Multi-source data fusion prediction method, device, equipment, storage medium and product

Abstract

This application discloses a multi-source data fusion prediction method, apparatus, device, storage medium, and product, relating to the field of data prediction technology. The method includes: acquiring multi-source data of financial objects, including time-series structured data, untime-series structured data, and unstructured text data; constructing a semantic factor sequence and generating text semantic vectors based on knowledge enhancement using a large model on the unstructured text data; encoding the time-series structured data and / or untime-series structured data to obtain corresponding structured time-series representations and / or static context representations; dynamically fusing at least one of the semantic factor vectors, text semantic vectors, structured time-series representations, and static context representations from the semantic factor sequence through a gating mechanism to obtain a fused representation; and inputting the fused representation into a multi-task prediction output module to obtain the result of at least one prediction task. This reduces real-time inference costs and improves system prediction efficiency, stability, and scalability.

Description

Technical Field

[0001] This application relates to the field of data prediction technology, and in particular to a method, apparatus, device, storage medium and product for multi-source data fusion prediction. Background Technology

[0002] Currently, financial target prediction tasks primarily rely on structured data modeling, such as features like customer behavior, account status, transaction history, product attributes, and market trends. With the development of large-scale modeling technology, some solutions have begun to incorporate textual information to assist in risk identification, customer management, product recommendation, or market trend prediction. These solutions typically transform unstructured text into text summaries, sentiment scores, keyword statistics, or text vectors, and then simply concatenate or statically weight them with structured features to improve predictive performance.

[0003] However, existing methods still have significant shortcomings: First, the use of textual semantics is superficial, making it difficult to form interpretable predictive factors and clearly express semantic content such as customer intent, risk signals, or event impact, which is not conducive to auditing and interpretation. Second, there is a lack of effective alignment between textual information and the business timeline; irregular release times and durations of impact of text are difficult to map to the correct time steps or business stages, easily leading to mismatches between semantic and temporal features. Third, the fusion of structured features and semantic features is crude. Fourth, there is insufficient support for multi-task and multi-time-window prediction, often requiring separate modeling for different tasks, resulting in system complexity and redundant resource consumption. Fifth, relying entirely on large-scale end-to-end prediction models presents problems such as high training costs, large inference latency, and complex deployment.

[0004] Therefore, existing technologies suffer from problems such as shallow utilization of text semantics, insufficient alignment of business time, coarse fusion of heterogeneous features, fragmented multi-task prediction, and high deployment costs of large models. It is necessary to propose a technical solution to address the issue of how text semantics can participate in financial prediction in a stable and interpretable manner. Summary of the Invention

[0005] The main purpose of this application is to provide a multi-source data fusion prediction method, apparatus, device, storage medium and product, which aims to solve the problem of how text semantics can be stably and interpretably involved in financial prediction, reduce real-time inference costs and improve system prediction efficiency, stability and scalability.

[0006] To achieve the above objectives, this application proposes a multi-source data fusion prediction method, the method comprising: Obtain multi-source data of financial objects, including time-series structured data, non-time-series structured data, and unstructured text data; Based on the unstructured text data, a semantic factor sequence is constructed using a large model based on knowledge enhancement, and a text semantic vector is generated. The time-series structured data and / or non-time-series structured data are encoded to obtain the corresponding structured time-series representation and / or static context representation; The semantic factor vector, the text semantic vector, the structured temporal representation, and the static context representation in the semantic factor sequence are dynamically fused through a gating mechanism to obtain a fused representation; The fused representation is input into the multi-task prediction output module to obtain the result of at least one prediction task.

[0007] In one embodiment, the step of constructing a semantic factor sequence based on knowledge enhancement using a large model based on the unstructured text data includes: By leveraging large models and knowledge from the financial field, knowledge-enhanced semantic factor records are extracted from the unstructured text data. Each semantic factor record includes at least the object identifier, factor category, semantic direction, factor strength, confidence level, occurrence time, and effective time range. Using the business timeline of the financial object as an index, the semantic factor records are mapped to a semantic factor sequence with the same granularity as the structured time-series data, wherein the semantic factor vector of each time step is obtained by weighted summation of the relevant semantic factor records.

[0008] In one embodiment, the step of mapping the semantic factor records to a semantic factor sequence consistent with the granularity of structured time-series data, using the business timeline of the financial object as an index, wherein the semantic factor vector of each time step is obtained by weighted summation of relevant semantic factor records includes: For any financial object and time step, candidate semantic factors related to the financial object, time step and prediction task are screened, and the role weights are calculated based on time relationship, object relationship, source credibility, factor confidence and historical business response calibration parameters. The candidate semantic factors are weighted and aggregated according to their respective weights to obtain the semantic factor vector corresponding to the time step. Based on the semantic factor vector, a semantic factor sequence corresponding one-to-one with each time step within the observation window is obtained.

[0009] In one embodiment, the step of encoding the temporal structured data and / or non-temporal structured data to obtain the corresponding structured temporal representation and / or static context representation includes: For the aforementioned temporal structured data, a lightweight temporal coding network is used to generate a structured temporal representation. The lightweight temporal coding network includes at least one of the following: numerical projection layer, temporal embedding layer, self-attention layer, convolutional layer, and recurrent layer. For the non-temporally structured data, a static context representation is generated through an embedding layer or a multilayer perceptron.

[0010] In one embodiment, the step of dynamically fusing at least one of the semantic factor vectors in the semantic factor sequence, the text semantic vectors, the structured temporal representation, and the static context representation through a gating mechanism to obtain the fused representation includes: For each time step, the structured temporal representation, semantic factor vectors in the semantic factor sequence, text semantic vectors, and static context representation are received through the gating fusion module, and corresponding gating weights are generated by combining the task identifier and prediction window. Based on the gating weights, the semantic factor vector, the structured temporal representation, the text semantic vector, and the static context representation are weighted and fused to obtain a fused representation.

[0011] In one embodiment, the same semantic factor vector has different fusion weights under different task identifiers or different prediction window lengths.

[0012] In one embodiment, the multi-task prediction output module includes a shared prediction representation layer and multiple task-specific output heads; different output heads are used for classification, regression, sorting, prediction of multiple future time points, or prediction of future time window statistics, respectively. The multi-time point prediction output head is used to output continuous predicted values ​​for multiple future time steps; the time window statistical prediction output head is used to output the prediction results of the maximum value, mean, quantile, or whether a threshold is exceeded within a specified future time window.

[0013] In one embodiment, the method further includes the following during the training phase: During the training phase, the large model is jointly trained by multi-task supervised loss combined with semantic factor consistency loss, gating constraints, and robustness loss; the semantic factor consistency loss is used to constrain the correspondence between semantic factors and historical business responses.

[0014] In one embodiment, the step of acquiring multi-source data of financial objects further includes: Perform at least one of the following preprocessing steps on the time-series structured data: missing value imputation, outlier handling, normalization, bucketing encoding, sliding window construction, and time position encoding, to obtain structured time-series features; Perform category encoding, normalization, or embedding mapping on the non-temporal structured data to obtain static structured features; Perform at least one of the following preprocessing steps on the unstructured text data: cleaning, deduplication, sentence segmentation, timestamp parsing, object identification, source tagging, and anonymization, to obtain a set of text fragments.

[0015] In one embodiment, the method further includes: In the offline phase, a large model is used to extract semantic factor records from unstructured text data, generate semantic factor sequences, and store semantic factor records, semantic factor sequences, and text semantic vectors. During the online phase, the latest structured features and generated semantic factor sequences are read, and lightweight encoding, gated fusion, and multi-task prediction output are performed.

[0016] In one embodiment, the large model is at least one of a general large language model, a financial domain large model, an encoder-based pre-trained model, a rule extraction model, and a knowledge graph reasoning model.

