Product recommendation method and device, equipment and medium

Abstract

The invention belongs to the field of artificial intelligence, and relates to a product recommendation method and device, equipment and a medium, and the method comprises the steps: obtaining dialogue data of each round of dialogue of a user, inputting a data code to a user side in the dialogue data of the current round, and obtaining an intention vector of the current round; and calculating the intention drift degree of the intention vector of the current round and the intention vector of the previous round, and if the intention drift degree is smaller than a drift degree threshold value, extracting scene association features of the dialogue data of the current round as dynamic context features. Meanwhile, the emotion score of the user in the current round is obtained, and the risk bearing level is updated in combination with the initial risk reference value and the dynamic context features. And finally, according to the intention vector, the risk tolerance level and the cross-domain product knowledge graph of the current round, determining a product recommendation list of the current round. The method can be applied to business fields such as financial science and technology, and the accuracy of product recommendation can be improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology and is applied to online processing business scenarios such as fintech, and in particular to a product recommendation method, apparatus, device and medium. Background Technology

[0002] In the financial and insurance sector, with the diversification of financial businesses, many financial institutions encompass various business types such as life insurance, property insurance, banking, asset management, and health insurance, resulting in increasingly large and complex product portfolios. Against this backdrop, intelligent recommendation systems have become a key tool to assist users in selecting suitable products and improving business efficiency.

[0003] However, existing intelligent recommendation systems in the industry have many shortcomings. On the one hand, most recommendation systems are based on static user profiles, such as age, income, and historical purchase records, combined with collaborative filtering algorithms. This approach lacks a deep understanding of real-time conversational contexts, often assuming customer needs are fixed. In reality, during multiple rounds of customer consultation, their intentions may waver, change, or drift. Existing systems cannot capture these dynamic changes, leading to a mismatch between recommended products and actual user needs. On the other hand, financial product recommendations must strictly adhere to the regulatory principle of "appropriateness," ensuring that product risk matches the customer's risk tolerance. However, existing systems often use simple questionnaires or static models to calculate customer risk tolerance, failing to adjust the customer's risk boundaries in a timely manner based on dynamic information revealed during conversations, thus easily triggering compliance risks. Furthermore, traditional recommendation logic is extremely inefficient when faced with complex and large product matrices. Intelligent customer service typically only recommends preset solutions, failing to meet diverse user needs.

[0004] In summary, existing intelligent recommendation systems lack a deep understanding of real-time dialogue contexts, cannot capture shifts in user needs and intentions, cannot adjust risk boundaries based on dynamic information when calculating customer risk tolerance, and struggle to achieve real-time customized screening across business lines when faced with complex product matrices. Summary of the Invention

[0005] The purpose of this application is to propose a product recommendation method, apparatus, computer equipment, and storage medium to solve the problems of existing intelligent recommendation systems, such as difficulty in deeply understanding real-time dialogues, capturing user intent drift, adjusting user risk boundaries based on dynamic information, and difficulty in customizing product screening across business lines in real time.

[0006] Firstly, a product recommendation method is provided, which adopts the following technical solution: The system acquires dialogue data from each round of conversation with the target user, including user input data. It encodes the current round's user input data to obtain the intent vector for that round. It calculates the intent drift between the current and previous rounds' intent vectors. The intent drift is compared to a preset drift threshold; if the drift is less than the threshold, feature extraction is performed on the current round's dialogue data to obtain scene-related features, which are then used as dynamic context features. The system obtains the target user's sentiment score for the current round and updates the target user's risk tolerance level based on the initial risk baseline, dynamic context features, and sentiment score. Finally, it determines the product recommendation list for the current round based on the current round's intent vector, risk tolerance level, and a preset cross-domain product knowledge graph.

[0007] Secondly, a product recommendation device is provided, which adopts the following technical solution: The acquisition module is used to acquire dialogue data for each round of dialogue of the target user. The dialogue data includes user input data. The encoding module is used to encode the user input data for the current round to obtain the intent vector for the current round. The calculation module is used to calculate the intent drift between the intent vector of the current round and the intent vector of the previous round. The comparison module is used to compare the intent drift degree with the preset drift degree threshold. If the intent drift degree is less than the drift degree threshold, the current round of dialogue data is used to extract features to obtain scene-related features, which are then used as dynamic context features. The update module is used to obtain the target user's sentiment score in the current round, and update the target user's risk tolerance level based on the target user's initial risk benchmark, dynamic context features, and sentiment score; The determination module is used to determine the product recommendation list for the current round based on the intent vector of the current round, the risk tolerance level, and the preset cross-domain product knowledge graph.

[0008] Thirdly, a computer device is provided, which adopts the following technical solution: The system acquires dialogue data from each round of conversation with the target user, including user input data. It encodes the current round's user input data to obtain the intent vector for that round. It calculates the intent drift between the current and previous rounds' intent vectors. The intent drift is compared to a preset drift threshold; if the drift is less than the threshold, feature extraction is performed on the current round's dialogue data to obtain scene-related features, which are then used as dynamic context features. The system obtains the target user's sentiment score for the current round and updates the target user's risk tolerance level based on the initial risk baseline, dynamic context features, and sentiment score. Finally, it determines the product recommendation list for the current round based on the current round's intent vector, risk tolerance level, and a preset cross-domain product knowledge graph.

