Financial user portrait intelligent decision-making system based on multi-modal fusion

Abstract

The invention discloses a financial user portrait intelligent decision-making system based on multi-modal fusion, and relates to the technical field of artificial intelligence data processing. According to the system, a data processing module collects financial service data and carries out preprocessing and fusion to generate unified user portrait features; a label generation module generates a user value label and label text interpretation based on the features; the portrait construction module combines value labels according to scene configuration rules to generate scene scores and construct financial user portraits; the intention portrait module generates a prediction window intention label write-in portrait based on portrait features and a historical behavior sequence; and the intelligent decision-making module fuses the real-time behavior characteristics and portrait reasoning to generate a decision-making result, and writes back and updates the portrait, the label and the explanation. Through multi-modal unified feature fusion, scene rule configurable portrait construction and intention prediction, and real-time decision-driven portrait and label closed-loop updating, a traceable and evolvable user portrait intelligent decision-making system for financial services is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence data processing, in particular to a financial user portrait intelligent decision system based on multi-modal fusion. BACKGROUND

[0002] With the promotion of digital transformation of financial institutions, a large amount of data is deposited in business systems in the aspects of transaction, credit, channel interaction and customer service, including structured data such as account and transaction flow, and unstructured data such as communication text, voice call, image of certificate or voucher. How to uniformly represent multi-source heterogeneous data and form an updateable user cognitive basis has become a key support for precise marketing, risk control and service operation.

[0003] Existing user portrait schemes are mostly built on big data platforms and static label systems, usually using batch data integration and offline feature processing, and the processing cycle is mainly daily or longer. For unstructured data such as text and image, the processing is usually limited to keyword extraction and character recognition, and the fusion is mostly table-level aggregation or association splicing, which is difficult to achieve uniform expression and consistency constraint at the feature level.

[0004] At the label system level, existing technologies generally rely on artificial rules or fixed models to generate labels, with long update cycle, weak scalability and insufficient adaptability to business changes. In marketing and risk control scenarios, the static label system cannot support users to present differentiated features in different business scenarios, so traditional portrait systems are generally in the state of "one face for thousands of people".

[0005] At the application level, the business scenarios of financial institutions are complex and diverse, including wealth management, loan approval, anti-fraud, enterprise due diligence, customer churn prediction, etc., but existing portrait systems cannot dynamically assemble user portraits according to different scenarios. Lack of real-time and flexibility, unable to meet the requirements of generating scenario portraits triggered by business events, and unable to support users to present differentiated portraits among multiple scenarios.

[0006] Therefore, there is an urgent need for a new decision system that can realize multi-modal data deep fusion, dynamic label generation, scenario-based portrait construction, intent recognition and real-time decision linkage, supporting the intelligent, real-time and fine development of financial business. SUMMARY

[0007] Based on the shortcomings of the existing technology described above, the purpose of the present application is to provide a financial user portrait intelligent decision system based on multi-modal fusion to solve the above technical problems.

[0008] To achieve the above purpose, the present application provides the following technical scheme: a financial user portrait intelligent decision system based on multi-modal fusion, comprising: The data processing module is configured to collect structured data, text data, image data and voice data of the financial service, extract features from the data after preprocessing, and generate a unified user portrait feature by fusion. The label generation module is configured to generate a user value label and a label text explanation according to the unified user portrait feature. The portrait construction module is configured to combine the user value label according to a scene configuration rule, generate a scene score, and construct a financial user portrait. The intention portrait module is configured to generate an intention label facing a prediction time window according to the unified user portrait feature and a collected user historical behavior sequence, and write the intention label into the financial user portrait. The intelligent decision module is configured to input the obtained user real-time behavior feature data and the financial user portrait into a preset decision model to generate a service decision result, and update the financial user portrait, the user value label and the label text explanation.

[0009] The application further provides that the structured data, text data, image data and voice data of the financial service are collected, and the feature extraction after preprocessing of the data includes: The structured data, text data, image data and voice data are organized into an event record set according to user identification and time identification; The event record set is subjected to data regularization to form preprocessed data, wherein the voice data is converted into text data by voice recognition; Feature extraction is performed based on the preprocessed data to form structured features, text semantic features and image features, wherein the image data generates image-related text and is incorporated into the text semantic features when the text recognition condition is met.

[0010] The application further provides that the unified user portrait feature is generated by fusion, including: Linear mapping is performed on the structured features, text semantic features and image features according to the dimension setting of the unified feature space to form intermediate features corresponding to the unified feature space, and the intermediate features are processed to generate multi-modal unified features; The modal level original attention score is constructed according to the multi-modal unified features; The multi-modal unified features are weighted and normalized according to the original attention score to obtain the attention weights of the text, image and structured data features; The multi-modal unified features are weighted and fused according to the attention weights to generate the unified user portrait feature.

[0011] The application further provides that the user value label is generated according to the unified user portrait feature, including: Based on unified user profile features and user identifiers and time identifiers associated with unified user profile features, unified user profile features are normalized and modeled through preset feature constraint rules to generate modeled user profile features. Based on the features of the user profile, a user value score is generated through a preset value scoring mapping rule. Based on user value scores, user value tags are generated through a set of preset value level thresholds, resulting in tag records that include user identifier, update time identifier, tag version identifier, user value score, and user value tag.

[0012] The present invention is further configured such that the generated tag text interpretation includes: Based on the input profile features and the preset value scoring mapping rules, an explanatory element set is generated through feature contribution relationship analysis. The explanatory element set includes explanatory element identifier, explanatory element name mapping information, and contribution direction identifier. Based on the set of explanatory elements, the system generates tag text explanations corresponding to user value tags through preset text templates, generating explanation records including user identifier, update time identifier, tag version identifier, user value tags and tag text explanations; When user value scores and user value tags are updated, tag text explanations corresponding to the updated user value tags are generated simultaneously.

[0013] The present invention is further configured such that combining user value tags according to scenario configuration rules includes: The corresponding scene configuration rules are obtained based on the scene identifier. The scene configuration rules include a set of rule entries and logical combination relationships. The set of rule entries includes label identifiers, weight information, threshold information and comparison operator information. The set of scene reference tags is determined based on the set of rule entries, and the user value tag and update time tag corresponding to the user identifier are obtained from the tag record of the set of scene reference tags. Based on logical combination relationships, the set of rule entries is subjected to rule constraint processing to form an executable rule set.

[0014] The present invention is further configured such that generating scenario scores and constructing financial user profiles includes: Rule satisfaction information is generated based on the set of executable rules and user value tags. The rule satisfaction information corresponds one-to-one with the weight information, threshold information and comparison operator information. A scenario score is generated based on rule fulfillment information and weight information, and the scenario score is associated with user identifier, scenario identifier, update time identifier and rule version identifier. Financial user profiles are constructed based on scenario scoring. These profiles include scenario scores and constituent information generated from a set of rule entries. The constituent information includes the tags, weights, and corresponding rule fulfillment information for each scenario calculation.

