Method and system for providing financial product / targeted advertisement content recommendation by analyzing user conversation history in large language model

Abstract

A method for recommending financial products and advertising content through user conversation history analysis is disclosed, comprising the steps of: collecting conversation history with a user's large-scale language model (LLM) generated from a conversational interface of a user device; generating a multidimensional propensity vector (DV) including the user's intrinsic propensity and interests from the conversation history using a pre-trained artificial intelligence model; constructing a knowledge graph including product vectors (PV) representing the attributes of each financial product or advertising content using a graph neural network (GNN); calculating a personal fit score (PRS) based on the multidimensional propensity vector and product vectors to select one or more recommendation candidates for financial products and advertising content; verifying financial regulatory compliance based on the multidimensional propensity vector for one or more recommendation candidates to determine the final recommended financial product or advertising content; and transmitting the final recommended financial product or advertising content to the user device.

Description

Technology Field

[0001] The present invention relates to a method and system for recommending financial products / advertising content through analysis of user conversation history based on a large-scale language model (LLM), and more specifically, to a technology for recommending customized financial products or advertising content based on a user's multidimensional propensity vector and product vector. Background Technology

[0002] With the advancement of digital finance, hyper-personalization—recommending financial products optimized for individual customers and providing customized information—has emerged as a core competitive advantage for financial platforms. Existing financial applications have utilized transaction history, asset information, and demographic data to achieve this hyper-personalization.

[0003] However, this approach faces several inherent limitations. First, customers' frequency of using financial apps is very low, averaging only 2 to 3 times per month, making it absolutely impossible to build meaningful personas by identifying customers' current interests or psychological states in real time due to an absolute lack of data and poor timeliness.

[0004] Second, while the introduction of the MyData system has enabled integrated inquiry of personal credit information, it presupposes complex and cumbersome consent procedures for users. Customers experience fatigue regarding the provision of information and anxiety about personal data leaks, while service providers incur massive technical and administrative costs for data encryption, distributed storage, and access control to ensure regulatory compliance.

[0005] Third, existing product recommendation technologies often rely on surveys in which users directly evaluate their own investment propensity. Such self-report data has limitations, as it may differ from actual tendencies due to users' cognitive biases or intentional distortions, and results can vary significantly depending on the design of the survey items.

[0006] Accordingly, there is a need for technology that matches financial products and advertising content based on propensity vectors generated without the additional collection of personal information, while verifying compliance with complex financial regulations, such as the Capital Markets Act and the Financial Consumer Protection Act, in real time and providing users only those results that have passed. Prior art literature

[0007] Korean Registered Patent No. 10-1712591 (February 27, 2017) The problem to be solved

[0008] The present invention aims to provide a method and system for recommending financial products / advertising content through analysis of user conversation history based on a large-scale language model (LLM). means of solving the problem

[0009] As a technical means for achieving the above-mentioned task, the first aspect of the present invention may provide a method for recommending financial products and advertising content through user conversation history analysis, comprising: collecting conversation history with a large-scale language model (LLM) of a user generated from a conversational interface of a user device; generating a multidimensional propensity vector (DV) including the user's intrinsic propensity and interests from the conversation history using a pre-trained artificial intelligence model; constructing a knowledge graph including product vectors (PV) representing the attributes of each financial product or advertising content using a graph neural network (GNN); calculating a personal fit score (PRS) based on the multidimensional propensity vector and product vectors to select one or more recommendation candidates for financial products and advertising content; verifying financial regulatory compliance based on the multidimensional propensity vector for one or more recommendation candidates to determine the final recommended financial product or advertising content; and transmitting the final recommended financial product or advertising content to the user device.

[0010] In addition, the multidimensional propensity vector may include MBTI (Meyers-Briggs Type Indicator) personality types, Big-5 personality traits, investment risk preferences, and time preferences.

[0011] In addition, the multidimensional propensity vector is a high-density floating-point vector of 32 dimensions or more, and the multidimensional propensity vector may be stored by mapping it to a user identifier.

[0012] In addition, the individual fit score is calculated by mathematical formula 1, and

[0013] Mathematical formula 1 is

[0014]

[0015] (Note: PRS: Individual Fit Score,

[0016] Sim(DV, PV): Cosine similarity between the multidimensional propensity vector and the product vector,

[0017] IntentScore: Bonus points awarded to financial products or advertising content related to specific keywords within conversation history,

[0018] HistoryScore: Bonus points for a user's history of clicking on financial products or advertising content or showing a positive response in the past.