[0017] In one embodiment, the step of constructing a semantic factor sequence based on knowledge enhancement using a large model based on the unstructured text data further includes at least one of the following: A hard alignment method based on fixed time granularity is adopted; By learning a soft alignment method for weight allocation across multiple time steps; The persistent impact of semantic factors can be controlled by using decay kernel functions, validity windows, or historical response calibration parameters.

[0018] Furthermore, to achieve the above objectives, this application also proposes a multi-source data fusion prediction device, the device comprising: The data acquisition module is used to acquire multi-source data of financial objects, wherein the multi-source data includes at least one of time-series structured data, non-time-series structured data, and unstructured text data; The semantic factor sequence construction module is used to construct a semantic factor sequence based on knowledge enhancement using a large model based on the unstructured text data, and generate text semantic vectors. The structured feature encoding module is used to encode the temporal structured data and / or non-temporal structured data to obtain the corresponding structured temporal representation and / or static context representation; The gating fusion module is used to dynamically fuse at least one of the semantic factor vectors, the text semantic vectors, the structured temporal representation, and the static context representation in the semantic factor sequence through a gating mechanism to obtain a fused representation; A multi-task prediction output module is used to perform multi-task prediction based on the fused representation to obtain the result of at least one prediction task.

[0019] In addition, to achieve the above objectives, this application also proposes a multi-source data fusion prediction device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the multi-source data fusion prediction method as described above.

[0020] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the multi-source data fusion prediction method described above.

[0021] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the multi-source data fusion prediction method described above.

[0022] One or more technical solutions proposed in this application have at least the following technical effects: By acquiring multi-source data of financial objects, the multi-source data includes time-series structured data, untime-series structured data, and unstructured text data; based on the unstructured text data, a semantic factor sequence is constructed using a large model based on knowledge enhancement, and a text semantic vector is generated; the time-series structured data and / or untime-series structured data are encoded to obtain corresponding structured time-series representations and / or static context representations; at least one of the semantic factor vectors, text semantic vectors, structured time-series representations, and static context representations in the semantic factor sequence is dynamically fused through a gating mechanism to obtain a fused representation; the fused representation is input into a multi-task prediction output module to obtain the result of at least one prediction task. This application targets unstructured text data and utilizes a large model to construct a semantic factor sequence based on knowledge enhancement. This enables text information to participate in prediction in an interpretable, alignable, and trainable form. Simultaneously, it encodes temporal structured data and / or untemporal structured data, and dynamically fuses the semantic factor vectors, the structured temporal representation, and the static context representation in the semantic factor sequence through a gating mechanism. This allows the model to dynamically allocate feature weights according to different tasks and prediction windows. This application retains the ability of the large model to understand complex text and domain knowledge, while entrusting the final prediction process to a lightweight structured encoding network and a gating fusion network, thereby reducing real-time inference costs and improving the system's prediction efficiency, stability, and scalability. Attached Figure Description

[0023] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0024] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a flowchart illustrating the first embodiment of the multi-source data fusion prediction method of this application. Figure 2 This is a schematic diagram of the overall architecture of the multi-source data fusion prediction method embodiment provided in this application; Figure 3 This is a schematic diagram of the module structure of the multi-source data fusion prediction device according to an embodiment of this application; Figure 4 This is a schematic diagram of the device structure of the hardware operating environment involved in the multi-source data fusion prediction method in the embodiments of this application.

[0026] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0027] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0028] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0029] The main solution of this application embodiment is as follows: First, acquire multi-source data of financial objects, including time-series structured data, un-time-series structured data, and unstructured text data. Second, based on the unstructured text data, construct a semantic factor sequence using a large model with knowledge enhancement, and generate a text semantic vector. Third, encode the time-series structured data and / or un-time-series structured data to obtain corresponding structured time-series representations and / or static context representations. Fourth, dynamically fuse at least one of the semantic factor vectors, text semantic vectors, structured time-series representations, and static context representations in the semantic factor sequence through a gating mechanism to obtain a fused representation. Fifth, input the fused representation into a multi-task prediction output module to obtain the result of at least one prediction task. This application targets unstructured text data and utilizes a large model to construct a semantic factor sequence based on knowledge enhancement. This enables text information to participate in prediction in an interpretable, alignable, and trainable form. Simultaneously, it encodes temporal structured data and / or untemporal structured data, and dynamically fuses the semantic factor vectors, the structured temporal representation, and the static context representation in the semantic factor sequence through a gating mechanism. This allows the model to dynamically allocate feature weights according to different tasks and prediction windows. This application retains the ability of the large model to understand complex text and domain knowledge, while entrusting the final prediction process to a lightweight structured encoding network and a gating fusion network, thereby reducing real-time inference costs and improving the system's prediction efficiency, stability, and scalability.

[0030] This application's embodiments take into account that currently, financial object prediction tasks mainly rely on structured data modeling, such as features like customer behavior, account status, transaction history, product attributes, and market trends. With the development of large-scale modeling technology, some solutions have begun to incorporate textual information to assist in risk identification, customer management, product recommendation, or market trend prediction. However, most existing methods still remain at the level of text summarization, sentiment scoring, keyword statistics, or text vector concatenation, and have not yet fully addressed the issue of how text semantics can be stably and interpretably incorporated into financial prediction. Specifically, current similar technologies mainly suffer from the following shortcomings: 1) The use of textual semantics is superficial and it is difficult to form interpretable predictive factors. Existing solutions usually convert the entire text into a vector or extract a small number of sentiment, theme, and keyword features, which is difficult to clearly express the semantic content of customer intent, risk signals, product preferences, event impact, etc. in the text, and is also not convenient for subsequent auditing and interpretation.

[0031] 2) The text information lacks effective alignment with the business timeline. Unstructured text often has irregular release times and durations of impact. Existing solutions often use simple splicing or fixed window aggregation, making it difficult to determine which time step, which business stage, and how long a piece of text information should be applied to. This can easily lead to a mismatch between semantic information and structured temporal features.

[0032] 3) Insufficient ability to fuse structured features and semantic features. Traditional fusion methods are mostly direct splicing or static weighting, which makes it difficult to dynamically adjust the strength of semantic information according to different prediction tasks, different prediction windows and different object states, resulting in limited adaptability of the model to complex financial scenarios.

[0033] 4) Insufficient support for multi-task prediction and multi-time window prediction. Existing systems often train models separately for different tasks, making it difficult to support classification, regression, ranking and multi-time window prediction simultaneously under a unified framework. This results in complex model maintenance, redundant consumption of computing resources, and is not conducive to sharing effective features between different tasks.

[0034] 5) Large-scale end-to-end prediction is costly. Relying entirely on large models for prediction typically results in high training costs, large inference latency, and complex deployment. On the other hand, relying solely on lightweight small models makes it difficult to fully utilize textual semantics and domain knowledge. Existing solutions struggle to achieve a balance between semantic understanding capabilities, prediction efficiency, and engineering deployability.

[0035] Therefore, this application provides a solution to address the problems of shallow text semantic utilization, insufficient business time alignment, coarse heterogeneous feature fusion, fragmented multi-task prediction, and high deployment cost of large models in existing technologies. It proposes a knowledge-enhanced financial multi-source data fusion prediction method, which transforms unstructured text into a semantic factor sequence with clear objects, time, direction, intensity, confidence, and validity period. This allows for precise alignment with structured features along the business timeline, and dynamic fusion is achieved through a task-time window conditional gating mechanism. This supports classification, regression, ranking, multi-time point prediction, and time window statistical prediction within a unified framework, while also taking into account the semantic understanding capabilities of large models and the online deployment efficiency of lightweight networks.

[0036] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device, multi-source data fusion prediction device, or system capable of realizing the above functions. The following description uses a multi-source data fusion prediction system as an example to illustrate the various embodiments.