[0015] The present invention is further configured such that the intent profiling module includes: User context features are obtained based on unified user profile features and time stamps are added; Based on the collected transaction records and interaction logs, historical behavior sequences are organized according to user identifiers and time identifiers, and the historical behavior sequences are converted into event semantic feature sequences. The event semantic feature sequence is enhanced based on user context features. The enhanced sequence is then input into a bidirectional temporal coding network and generated through attention aggregation to produce an intent representation. Intent label probability set is generated based on intent representation and associated with prediction time window; intent label is generated based on intent threshold set. Intent tags, intent tag probability sets, prediction time window identifiers, generation time identifiers, evidence summary information, and model version identifiers are written into the financial user profile.

[0016] The present invention is further configured such that, after the prediction time window expires, the real feedback event corresponding to the intent tag is obtained, and a supervision sample is formed based on the real feedback event and the historical behavior sequence and associated with the sample source identifier; Based on supervised samples, incremental updates are performed on the bidirectional temporal coding network and the intent output layer to generate an updated version identifier and an update time identifier. Based on the updated model, the intent label probability set and intent label are regenerated, the intent label data associated with the prediction time window in the financial user profile are updated, and the evidence summary information is updated synchronously. The updated version identifier, update time identifier, and sample source identifier will be written into the intent profile record of the financial user profile.

[0017] The present invention is further configured such that the intelligent decision-making module includes: Real-time user behavior feature data is generated based on real-time user behavior events. Real-time behavior events include user identifiers, scene identifiers, and time identifiers. Real-time user behavior feature data is associated with user identifiers and time identifiers. Financial user profile data is retrieved based on user identifiers. The financial user profile data includes unified user profile features, scenario profile records, user value tags, tag text explanations, and intent tag data. The financial user profile data is assembled with real-time user behavior feature data to form decision input data. Based on the decision input data, a preset decision model is invoked to generate business decision results. The business decision results are associated with the model version identifier and time identifier to form a decision record. Based on business decision results, write-back updates are performed on financial user profile data. Write-back updates include updating scenario profile records and decision summary information, triggering user value tag updates and synchronously updating tag text explanations, and updating intent tag related records.

[0018] This invention provides a multimodal fusion-based intelligent decision-making system for financial user profiling. The system comprises a data processing module for collecting structured data, text data, image data, and voice data from financial transactions; preprocessing the data to extract features and fusing them to generate unified user profile features; a tag generation module for generating user value tags and tag text explanations based on the unified user profile features; a profile construction module for combining user value tags according to scenario configuration rules to generate scenario scores and construct financial user profiles; an intent profiling module for generating intent tags oriented towards prediction time windows based on the unified user profile features and collected user historical behavior sequences, and writing them into the financial user profile; and an intelligent decision-making module for inputting the acquired real-time user behavior feature data and the financial user profile into a preset decision model to generate business decision results, updating the financial user profile, user value tags, and tag text explanations. The beneficial effects include: 1. Unified profile features connect multi-source data links: Multi-source structured, text, image and voice data are preprocessed, feature extracted and fused to form unified user profile features, and user identifiers and time identifiers are used to connect tag generation, scene profile and intent profile, ensuring data transmission closure and consistency in profile construction; 2. Scenario-based tag factory enables configurable combination and traceable profiles: Based on scenario configuration rules, user value tags are combined to generate scenario scores and build financial user profile records containing constituent information. It supports flexible configuration of rule items, weights and thresholds according to scenario identifiers to form versionable and traceable scenario profile outputs. 3. Intent prediction and real-time decision-making are linked to form a closed-loop update: Based on unified profile features and historical behavior sequences, intent tags for prediction time windows are generated and written into the profile. The intelligent decision-making module integrates real-time behavior features and financial user profiles to generate business decision results, and simultaneously writes back and updates the profile, value tags and tag explanations to achieve a decision-driven continuous update chain.

[0019] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 The flowchart illustrates an exemplary embodiment of the present invention: a financial user profiling intelligent decision-making system based on multimodal fusion. Detailed Implementation

[0021] The embodiments of the present invention will be described below with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred embodiments are only for illustrating the present invention and not for limiting the scope of protection of the present invention.

[0022] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0023] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0024] Example 1: A financial user profiling intelligent decision-making system based on multimodal fusion, such as... Figure 1 As shown, it includes: The data processing module is used to collect structured data, text data, image data, and voice data from financial business, and after preprocessing the data, extract features and fuse them to generate unified user profile features. The tag generation module is used to generate user value tags and tag text explanations based on unified user profile characteristics. The profile building module is used to combine user value tags according to scenario configuration rules to generate scenario scores and build financial user profiles. The intent profiling module is used to generate intent labels for prediction time windows based on unified user profile features and collected user historical behavior sequences, and write them into the financial user profile. The intelligent decision-making module is used to input the acquired real-time user behavior feature data and financial user profiles into a preset decision-making model to generate business decision results and update the financial user profiles, user value tags, and tag text explanations.

