss function to eliminate semantic bias between text descriptions, chart data and numerical sequences, map them to a unified semantic space and generate a joint representation vector.

[0088] In this step, see Figure 3As shown, constructing a cross-modal dynamic alignment matrix and performing cross-modal dynamic alignment includes the following steps:

[0089] Step S301: Use a multi-granularity attention mechanism for the text modality to identify the weight between each word or sentence in the text, obtain the text information most relevant to user needs, and classify them according to the intensity of supervision.

[0090] Step S302: Construct a morphological parser for the image modality, extract the implicit patterns of technical indicators in the candlestick chart, analyze the features of the image, and extract technical indicators related to market trends; wherein, the present invention uses standard candlestick chart patterns for feature extraction, which is used for technical analysis to identify market trends.

[0091] Step S303: Use the time-series causal discovery algorithm on the numerical model to explore the lagged relationship between user trading behavior and macroeconomic indicators, and the time delay relationship between trading behavior and economic variables. The time-series causal discovery algorithm adopts Granger causality test, and the specific steps include: 1) performing stationarity test on the time series of user trading behavior and economic indicators; 2) constructing a VAR model; 3) judging the lagged relationship through F test. If the p value < 0.05, it is considered that there is a causal relationship.

[0092] In this embodiment, the Granger causality test uses a time-series causal discovery algorithm to mine user transaction behavior. With macroeconomic indicators The lag relationship. The specific formula for the Granger causality test is based on the vector autoregression (VAR) model:

[0093] Perform an F-test: ;in, and These are the sums of squared residuals for the constrained and unconstrained models, respectively. It is the lag order. It is the length of the time series. If the calculated... The statistic corresponds to If the value is less than the significance level, such as 0.05, then the null hypothesis is rejected, and X is considered to be a Granger cause of Y.

[0094] Step S304: Map the features of different modalities in text, images and numerical data to a unified semantic space, construct a cross-modal alignment matrix, and fuse the features of text, image and numerical data.

[0095] This invention effectively maps heterogeneous data such as text, images, and numerical data to a unified semantic space by constructing a cross-modal dynamic alignment matrix. Cross-modal data fusion of information carried by different modalities (such as transaction behavior, news articles, and financial charts) enhances the expressive power and accuracy of the data, avoids the limitations of traditional single-modal methods, and improves the multi-dimensional expressive power and accuracy of information.

[0096] In this embodiment, the InfoNCE loss function is used to construct a cross-modal dynamic alignment matrix, and the contrastive learning loss function is used for a set of text... and images For positive sample pairs, the loss function is:

[0097] ;

[0098] in, A text modality encoder is a neural network that maps text input to semantic vectors. An image modality encoder is a convolutional network that maps image input to feature vectors. is the cosine similarity function, used to calculate the degree of similarity between two vectors in semantic space; Temperature parameters The setting range is 0.05-0.1; The batch size is the number of samples included in a single training session; , These are positive sample pairs, used to describe text-image pairing data with the same semantic meaning; Negative samples represent image data that does not match the current text.

[0099] For example, suppose the positive sample pairs are:

[0100] Text modality: "Group A's stock price broke through 2,000 yuan, and its market value exceeded 2.5 trillion yuan";

[0101] Image modality: The candlestick chart of Group A shows the trend of its stock price breaking through 2000 yuan;

[0102] Negative sample pairs are:

[0103] Text modality: "B Group releases new battery technology, achieving a range of over 1000 kilometers";

[0104] Image modality: A Group candlestick chart.

[0105] When calculating the loss function, the similarity of positive samples is: Negative sample similarity: Temperature parameters The loss value is:

[0106] .

[0107] Step S40: Obtain real-time search keywords and click behavior to form short-term intent, analyze historical cross-platform behavior trajectories to form long-term preferences, and construct a two-layer intent recognition network based on short-term intent and long-term preferences.

[0108] The dual-layer intent recognition network consists of a short-term intent recognition network (LSTM-based) and a long-term preference recognition network (GNN-based). It fuses short-term and long-term features through an attention mechanism to output a comprehensive intent prediction. The network structure includes an input layer, a fusion layer, and an output layer. In this embodiment, the short-term intent feature vector... With long-term preference feature vector A comprehensive intent vector is generated by fusion through an attention mechanism. The formula for fusion of attention weights is: ;in, and This represents the normalized attention weights.