[0009] Fourthly, a computer-readable storage medium is provided, which adopts the following technical solution: The system acquires dialogue data from each round of conversation with the target user, including user input data. It encodes the current round's user input data to obtain the intent vector for that round. It calculates the intent drift between the current and previous rounds' intent vectors. The intent drift is compared to a preset drift threshold; if the drift is less than the threshold, feature extraction is performed on the current round's dialogue data to obtain scene-related features, which are then used as dynamic context features. The system obtains the target user's sentiment score for the current round and updates the target user's risk tolerance level based on the initial risk baseline, dynamic context features, and sentiment score. Finally, it determines the product recommendation list for the current round based on the current round's intent vector, risk tolerance level, and a preset cross-domain product knowledge graph.

[0010] Compared with existing technologies, the embodiments of this application have the following main advantages: By acquiring dialogue data from each round of the target user, encoding the current round's user input data using a semantic coding model to obtain an intent vector, and calculating the drift degree compared to the previous round's intent vector, the dynamic changes in user needs and intents across multiple rounds of consultation can be accurately captured, avoiding mismatches between recommended products and actual needs due to demand drift. When the intent drift degree is less than a threshold, scenario-related features are extracted as dynamic context features, and the risk tolerance level is updated in conjunction with the user's emotion score and initial risk benchmark value. This allows for timely adjustment of risk boundaries based on dynamic information in the dialogue, reducing compliance risks. Simultaneously, by determining the product recommendation list based on the current round's intent vector, the updated risk tolerance level, and the cross-domain product knowledge graph, real-time customized filtering across business lines can be achieved, overcoming the limitations of traditional recommendation logic's low efficiency and inability to recommend only preset solutions. This meets diverse user needs and improves the accuracy of product recommendations. Attached Figure Description

[0011] To more clearly illustrate the solutions in this application, the accompanying drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is an exemplary system architecture diagram to which this application can be applied; Figure 2 A flowchart of an embodiment of the product recommendation method according to this application; Figure 3 This is a schematic diagram of the structure of one embodiment of the product recommendation device according to this application; Figure 4 This is a schematic diagram of the structure of one embodiment of the computer device according to this application. Detailed Implementation

[0013] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein in the specification of the application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing drawings of this application, are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or foregoing drawings of this application are used to distinguish different objects, not to describe a particular order.

[0014] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0015] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0016] like Figure 1As shown, system architecture 100 may include terminal device 101, network 102, and server 103. Terminal device 101 may be a laptop 1011, tablet 1012, or mobile phone 1013. Network 102 is used as a medium to provide a communication link between terminal device 101 and server 103. Network 102 may include various connection types, such as wired, wireless communication links, or fiber optic cables.

[0017] Users can use terminal device 101 to interact with server 103 via network 102 to receive or send messages, etc. Various communication client applications can be installed on terminal device 101, such as web browser applications, shopping applications, search applications, instant messaging tools, email clients, social media platform software, etc.

[0018] Terminal device 101 can be various electronic devices with a display screen and support web browsing. In addition to laptops 1011, tablets 1012, or mobile phones 1013, terminal device 101 can also be e-book readers, MP3 players (Moving Picture Experts Group Audio Layer III), MP4 players (Moving Picture Experts Group Audio Layer IV), laptops, and desktop computers.

[0019] Server 103 can be a server that provides various services, such as a backend server that provides support for the pages displayed on terminal device 101.

[0020] It should be noted that the product recommendation method provided in this application embodiment is generally executed by a server / terminal device, and correspondingly, the product recommendation device is generally set in the server / terminal device.

[0021] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0022] Continue to refer to Figure 2 A flowchart of an embodiment of a product recommendation method according to this application is shown. The product recommendation method includes the following steps: Step S201: Obtain dialogue data for each round of dialogue of the target user. The dialogue data includes user input data.

[0023] The target users refer to natural or legal entities that engage in multiple rounds of dialogue with the intelligent recommendation system and have needs for product consultation, selection, or purchase. These target users are the core service recipients for the system in collecting dialogue data, identifying intent, assessing risk levels, and recommending products.

[0024] Dialogue data refers to the entire data set generated during multiple rounds of interaction between the target user and the intelligent recommendation system. This includes not only user-side input data actively entered by the target user, but also auxiliary information such as interaction timestamps, dialogue round identifiers, and system output data from the recommendation system in response to the target user. User-side input data is the core component of dialogue data, referring to the raw information actively entered by the target user into the recommendation system through intelligent terminals (such as mobile apps, web clients, and intelligent customer service interfaces), directly reflecting the user's current consultation needs, scenario descriptions, or emotional expressions.

[0025] Step S202: Encode the user input data for the current round to obtain the intent vector for the current round.