[0025] The present invention is further configured such that, after preprocessing the structured data, text data, image data, and voice data collected from financial transactions, feature extraction includes: Structured data, text data, image data, and voice data are organized into event record sets based on user identifiers and time identifiers. Specifically, structured data includes transaction records and account information; text data includes communication logs and social media texts; image data includes ID photos and business vouchers; and voice data includes call recordings, object addresses, and metadata. On the data access side, multi-source data is accessed from the core transaction system, credit system, customer relationship management system, mobile terminal tracking, call center recording library, and video platform. Structured data can be accessed via database connections, interfaces, or message queues, while unstructured data is accessed via file objects or object storage addresses. Upon access, all data is uniformly mapped to event records with "user identifier + time identifier." For each piece of raw data, the user identifier and time identifier are first parsed out. If a user identifier is missing, it is filled in and matched according to the priority of "phone number / ID number / device fingerprint + business transaction number," and records with a matching confidence level below a threshold are marked as pending verification. The event record set is processed to form preprocessed data, including speech data which is transcribed into text data through speech recognition. Specifically, the event record set is uniformly processed to form preprocessed data that can be used for feature extraction. The processing includes three sub-processes: structured cleaning, unstructured normalization, and speech transcription. Structured data processing and cleaning include missing value handling: distinguishing by field type—numerical fields are filled with "median value of the nearest window of the same user / default value of business rules" and the filling flag is recorded; category fields are placed with "unknown class" and the original missing reason is retained; outlier handling: heavy-tailed fields such as amount and frequency are truncated (as above truncated) and the truncation ratio is recorded; non-compliant enumerated values ​​are corrected by dictionary mapping. Standardized scaling: Numerical fields are standardized (e.g., using Z-Score normalization), with the mean and variance derived from historical baseline statistics of the same product line and channel to avoid cross-business line scale contamination; Category coding: One-hot or embedded coding is performed on category fields such as "product type, city, and occupation" for input preparation; Text data normalization includes word segmentation and standardization: Chinese text is segmented, stop word removed, and normalized (synonym normalization, emoticon and special character normalization), and long texts are segmented into sentences to adapt to the input length of subsequent semantic models; Speech transcription and text integration includes speech preprocessing: Recordings are segmented for silence and denoised, and sampling rate and channels are unified; Multi-channel calls are processed... Speaker channel markers are preserved; Automatic speech recognition: The automatic speech recognition engine is invoked to transcribe the speech into text and outputs sentence-by-sentence timestamps and confidence scores; Segments with confidence scores below the threshold are entered into a "low-confidence segment queue" for manual quality inspection or secondary transcription; Structured incorporation: The transcribed text is written back to the speech event record as new text data, so that it uses the same semantic feature extraction path as the original text in subsequent steps; After the above processing, the preprocessed data is output in an "event-level unified format": Each record simultaneously contains a (can be empty) set of structured fields, regularized text, regularized image objects, and transcribed text, and all carry user identifiers, time identifiers, and versioned metadata to meet the requirements of traceability and recalculation; Based on preprocessed data, feature extraction is performed to form structured features, textual semantic features, and image features. Specifically, image data that meets the conditions for text recognition generates image-associated text, which is then incorporated into the textual semantic features. In the feature extraction stage, features are calculated for each of the three modalities, and image-to-text recognition is triggered on images that meet the text recognition conditions. The recognized text is then incorporated into the input corpus for textual semantic features, forming a more complete semantic representation. Structured feature extraction includes using standardized numerical fields for numerical features, grouping and concatenating them according to "account dimension, transaction dimension, and behavioral statistics dimension," while simultaneously calculating time windows. Statistical features (such as frequency in the last 7 / 30 days, aggregated amounts, cross-regional frequency, etc.) are recorded, along with the window length and statistical scope for each type of statistical feature. For categorical features, the categorical field is mapped to a one-hot vector or an embedded index. If embedding is used, the required categorical ID for the embedding input is output along with the vocabulary version to ensure consistency between online and offline data. Text semantic feature extraction includes corpus merging: concatenating the "original text + speech-to-text + (if triggered) image recognition text" of the same user within the same time window in chronological order, while retaining source markers (customer service, APP, branch, images, etc.); semantic vectorization: the merged data is then... The text input is pre-trained into a language model, and semantic convergence vectors are taken as text semantic features. For example, BERT is used and specific convergence positions are taken to form a fixed-dimensional vector. Image feature extraction (and determination of whether text recognition is triggered) includes image preprocessing: performing size normalization, cropping, and pixel normalization on ID photo / certificate images to ensure stable input distribution; visual feature vectorization: extracting visual vectors as image features from the preprocessed image input to the backbone of a convolutional neural network; for the determination and processing of "meeting the conditions for text recognition": the condition determination logic is based on image resolution and sharpness (which can be determined by sharpness indicators such as Laplacian variance). The text region proportion (estimated by a lightweight text detection network) and tilt angle (estimated by layout detection) are comprehensively judged. When the comprehensive score exceeds the threshold, OCR is triggered to generate image-related text: text recognition is performed on the triggered voucher / document image to obtain image-related text; it is then incorporated into the aforementioned text corpus merging process as "image source text" and follows the same semantic vectorization path to achieve "semanticization of image information into text semantic features"; finally, three types of feature results are output: structured features, text semantic features, and image features, which are stored in the database (or written to feature storage) at the event level and user level, respectively.

[0026] The present invention is further configured such that the fusion of features to generate a unified user profile includes: Based on the dimensionality setting of the unified feature space, linear mapping is performed on structured features, text semantic features, and image features respectively to form intermediate features corresponding to the unified feature space. These intermediate features are then processed to generate multimodal unified features. Specifically, structured features, text semantic features, and image features are each input into their respective fully connected projection layers to complete the linear transformation from the original dimension to the unified dimension. Specifically, text semantic features are projected from the original 768 dimensions to 512 dimensions, image features from the original 2048 dimensions to 512 dimensions, and structured features from the concatenated combined dimension to 512 dimensions. After the linear transformation, layer normalization and nonlinear activation are sequentially performed on each intermediate feature to suppress differences in numerical scales between different modalities and enhance expressive power. The processed three intermediate features are then combined into a multimodal unified feature set according to modal order. Modality-level raw attention scores are constructed based on multimodal unified features. Specifically, the multimodal unified feature set is input into an attention network to generate modality-level raw attention scores. This attention network adopts a shallow self-attention encoding structure: first, internal interaction modeling is performed on each modality feature, and then the raw importance score corresponding to each modality is output. The calculation process of modality-level raw attention scores includes two parts: feature correlation estimation and score mapping. Correlation estimation is used to characterize the relative contribution of text, images, and structured information to user characterization under the current business event. Score mapping is used to convert the correlation results into comparable scalar scores and retain the network version and parameter version information at the time of score generation for traceability. Based on the original modal-level attention scores, the weights of the multimodal unified features are normalized to obtain the attention weights for text, image, and structured data features. Specifically, the output original modal-level attention scores are normalized by exponentialization followed by summation, ensuring that the weights of each modality fall within the range of zero to one and their sum is one. This normalization constraint is used to ensure that the contributions of different modalities during fusion are controllable and to avoid drastic changes in the scale of the fusion result. The normalized weights correspond to text weights, image weights, and structured data weights, respectively, and together with the current user identifier and time identifier, they form a weight record. Weighted fusion of multimodal unified features is performed based on attention weights to generate unified user profile features. Specifically, based on three types of modal weights—text weight, image weight, and structure weight—the three types of intermediate features are weighted and summed to obtain unified user profile features. The weighting process involves scaling each intermediate feature according to its corresponding weight, and the summing process involves merging the scaled features element-wise under a unified dimension. Finally, a unified user profile feature record is output. This record uses the user's unique identifier as the key and contains a unified feature array, feature source and version time information, and the last update time.