[0109] In this step, see Figure 4 As shown, obtaining real-time search keywords and click behavior constitutes short-term intent, and analyzing historical cross-platform behavior patterns constitutes long-term preferences, including the following steps:

[0110] Step S401: Obtain the user's real-time search keywords through the integrated search engine, process and clean the search keywords, and remove noise information;

[0111] Step S402: Extract information from the user's real-time click behavior data, record the time, page, and clicked object of each click, and generate a click log by tracking the user's click behavior;

[0112] Step S403: Extract features from real-time search keywords and click behavior data, extract short-term intent features, and use a short-term intent recognition network for intent classification; wherein, the short-term intent recognition network adopts an LSTM model, the input is the embedding vector of real-time search keywords and click behavior, two layers of LSTM units capture temporal dependencies, and the output layer uses Softmax for intent classification (such as investment, query, transaction, etc.).

[0113] Step S404: Predict the user's long-term preferences based on historical behavior trajectories, and model them using a long-term preference recognition network. The network combines short-term and long-term intentions to output intent recognition. The long-term preference recognition network employs a graph neural network (GNN), with the input being a knowledge graph constructed from historical behavior. It aggregates neighbor node information through graph convolutional layers and outputs a user's long-term preference vector. The network structure includes a 3-layer GCN with ReLU activation.

[0114] This invention combines short-term intent recognition with long-term preference recognition to form a two-layer intent recognition network, enabling the system to more accurately capture users' immediate needs and potential preferences. Short-term intents, such as real-time search keywords and click behavior, reflect users' current needs; while long-term preferences, based on the analysis of historical behavioral patterns, can reveal users' potential interests and long-term investment strategies. Through the application of the two-layer intent recognition network, the system can combine the features of both to output a comprehensive prediction of users' intent, providing more accurate personalized and dynamic recommendation services.

[0115] Step S50: Input the semantic parsing and transaction features of the heterogeneous data into a two-layer intent recognition network to capture the implicit relationships between financial entities, drive the generation of cross-cycle recommendation strategies, and output dynamic recommendation results that integrate business objectives and user satisfaction indicators.

[0116] This invention employs unsupervised learning to improve the system's generalization ability and adaptability. It can automatically discover potential relationships from a large amount of heterogeneous data without relying on manually labeled data, enabling the system to adapt to the ever-changing market environment and exhibiting strong robustness on diverse financial data. It effectively avoids the time and labor costs of traditional supervised learning in acquiring labeled data.

[0117] This invention analyzes users' short-term intentions and long-term preferences, and combines them with factors such as seasonal changes and cyclical fluctuations in the market, to generate cross-cycle recommendation strategies. This ensures that the recommendation results are not limited to the user's needs at the current moment, but also provide users with long-term and stable investment advice based on market trends, meeting the needs of users in different time dimensions and improving the applicability and timeliness of the recommendations.

[0118] It should be noted that the above figures are merely illustrative of the processes included in the method according to exemplary embodiments of the present invention, and are not intended to be limiting. It is readily understood that the processes shown in the above figures do not indicate or limit the temporal order of these processes. Furthermore, it is readily understood that these processes may be executed synchronously or asynchronously in multiple modules.

[0119] It should be understood that although the above description follows a certain order, these steps are not necessarily executed in that order. Unless otherwise expressly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, some steps in this embodiment may include multiple steps or multiple stages, which are not necessarily completed at the same time, but may be executed at different times. The execution order of these steps or stages is not necessarily sequential, but may be performed alternately or in turn with other steps or at least a portion of steps or stages in other steps.

[0120] In a second aspect, the present invention also provides a financial information aggregation and recommendation system based on unsupervised cross-modal learning, comprising:

[0121] The data integration layer includes a multi-channel crawler engine and an incremental crawling mechanism, used to acquire heterogeneous data from cross-platform data sources and update the data sources regularly;

[0122] The data processing layer includes a semantic engine and a numerical engine. By deploying a heterogeneous dual-engine architecture of semantic engine and numerical engine, it performs semantic parsing on the acquired heterogeneous data and extracts transaction features.