[0019] RiskPenalty: May be a penalty imposed when the risk rating of a financial product or advertising content exceeds the user's risk tolerance.

[0020] Additionally, the method may further include the step of generating an audit log containing information about the final recommended financial product or advertising content transmitted, and the step of connecting the audit logs in a hash-chain manner to record them on an immutable ledger that cannot be tampered with.

[0021] In addition, the audit log may include at least one of the following: a timestamp at the time the final recommended financial product or advertising content was transmitted, an identifier of the final recommended financial product or advertising content, an identifier of the user, a multidimensional propensity vector, and an identifier of a financial regulatory rule.

[0022] A second aspect of the present invention may provide a computer-readable storage medium having a program recorded thereon for executing the method of the first aspect on a computer. Effects of the invention

[0023] Users can receive a hyper-personalized financial experience without separate MyData or consent for the collection of personal information, as their current interests and unconscious tendencies are captured in real time solely through conversations with chatbots or LLMs, without the need for separate surveys or authentication procedures. This has the advantage of dramatically reducing friction in the customer onboarding process and inducing a high level of satisfaction and engagement with the platform.

[0024] In addition, the present invention automatically verifies suitability for every recommendation / advertisement, thereby fundamentally blocking the possibility of mis-selling, and has the advantage of increasing the agility of companies regarding regulatory changes by enabling rapid response through rule set updates even if the guidelines of financial authorities change.

[0025] In addition, all recommendation history is recorded in an immutable ledger, ensuring transparency and thereby gaining the trust of users and regulatory authorities.

[0026] Through a feedback loop that utilizes conversion data, such as user ad clicks and product subscriptions, as retraining data, the system has the technical effect of enabling self-supervised learning, allowing it to improve its performance over time. Brief explanation of the drawing

[0027] FIG. 1 is a drawing illustrating an example of a system that recommends financial products and advertising content according to one embodiment. FIG. 2 is a flowchart illustrating a method for recommending financial products and advertising content according to one embodiment. FIG. 3 is a diagram showing an example of a knowledge graph structure according to one embodiment. FIGS. 4a and 4b are drawings illustrating an example of recommending financial products based on a user's propensity analysis according to one embodiment. FIG. 5 is a block diagram illustrating the configuration of a system according to one embodiment. Specific details for implementing the invention

[0028] The invention will be described in detail below by way of exemplary embodiments with reference to the attached drawings. The following embodiments are intended only to embody the invention and do not limit or restrict the scope of the invention. Anything that can be easily inferred by a person skilled in the art to which the invention pertains from the detailed description and embodiments shall be interpreted as falling within the scope of the invention.

[0029] Terms such as 'composed' or 'comprising' as used in this specification should not be interpreted as necessarily including all of the various components or steps, and should be interpreted as some of the components or steps may not be included, or additional components or steps may be included.

[0030] The terms used in this specification are described in their current, general form considering the functions mentioned herein; however, they may refer to various other terms depending on the intent of those skilled in the art, case law, the emergence of new technologies, etc. Therefore, the terms used in this specification should not be interpreted solely by their names, but should be interpreted based on the meaning of the terms and the overall content of this specification. Furthermore, singular expressions include a plural meaning unless the context clearly indicates a singular meaning.

[0031] The terms “above” and similar designations used in this specification (particularly in the claims) may indicate both singular and plural forms. Furthermore, unless there is a description explicitly specifying the order of the steps describing the method according to this specification, the described steps may be performed in a suitable order. The present invention is not limited by the order in which the described steps are described.

[0032] Phrases such as "in one embodiment" appearing in various places in this specification do not necessarily refer to the same embodiment.

[0033] Some embodiments of this specification may be represented by functional block configurations and various processing steps. Some or all of these functional blocks may be implemented by various numbers of hardware and / or software configurations that execute specific functions.

[0034] Furthermore, the connecting lines or connecting members between the components depicted in the drawings are merely illustrative of functional connections and / or physical or circuit connections. In the actual device, connections between components may be represented by various alternative or added functional connections, physical connections, or circuit connections.

[0035] The embodiments relate to a method and system for recommending financial products / advertising content through analysis of user conversation history based on a Large Language Model (LLM). Detailed descriptions of matters widely known to those skilled in the art to which the following embodiments belong are omitted. The present invention will be described in detail below with reference to the attached drawings.

[0036] Preferred embodiments of the present invention will be described in detail below with reference to the drawings.

[0037] FIG. 1 is a diagram illustrating an example of a system for recommending financial products and advertising content according to one embodiment. Referring to FIG. 1, the system (100) may be connected to a user device, a compliance gateway, an advertising server, and a Kubernetes cluster.