[0027] The present invention is further configured such that generating user value tags based on unified user profile features includes: Based on unified user profile features and their associated user and time identifiers, the unified user profile features are normalized and modeled using preset feature constraint rules to generate modeled user profile features. Specifically, the unified user profile feature record is retrieved by user identifier in the unified user profile feature view, and the associated time identifier and feature version information are read. After consistency verification, the modeling process begins. The preset feature constraint rules are fixed in a rule configuration file, which is generated in four stages before the system goes live: feature list freezing, threshold initialization, rule verification, and version release. First, a whitelist of usable features is formed by filtering from the structured, text, image, and voice feature production chains, and the field source, update frequency, and sensitivity level are marked. Then, based on historical sample distribution and business hard boundaries, value ranges, missing value strategies, anomaly identification conditions, and update time freshness thresholds are set for each feature. Finally, rule syntax verification and cross-field logical consistency verification are performed, and rule version identifiers are generated and released. During the model normalization process, out-of-bounds features are truncated by interval and a truncation marker is written; missing features are filled according to the missing strategy and the reason for the missing feature is written; scale fluctuation features are scaled according to the training baseline; when a modality source is found to be marked as unavailable, a placeholder value and a reliability marker are written for the corresponding dimension of that modality. After completion, the model profile features are generated and associated with the user identifier, time identifier, and rule version identifier. Based on the user profile features, a user value score is generated through preset value scoring mapping rules. Specifically, the preset value scoring mapping rules map the user profile features to continuous value scores, and the preset method adopts an "offline calibration and online solidification" path: In the offline stage, training samples and calibration sets are constructed using historical business results to determine the score output range and stability constraints; at the mapping rule level, a lightweight scoring model or rule-based mapping structure that matches the low-latency online deployment is selected, and the model parameters, input field order, scale constraints, and version information are solidified into the mapping rule configuration; before going live, backtesting is used to verify coverage, stability, and consistency, and a mapping rule version identifier is generated. During online calculation, the user profile features are assembled into model input samples according to the mapping rule configuration, and inference is performed to obtain the value score; at the same time, the completeness and reliability of the inference input fields are checked for consistency. When a degradation condition is triggered, a conservative score is generated according to the degradation branch in the mapping rule, and the degradation reason is written. Finally, a user value score is generated, and the score is associated with the user identifier, time identifier, and mapping rule version identifier. Based on user value scores, user value tags are generated through a preset set of value level thresholds. This generates tag records including user identifier, update time identifier, tag version identifier, user value score, and user value tag. Specifically, the preset value level threshold set discretizes continuous value scores into user value tags. The preset method adopts a "joint setting of distribution quantiles and business capacity": during the threshold initialization phase, candidate segmentation points are determined based on historical score distribution, and then adjusted according to business reach capacity, risk preference, and strategy caliber, forming a multi-level threshold set; during the release phase, monotonicity verification, interval coverage verification, and version solidification are performed on the threshold set, generating a threshold set version identifier and establishing a connection with the score mapping rule version. During online tag generation, the threshold set matching the current mapping rule version is read, and user value tags are generated if the user value score falls into the corresponding threshold interval; when multiple tag calibers exist, a unique mapping is performed according to priority to avoid conflicts. Subsequently, tag records are generated and written to the tag storage. The tag record includes at least the user identifier, update time identifier, tag version identifier, user value score, and user value tag. The update time identifier is determined by the time identifier of the unified user profile feature and the current generation time. The tag version identifier is generated by a combination of the feature constraint rule version, the score mapping rule version, and the threshold set version.

[0028] The present invention is further configured such that the generated tag text interpretation includes: Based on the user profile features and preset value scoring mapping rules, an explanatory element set is generated through feature contribution relationship analysis. This set includes explanatory element identifiers, explanatory element name mapping information, and contribution direction identifiers. Specifically, after generating user value scores and user value tags, the user profile features corresponding to this score are read according to the user identifier and update time identifier, along with the version information of the value scoring mapping rule used. The generation of the explanatory element set is driven by "feature contribution relationship analysis configuration": a list of candidate explanatory features, a mapping table from features to explanatory element names, direction determination criteria, and an upper limit on the number of explanatory elements are pre-defined in the explanatory configuration file, and a binding relationship is established with the value scoring mapping rule version to ensure consistency between the explanatory output and the scoring criteria. At runtime, the contribution strength of each candidate feature to the scoring result is calculated within the candidate feature list. A consistent perturbation evaluation method is used to calculate the contribution strength: a limited perturbation is applied to each candidate feature while keeping the other features unchanged, and the change in score before and after the perturbation is obtained. The absolute value of the score change represents the contribution strength of the feature. When the mapping rule uses a tree model or a linear model, its built-in feature contribution interface is called first to obtain the contribution ranking result, and the contribution source identifier is retained in the output. Subsequently, several explanatory elements ranked by contribution strength are selected to form an explanatory element set. Each explanatory element contains at least an explanatory element identifier, explanatory element name mapping information, and a contribution direction identifier. The contribution direction identifier is determined by the positive or negative sign of the score change, and the name mapping information is obtained by converting internal feature items into business-readable explanatory element names using a mapping table. Based on the set of explanatory elements, tag text explanations corresponding to user value tags are generated through preset text templates. This generates explanation records including user identifier, update time identifier, tag version identifier, user value tag, and tag text explanation. Specifically, after obtaining the set of explanatory elements, a preset set of text templates matching the user value tag level is read. This set of text templates is preset in configuration files before system deployment and undergoes syntax and placeholder validation. The templates contain optional structures such as tag level description sections, evidence element filling sections, and risk warning or business warning sections, while also constraining output length and sensitive word filtering rules. When generating text explanations, the names, contribution directions, and contribution strength levels of the explanatory elements in the set of explanatory elements are filled in according to the placeholder requirements in the templates. The process is as follows: Contribution strength levels are mapped from their ranking position to segment thresholds, expressed using phrases like "primary driving factor" and "secondary factor." Contribution direction is expressed using phrases like "positive support" and "negative inhibition." When the number of explanatory elements is insufficient or low-reliability markers exist, the template-based downgrade branch outputs the qualifier "generated based on limited features" and includes a downgrade reason identifier. After completion, a tag text explanation corresponding to the user value tag is generated, and an explanation record is written to the explanation storage. The explanation record includes at least the user identifier, update time identifier, tag version identifier, user value tag, and tag text explanation. The tag version identifier is associated with the version of the current scoring mapping rule, the explanation configuration version, and the text template version. When user value scoring and user value tag updates, the corresponding tag text explanations are generated synchronously. Specifically, when subsequent links trigger user value scoring or user value tag updates, such as unified user profile feature updates, scoring mapping rule version releases, threshold set adjustments, or intelligent decision-making module write-back updates, the synchronous update mechanism first locates the tag and explanation records of the previous version based on the user identifier and reads the new scoring and tag results. Then, it determines the update type: if the score value changes but the tag level remains the same, the original template is reused and the contribution ranking of explanation elements is recalculated, updating the evidence segments in the explanation element set and explanation text; if the tag level changes, it switches to the template corresponding to the new level and reconstructs the explanation text according to the new template, while refreshing the direction and strength levels of the explanation element set; if the scoring mapping rule version or explanation configuration version changes, the explanation element set and tag text explanations are regenerated according to the new version, and the cross-version migration identifier is recorded. After the synchronous update is completed, the update time identifier and tag version identifier in the explanation record are updated, and the new explanation record is associated with the new tag record.