[0123] The intelligent processing layer includes a heterogeneous computing module, a cross-modal dynamic alignment matrix, and a knowledge graph service layer. It is used to eliminate semantic biases between text descriptions, chart data, and numerical sequences by using a contrastive learning loss function, mapping them to a unified semantic space and generating joint representation vectors.

[0124] The intent recognition layer is used to obtain real-time search keywords and click behaviors to form short-term intents, analyze historical cross-platform behavior trajectories to form long-term preferences, and construct a two-layer intent recognition network based on short-term intents and long-term preferences.

[0125] The service output layer is used to input the semantic parsing and transaction features of the heterogeneous data into a two-layer intent recognition network, capture the implicit relationships between financial entities, drive the generation of cross-cycle recommendation strategies, and output dynamic recommendation results that integrate business objectives and user satisfaction indicators.

[0126] The semantic engine is responsible for semantic parsing of text data, using the BERT model from the financial field for text semantic parsing, and combining it with the ResNet-50 network to extract features from image data; the numerical engine uses a temporal autoencoder to extract transaction features and analyze users' transaction behavior patterns.

[0127] In this embodiment, the heterogeneous computing module includes an FPGA-accelerated semantic engine and a GPU cluster-supported numerical engine. The cross-modal dynamic alignment matrix includes a matrix operation unit that supports real-time alignment of multimodal data and a built-in configurable contrastive loss function library. The knowledge graph service layer includes a graph database that dynamically updates financial entity relationships and an integrated GNN inference interface. The matrix operation unit performs feature mapping operations, and the built-in loss function library implements contrastive learning to ensure that text, image, and numerical features are aligned in a unified space.

[0128] Through the detailed steps described above, the financial information aggregation and recommendation system based on unsupervised cross-modal learning of the present invention is used to execute the steps of the financial information aggregation and recommendation method based on unsupervised cross-modal learning in the above embodiments, which will not be repeated here.

[0129] In summary, the financial information aggregation and recommendation method and system based on unsupervised cross-modal learning of this invention enhances the expressive power of financial information through data fusion and semantic alignment, enabling multi-faceted analysis of user needs. Unsupervised learning strengthens the system's adaptability and generalization capabilities, reducing the need for manual annotation. Two-layer intent recognition improves the accuracy of identifying short-term and long-term needs, precisely matching users' investment goals. Cross-period recommendation strategies meet changing user needs across different time dimensions. A dynamic recommendation mechanism continuously optimizes user experience and supports the achievement of business objectives. Efficient data collection and processing mechanisms ensure real-time performance and data quality, enhancing the flexibility and applicability of the recommendation system. Supporting multiple types of financial data, the system can quickly adapt to changes in user behavior and automatically provide optimal recommendations.

[0130] A third aspect of the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program that, when executed by the processor, implements the method of any of the above embodiments.

[0131] The computer device includes a processor and a memory, and may also include an input system and an output system. The processor, memory, input system, and output system can be connected via a bus or other means. The input system can receive input digital or character information and generate signal inputs related to the transfer of financial information aggregation recommendations based on unsupervised cross-modal learning. The output system may include display devices such as a screen.

[0132] Memory, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules, such as the program instructions / modules corresponding to the financial information aggregation and recommendation method based on unsupervised cross-modal learning in the embodiments of this application. Memory may include a program storage area and a data storage area, wherein the program storage area may store the operating system and at least one application program required for a function; the data storage area may store data created by using the financial information aggregation and recommendation method based on unsupervised cross-modal learning, etc. Furthermore, memory may include high-speed random access memory and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the local module via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0133] In some embodiments, the processor may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip. The processor is typically used to control the overall operation of a computer device. In this embodiment, the processor is used to run program code stored in memory or process data. In this embodiment, the processors of multiple computer devices execute various server functions and data processing by running non-volatile software programs, instructions, and modules stored in memory, thereby implementing the steps of the financial information aggregation and recommendation method based on unsupervised cross-modal learning described in the above method embodiments.

[0134] It should be understood that, where there is no conflict, all the embodiments, features and advantages described above for the financial information aggregation and recommendation method based on unsupervised cross-modal learning according to the present invention are equally applicable to the financial information aggregation and recommendation and storage medium based on unsupervised cross-modal learning according to the present invention.