[0038] The system (100) may be a computing device such as a server or a micro server, but is not limited thereto, and may include any type of device capable of transmitting and receiving information to and from other devices or systems to generate multidimensional propensity vectors and recommend financial products or advertising content.

[0039] In one embodiment, the system (100) is a set of microservices running on a Kubernetes cluster and may include or be connected to each other to operate the following components.

[0040] A Data Collection Unit (DCU) represents a data ingestion pipeline that asynchronously collects user query text, voice data streams, and session metadata (e.g., query time, usage time, device information, etc.) originating from a user device's conversational interface (e.g., chat, voice assistant, etc.) and securely delivers them to a specific topic in a message queue system such as Apache Kafka.

[0041] The Propensity Profiling Engine (PPE) consumes user conversation data from a message queue, extracts text and speech embeddings based on pre-trained large-scale language models (e.g., KoBERT, GPT, T5, etc.), and inputs them into an inference model composed of a Multi-Layer Perceptron (MLP) and a Multi-Head Attention mechanism to produce a Disposition Vector (DV), which is a high-density floating-point vector of 32 dimensions or more representing the user's disposition, interests, emotional state, etc. This multidimensional disposition vector can be mapped to a user identifier and stored in a vector database in an encrypted state.

[0042] The Item-regulation Graph Constructor (IGC) constructs and manages a Knowledge Graph that represents financial products (e.g., funds, ELS, loans, etc.), affiliate advertising content, and related regulatory and risk attributes (e.g., investor suitability, risk rating, advertising approval status, etc.) as multiple types of nodes and edges. The Knowledge Graph is stored in a graph database such as Neo4j, and a Graph Neural Network (GNN, e.g., GraphSAGE, GAT, etc.) generates Product Vectors (PVs) representing the characteristics of each product / ad node and projects them into the same vector space as propensity vectors.

[0043] Upon a user request, the Personalized Matching & Recommendation Module (PAMM) calculates the cosine similarity or dot product between the user's Disposition Vector (DV) and multiple Product Vector (PV) within the Product / Ad Knowledge Graph. It then calculates a Personal Relevance Score (PRS) by applying a weighted sum to this, along with real-time intent signals (e.g., keywords used in the current conversation) and historical interaction metrics (clicks, conversion rates), and selects the top N candidates with the highest scores.

[0044] The Real-time Compliance Gateway (CGW) verifies the regulatory compliance between the user and the product / advertisement in real time for the Top-N candidates selected by the personalized matching and recommendation module. This process checks (1) the suitability and appropriateness principles under the Financial Consumer Protection Act (e.g., age, investment experience, risk tolerance), (2) sales qualifications for specific products, country of residence restrictions, and (3) Chinese Wall regulations to prevent conflicts of interest between group companies, by combining a rule engine and a deep learning-based anomaly detection model in a hybrid manner, and filters out unsuitable candidates from the final results.

[0045] The Immutable Audit Ledger (IAL) generates log records for all recommendation and ad serving events that pass through the compliance gateway and are transmitted to user devices. These records include timestamps, user anonymous IDs, exposed content IDs, applied disposition vector (DV) snapshots, and passed compliance rule IDs. The IAL then records these records in a hash-chain structured ledger that links the cryptographic hash value of the record to the hash value of the previous record. The ledger can be stored on a private blockchain, such as Hyperledger Fabric, or on an immutable database, such as AWS QLDB, to ensure data integrity.

[0046] The Platform Orchestrator deploys and manages all the aforementioned components (microservices) as containers on a Kubernetes cluster. It dynamically adjusts the number of pods based on the load of each service through the Horizontal Pod Autoscaler (HPA), ensures complete isolation and stable service levels by allocating unique namespaces and resource quotas to each tenant, and controls and monitors communication between services through a Service Mesh such as Istio.

[0047] In addition, the system (100) can provide a flexible and resilient architecture that ensures that data of each tenant is completely isolated from other tenants in a multi-tenant cloud environment where multiple client companies (e.g., banks, securities firms, asset management companies) use the same platform, and allows related microservices to be independently horizontally scaled as traffic increases.

[0048] FIG. 2 is a flowchart illustrating a method for recommending financial products and advertising content according to one embodiment.

[0049] In step S210, the history of conversations between a user and a large-scale language model (LLM) occurring from a conversational interface of a user device can be collected. For example, queries entered by a user through a conversational interface on various client terminals, such as smartphones, web browsers, and AI speakers, can be collected.