[0037] Based on this, embodiments of this application provide a multi-source data fusion prediction method, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the multi-source data fusion prediction method of this application.

[0038] In this embodiment, the multi-source data fusion prediction method includes steps S10 to S50. The following provides a detailed explanation of each step.

[0039] like Figure 1 As shown, the first embodiment of this application proposes a multi-source data fusion prediction method, the method comprising: Step S10: Obtain multi-source data of financial objects, including time-series structured data, non-time-series structured data, and unstructured text data; Among them, the financial targets are customers of financial systems such as banks, such as individual credit card customers.

[0040] This example illustrates a bank's task of predicting the "risk of overdue payments in the next 30 days" for credit card customers.

[0041] The overall architecture for multi-source data fusion prediction in this embodiment can be referred to as follows. Figure 2 As shown.

[0042] like Figure 1As shown, the overall architecture for multi-source data fusion prediction in this embodiment may include a data preprocessing module, a knowledge-enhanced semantic factor extraction module, a semantic factor sequence construction module, a structured feature encoding module, a gating fusion module, and a multi-task prediction output module. These modules are connected sequentially to form a complete process from multi-source data input to fusion prediction output.

[0043] First, obtain multi-source data on financial entities.

[0044] The multi-source data includes time-series structured data, non-time-series structured data, and unstructured text data.

[0045] As one implementation method, samples are constructed using an observation window [t0-L+1, t0] preceding the prediction time point t0. The input data includes three categories: time-series structured data, untime-series structured data, and unstructured text data.

[0046] The time-series structured data includes numerical or categorical features recorded over time, such as transactions, behaviors, accounts, and market data; the non-time-series structured data includes features such as object attributes, product attributes, risk attributes, and category tags; and the unstructured text data includes textual information such as communication records, feedback texts, product descriptions, announcements, reports, and public opinion. Taking credit card customers as an example, time-series structured data includes: daily spending, repayment amounts, changes in available credit, and transaction frequency recorded chronologically over the past 6 months (observation window). Non-time-series structured data includes: static attributes that do not change over time, such as the customer's age, occupation, annual income, card level, and account opening date. Unstructured text data includes: communication records between the customer and the bank's customer service over the past 6 months (such as recorded phone conversations, online customer service chat logs), emails sent to the bank, and public statements on social media regarding their financial situation.

[0047] By comprehensively acquiring multi-dimensional information reflecting customer status and behavior, a foundation is laid for subsequent integrated prediction.

[0048] Step S20: Based on the unstructured text data, construct a semantic factor sequence using a large model based on knowledge enhancement, and generate a text semantic vector; In this embodiment, instead of simply concatenating text vectors with structured features, the unstructured text is first transformed into an interpretable, alignable, and decaying sequence of semantic factors using a large model and domain knowledge. Then, a semantic factor sequence is formed according to the business timeline, simultaneously generating text semantic vectors, which are finally fused with structured features for prediction.

[0049] Specifically, the customer's unstructured text data is input into a large language model fine-tuned with financial knowledge (such as credit policies, risk events, and consumption behavior patterns). The large model extracts structured semantic factor records from the text and constructs a semantic factor sequence; and generates a text semantic vector. This semantic factor record includes at least the object identifier, factor category, semantic direction, factor strength, confidence level, occurrence time, and effective time range.

[0050] For example, from a text that reads "a customer complains about a late salary payment and inquires about installment repayment policies," the large model can extract semantic factors such as "negative income signals" and "repayment pressure" and form a time sequence. On the other hand, it can generate a dense vector with fixed dimensions that comprehensively reflects the deep semantics of the entire text, namely, the text semantic vector.

[0051] This step transforms ambiguous textual information into structured temporal features with clear business meaning, quantifiable intensity, and precise alignment to specific dates, resolving the issues of shallow text utilization and time mismatch. Simultaneously, it creates a dual-channel text representation of "semantic factor sequences" (interpretable and temporally sequential) and "textual semantic vectors" (comprehensive and global), enriching the ways textual information can be utilized.

[0052] Step S30: Encode the temporal structured data and / or non-temporal structured data to obtain the corresponding structured temporal representation and / or static context representation; Specifically, as one implementation method, time-series structured data (such as daily consumption) is encoded through a lightweight temporal convolutional network to obtain a structured temporal representation for each time step, which can capture short-term patterns and trends in customer behavior.

[0053] Non-time-series structured data (age, occupation, etc.) are encoded through an embedding layer (processing categorical data) and a fully connected layer (processing numerical data) to obtain a unified static context representation that represents the customer's long-term inherent attributes.

[0054] This step transforms the raw structured data into a high-order feature representation that can be directly processed by a deep learning model.

[0055] Step S40: Dynamically fuse at least one of the semantic factor vector, the text semantic vector, the structured temporal representation, and the static context representation in the semantic factor sequence through a gating mechanism to obtain a fused representation; In this embodiment, the task-time window conditional gating fusion mechanism enables the model to dynamically adjust the contribution of various features according to different prediction tasks, different prediction windows, and different object states.

[0056] Specifically, a gating fusion module can be designed. For the prediction point in time, this module receives the structured time-series representation, semantic factor vector, text semantic vector, and static context representation for that day. Simultaneously, the module receives the task identifier (e.g., a classification task of "whether it is overdue within 30 days") and the prediction window length (e.g., 30 days). Based on these inputs, the gating fusion module dynamically generates corresponding gating weights, corresponding to the four types of features: structured time-series representation, semantic factor vector, text semantic vector, and static context representation. For example, the gating weights might favor structured time-series representation and text semantic vector (reflecting immediate behavior and intent); for long-term risk prediction, the weights might favor semantic factor sequence (reflecting long-term trend signals) and static context representation. Then, the four types of feature representations are weighted and summed according to their weights to obtain the final fused representation.

[0057] This step enables the dynamic and conditional fusion of dual-channel text representation and structured features, allowing the model to more intelligently integrate all available information, intelligently adjust the emphasis on different information sources, and improve the flexibility of prediction.

[0058] Step S50: Input the fused representation into the multi-task prediction output module to obtain the result of at least one prediction task.

[0059] Specifically, a multi-task prediction output module will be integrated to represent the input. This multi-task prediction output module includes a shared prediction representation layer and multiple task-specific output heads. Different output heads are used for classification, regression, ranking, prediction at multiple future time points, or prediction of future time window statistics, respectively. The multi-time point prediction output head is used to output continuous predicted values ​​at multiple future time steps; the time window statistical prediction output head is used to output the prediction results of the maximum value, mean, quantile, or whether a threshold is exceeded within a specified future time window.

[0060] For example, if you need to predict the "probability curve of delinquency for each day in the next 7 days", activate the "Future Multi-Time Point Prediction" output header; if you need to predict the "maximum delinquency amount in the next 30 days", activate the "maximum value" prediction function in the "Time Window Statistical Prediction" output header.

[0061] Therefore, a unified framework can support multiple prediction tasks, improve system efficiency and reusability, and output a wide variety of prediction results to meet different business needs.

[0062] Through the above steps, information can be extracted deeply and interpretably from multi-source heterogeneous data (especially unstructured text), and accurate and flexible predictions of the future state of financial objects (customers) can be achieved through dynamic fusion and a unified multi-task framework. This embodiment as a whole realizes the technical effect of constructing a dual-channel text representation and dynamically fusing it with structured features to support rich multi-task prediction.