[0029] The present invention is further configured such that combining user value tags according to scenario configuration rules includes: The corresponding scenario configuration rules are obtained based on scenario identifiers. Scenario configuration rules include a set of rule entries and logical combinations. The set of rule entries includes tag identifiers, weight information, threshold information, and comparison operator information. Specifically, when a business event triggers profile construction, the scenario identifier is parsed from the request context or event data, and the scenario identifier is used as the search key to read the currently enabled version of the scenario configuration rule file from the scenario rule storage. Scenario configuration rules are generated by business personnel selecting target tags from the basic tag library in a graphical configuration interface. During configuration, logical operation relationships and parameters such as the weight and threshold of each tag are set simultaneously, forming a versionable rule configuration record. Syntax and logic validation are completed before being stored in the database. The rule entry set refers to a collection of multiple "tag constraint entries" used to characterize users within the same scenario. Each entry contains at least a tag identifier, weight information, threshold information, and comparison operator information: the tag identifier uniquely points to a specific user tag in the basic tag library; the weight information characterizes the tag's contribution ratio in the scenario calculation and is subject to weight rationality verification; the threshold information provides the threshold for the tag value to meet the conditions or participate in the calculation; the comparison operator information defines the comparison relationship between the tag value and the threshold, such as greater than, equal to, greater than or equal to, etc., and allows the configuration target value in the equal value comparison scenario. The logical combination relationship refers to the logical structure that organizes the conditional relationships of each entry within the rule entry set, used to express combinations such as "AND," "OR," and "NOT," enabling multiple tag constraints to form an executable conditional expression structure according to business intent. After the scenario configuration rules are loaded, the system performs configuration verification to ensure that the rule entry fields are complete, the comparison operator matches the threshold type, the total weight is within a reasonable range, and there are no rule conflicts, thereby obtaining a rule version and rule element context that can enter the runtime state. A set of scenario reference tags is determined based on a set of rule entries. The user value tag and update time identifier corresponding to the user identifier are then retrieved from the tag records in this set. Specifically, after the set of rule entries is ready, tag identifiers are extracted from each rule entry and deduplicated to form the scenario reference tag set. Subsequently, using the user identifier as an index, the tag records for that user under the scenario reference tag set are read in parallel from the tag storage or cache. Each tag record contains at least the tag value and its update time identifier. The read results are then backfilled into the corresponding rule entries according to the tag identifiers, forming a one-to-one binding relationship between rule entries and tag records. During this process, a validity check is performed on the update time identifier. If an expired or missing tag record is found, a pre-defined degradation strategy is used to trigger tag recalculation or select the previous available version tag. The data source and reliability identifier are recorded in the entry context to ensure that the input is closed and traceable during subsequent rule execution. Based on logical combination relationships, rule constraint processing is performed on the set of rule entries to form an executable rule set. Specifically, after binding the rule entries with the tag records, rule constraint processing is performed on the set of rule entries according to logical combination relationships to form a set of directly executable rules and an execution plan. The rule constraint processing first performs type consistency checks on the comparison operator and threshold information of each rule entry to ensure that numerical comparisons, Boolean comparisons, and other methods are not mixed. Second, it performs conflict detection for cases where the same label is referenced multiple times in the same scene rule, identifying mutually exclusive conditions, duplicate conditions, or unsatisfactory conditions, and merging, removing, or marking them as unexecutable according to the configured strategy to avoid non-deterministic results at runtime. Third, it performs constraint processing on the weight information, checking whether the total weight is within the reasonable range of the configuration check. If not, it performs proportional correction or rollback to the default weight template according to the correction strategy defined in the rule version, and records the correction flag for auditing. Finally, it compiles the logical combination relationship into an executable logical structure, encapsulates each rule entry into a condition node, and establishes binding relationships between the node and the label value, comparison operator, threshold, and weight, generating an executable rule set. This executable rule set carries a scene identifier and a rule version identifier and can be cached and reused to reduce the overhead of repeated parsing.

[0030] The present invention is further configured such that generating scenario scores and constructing financial user profiles includes: Rule satisfaction information is generated based on the executable rule set and user value tags. The rule satisfaction information corresponds one-to-one with the weight information, threshold information, and comparison operator information. Specifically, after completing the parsing of the scenario configuration rules and forming the executable rule set, for each rule entry of the same user in the current scenario, the user value tag record bound to the rule entry is read one by one to obtain the tag value and update time identifier, and then the consistency is checked with the comparison operator information and threshold information in the rule entry. The comparison operator information is used to define the type of comparison relationship, the threshold information is used to give the threshold for satisfying the conditions, the weight information is used to give the contribution ratio of the entry in the scenario score, and the tag identifier is used to uniquely point to the target tag in the basic tag library. The generation of rule satisfaction information is performed item by item: first, the comparison method is determined based on the comparison operator information; then, the label value is compared with the threshold information to obtain the satisfaction status; and in addition to the satisfaction status, "satisfaction degree" and "validity mark" are generated. The satisfaction degree quantifies the degree to which the label value is far from the threshold, using interval mapping to transform "just satisfied, significantly satisfied, not satisfied but close, significantly dissatisfied" into aggregateable level values, avoiding scenario score jumps caused by using only binary judgments. The validity mark is obtained by verifying the label update time and the effective time of the rule version, used to indicate whether the item is expired, missing, or has a downgraded source when participating in the calculation. Finally, a set of rule satisfaction information is formed, establishing a one-to-one correspondence between each rule satisfaction information and the corresponding item's weight information, threshold information, and comparison operator information. Scenario scores are generated based on rule fulfillment and weight information, and are associated with user identifier, scenario identifier, update time identifier, and rule version identifier. Specifically, after the rule fulfillment information set is generated, the satisfaction of each item is weighted and aggregated according to the aggregation method defined in the scenario configuration rules to obtain the user's scenario score. When the aggregation method is configured as weighted summation, the satisfaction of each item is proportionally converted according to the weight information and accumulated to obtain the final score. During the score generation stage, the weight rationality is checked again to ensure that the total weight is within a reasonable range that has passed the configuration verification. For items marked as downgraded or missing, the contribution of the item is limited to zero contribution or conservative contribution according to the downgrade criteria of the scenario rules, and the downgrade reason is written into the scoring context to ensure that the score output is consistent with the input data quality. After the scenario score is generated, it is associated and solidified with user identifier, scenario identifier, update time identifier, and rule version identifier. The update time identifier is generated by normalizing the set of update time identifiers of the tag records involved in the calculation and is used to identify the profile time segment corresponding to this scenario score. Financial user profiles are constructed based on scenario scoring. These profiles include scenario scores and constituent information generated from a set of rule entries. The constituent information includes the tags, weights, and corresponding rule fulfillment information involved in the scenario calculation. Specifically, after obtaining the scenario score, a financial user profile corresponding to that scenario is constructed. This profile record includes at least the scenario score and constituent information generated from the set of rule entries. The constituent information is used to structurally solidify "which tags participate in the scenario score, their respective weights, and the fulfillment status," to meet the traceability requirements of scenario-based profiling. The generation process of the constituent information is as follows: Tag identifiers and weights involved in the calculation are extracted from the set of executable rules and aligned with the rule fulfillment information item by item to form entry-level constituent fragments. Each constituent fragment includes a tag identifier, weight information, fulfillment status, fulfillment level, and validity marker, and may include comparison information and threshold version references for auditing. Subsequently, the scenario score, constituent information, user identifier, scenario identifier, update time identifier, and rule version identifier are written together into the profile storage or cache.