[0050] Korean text queries first undergo a preprocessing stage. This preprocessing may include (i) tokenization using a morphological analyzer (e.g., MeCab, KoNLPy), (ii) spelling and spacing correction, and (iii) removal of meaningless stopwords. Subsequently, subword tokenizers such as SentencePiece or BPE (Byte Pair Encoding) can be applied to transform the text to fit the input format of a large-scale language model.

[0051] Speech queries are converted into text through a Speech-to-Text (STT) engine, and at the same time, prosody features of the speech signal itself can be extracted. For example, variation patterns such as speech rate, pitch, and intensity can be converted into a Mel-Spectrogram and passed through a deep learning model like CNN-LSTM to extract prosody embedding vectors representing sentiment, confidence, urgency, etc. These prosody embeddings can be combined with text embeddings to improve the accuracy of propensity inference.

[0052] It detects patterns of personally identifiable information, such as resident registration numbers, phone numbers, addresses, and credit card numbers, in text converted from input or voice using regular expression matching, and immediately replaces them with special tokens like [SSN], [PHONE], and [ADDRESS] to fundamentally prevent the leakage of sensitive information. The preprocessed data can be serialized in JSON format and transmitted to a specific topic.

[0053] In step S220, a pre-trained artificial intelligence model can be used to generate a multidimensional propensity vector (DV) containing the user's intrinsic propensity and interests from the conversation history. In one embodiment, the multidimensional propensity vector may include MBTI personality types, Big-5 personality traits, investment risk preferences, and time preferences. Additionally, the multidimensional propensity vector may be a high-density floating-point vector of 32 dimensions or more, and the multidimensional propensity vector may be mapped to the user's identifier and stored in a vector database.

[0054] In one embodiment, after subscribing to topics to retrieve user conversation data in real time, propensity vectors can be generated through a pre-trained artificial intelligence model. The foundation model for this purpose may be a large-scale Korean language model (e.g., KoBERT tuned in collaboration with the National Institute of Korean Language, or Fount-LLM-13B developed in-house). The foundation model is further trained (domain-adaptive pre-training) using hundreds of millions of publicly available documents related to finance and the economy, such as minutes of the Trade, Industry and Energy Committee and the National Assembly, press releases from the Financial Supervisory Service, terms and conditions of financial products, and posts from investment advice communities, to ensure a deep understanding of the financial domain.

[0055] Then, the underlying model is fine-tuned using a Multi-Task Learning method. To this end, it learns multiple tasks simultaneously, such as (i) classifying MBTI tendencies by learning from hundreds of thousands of sentence samples (labeled via crowdsourcing) representing the characteristics of the 16 MBTI types, (ii) learning from actual customer data classifying investment risk preferences into 5 levels from 'stability-seeking' to 'aggressive investment', and (iii) predicting time preferences by learning contexts containing keywords such as 'short-term profit' and 'long-term growth'.

[0056] A high-dimensional vector output from the last hidden layer of a fine-tuned base model is input, passed through a multilayer perceptron (MLP) and an attention layer to finally generate a 32-dimensional disposition vector (DV). For example, this 32-dimensional vector can be composed of interpretable sub-dimensions such as {Extroversion (E / I), Intuition (N / S), Thinking (T / F), Judgment (J / P) - 4 bits each}, {Risk Tolerance - 4 bits}, {Time Preference - 4 bits}, {Socially Responsible Investment (ESG) Interest - 4 bits}, {Tech Stock Preference - 4 bits}, etc.

[0057] The generated multidimensional propensity vector is stored in a vector database (e.g., Faiss, Milvus) along with the user's anonymous ID to support fast similarity searches. As user conversations accumulate, the multidimensional propensity vector is designed to be continuously updated in a manner similar to an Exponential Moving Average to better reflect recent interests.

[0058] In step S230, a knowledge graph containing Product Vectors (PVs) representing the attributes of each financial product or advertising content can be constructed using a Graph Neural Network (GNN).

[0059] In one embodiment, product vectors can be generated using an S-BERT (Sentence-BERT) model. Standard BERT models generate context-dependent word embeddings but are not optimized for deriving a single vector that captures the meaning of an entire sentence. S-BERT generates embeddings by modifying the BERT architecture, primarily using a Siamese network structure, so that semantically similar sentences are located close together in the vector space. First, structured and unstructured text data related to products, such as product names, descriptions, categories, specifications, and user reviews, are collected as input data for the S-BERT model. Then, a pre-trained S-BERT model, such as all-MiniLM-L6-v2, is further fine-tuned on a domain-specific corpus, such as product descriptions, to maximize relevance. When each text fragment is input into the S-BERT model, the model outputs a fixed-length, high-dimensional vector (e.g., 384 or 768 dimensions). The output vector is the product vector, which can capture the deep semantic essence of a product beyond simple keywords.