[0063] This embodiment's technical solution involves acquiring multi-source data of financial objects, including time-series structured data, untime-series structured data, and unstructured text data. Based on the unstructured text data, a semantic factor sequence is constructed using a large model with knowledge enhancement, and a text semantic vector is generated. The time-series structured data and / or untime-series structured data are encoded to obtain corresponding structured time-series representations and / or static context representations. At least one of the semantic factor vectors, text semantic vectors, structured time-series representations, and static context representations from the semantic factor sequence is dynamically fused using a gating mechanism to obtain a fused representation. The fused representation is then input into a multi-task prediction output module to obtain the result of at least one prediction task. This application targets unstructured text data and utilizes a large model to construct a semantic factor sequence based on knowledge enhancement. This enables text information to participate in prediction in an interpretable, alignable, and trainable form. Simultaneously, it encodes temporal structured data and / or untemporal structured data, and dynamically fuses the semantic factor vectors, the structured temporal representation, the text semantic vectors, and the static context representation in the semantic factor sequence through a gating mechanism. This allows the model to dynamically allocate feature weights according to different tasks and prediction windows. This application retains the ability of the large model to understand complex text and domain knowledge, while entrusting the final prediction process to a lightweight structured encoding network and a gating fusion network, thereby reducing real-time inference costs and improving the system's prediction efficiency, stability, and scalability.

[0064] Based on the first embodiment described above, a second embodiment of this application is proposed. For content identical to the first embodiment, please refer to the first embodiment above, and it will not be repeated here. The second embodiment of this application proposes a multi-source data fusion prediction method. In this embodiment, after the step of obtaining multi-source data of financial objects, the method further includes: Step S101: Preprocess the multi-source data of the financial object.

[0065] like Figure 2 As shown, the multi-source data of the financial object can be preprocessed through the data preprocessing module.

[0066] Specifically, the preprocessing of the time-series structured data includes performing at least one of the following preprocessing methods on the time-series structured data: missing value imputation, outlier handling, normalization, bucketing encoding, sliding window construction, and time position encoding, to obtain structured time-series features.

[0067] As mentioned earlier, time-series structured data includes numerical or categorical features recorded over time, such as transactions, behaviors, accounts, and market data. Examples include a customer's daily spending and repayment amounts.

[0068] The sliding window construction refers to taking the prediction time point t0 as the benchmark and extracting historical data of a fixed length L (such as 180 days) to form a sample within the observation window [t0-L+1,t0].

[0069] Missing value imputation, for example, refers to filling in the missing "spending amount" for a certain day with the average of the past 7 days.

[0070] Outlier handling refers to identifying and correcting data that is clearly illogical, such as a single transaction exceeding the credit limit by 10 times.

[0071] Normalization scales numerical features such as "consumption amount" and "repayment amount" to the [0,1] range to eliminate the influence of dimensions.

[0072] Bucket coding divides the "number of transactions" into "low frequency (0-2)", "medium frequency (3-5)" and "high frequency (>5)" and performs one-hot coding.

[0073] Time location encoding refers to adding a periodic code (such as day of the week, whether it is the beginning or end of the month) to each day within a window, so that the model can understand the time pattern.

[0074] Therefore, by preprocessing multi-source data such as temporal structured data, untemporal structured data, and unstructured text data, high-quality and model-readable structured temporal features are obtained, providing clean and standardized input for temporal coding networks.

[0075] The preprocessing of the non-temporally structured data includes performing category encoding, standardization, or embedding mapping on the non-temporally structured data to obtain static structured features. As mentioned earlier, the non-temporally structured data includes features such as object attributes, product attributes, risk attributes, and category labels.

[0076] Taking credit card customers as an example, non-time-series structured data can refer to a customer's age, occupation, annual income, and card level. Categorical encoding refers to using labels to encode categorical features such as "occupation." Standardization refers to Z-score standardization of "annual income." Embedding mapping refers to learning a low-dimensional, dense embedding vector for high-cardinality ordered categories such as "card level" (e.g., standard card, gold card, platinum card).

[0077] This transforms various types of static data into unified numerical static structured features, making them easier for neural networks to process.

[0078] The preprocessing of the unstructured text data includes performing at least one of the following preprocessing steps: cleaning, deduplication, sentence segmentation, timestamp parsing, object identification, source marking, and anonymization on the unstructured text data to obtain a set of text fragments.

[0079] As mentioned earlier, unstructured text data includes textual information such as communication records, feedback texts, product descriptions, announcements, reports, and public opinion. This includes, for example, customer service records and customer emails. Data cleaning refers to removing irrelevant characters and garbled text. Deduplication involves merging customer service records with identical content. Sentence segmentation involves breaking long paragraphs into independent semantic sentences. Timestamp parsing involves extracting the precise generation time of each text entry from the logs, which is crucial for subsequent alignment with the business timeline. Object identification can involve using named entity recognition technology to confirm the customer ID or card number mentioned in the text. Source tagging can involve marking the source of the text as "telephone customer service," "online customer service," or "email," which can be used to assess the credibility of the information later. Anonymization involves hiding personal identification information (such as ID card numbers and mobile phone numbers) in the text.

[0080] Therefore, by preprocessing unstructured text data, a clean, segmented set of text fragments with temporal and object labels is obtained, laying the foundation for accurate extraction of semantic factors in subsequent large-scale models.

[0081] Through the above steps, the original multi-source heterogeneous data was cleaned, transformed, and standardized, providing high-quality, standardized input for subsequent feature extraction, semantic understanding, and fusion prediction modules. This is a fundamental step in ensuring the prediction accuracy and stability of the entire system. This embodiment achieves the technical effects of data standardization and lossless information preservation, clearing obstacles for downstream deep processing.

[0082] Based on the first or second embodiment described above, a third embodiment of this application is proposed. For content that is the same as the first or second embodiment, please refer to the first or second embodiment described above, and it will not be repeated here. The third embodiment of this application proposes a multi-source data fusion prediction method, which refines the above step S20.

[0083] In one feasible implementation, step S20, which involves constructing a semantic factor sequence based on knowledge enhancement using a large model based on the unstructured text data, includes steps S201-S202: Step S201: Using a large model combined with financial domain knowledge, extract knowledge-enhanced semantic factor records from the unstructured text data. Each semantic factor record includes at least object identifier, factor category, semantic direction, factor strength, confidence level, occurrence time, and effective time range. Specifically, this step can be performed through the knowledge-enhanced semantic factor extraction module.

[0084] As one implementation method, the knowledge-enhanced semantic factor extraction module includes a domain knowledge constraint unit, a large-scale model semantic understanding unit, and a semantic factor generation unit. The domain knowledge constraint unit provides constraints such as factor categories, object relationships, business timelines, task objectives, and factor value ranges; the large-scale model semantic understanding unit analyzes text fragments; and the semantic factor generation unit organizes the output of the large-scale model into structured semantic factor records.

[0085] For each text segment, the large model generates a semantic factor record under the constraints of domain knowledge. This record includes at least the object identifier, factor category, semantic direction, factor strength, confidence level, occurrence time, effective time range, source identifier, evidence text, and semantic vector.

[0086] It should be noted that fields such as direction, intensity, confidence, and validity period in semantic factor records can be generated by a large model under domain knowledge constraints, or calibrated by combining rules, knowledge graphs, or historical business tags. This application's embodiments do not limit the specific extraction model or the number of fields.

[0087] As a result, textual information is no longer compressed into black-box vectors, but forms semantic factor records that can participate in temporal modeling.

[0088] Step S202: Using the business timeline of the financial object as an index, the semantic factor records are mapped to a semantic factor sequence with the same granularity as the structured time-series data, wherein the semantic factor vector of each time step is obtained by weighted summation of the relevant semantic factor records.

[0089] Specifically, this step can be performed using the semantic factor sequence building module.

[0090] In one implementation method, the semantic factor sequence construction module uses financial objects and business timelines as indexes to map discrete semantic factor records into a semantic factor sequence with the same granularity as structured time-series data. The business timeline can be set to calendar day, transaction day, payment period, customer lifecycle stage, product launch period, marketing outreach cycle, or other business time granularities, and its function is to provide a clear time position and effective scope for text semantics.