[0135] Those skilled in the art will also understand that the various exemplary logic blocks, modules, circuits, and algorithm steps described in conjunction with the disclosure herein can be implemented as electronic hardware, computer software, or a combination of both. To clearly illustrate this interchangeability between hardware and software, the functionality of various illustrative components, blocks, modules, circuits, and steps has been generally described. Whether this functionality is implemented as software or as hardware depends on the specific application and the design constraints imposed on the system as a whole. Those skilled in the art can implement the functionality in various ways for each specific application, but such implementation decisions should not be construed as departing from the scope of the embodiments disclosed herein.

[0136] Finally, it should be noted that the computer-readable storage medium (e.g., memory) described herein can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. By way of example, and not limitation, non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM), which can act as external cache memory. By way of example, and not limitation, RAM can be obtained in various forms, such as synchronous RAM (DRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), and direct Rambus RAM (DRRAM). The storage devices disclosed herein are intended to include, but are not limited to, these and other suitable types of memory.

[0137] The various exemplary logic blocks, modules, and circuits described herein can be implemented or performed using the following components designed to perform the functions herein: general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic, discrete hardware components, or any combination of these components. A general-purpose processor may be a microprocessor, but alternatively, the processor may be any conventional processor, controller, microcontroller, or state machine. The processor may also be implemented as a combination of computing devices, such as a combination of a DSP and a microprocessor, multiple microprocessors, one or more microprocessors combined with a DSP, and / or any other such configuration.

[0138] The above are exemplary embodiments disclosed in this invention. However, it should be noted that various changes and modifications can be made without departing from the scope of the embodiments of this invention as defined by the claims. The functions, steps, and / or actions of the methods according to the disclosed embodiments described herein do not need to be performed in any particular order. Furthermore, although the elements disclosed in the embodiments of this invention may be described or claimed individually, they may be understood as multiple unless explicitly limited to a singular number.

[0139] It should be understood that, as used herein, the singular form "a" is intended to include the plural form as well, unless the context clearly supports an exception. It should also be understood that, as used herein, "and / or" refers to any and all possible combinations of one or more of the associatedly listed items. The embodiment numbers disclosed above are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0140] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention (including the claims) is limited to these examples. Within the framework of the invention, technical features of the above embodiments or different embodiments can be combined, and many other variations of different aspects of the invention exist, which are not provided in the details for the sake of brevity. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the protection scope of the invention.

Claims

1. A financial information aggregation recommendation method based on unsupervised cross-modal learning, characterized in that, The method includes the following steps: Step S10: Build a crawler engine for multi-channel data collection and use dynamic crawlers to obtain heterogeneous data from cross-platform data sources; Step S20: Deploy a heterogeneous dual-engine architecture of semantic engine and numerical engine to perform semantic parsing on the acquired heterogeneous data and extract transaction features; wherein, the semantic engine uses a dynamically pruned semantic parsing model to process text and image data, and the numerical engine extracts transaction features through a time-series autoencoder. Step S30: Construct a cross-modal dynamic alignment matrix, use the contrastive learning loss function to eliminate semantic bias between text descriptions, chart data and numerical sequences, map them to a unified semantic space and generate a joint representation vector; Step S40: Obtain real-time search keywords and click behavior to form short-term intent, analyze historical cross-platform behavior trajectories to form long-term preferences, and construct a two-layer intent recognition network based on short-term intent and long-term preferences; Step S50: Input the semantic parsing and transaction features of the heterogeneous data into a two-layer intent recognition network to capture the implicit relationships between financial entities, drive the generation of cross-cycle recommendation strategies, and output dynamic recommendation results that integrate business objectives and user satisfaction indicators. In step S40, obtaining real-time search keywords and click behavior constitutes short-term intent, and analyzing historical cross-platform behavior patterns constitutes long-term preferences, including the following steps: The system obtains users' real-time search keywords through an integrated search engine, processes and cleans these keywords, and removes noisy information. Information is extracted from real-time user click behavior data, recording the time, page, and clicked object of each click, and a click log is generated by tracking user click behavior. Feature extraction is performed on real-time search keyword and click behavior data. Short-term intent features are extracted and intent classification is performed using a short-term intent recognition network. Predict users' long-term preferences based on historical behavior trajectories, and model them using a long-term preference recognition network. Combine short-term intentions with long-term preferences to output intention recognition.