[0060] The generated product vectors can serve as the foundation nodes of a knowledge graph (KG). Knowledge graph embeddings represent both entities (e.g., product, brand, etc.) and relationships (is_a, compatible_with, etc.) in a low-dimensional vector space to predict missing links. This can directly solve the problem of data sparsity, a chronic issue in recommendation systems. In one embodiment, TransR, among translational models, may be adopted to construct the knowledge graph. TransR learns a unique projection matrix for each relationship type to project entities into a relation-specific space. Through this, the entity "Samsung" can have different representations in the context of the "manufactures mobile phones" relationship and the "manufactures TVs" relationship. By separating the entity space and the relationship space in this way, various relationships in the knowledge graph can be accurately modeled. The structure of the knowledge graph will be described later with reference to Fig. 3.

[0061] In step S240, a Personal Fit Score (PRS) is calculated based on multidimensional propensity vectors and product vectors to select one or more recommendation candidates for financial products and advertising content. For example, when there is a recommendation request from a user, an optimal candidate group can be selected using the user's latest propensity vector and the product vector (PV) of the product / ad knowledge graph.

[0062] In one embodiment, the individual fit score can be calculated by the following mathematical formula 1.

[0063]

[0064] PRS: Individual Fit Score.

[0065] Sim(DV, PV): Cosine similarity between the multidimensional propensity vector and the product vector. Represents the fundamental agreement between propensity and product characteristics.

[0066] IntentScore: A score awarded to products related to specific keywords, such as "dividend stocks" or "retirement preparation," when those keywords appear in the current or recent conversation. The relevance between keyword vectors and product descriptions can be calculated using algorithms such as the BM25 algorithm.

[0067] HistoryScore: Bonus points for a user's history of clicking on similar products or showing positive reactions in the past.

[0068] RiskPenalty: A penalty imposed when the risk rating of a financial product or advertising content exceeds the user's risk tolerance.

[0069] Based on the calculated personal suitability scores, the top N (e.g., 20) financial products / advertisement contents can be selected as a primary candidate group and sent to the compliance gateway.

[0070] In step S250, the final recommended financial product or advertising content can be determined by verifying financial regulatory compliance based on multidimensional propensity vectors for one or more recommendation candidates. The Compliance Gateway (CGW) may sequentially perform the following verifications on the primary candidate group to eliminate legal and regulatory risks.

[0071] 1. Suitability Principle: Do the user's age, investment experience, and risk tolerance (all inferred from the propensity vector) satisfy the minimum requirements of the product? (e.g., derivatives are permitted only for investors capable of accepting 'ultra-high risk')

[0072] 2. Appropriateness Principle: Does the user possess the minimum level of knowledge to understand the risks of the product? (Judgmented by the 'financial literacy' dimension of the propensity vector)

[0073] 3. Advertising Review Regulations: Has the relevant advertising content passed pre-review? Does it not contain prohibited expressions (e.g., 'guaranteed return,' 'principal guarantee,' etc.)?

[0074] 4. Other restrictions: Check other business rules such as nationality (products not available for overseas residents to invest in), affiliated group (Chinese Wall), etc.

[0075] 5. Dynamic rule updates: Changes in regulations by financial authorities can be reflected in the rule engine in real time.

[0076] Only the final pool of recommended financial products or advertising content candidates that have passed all these filters are confirmed as targets to be displayed to users.

[0077] In step S260, the final recommended financial product or advertising content may be transmitted to the user device. The system and the user device may communicate through a predetermined network. In this case, the network includes a Local Area Network (LAN), a Wide Area Network (WAN), a Value Added Network (VAN), a mobile radio communication network, a satellite communication network, and combinations thereof; it is a data communication network in a comprehensive sense that enables each network constituent entity to communicate smoothly with one another, and may include wired internet, wireless internet, and mobile wireless communication networks. Wireless communication may include, for example, Wi-Fi, Bluetooth, Bluetooth Low Energy, Zigbee, Wi-Fi Direct (WFD), Ultra Wideband (UWB), Infrared Data Association (IrDA), Near Field Communication (NFC), but is not limited thereto.

[0078] After step S260, an audit log containing information about the transmitted final recommended financial product or advertising content can be generated, and the audit log can be linked in a hash-chain manner and recorded on an immutable ledger that cannot be tampered with.