[0091] In one implementation method, for any financial object and time step, the semantic factor sequence construction module filters candidate semantic factors related to the financial object, time step, and prediction task, and calculates the impact weights based on time relationships, object relationships, source credibility, factor confidence, and historical business response calibration parameters; the candidate semantic factors are weighted and summarized according to the impact weights to obtain the semantic factor vector corresponding to the time step; and a semantic factor sequence corresponding to each time step within the observation window is obtained based on the semantic factor vector.

[0092] More specifically, for any object and time step, the system filters candidate factors related to the object, time step, and prediction task, and calculates the impact weights based on temporal relationships, object relationships, source credibility, factor confidence, and historical business response calibration parameters. Subsequently, the candidate semantic factors are weighted and aggregated according to their impact weights to obtain the semantic factor vector corresponding to that time step. Through this step, the system obtains a semantic factor sequence that corresponds one-to-one with each time step within the observation window. This transforms discrete, irregularly timed text events into continuous temporal features strictly aligned with the timeline of structured data, solving the core problem of temporal mismatch between text and structured data.

[0093] This embodiment, through the above-described scheme, achieves the precise transformation of free text into structured predictive factors with rich attributes (direction, intensity, and time) using the external knowledge of a large model, and then organizes it into a time-series according to business rhythm. Overall, this embodiment achieves the technical effects of structuring, temporalizing, and interpreting text information.

[0094] Furthermore, the scheme of constructing semantic factor sequences based on knowledge enhancement using a large model based on the unstructured text data may also include at least one of the following: A hard alignment method based on fixed time granularity is adopted; By learning a soft alignment method for weight allocation across multiple time steps; The persistent impact of semantic factors can be controlled by using decay kernel functions, validity windows, or historical response calibration parameters.

[0095] Among them, hard alignment, soft alignment, and decay kernel function are all methods that define a semantic factor record within the observation window [t0-L+1,t0] to determine how a semantic factor record affects the semantic factor vector several time steps before and after its occurrence.

[0096] Specifically, a hard alignment method using fixed-time granularity mapping is adopted, meaning that semantic factor records are strictly mapped to the specific time step to which their occurrence time belongs, according to a fixed time granularity (e.g., "day"). For example, a semantic factor record occurring at 14:30 on 2025-05-20 only contributes to the semantic factor vector for the day 2025-05-20. Its advantages are simplicity and strict temporal alignment.

[0097] Soft alignment, which involves learning weights to allocate across multiple time steps, refers to the ability to learn weights to distribute the impact of a semantic factor record (such as "expected project payment next quarter") across multiple time steps. For example, a model can automatically learn through a learnable attention mechanism that this "expected payment" record contributes not only to the month of the "occurrence quarter" but also to the following two months with weights of 0.6, 0.3, and 0.1, respectively. Its purpose is to more flexibly model events with continuous or delayed effects, with the alignment learned through data-driven learning.

[0098] The use of decay kernel functions, validity windows, or historical business response calibration parameters to control the sustained impact of semantic factors refers to the following: To model the decay effect of semantic factors, decay kernel functions (such as exponential decay) can be used to reduce the factor strength over time. For example, the impact factor strength of a negative event can be set to decay by 5% per day, or its scope of influence can be clearly defined using a validity window. Furthermore, historical business response calibration parameters can be used to learn the actual impact patterns of different types of semantic factors (such as an initial increase followed by a decrease) from real business data, which can then be used to dynamically adjust their weights at different lag periods. This allows for a more accurate characterization of the lifecycle and dynamic impact patterns of textual information, improving the time-series accuracy of predictions.

[0099] Through the above steps, a variety of flexible schemes for constructing semantic factor sequences are provided, allowing the selection of the most suitable alignment and impact modeling method based on the complexity (instantaneous, continuous, delayed, periodic) of event impact patterns in specific business scenarios. This embodiment as a whole achieves the technical effect of diversity and adaptability in semantic factor temporal modeling methods.

[0100] In one feasible implementation, step S30, which encodes the temporal structured data and / or non-temporal structured data to obtain the corresponding structured temporal representation and / or static context representation, includes steps S301-S302: Step S301: For the temporal structured data, a lightweight temporal coding network is used to generate a structured temporal representation. The lightweight temporal coding network includes at least one of a numerical projection layer, a temporal embedding layer, a self-attention layer, a convolutional layer, and a recurrent layer. Specifically, the time-series structured data and / or non-time-series structured data can be encoded using a structured feature encoding module.

[0101] For example, when processing daily feature sequences (such as spending, repayment amount, number of transactions, etc.) from the past 6 months, a specific encoding network instance is as follows: Numerical projection layer: The preprocessed numerical features are mapped to a higher-dimensional vector space through a fully connected layer, thereby enhancing the representation capability.

[0102] Temporal embedding layer: The time location encoding (such as day of the week, month of the month identifier) ​​is also converted into a vector and added to the numerically projected features to inject temporal information.

[0103] Convolutional layer: Uses a one-dimensional convolutional kernel to perform convolution operations on the sequence to extract local patterns (such as consumption trends over several consecutive days).

[0104] Self-attention layer: allows features from any two time steps in the sequence to interact, capturing long-term dependencies (such as the association between monthly consumption and end-of-month repayment).

[0105] Recurrent layers (such as GRU): process sequences in time step order and model the sequential dependencies of sequences.

[0106] Ultimately, the network outputs a structured temporal representation for each time step, which is a feature vector that condenses the information of that moment and its historical context. This step allows for the automatic learning and extraction of deep temporal patterns from the behavioral sequences of financial entities.

[0107] Step S302: For the non-temporally structured data, a static context representation is generated through an embedding layer or a multilayer perceptron.

[0108] Specifically, the static attribute vectors of customers (such as encoded occupation, standardized annual income, and credit card level embedding vectors) are encoded using a multilayer perceptron (MLP, a network composed of multiple fully connected layers and nonlinear activation functions). The MLP performs nonlinear transformation and compression on the concatenated static feature vectors, ultimately outputting a fixed-dimensional static context representation. This transforms the static profile of a financial object into an information-rich context vector, providing stable background information for dynamic time-series prediction.

[0109] This embodiment achieves efficient and in-depth feature learning for both temporal dynamic features and static background features through the above-described scheme, transforming them into high-quality representations suitable for subsequent fusion module processing. Overall, this embodiment achieves the technical effect of automatic extraction of deep features from structured data.

[0110] In one feasible implementation, step S40, which dynamically fuses at least one of the semantic factor vectors in the semantic factor sequence, the text semantic vectors, the structured temporal representation, and the static context representation through a gating mechanism to obtain a fused representation, includes steps S401-S402: Step S401: For each time step, the structured temporal representation, semantic factor vector in the semantic factor sequence, text semantic vector and static context representation are received through the gating fusion module, and the corresponding gating weights are generated by combining the task identifier and prediction window. The gating fusion module dynamically determines the contribution of different feature sources under different tasks and prediction windows. Specifically, it dynamically allocates multi-source feature weights based on the task and prediction window. The fusion weights of this module are influenced by the task, time window, and object state.

[0111] Specifically, for the current prediction time step t0, the gated fusion module receives four sets of inputs: the structured time series representation at that moment, the semantic factor vector, the text semantic vector, and the static context representation. Simultaneously, the gated fusion module also receives two conditional signals: a task identifier (e.g., whether to predict "whether to trade tomorrow" or "whether to overdue next month") and a prediction window length (e.g., 1 day or 30 days).

[0112] The gating fusion module contains a small neural network (such as a fully connected layer) that takes four types of feature representations and conditional signals as input, calculates and outputs four sets of gating weights (usually normalized by Softmax, with the weights summing to 1).