[0031] The present invention is further configured such that the intent profiling module includes: User context features are obtained based on unified user profile features and time stamps are attached. Specifically, when intent profile calculation is triggered, the latest unified user profile feature record is read from the in-memory database using the user identifier, and the update time information of the record is read synchronously as the time stamp. Then, according to the intent model input specification, context-related components are extracted from the unified user profile features to form user context features. The context components cover relatively stable information such as asset and product holding overview, recent risk status and interaction preferences, and the time stamp and feature version identifier are bound together. Based on the collected transaction records and interaction logs, a historical behavior sequence is organized according to user identifiers and time identifiers, and then converted into an event semantic feature sequence. Specifically, the user's transaction records and interaction logs are collected from the transaction system and channel interaction system, associated with user identifiers and time identifiers, and organized into a historical behavior sequence in chronological order. The sequence length is configured to the most recent one hundred behaviors, and the maximum backtracking period is limited to thirty days; events exceeding the range are truncated by time. In the sequence conversion stage, a semantic event description is uniformly generated for each event. The semantic event description is mapped from fields such as event type, channel, amount range, location range, device, and session characteristics, and is converted into an event semantic feature sequence that can be input into a time-series network through behavioral semantic encoding rules, so that the transaction time series and application clickstream can be modeled in the same sequence space. Context enhancement is performed on the event semantic feature sequence based on user context features. The context-enhanced sequence is then input into a bidirectional temporal coding network and intent representation is generated through attention aggregation. Specifically, user context features are used as conditional information to perform context enhancement on the event semantic feature sequence. The enhancement method involves injecting a user state summary related to the event into each event semantic feature and applying higher context matching weights to "key behavior segments," enabling the model to distinguish the meaning differences of the same behavior under different user states. Subsequently, the enhanced sequence is input into a bidirectional temporal coding network for sequence modeling. The network structure adopts a two-layer bidirectional long short-term memory network with 256 hidden units. After obtaining the bidirectional encoding output, intent representation is obtained through attention aggregation. Attention aggregation is used to select the behavior segments that contribute most to future intent across the entire sequence and complete information aggregation. Intent label probability sets are generated based on intent representations and associated with prediction time windows, and intent labels are generated based on intent threshold sets. Specifically, intent representations are input into a classifier to obtain intent label probability sets associated with prediction time windows. Each item in the probability set corresponds to a configurable intent label, such as "apply for a credit card in the next 15 days" or "high-frequency transactions may occur in the next 7 days." The output is the probability value of the corresponding intent. The prediction time window is determined by the scenario configuration and written into the current inference context. Subsequently, the intent threshold set bound to the current model version is read, and each intent probability is matched with the threshold to generate intent labels. When multiple intents simultaneously meet the thresholds, the final label set is selected according to intent priority and mutual exclusion constraint rules, and the threshold version and selection criteria are recorded. The intent tag, intent tag probability set, prediction time window identifier, generation time identifier, evidence summary information, and model version identifier are written into the financial user profile. Specifically, after the intent tag is generated, the intent tag, intent tag probability set, prediction time window identifier, generation time identifier, and model version identifier are written into the financial user profile. At the same time, evidence summary information is generated and associated with this intent record. The formation process of the evidence summary information involves selecting several key behavioral fragments with the highest contribution from the attention aggregation process, extracting their event type, occurrence time, channel, and summary fields, and compressing and merging them to form an auditable "key evidence list" to support the traceability of the basis when the profile is invoked and linked to decision-making. After the writing is completed, an association index is established between this intent profile record and the main user profile record, enabling the real-time decision-making chain to directly read the latest intent tag and its evidence summary when the profile is invoked at the millisecond level.

[0032] The invention is further configured such that, after the prediction time window expires, the actual feedback event corresponding to the intent tag is obtained, and a supervision sample is formed based on the actual feedback event and the historical behavior sequence, and the sample source identifier is associated. Specifically, when the intent profile is written, the prediction time window identifier, the generation time identifier, the intent tag, and the intent tag probability set are fixed into the intent profile record. When the prediction time window expires, the incremental learning task scans the expired record according to the time index and obtains the actual feedback event according to the intent tag and feedback event mapping table. The feedback event mapping table is used to define the business empirical criteria corresponding to each type of intent tag. For example, "application intent" is mapped to credit application or product application events, "purchase intent" is mapped to subscription or transaction confirmation events, and "churn intent" is mapped to status events such as continuous period of no login and no transactions. When obtaining the event, the corresponding event is retrieved from the transaction flow, business acceptance ledger, and interaction log to see if it occurred within the window period, and the retrieved event summary, occurrence time, and event type are merged into the actual feedback event record. The actual feedback event records are then paired with the historical behavior sequences and user context features corresponding to the original intention to form a supervision sample package. In addition to the paired sequence and context information, the supervision sample package also contains a sample source identifier. The sample source identifier is generated by concatenating the data domain identifier, the collection link identifier, the window range identifier, and the rule version identifier. It is used to locate the source system and generation criteria of the sample during auditing or recalculation. Incremental updates are performed on the bidirectional temporal coding network and intent output layer based on supervised samples, generating update version identifiers and update time identifiers. Specifically, the incremental update phase uses supervised sample packages as input, performs bucket sampling according to intent label type, constructs small batch training data, and merges it with historical replay samples to avoid model forgetting due to using only the latest samples. During the training process, updates are only available to the bidirectional temporal coding network and intent output layer; the configuration of other features related to unified profile production remains unchanged, thus ensuring stable input. During updates, parameter iterations are performed according to a preset learning rate and update steps, and the update process is constrained by validation set monitoring thresholds and drift thresholds: when the online performance index does not reach the minimum change or the drift index triggers the risk threshold, the update is terminated and reverted to the previous stable version; when the update passes the threshold verification, this update is solidified as the new version and written to the model registry. The update version identifier is generated using the method of "baseline version number plus incremental sequence number" and is bound to the sample source identifier set of the training data and the training time interval; the update time identifier records the effective time of this version release and is used for version association of intent profile records. Based on the updated model, the intent label probability set and intent labels are regenerated, and the intent label data associated with the prediction time window in the financial user profile is updated. Evidence summary information is updated synchronously. Specifically, after the model releases a new version, a re-push is performed on intent profile records that are still valid or need to be backfilled: the latest unified user profile features and the latest historical behavior sequences are retrieved using the user identifier; event semantic feature sequences are generated according to the established intent profile input specifications, and context enhancement is performed; then, the updated model is called to output a new intent label probability set, and new intent labels are generated based on the intent threshold set. During the re-push process, evidence summary information is updated synchronously: based on the model's internal attention aggregation output, the key behavior segments with the highest contribution are selected and summary fields are generated. The summary fields must at least include the behavior type, occurrence time, channel, and brief description, and retain the model version information at the time of segment selection; when behavior segments are missing or have data quality degradation markers, the degradation reason is written in the summary to ensure traceability. After the update is completed, the new intent labels, intent label probability sets, evidence summary information, and generation time identifier are written back to the intent profile data location associated with the prediction time window in the financial user profile, ensuring that the intent results in the same window are consistent with the current model version. The updated version identifier, update time identifier, and sample source identifier are written into the intent profile record of the financial user profile. Specifically, during the intent profile write-back stage, the updated version identifier, update time identifier, and sample source identifier are written into the corresponding intent profile record, and an association index is established with the prediction time window identifier. This allows the profiler to simultaneously trace "which model version generated the intent conclusion, which source samples drove the update of this version, and when it took effect." At the same time, the statistical summary of the supervision samples and the summary of the playback samples involved in this update are written into the audit storage. The audit information includes at least the sample source identifier set, window range, rule version, and data domain identifier. This information is used for review and recalculation by version in case of subsequent strategy disputes, model decay, or caliber adjustments, thus forming a closed-loop implementation path of "prediction—feedback—incremental update—profile write-back—audit traceability."