[0079] For example, the audit log may include at least one of the following: a timestamp at the time the final recommended financial product or advertising content was transmitted, an identifier of the final recommended financial product or advertising content, a user identifier, a multidimensional propensity vector, and an identifier of financial regulatory rules. Additionally, the audit log may include the reason why a specific financial product or advertising content was exposed to a specific user. That is, the audit log may be composed of fields such as {record_id, timestamp, user_pseudo_id, item_id, dv_snapshot, cgw_rules_passed[], previous_hash, current_hash}. current_hash is the hash value of all data in the current record, and this can be included in the previous_hash field of the next record to form a chain structure. This immutable ledger can be utilized as decisive evidence when a user inquires about the basis for a recommendation, when the Financial Supervisory Service audits whether a specific transaction was mis-sold, and when debugging the system's recommendation logic. Meanwhile, data on users clicking recommended ad content or converting into actual transactions is collected again via Kafka and can be used to evaluate the performance of AI models that generate propensity vectors, S-BERT models that generate product vectors, and TransR models that build knowledge graphs, and to retrain them periodically (e.g., at 3 AM every day). For example, if model drift is detected where the model's prediction distribution differs significantly from the actual user response distribution (e.g., Population Stability Index > 0.1), irregular retraining may be triggered.

[0080] FIG. 3 is a diagram showing an example of a knowledge graph structure according to one embodiment.

[0081] Referring to Fig. 3, a node-link diagram visually representing the structure of a knowledge graph can be constructed as follows. Nodes are represented by circles or ellipses with labels such as "Galaxy S24", "Samsung", "Smartphone", "Android OS", and "USB-C Charger". Typed edges are represented by directed arrows with labels indicating relationship types such as has_brand, in_category, uses_os, and compatible_with. Fig. 3 highlights a multi-hop path such as "Galaxy S24" → has_brand → "Samsung" → also_manufactures → "Galaxy Watch" to show how the knowledge graph can discover non-obvious relationships for recommendations.

[0082] Figure 3 illustrates the conceptual structure of an intelligent recommendation knowledge graph according to one embodiment, where each node represents an entity such as a product (e.g., Galaxy S24), a brand (e.g., Samsung), a category (e.g., a smartphone), or an attribute (e.g., Android OS). These nodes can be connected by directed edges representing semantic relationships such as has_brand, in_category, and compatible_with. Through this structure, the system can understand and utilize complex and multi-layered relationship networks between entities, going beyond the attributes of a single product. For example, if a user shows interest in "Galaxy S24," the system can generate multifaceted and in-depth recommendations, going beyond simply recommending similar smartphones to recommending a "Galaxy Watch" manufactured by the same brand "Samsung," or a "USB-C charger" that has a compatible_with relationship with "Galaxy S24."

[0083] Through this two-step process (S-BERT for semantic understanding and TransR for relational inference), it is possible to provide recommendations by understanding both the content of items and the context of their relationships. This enables recommendations that are contingent and explainable, going far beyond collaborative filtering such as "users who bought X also bought Y." The S-BERT model solves the semantic understanding problem to generate rich product vectors that recognize the similarity between "fast charging" and "fast power supply." However, this alone does not understand the relational network; therefore, TransR builds a relational structure on top of these semantic product vectors to learn that the compatible_with relationship is different from the is_successor_to relationship. Consequently, the system can generate multi-hop explainable recommendations, such as, "Since you like the Galaxy S24 with a high-end camera (S-BERT semantic understanding), you may be interested in this professional photo editing software that is highly rated by other owners of the same brand (TransR relational inference)."

[0084] FIGS. 4a and 4b are drawings illustrating an example of recommending financial products based on a user's propensity analysis according to one embodiment.

[0085] Referring to Fig. 4a, a user query such as “What do you think my MBTI and tendencies are for financial product recommendations?” may be received from a user device via a conversational interface. At this time, the system collects the conversation history between the user and the Large Language Model (LLM) generated from the conversational interface of the user device, and can estimate the user’s MBTI (Meyers-Briggs Type Indicator) personality type from the conversation history using a pre-trained artificial intelligence model. For example, based on observing the user’s behavior and analyzing conversation patterns, it can be estimated that there is a 99% probability that the user’s MBTI is ENTJ, a type that listens to the opinions of others but reaches final conclusions independently and keeps the 'strategy-execution-feedback' cycle short. Furthermore, the system can predict and present the likelihood of investment behavior based on the ENTJ MBTI personality type.