[0113] For example, when predicting "transaction fraud tomorrow," the weights might focus on "today's transaction behavior" (structured time-series representation) and "current session intent" (textual semantic vector); when predicting "customer churn next quarter," the weights might be more inclined towards "long-term risk signals" (semantic factor sequence) and "customer attributes" (static context representation). This enables the intelligent conditionalization of fusion strategies.

[0114] Step S402: Based on the gating weights, the semantic factor vector, the structured temporal representation, the text semantic vector, and the static context representation are weighted and fused to obtain a fused representation.

[0115] Specifically, the four feature representations are multiplied by their corresponding gating weights, and then summed to obtain the final fused representation. The same semantic factor vector has different fusion weights under different task labels or different prediction window lengths. For example, the semantic factor "customer inquires about installment payments for large purchases" might have a low weight in the task of predicting "short-term fraud risk" (because of its weak association with fraud); however, it might have a high weight in the task of predicting "willingness to make installments in the next month." Similarly, it might have a low weight in the task of predicting "whether to make a purchase tomorrow" but a high weight in the task of predicting "total consumption in the next year." Thus, a dynamically weighted unified feature representation can be generated, which selectively aggregates the most important information within the current prediction objective and time perspective, achieving refined and adaptive fusion.

[0116] Through the above steps, a dynamic and refined fusion of four types of feature representations based on the prediction task and window is achieved. This embodiment comprehensively achieves the technical effect of adaptively adjusting the information source weights according to the prediction target and time span, greatly enhancing the model's adaptability and prediction accuracy in complex multi-task scenarios.

[0117] In one feasible implementation, the method further includes the following during the training phase: Step S100: During the training phase, the large model is jointly trained by multi-task supervision loss and combined with semantic factor consistency loss, gating constraint and robustness loss; the semantic factor consistency loss is used to constrain the correspondence between semantic factors and historical business responses.

[0118] In the training or offline modeling phase, the large model is jointly trained by multi-task supervision loss combined with semantic factor consistency loss, gating constraints and robustness loss.

[0119] The multi-task supervision loss is the main loss function. For classification tasks (such as whether a payment is overdue), cross-entropy loss is used; for regression tasks (such as predicting the amount), mean squared error loss is used; and for ranking tasks, pairwise ranking loss is used. The total supervision loss is obtained by weighted summation of the losses from all tasks. This drives the model to achieve the basic prediction objective.

[0120] The semantic factor consistency loss is used to constrain the correspondence between semantic factors and historical business responses. For example, for a "negative" and high-intensity "income risk" semantic factor, if actual business data (such as repayment rates) does decline within the time period following its occurrence, the loss decreases; otherwise, the loss increases. This forces the model to learn semantic factors with realistic business predictive capabilities, enhancing its interpretability and reliability.

[0121] Among them, gating constraints can include constraints such as L1 regularization, which makes the gating weights tend to be sparse. That is, in some scenarios, the model can learn to rely entirely on one or two types of features, thereby improving the interpretability and stability of the gating mechanism and preventing overfitting.

[0122] For robustness loss, during training, some input features can be randomly masked (e.g., randomly discarding text or structured data from a few days), requiring the model to still make reasonable predictions. Robustness loss is used to improve the model's stability under conditions of missing text, text delay, or missing structured features. This improves the model's robustness to data loss, delay, or noise in real-world scenarios.

[0123] This embodiment, through the above-described scheme, achieves joint training of the model using multiple loss functions, which not only optimizes prediction accuracy but also simultaneously improves the quality of semantic factors, the interpretability of the gating mechanism, and the robustness of the entire system. Overall, this embodiment achieves the technical effects of multi-objective collaborative optimization of model training and enhanced generalization ability.

[0124] Furthermore, the method also includes: In the offline phase, a large model is used to extract semantic factor records from unstructured text data, generate semantic factor sequences, and store semantic factor records, semantic factor sequences, and text semantic vectors. During the online phase, the latest structured features and generated semantic factor sequences are read, and lightweight encoding, gated fusion, and multi-task prediction output are performed.

[0125] The large model mentioned herein is at least one of the following: a general large language model, a large financial model, an encoder-based pre-trained model, a rule extraction model, and a knowledge graph reasoning model. This embodiment does not specifically limit the specific model to this one.

[0126] As one implementation method, the technical solution of this embodiment is suitable for deployment combining offline and online methods. Text processing and semantic factor generation can be completed offline or semi-offline, while the online phase mainly performs lightweight feature encoding, fusion, and output, thus balancing the semantic capabilities of large models with the real-time requirements of financial transactions.

[0127] Specifically, in the offline phase, a large model is used to extract semantic factor records from unstructured text data, generating semantic factor sequences. In an offline environment, the system uses a large model to batch process massive amounts of historical and non-real-time streaming text data (such as customer emails, public opinion reports, and announcements), extracting semantic factor records. For example, it can run once every night to process all new text data added that day.

[0128] Simultaneously, semantic factor records, semantic factor sequences, and text semantic vectors are stored. Based on the business timeline, a historical semantic factor sequence is generated for each financial object. These semantic factor records, semantic factor sequences, and the original text semantic vectors generated by the large model are then stored together in a high-performance feature database or cache. This allows the computationally intensive and time-consuming deep learning text processing to be pre-processed, avoiding real-time online calls to large models.

[0129] During the online phase, the latest structured features and generated semantic factor sequences are read. When a prediction of a financial object is needed at a real-time prediction point t0, the online service obtains the latest structured features of the object (such as transaction flow up to t0) from the business system in real time. At the same time, the semantic factor sequence of the object, which has been calculated within the observation window [t0-L+1,t0], is quickly read from the feature storage.

[0130] Finally, lightweight encoding, gated fusion, and multi-task prediction output are performed. The data from these two parts is input into the lightweight encoding network, gated fusion module, and multi-task prediction output module deployed in the online service. These modules have a small number of parameters and fast inference speed, outputting prediction results within millisecond latency. For example, in a real-time credit card transaction anti-fraud scenario, when a transaction occurs, the system can complete the entire process from reading features to outputting a risk score within tens of milliseconds. Thus, the online service performs only lightweight computation, achieving low-latency, high-concurrency real-time prediction while balancing the semantic capabilities of large models with the real-time requirements of financial transactions.

[0131] This embodiment achieves a hierarchical collaborative architecture that utilizes the above-described scheme to enable deep offline understanding of a large model and efficient online prediction of a small model. Overall, this embodiment achieves the technical effect of meeting the low latency and high concurrency requirements of industrial-grade real-time prediction while maintaining powerful semantic understanding capabilities.

[0132] Compared with the prior art, the technical solution of this embodiment has the following advantages: 1) Clearer semantic expression This method transforms textual information into semantic factor records, rather than simply generating black-box text vectors. Semantic factors contain information such as object, time, direction, intensity, and confidence level, which facilitates model use, business interpretation, and result traceability.

[0133] 2) More accurate time alignment This method constructs a semantic factor sequence based on the business timeline, enabling unstructured text to be mapped to the corresponding time step or business stage, thereby reducing the risk of mismatch between text information and structured temporal features.

[0134] 3) More flexible integration methods This method integrates semantic factors, text semantic vectors, and structured features through a task-time window conditional gating mechanism, enabling the model to dynamically adjust the contribution of various features according to different task objectives and prediction windows.

[0135] 4) Enhanced multi-tasking support capabilities This method supports classification, regression, ranking, multi-time point prediction, and time window statistical prediction within a unified framework, avoiding redundant modeling for different tasks and improving system efficiency and task coordination capabilities.