[0033] The present invention is further configured such that the intelligent decision-making module includes: Real-time user behavior feature data is generated based on real-time user behavior events. These events include user identifiers, scene identifiers, and time identifiers, and the user behavior feature data is associated with the user identifier and time identifier. Specifically, when a real-time business link is triggered, the user's real-time behavior event is accessed from a message queue or interface gateway. The event carries the user identifier, scene identifier, and time identifier, as well as key data fields of this interaction (e.g., transaction request elements, device fingerprint, channel identifier, session identifier, geographic location summary, and interaction action sequence summary). Subsequently, in the streaming computing job, two types of real-time windows are constructed: short windows and long windows, using the user identifier as the primary key and the time identifier as the order key. Feature shaping operators are executed on the events within each window: the short window side generates behavior speed and mutation features (e.g., number of requests per unit time, number of consecutive failures, intensity of short-term amount changes, and cross-regional jump markers), while the long window side generates stability and deviation features (e.g., recent period activity segmentation, recent transaction structure ratio, and deviation level from historical baseline). Fields that cannot be directly generated are backfilled according to a preset missing strategy and a reliability flag is written, ultimately forming a record of real-time user behavior characteristics, and the record is bound together with the user identifier and time identifier; Financial user profile data is retrieved based on user identifiers. This data includes unified user profile features, scenario profile records, user value tags, tag text explanations, and intent tag data. The financial user profile data is then assembled with real-time user behavior feature data to form decision input data. Specifically, after obtaining real-time user behavior feature data, the financial user profile data is retrieved from profile storage or cache using the user identifier. This financial user profile data includes unified user profile features, scenario profile records, user value tags, tag text explanations, and intent tag data. Simultaneously, based on the scenario identifier, a scenario profile record version matching the current scenario is selected, and the update time identifiers of various profile elements are checked for freshness. If a profile element is found to be outdated or in a degraded state, the previous available version is selected according to preset rollback rules, or a recalculation request is triggered. The version and source identifier are written into the assembly context. Subsequently, the financial user profile data and real-time user behavior feature data are assembled according to unified field specifications to form a decision input data package. The assembly process includes field alignment, scale consistency, and caliber consistency checks, and feature version identifiers, rule version identifiers, and profile update time identifiers are written into the data package. Based on the decision input data, a preset decision model is invoked to generate business decision results. These results are then associated with the model version identifier and time identifier to form a decision record. Specifically, the decision input data package is submitted to the online inference service. The inference service loads the preset decision model version bound to the current business scenario, parses the decision input data according to the model input specifications, and executes the inference output to obtain the business decision result. The business decision result can be a structured result such as approval or disapproval, risk level, recommended action type, or credit limit range, carrying decision confidence and strategy routing markers. After inference is completed, a decision record is generated, and the business decision result is associated with and solidified using the model version identifier and time identifier. Simultaneously, the profile update time identifier, rule version identifier, and key input summary of this call are written into the decision record to ensure that the "input snapshot and model version corresponding to the same decision result" can be located. After the decision record is written to the audit storage, the business decision result and decision record index are passed to the write-back update stage. Based on business decision results, a write-back update is performed on the financial user profile data. This write-back update includes updating scenario profile records and decision summary information, triggering user value tag updates and simultaneously updating tag text explanations, and updating intent tag-related records. Specifically, during the write-back phase, the main financial user profile record is located using the user identifier. First, the summary information of the current business decision result is written into the scenario profile record corresponding to the current scenario identifier, along with the decision record index and update time identifier, forming a traceable scenario profile update trajectory. Then, the user value tag update process is triggered based on the decision result type. When the decision result carries new risk status or operational status change markers, a recalculation request or incremental update request is submitted to the tag generation module. After the tag update is completed, the corresponding tag text explanation record is updated synchronously to ensure consistency between the tag value and the explanation text. For intent tag-related records, the evidence summary information and latest update time identifier of the intent profile record are updated based on the behavioral facts or feedback event summary reflected in the decision result. When necessary, the intent model's re-inference is triggered, or a feedback loop update is initiated after the prediction time window expires. After completing the above write-back, the version identifier, update time identifier, and decision record index of this write-back update will be fixed to the profile-side audit field to ensure that subsequent scenario-based profile calls and online decision-making links can read consistent, up-to-date, and traceable user profile data.

[0034] It should be noted that the specific methods of operation of each module in the intelligent decision-making system for financial user profiling based on multimodal fusion provided in the above embodiments have been described in detail in Embodiment 1, and will not be repeated here. In practical applications, the intelligent decision-making system for financial user profiling based on multimodal fusion provided in the above embodiments can be assigned to different functional modules as needed, that is, the internal structure of the system can be divided into different functional modules to complete all or part of the functions described above, and this is not a limitation.