2. The financial information aggregation recommendation method based on unsupervised cross-modal learning according to claim 1, wherein, Step S10 is as follows: By using a web crawler engine to identify the structural features of the target website or platform in the data source, dynamic web crawlers are used to simulate user behavior, process dynamically loaded web pages, and extract heterogeneous data. Remove HTML tags and duplicate content from the extracted heterogeneous data and convert the encoding format to structure the data into a standard format; The processed heterogeneous data is stored, and an incremental crawling mechanism is set up to periodically update the heterogeneous data in the database based on the data update timestamp.

3. The financial information aggregation recommendation method based on unsupervised cross-modal learning according to claim 2, wherein, The semantic engine uses a pre-trained BERT model from the financial domain as a semantic parsing model to perform text semantic parsing, and combines it with a ResNet-50 network to extract image features.

4. The financial information aggregation recommendation method based on unsupervised cross-modal learning according to claim 1, wherein, In step S30, a cross-modal dynamic alignment matrix is ​​constructed, and cross-modal dynamic alignment is performed, including: A multi-granularity attention mechanism is used for text modalities to identify the weight between each word or sentence in the text, obtain the text information most relevant to user needs, and classify them according to regulatory intensity; A morphological parser is constructed for image modalities to extract the implicit patterns of technical indicators in candlestick charts. The features of the images are analyzed to extract technical indicators related to market trends. We use a time-series causal discovery algorithm on numerical modes to explore the lagged relationship between user transaction behavior and macroeconomic indicators, as well as the time delay relationship between transaction behavior and economic variables. By mapping features from different modalities in text, images, and numerical data to a unified semantic space, a cross-modal alignment matrix is ​​constructed to fuse the features of text, image, and numerical data. 5.A financial information aggregation recommendation system based on unsupervised cross-modal learning, characterized in that, The system is used to execute the financial information aggregation and recommendation method based on unsupervised cross-modal learning as described in any one of claims 1-4, the system comprising: The data integration layer includes a multi-channel crawler engine and an incremental crawling mechanism, used to acquire heterogeneous data from cross-platform data sources and update the data sources regularly; The data processing layer includes a semantic engine and a numerical engine. By deploying a heterogeneous dual-engine architecture of semantic engine and numerical engine, it performs semantic parsing on the acquired heterogeneous data and extracts transaction features. The intelligent processing layer includes a heterogeneous computing module, a cross-modal dynamic alignment matrix, and a knowledge graph service layer. It is used to eliminate semantic biases between text descriptions, chart data, and numerical sequences using a contrastive learning loss function, mapping them to a unified semantic space and generating joint representation vectors. The heterogeneous computing module includes an FPGA-accelerated semantic engine and a GPU cluster-supported numerical engine. The knowledge graph service layer includes a graph database that dynamically updates the relationships between financial entities and an integrated GNN inference interface. The intent recognition layer is used to obtain short-term intents by real-time search keywords and click behaviors, and to analyze long-term preferences by analyzing historical cross-platform behavior trajectories. A two-layer intent recognition network is constructed based on short-term intents and long-term preferences. The service output layer is used to input the semantic parsing and transaction features of the heterogeneous data into a two-layer intent recognition network, capture the implicit relationships between financial entities, drive the generation of cross-cycle recommendation strategies, and output dynamic recommendation results that integrate business objectives and user satisfaction indicators.

6. The financial information aggregation recommendation system based on unsupervised cross-modal learning of claim 5, wherein, The semantic engine is responsible for semantic parsing of text data. It uses the BERT model from the financial field for text semantic parsing and combines it with the ResNet-50 network to extract features from image data. The numerical engine uses a temporal autoencoder to extract transaction features and analyze users' transaction behavior patterns.

7. The financial information aggregation recommendation system based on unsupervised cross-modal learning of claim 6, wherein, The cross-modal dynamic alignment matrix includes: Matrix operation unit that supports real-time alignment of multimodal data; Built-in configurable contrastive loss function library.

Images