[0136] 5) Higher deployment efficiency This method adopts a hierarchical structure of "large model responsible for semantic factor generation and small model responsible for fusion prediction", which can cache text processing and semantic factor generation in advance, and only perform lightweight encoding and prediction in the online stage, reducing the cost of real-time inference.

[0137] It should be noted that, to meet different application scenarios, this application embodiment can provide a variety of alternative implementation methods: 1) Regarding semantic factor extraction, the large model can be replaced by a general large language model, a large financial domain model, an encoder-based pre-trained model, a rule extraction model, a knowledge graph reasoning model, or a combination of the above models. Semantic factor fields can be added or removed according to task needs, as long as basic information such as object, time, factor meaning, and influence strength are retained.

[0138] 2) In terms of semantic factor sequence construction, a hard alignment method with fixed time granularity mapping can be adopted, or a soft alignment method with learnable weights distributed to multiple time steps can be adopted. Alternatively, a decay kernel function, validity window, or historical response calibration parameter can be used to control the continuous influence of semantic factors.

[0139] 3) In terms of structured feature encoding and fusion, lightweight temporal coding networks can adopt Transformer, temporal convolutional network, recurrent neural network, state space model, multilayer perceptron or combination thereof; task-time window conditional gating network can be replaced by cross attention network, expert hybrid network, feature selection network or other learnable fusion structure.

[0140] 4) In terms of output and deployment, this application can output a single prediction result or a multi-task result; it can call a large model to generate semantic factors in real time or generate and cache semantic factors offline; when the text is missing, it can use only structured features or use historical semantic factors or default semantic factors to supplement it.

[0141] The above alternatives do not change the core idea of ​​this application, which is to use large models and domain knowledge to transform unstructured text into semantic factor sequences, and to use them together with structured financial features through a learnable fusion mechanism to complete multi-task, multi-time-window prediction of financial objects.

[0142] Furthermore, to achieve the above objectives, this application also proposes a multi-source data fusion prediction device, such as... Figure 3 As shown, the device includes: The data acquisition module 10 is used to acquire multi-source data of financial objects, wherein the multi-source data includes at least one of time-series structured data, non-time-series structured data and unstructured text data; The semantic factor sequence construction module 20 is used to construct a semantic factor sequence based on knowledge enhancement using a large model based on the unstructured text data, and generate a text semantic vector. The structured feature encoding module 30 is used to encode the temporal structured data and / or non-temporal structured data to obtain the corresponding structured temporal representation and / or static context representation; The gating fusion module 40 is used to dynamically fuse at least one of the semantic factor vector, the text semantic vector, the structured temporal representation, and the static context representation in the semantic factor sequence through a gating mechanism to obtain a fused representation; The multi-task prediction output module 50 is used to perform multi-task prediction based on the fused representation to obtain the result of at least one prediction task.

[0143] The multi-source data fusion prediction device provided in this application employs the multi-source data fusion prediction method described in the above embodiments. This method retains the ability of large models to understand complex text and domain knowledge while delegating the final prediction process to a lightweight structured coding network and a gated fusion network. This reduces real-time inference costs and improves system prediction efficiency, stability, and scalability. Compared with existing technologies, the beneficial effects of the multi-source data fusion prediction device provided in this application are the same as those of the multi-source data fusion prediction method described in the above embodiments. Furthermore, other technical features of the multi-source data fusion prediction device are the same as those disclosed in the methods of the above embodiments, and will not be elaborated upon here.

[0144] In addition, to achieve the above objectives, this application also proposes a multi-source data fusion prediction device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the multi-source data fusion prediction method as described above.

[0145] The following is for reference. Figure 4 This document illustrates a structural schematic diagram of a multi-source data fusion prediction device suitable for implementing embodiments of this application. The multi-source data fusion prediction device in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 4 The multi-source data fusion prediction device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0146] like Figure 4As shown, the multi-source data fusion prediction device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory 1002 or a program loaded from a storage device 1003 into a random access memory 1004. The random access memory 1004 also stores various programs and data required for the operation of the multi-source data fusion prediction device. The processing unit 1001, the read-only memory 1002, and the random access memory 1004 are interconnected via a bus 1005. An input / output interface 1006 is also connected to the bus. Typically, the following systems can be connected to the interface 1006: input devices 1007 including, for example, a touch screen, touchpad, keyboard, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 1003 including, for example, magnetic tape, hard disk, etc.; and communication devices 1009. Communication device 1009 allows the multi-source data fusion prediction device to communicate wirelessly or wiredly with other devices to exchange data. While the figure shows multi-source data fusion prediction devices with various systems, it should be understood that implementation or possession of all the systems shown is not required. More or fewer systems may be implemented alternatively.

[0147] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0148] The multi-source data fusion prediction device provided in this application employs the multi-source data fusion prediction method described in the above embodiments. This method retains the ability of large models to understand complex text and domain knowledge while delegating the final prediction process to a lightweight structured coding network and a gated fusion network. This reduces real-time inference costs and improves system prediction efficiency, stability, and scalability. Compared with existing technologies, the beneficial effects of the multi-source data fusion prediction device provided in this application are the same as those of the multi-source data fusion prediction method described in the above embodiments. Furthermore, other technical features of this multi-source data fusion prediction device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0149] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0150] In addition, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the multi-source data fusion prediction method described above.

[0151] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0152] The aforementioned computer-readable storage medium may be included in the multi-source data fusion prediction device; or it may exist independently and not assembled into the multi-source data fusion prediction device.

[0153] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by the multi-source data fusion prediction device, the multi-source data fusion prediction device: acquires multi-source data of financial objects, including time-series structured data, untime-series structured data, and unstructured text data; constructs a semantic factor sequence based on knowledge enhancement using a large model based on the unstructured text data, and generates a text semantic vector; encodes the time-series structured data and / or untime-series structured data to obtain corresponding structured time-series representations and / or static context representations; dynamically fuses at least one of the semantic factor vectors, text semantic vectors, structured time-series representations, and static context representations in the semantic factor sequence through a gating mechanism to obtain a fused representation; and inputs the fused representation into a multi-task prediction output module to obtain the result of at least one prediction task. This application targets unstructured text data and utilizes a large model to construct a semantic factor sequence based on knowledge enhancement. This enables text information to participate in prediction in an interpretable, alignable, and trainable form. Simultaneously, it encodes temporal structured data and / or untemporal structured data, and dynamically fuses the semantic factor vectors, the structured temporal representation, and the static context representation in the semantic factor sequence through a gating mechanism. This allows the model to dynamically allocate feature weights according to different tasks and prediction windows. This application retains the ability of the large model to understand complex text and domain knowledge, while entrusting the final prediction process to a lightweight structured encoding network and a gating fusion network, thereby reducing real-time inference costs and improving the system's prediction efficiency, stability, and scalability.

[0154] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0155] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0156] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0157] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., computer programs) for executing the above-described multi-source data fusion prediction method. This preserves the ability of large models to understand complex text and domain knowledge while delegating the final prediction process to lightweight structured coding networks and gated fusion networks, thereby reducing real-time inference costs and improving system prediction efficiency, stability, and scalability. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the multi-source data fusion prediction method provided in the above embodiments, and will not be repeated here.

[0158] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the multi-source data fusion prediction method described above.

[0159] The computer program product provided in this application retains the ability of large models to understand complex text and domain knowledge, while delegating the final prediction process to a lightweight structured coding network and a gated fusion network, thereby reducing real-time inference costs and improving the system's prediction efficiency, stability, and scalability. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the multi-source data fusion prediction method provided in the above embodiments, and will not be repeated here.