[0035] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A financial user profiling intelligent decision-making system based on multimodal fusion, characterized in that, include: The data processing module is used to collect structured data, text data, image data, and voice data from financial business, and after preprocessing the data, extract features and fuse them to generate unified user profile features. The tag generation module is used to generate user value tags and tag text explanations based on unified user profile characteristics. The profile building module is used to combine user value tags according to scenario configuration rules to generate scenario scores and build financial user profiles. The intent profiling module is used to generate intent labels for a prediction time window based on unified user profile features and collected user historical behavior sequences, and write them into the financial user profile. The intelligent decision-making module is used to input the acquired real-time user behavior feature data and financial user profiles into a preset decision-making model to generate business decision results and update the financial user profiles, user value tags, and tag text explanations.

2. The intelligent decision-making system for financial user profiling based on multimodal fusion according to claim 1, characterized in that, The process involves collecting structured data, text data, image data, and voice data from financial transactions, and then preprocessing the data to extract features, including: Structured data, text data, image data, and voice data are organized into event record sets according to user identifiers and time identifiers; The event log set is processed to form preprocessed data, where the voice data is transcribed into text data through speech recognition. Based on the preprocessed data, feature extraction is performed to form structured features, text semantic features, and image features. Among them, when the image data meets the conditions for text recognition, image-related text is generated and incorporated into the text semantic features.

3. The intelligent decision-making system for financial user profiling based on multimodal fusion according to claim 2, characterized in that, The features used to generate a unified user profile include: Based on the dimensional setting of the unified feature space, linear mapping is performed on the structured features, text semantic features and image features respectively to form intermediate features corresponding to the unified feature space. The intermediate features are then processed to generate multimodal unified features. Construct modality-level raw attention scores based on multimodal unified features; The attention weights of text, image, and structured data features are obtained by normalizing the weights of the multimodal unified features based on the original attention scores. Weighted fusion of multimodal unified features is performed based on attention weights to generate unified user profile features.

4. The intelligent decision-making system for financial user profiling based on multimodal fusion according to claim 1, characterized in that, User value tags generated based on unified user profile features include: Based on unified user profile features and user identifiers and time identifiers associated with unified user profile features, unified user profile features are normalized and modeled through preset feature constraint rules to generate modeled user profile features. Based on the features of the user profile, a user value score is generated through a preset value scoring mapping rule. Based on user value scores, user value tags are generated through a set of preset value level thresholds, resulting in tag records that include user identifier, update time identifier, tag version identifier, user value score, and user value tag.

5. The intelligent decision-making system for financial user profiling based on multimodal fusion according to claim 4, characterized in that, The generated tag text explanation includes: Based on the input profile features and the preset value scoring mapping rules, an explanatory element set is generated through feature contribution relationship analysis. The explanatory element set includes explanatory element identifier, explanatory element name mapping information, and contribution direction identifier. Based on the set of explanatory elements, the system generates tag text explanations corresponding to user value tags through preset text templates, generating explanation records including user identifier, update time identifier, tag version identifier, user value tags and tag text explanations; When user value scores and user value tags are updated, tag text explanations corresponding to the updated user value tags are generated simultaneously.

6. The intelligent decision-making system for financial user profiling based on multimodal fusion according to claim 1, characterized in that, The user value tags are combined according to the scenario configuration rules, including: The corresponding scene configuration rules are obtained based on the scene identifier. The scene configuration rules include a set of rule entries and logical combination relationships. The set of rule entries includes label identifiers, weight information, threshold information and comparison operator information. The set of scene reference tags is determined based on the set of rule entries, and the user value tag and update time tag corresponding to the user identifier are obtained from the tag record of the set of scene reference tags. Based on logical combination relationships, the set of rule entries is subjected to rule constraint processing to form an executable rule set.

7. The intelligent decision-making system for financial user profiling based on multimodal fusion according to claim 6, characterized in that, Generating scenario scores and building financial user profiles includes: Rule satisfaction information is generated based on the set of executable rules and user value tags. The rule satisfaction information corresponds one-to-one with the weight information, threshold information and comparison operator information. A scenario score is generated based on rule fulfillment information and weight information, and the scenario score is associated with user identifier, scenario identifier, update time identifier and rule version identifier. Financial user profiles are constructed based on scenario scoring. These profiles include scenario scores and constituent information generated from a set of rule entries. The constituent information includes the tags, weights, and corresponding rule fulfillment information for each scenario calculation.

8. The intelligent decision-making system for financial user profiling based on multimodal fusion according to claim 1, characterized in that, The intent profiling module includes: User context features are obtained based on unified user profile features and time stamps are added; Based on the collected transaction records and interaction logs, historical behavior sequences are organized according to user identifiers and time identifiers, and the historical behavior sequences are converted into event semantic feature sequences. The event semantic feature sequence is enhanced based on user context features. The enhanced sequence is then input into a bidirectional temporal coding network and generated through attention aggregation to produce an intent representation. Intent label probability set is generated based on intent representation and associated with prediction time window; intent label is generated based on intent threshold set. Intent tags, intent tag probability sets, prediction time window identifiers, generation time identifiers, evidence summary information, and model version identifiers are written into the financial user profile.

9. A financial user profiling intelligent decision-making system based on multimodal fusion according to claim 8, characterized in that, After the prediction time window expires, obtain the real feedback events corresponding to the intent tags, form supervision samples based on the real feedback events and historical behavior sequences, and associate the sample source identifiers; Based on supervised samples, incremental updates are performed on the bidirectional temporal coding network and the intent output layer to generate an updated version identifier and an update time identifier. Based on the updated model, the intent label probability set and intent label are regenerated, the intent label data associated with the prediction time window in the financial user profile are updated, and the evidence summary information is updated synchronously. The updated version identifier, update time identifier, and sample source identifier will be written into the intent profile record of the financial user profile.

10. The intelligent decision-making system for financial user profiling based on multimodal fusion according to claim 1, characterized in that, The intelligent decision-making module includes: Real-time user behavior feature data is generated based on real-time user behavior events. Real-time behavior events include user identifiers, scene identifiers, and time identifiers. Real-time user behavior feature data is associated with user identifiers and time identifiers. Financial user profile data is retrieved based on user identifiers. The financial user profile data includes unified user profile features, scenario profile records, user value tags, tag text explanations, and intent tag data. The financial user profile data is assembled with real-time user behavior feature data to form decision input data. Based on the decision input data, a preset decision model is invoked to generate business decision results. The business decision results are associated with the model version identifier and time identifier to form a decision record. Based on business decision results, write-back updates are performed on financial user profile data. Write-back updates include updating scenario profile records and decision summary information, triggering user value tag updates and synchronously updating tag text explanations, and updating intent tag related records.

Images