[0086] Referring to Fig. 4b, a customized recommended portfolio of financial products can be provided based on the user's estimated MBTI. For example, a portfolio can be recommended consisting of 40% Core Growth (Global Equity ETFs + K-Tech ETFs), 20% Alternative Alpha (VC Funds, Pre-IPO Funds, or Private Equity FoFs), 15% Tactical Satellite (Thematic ETFs such as AI, Robotics, and Clean Energy), 15% Stability & Liquidity (High-quality IG Bond ETFs + Short-term MMFs), and 10% Protection & Tax Optimization (Variable Annuities or ISA / IRPs). Additionally, the reason for providing the recommended portfolio can also be provided to the user.

[0087] In addition, you can see that tips for implementing the recommended portfolio are provided, along with the option to request official MBTI test results as the next step, or, if necessary, additional advice on gift and inheritance planning for family members is offered.

[0088] FIG. 5 is a block diagram illustrating the configuration of a system according to one embodiment.

[0089] As illustrated in FIG. 5, in one embodiment of the present invention, the system (100) may include a control unit (130), a communication unit (150), and a memory (170). However, not all of the illustrated components are essential components. The system (100) may be implemented with more components than illustrated, or with fewer components. The above components will be examined in turn below.

[0090] The control unit (130) typically controls the overall operation of the system (100). The control unit (130) may be equipped with at least one processor. Depending on its function and role, the control unit (130) may include multiple processors or a single processor in an integrated form. A processor may primarily refer to a central processing unit (CPU), an application processor (AP), a graphics processing unit (GPU), etc. Additionally, the CPU, AP, or GPU may include one or more cores within it, and the CPU, AP, or GPU may operate using an operating voltage and a clock signal. However, while the CPU or AP is composed of a few cores optimized for serial processing, the GPU may be composed of thousands of smaller and more efficient cores designed for parallel processing.

[0091] For example, the control unit (130) can control the output unit (not shown), the communication unit (150), etc., by executing programs stored in the memory (170). The control unit (130) can control the communication unit (150) to transmit and receive information with external devices (e.g., user device (not shown), secure storage (not shown), and financial institution server (not shown)).

[0092] The communication unit (150) may include one or more components that enable the system (100) to communicate with an external device. The external device may be a computing device such as the system (100), but is not limited thereto.

[0093] The memory (170) stores data that supports various functions of the control unit (130). The memory (170) can store a number of application programs (or applications) running on the control unit (130), data for the operation of the control unit (130), and instructions. At least some of these application programs may be downloaded from an external device or another server via wireless communication. Additionally, the application programs may be stored in the memory (170), installed on the control unit (130), and driven by a processor to perform the operation (or function) of the control unit (130).

[0094] The memory (170) may include at least one type of storage medium among flash memory type, hard disk type, SSD type (Solid State Disk type), SSD type (Silicon Disk Drive type), multimedia card micro type, card type memory (e.g., SD or XD memory, etc.), RAM (random access memory; RAM), SRAM (static random access memory), ROM (read-only memory; ROM), EEPROM (electrically erasable programmable read-only memory), PROM (programmable read-only memory), magnetic memory, magnetic disk, and optical disk. Additionally, the memory (170) may include web storage that performs storage functions over the internet.

[0095] A system (100) according to another embodiment of the present invention may be implemented to include a Data Collection Unit (DCU), a Propensity Profiling Engine (PPE), an Item-regulation Graph Constructor (IGC), a Personalized Matching & Recommendation Module (PAMM), and a vector database, as described above with reference to FIG. 1. The software module of the agent described above may correspond to the hardware module of FIG. 5. For example, the control unit (130) may include a Data Collection Unit (DCU), a Propensity Profiling Engine (PPE), an Item-regulation Graph Constructor (IGC), and a Personalized Matching & Recommendation Module (PAMM), and the memory (170) may include a vector database.

[0096] The present invention relates to a technology that generates propensity vectors representing the deep inherent tendencies of users by analyzing a previously unused data source, namely user conversation history with a large-scale language model (LLM). The system (100) of the present invention generates personalized recommendations by comparing the generated propensity vectors with a knowledge graph containing all characteristics and regulatory information regarding financial products and advertisements. In particular, all recommendations are delivered to users only after passing through a compliance gateway that verifies complex financial regulations, such as the Financial Consumer Protection Act, in real time. Furthermore, all recommendation and advertisement delivery history is recorded in an immutable audit ledger that cannot be tampered with, thereby ensuring complete transparency and accountability. Through a system operating on a Kubernetes-based multi-tenant architecture, financial platform operators can achieve excellent effects, such as maximizing customer engagement, minimizing regulatory risks, and conducting new trust-based target marketing and B2B solution businesses, without the burden of collecting separate personal information.