[0160] One or more technical solutions proposed in this application have at least the following technical effects: By acquiring multi-source data of financial objects, the multi-source data includes time-series structured data, untime-series structured data, and unstructured text data; based on the unstructured text data, a semantic factor sequence is constructed using a large model based on knowledge enhancement, and a text semantic vector is generated; the time-series structured data and / or untime-series structured data are encoded to obtain corresponding structured time-series representations and / or static context representations; at least one of the semantic factor vectors, text semantic vectors, structured time-series representations, and static context representations in the semantic factor sequence is dynamically fused through a gating mechanism to obtain a fused representation; the fused representation is input into a multi-task prediction output module to obtain the result of at least one prediction task. This application targets unstructured text data and utilizes a large model to construct a semantic factor sequence based on knowledge enhancement. This enables text information to participate in prediction in an interpretable, alignable, and trainable form. Simultaneously, it encodes temporal structured data and / or untemporal structured data, and dynamically fuses the semantic factor vectors, the structured temporal representation, and the static context representation in the semantic factor sequence through a gating mechanism. This allows the model to dynamically allocate feature weights according to different tasks and prediction windows. This application retains the ability of the large model to understand complex text and domain knowledge, while entrusting the final prediction process to a lightweight structured encoding network and a gating fusion network, thereby reducing real-time inference costs and improving the system's prediction efficiency, stability, and scalability.

[0161] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A multi-source data fusion prediction method, characterized in that, The method includes: Obtain multi-source data of financial objects, including time-series structured data, non-time-series structured data, and unstructured text data; Based on the unstructured text data, a semantic factor sequence is constructed using a large model based on knowledge enhancement, and a text semantic vector is generated. The time-series structured data and / or non-time-series structured data are encoded to obtain the corresponding structured time-series representation and / or static context representation; The semantic factor vector, the text semantic vector, the structured temporal representation, and the static context representation in the semantic factor sequence are dynamically fused through a gating mechanism to obtain a fused representation. The fused representation is input into the multi-task prediction output module to obtain the result of at least one prediction task.

2. The method as described in claim 1, characterized in that, The step of constructing a semantic factor sequence based on knowledge enhancement using a large model based on the unstructured text data includes: By leveraging large models and knowledge from the financial field, knowledge-enhanced semantic factor records are extracted from the unstructured text data. Each semantic factor record includes at least the object identifier, factor category, semantic direction, factor strength, confidence level, occurrence time, and effective time range. Using the business timeline of the financial object as an index, the semantic factor records are mapped to a semantic factor sequence with the same granularity as the structured time-series data, wherein the semantic factor vector of each time step is obtained by weighted summation of the relevant semantic factor records.

3. The method as described in claim 2, characterized in that, The step of mapping the semantic factor records to a semantic factor sequence with the same granularity as the structured time-series data, using the business timeline of the financial object as an index, wherein the semantic factor vector of each time step is obtained by weighted summation of relevant semantic factor records includes: For any financial object and time step, candidate semantic factors related to the financial object, time step and prediction task are screened, and the role weights are calculated based on time relationship, object relationship, source credibility, factor confidence and historical business response calibration parameters. The candidate semantic factors are weighted and aggregated according to their respective weights to obtain the semantic factor vector corresponding to the time step. Based on the semantic factor vector, a semantic factor sequence corresponding one-to-one with each time step within the observation window is obtained.

4. The method as described in claim 1, characterized in that, The step of encoding the temporal structured data and / or non-temporal structured data to obtain the corresponding structured temporal representation and / or static context representation includes: For the aforementioned temporal structured data, a lightweight temporal coding network is used to generate a structured temporal representation. The lightweight temporal coding network includes at least one of the following: a numerical projection layer, a temporal embedding layer, a self-attention layer, a convolutional layer, and a recurrent layer. For the non-temporally structured data, a static context representation is generated through an embedding layer or a multilayer perceptron.

5. The method as described in claim 1, characterized in that, The step of dynamically fusing at least one of the semantic factor vectors in the semantic factor sequence, the text semantic vectors, the structured temporal representation, and the static context representation through a gating mechanism to obtain the fused representation includes: For each time step, the structured temporal representation, semantic factor vectors in the semantic factor sequence, text semantic vectors, and static context representation are received through the gating fusion module, and corresponding gating weights are generated by combining the task identifier and prediction window. Based on the gating weights, the semantic factor vector, the structured temporal representation, the text semantic vector, and the static context representation are weighted and fused to obtain a fused representation.

6. The method as described in claim 5, characterized in that, The same semantic factor vector has different fusion weights under different task identifiers or different prediction window lengths.

7. The method as described in claim 1, characterized in that, The multi-task prediction output module includes a shared prediction representation layer and multiple task-specific output heads. Different output heads are used for classification, regression, ranking, prediction of multiple future time points, or prediction of future time window statistics, respectively. The multi-time point prediction output head is used to output continuous prediction values ​​for multiple future time steps. The time window statistical prediction output head is used to output the prediction results of the maximum value, mean, quantile, or whether the threshold is exceeded within a specified future time window.

8. The method as described in claim 1, characterized in that, The method further includes the following during the training phase: During the training phase, the large model is jointly trained by multi-task supervised loss combined with semantic factor consistency loss, gating constraints, and robustness loss; the semantic factor consistency loss is used to constrain the correspondence between semantic factors and historical business responses.

9. The method as described in claim 1, characterized in that, Following the step of obtaining multi-source data on financial objects, the following also includes: Perform at least one of the following preprocessing steps on the time-series structured data: missing value imputation, outlier handling, normalization, bucketing encoding, sliding window construction, and time position encoding, to obtain structured time-series features; Perform category encoding, normalization, or embedding mapping on the non-temporal structured data to obtain static structured features; Perform at least one of the following preprocessing steps on the unstructured text data: cleaning, deduplication, sentence segmentation, timestamp parsing, object identification, source tagging, and anonymization, to obtain a set of text fragments.

10. The method as described in claim 1, characterized in that, The method further includes: In the offline phase, a large model is used to extract semantic factor records from unstructured text data, generate semantic factor sequences, and store semantic factor records, semantic factor sequences, and text semantic vectors. During the online phase, the latest structured features and generated semantic factor sequences are read, and lightweight encoding, gated fusion, and multi-task prediction output are performed.

11. The method according to any one of claims 1-10, characterized in that, The large model is at least one of the following: general large language model, financial domain large model, encoder-based pre-trained model, rule extraction model, and knowledge graph reasoning model.

12. The method as described in claim 2 or 3, characterized in that, The step of constructing a semantic factor sequence based on knowledge enhancement using a large model based on the unstructured text data further includes at least one of the following: A hard alignment method based on fixed time granularity is adopted; By learning a soft alignment method for weight allocation across multiple time steps; The persistent impact of semantic factors can be controlled by using decay kernel functions, validity windows, or historical response calibration parameters.

13. A multi-source data fusion prediction device, characterized in that, The device includes: The data acquisition module is used to acquire multi-source data of financial objects, wherein the multi-source data includes at least one of time-series structured data, non-time-series structured data, and unstructured text data; The semantic factor sequence construction module is used to construct a semantic factor sequence based on knowledge enhancement using a large model based on the unstructured text data, and generate text semantic vectors. The structured feature encoding module is used to encode the temporal structured data and / or non-temporal structured data to obtain the corresponding structured temporal representation and / or static context representation; The gating fusion module is used to dynamically fuse at least one of the semantic factor vectors, the text semantic vectors, the structured temporal representation, and the static context representation in the semantic factor sequence through a gating mechanism to obtain a fused representation; A multi-task prediction output module is used to perform multi-task prediction based on the fused representation to obtain the result of at least one prediction task.

14. A multi-source data fusion prediction device, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the multi-source data fusion prediction method as described in any one of claims 1 to 10.

15. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the multi-source data fusion prediction method as described in any one of claims 1 to 10.

16. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the steps of the multi-source data fusion prediction method as described in any one of claims 1 to 10.

Images