[0097] The present invention described above can be modeled as computer-readable code on a medium on which a program is recorded. A computer-readable medium includes all types of recording devices in which data that can be read by a computer system is stored. Examples of computer-readable media include HDD (Hard Disk Drive), SSD (Solid State Disk), SSD (Silicon Disk Drive), ROM, RAM, CD-ROM, magnetic tape, floppy disk, optical data storage device, etc., and also include being modeled in the form of a carrier wave (e.g., transmission over the Internet).

[0098] Meanwhile, the configuration of the system (100) illustrated in FIG. 5 is an example, and each component of the system (100) may be integrated, added, or omitted according to the specifications of the system (100) being implemented. That is, as needed, two or more components may be combined into a single component, or a single component may be subdivided into two or more components. Furthermore, the functions performed by each component (or module) are intended to explain the examples, and the specific operations or devices thereof do not limit the scope of the present invention.

[0099] Meanwhile, the embodiments of the present invention described above can be written as a program executable on a computer and can be implemented on a general-purpose digital computer that operates the program using a computer-readable recording medium.

[0100] The above computer-readable recording media includes storage media such as magnetic storage media (e.g., ROM, floppy disk, hard disk, etc.) and optical reading media (e.g., CD-ROM, DVD, etc.).

[0101] Although embodiments of the present invention have been described above with reference to the attached drawings, those skilled in the art will understand that the present invention may be implemented in other specific forms without changing its technical concept or essential features. Therefore, the embodiments described above should be understood as illustrative in all respects and not restrictive. Explanation of the symbols

[0102] 100: System 130: Control unit 150: Communications Department 170: Memory

Claims

Claim 1 A method for recommending financial products and advertising content through user conversation history analysis, performed by a system comprising one or more processors, comprising: collecting conversation history with a user’s Large Language Model (LLM) generated from a conversational interface of a user device; generating a multidimensional disposition vector (DV) including the user’s intrinsic dispositions and interests from the conversation history using a pre-trained artificial intelligence model; constructing a knowledge graph including product vectors (PV) representing the attributes of each financial product or advertising content using a graph neural network (GNN); calculating a Personal Relevance Score (PRS) based on the multidimensional disposition vector and the product vectors to select one or more recommendation candidates for the financial products and advertising content; determining the final recommended financial product or advertising content by verifying financial regulatory compliance for each of the one or more recommendation candidates through a compliance gateway; and transmitting the final recommended financial product or advertising content to the user device. The method comprises the steps of: generating an audit log including at least one of a timestamp at the time when the final recommended financial product or advertising content was transmitted, an identifier of the final recommended financial product or advertising content, a user identifier, the multidimensional propensity vector, and an identifier of a financial regulatory rule; and recording the audit log in an immutable ledger that cannot be tampered with by linking it in a hash chain manner, wherein the determining step includes verifying whether the user's investment experience and risk tolerance inferred from the multidimensional propensity vector satisfy the minimum conditions required by the financial product corresponding to the recommendation candidate.A method comprising the step of verifying whether a user possesses a minimum level of knowledge to understand the risks of the financial product based on the financial literacy dimension of the multidimensional propensity vector.; Claim 2 A method according to claim 1, wherein the multidimensional propensity vector includes MBTI (Meyers-Briggs Type Indicator) personality types, Big-5 personality traits, investment risk preference, and time preference. Claim 3 A method according to paragraph 2, wherein the multidimensional propensity vector is a high-density floating-point vector of 32 dimensions or more, and the multidimensional propensity vector is stored by mapping to the anonymous identifier (ID) of the user. Claim 4 In claim 1, the individual fit score is calculated by mathematical formula 1, and the mathematical formula 1 is (wherein PRS: Personal Suitability Score, Sim(DV, PV): Cosine similarity between the multidimensional propensity vector and the product vector, IntentScore: Bonus points awarded to financial products or advertising content related to specific keywords within the conversation history, HistoryScore: Bonus points for a history of the user clicking on financial products or advertising content or showing a positive response, RiskPenalty: Penalty awarded when the risk rating of the financial product or advertising content exceeds the user's risk tolerance) a method Claim 5 delete Claim 6 delete Claim 7 A computer-readable storage medium having a program recorded thereon for executing the method of any one of paragraphs 1 through 4 on a computer.

Images