[0026] In one embodiment, the user input data for the current round can be encoded using a semantic encoding model to obtain the intent vector for the current round. The semantic encoding model refers to a pre-trained artificial intelligence model with natural language semantic understanding and vector transformation capabilities. Its core function is to transform unstructured user input data into a structured vector form to achieve a quantitative representation of user intent.

[0027] Encoding refers to the core data processing operation performed by the semantic encoding model. Specifically, it involves transforming unstructured natural language information (text or speech-to-text) input from the user into low-dimensional, dense vectors that can be processed by a computer through semantic parsing, feature extraction, and vector mapping within the model. The core objective of this process is to preserve the core semantic information of the user input while eliminating the ambiguity and redundancy of natural language, ensuring that the same intent expressed in different ways corresponds to similar vectors. For example, the user input "Are there any low-risk, retirement-friendly financial products?" and "I want to choose a stable financial product for retirement" can be encoded into two vectors with extremely high semantic similarity, ensuring consistency in intent recognition.

[0028] The intent vector for the current round refers to the low-dimensional dense vector obtained by encoding the user-end input data of the target user in the current round through a semantic encoding model, which is used to quantitatively represent the core needs and intents of the user in this round.

[0029] Step S203: Calculate the intention drift degree between the intention vector of the current round and the intention vector of the previous round.

[0030] The intent vector from the previous round refers to the intent vector obtained by encoding the user's input data in the most recent effective dialogue round before the current round using the same semantic encoding model. It has the same dimensions and data structure as the intent vector in the current round and serves as a benchmark for calculating intent drift. This vector records the core needs and intents of the target user in the previous round; by comparing it with the intent vector in the current round, it can be determined whether the user's needs have changed.

[0031] Among them, intent drift refers to a quantitative indicator used to quantify the semantic difference between the intent vector of the target user in the current round and the intent vector in the previous round. Its value directly reflects the magnitude of the change in the user's demand intent. The smaller the value, the more consistent the intent; the larger the value, the more significant the change in the user's demand.

[0032] Step S204: Compare the intent drift degree with a preset drift degree threshold. If the intent drift degree is less than the drift degree threshold, extract features from the dialogue data of the current round to obtain scene association features, and use the scene association features as dynamic context features.

[0033] The drift threshold is a pre-set critical value based on the statistical characteristics of historical dialogue data, business scenario requirements, and recommendation accuracy requirements. It is used to determine whether the degree of drift in the target user's intent is within an acceptable range and is a key parameter for triggering the dynamic context feature extraction logic. This threshold is usually obtained through training on historical dialogue data and needs to balance the sensitivity of intent changes with the stability of recommendations. For example, if the drift threshold is set to 0.3, when the calculated intent drift is 0.2 (less than the threshold), it indicates that the user's intent has not changed significantly, and scenario-related features can continue to be extracted; when the drift is 0.4 (greater than the threshold), it indicates that the intent has changed significantly, and the feature extraction strategy needs to be adjusted.

[0034] Feature extraction from the current round of dialogue data refers to the process where, when the intent drift is less than a drift threshold, the system uses a preset feature extraction algorithm to filter and extract key information strongly related to the user's needs from the current round of dialogue data. Scene-related features refer to structured feature information obtained after feature extraction from the current round of dialogue data that is directly related to the user's current product consultation scenario and reflects the user's dynamic needs. For example, information mentioned in the user's dialogue such as "just bought a house and need to repay the loan," "retirement savings," and "children's education funds," after feature extraction, forms "debt scenario," "retirement purpose," and "education funds," all of which are scene-related features.

[0035] Among them, dynamic context features refer to the core feature set that is determined by extracting scene-related features or scene-fusion features to represent the dynamic information of the user's current dialogue scene. This feature set will be dynamically updated as each round of dialogue progresses, and can reflect the changes in user demand scenarios in real time, which is different from static user profile features.

[0036] Step S205: Obtain the target user's sentiment score in the current round, and update the target user's risk tolerance level based on the target user's initial risk benchmark, dynamic context features, and sentiment score.

[0037] The current round's emotion score refers to a numerical indicator used to represent the user's emotional state in this round of dialogue after the target user's current round of dialogue data is quantified by a preset emotion analysis model. The value range can be [-1, 1]. The closer the value is to 1, the more positive the emotion; the closer it is to -1, the more negative the emotion; and 0 represents a neutral emotion.

[0038] The initial risk benchmark value refers to the user's initial risk preference benchmark value obtained after collecting static information such as age, income, asset status, and investment experience through a risk assessment questionnaire pre-set by the system before the target user has multiple rounds of dialogue with the system. This value is then standardized and quantitatively calculated. It serves as the basic reference value for subsequent updates to the risk tolerance level.

[0039] Among them, the risk tolerance level refers to the level indicator that is updated by weighted summation and other algorithms based on the user's initial risk benchmark value, dynamic context characteristics and the current round of sentiment score. It is used to characterize the upper limit of financial product risk that the user can accept at present and is the core compliance basis for product recommendation.

[0040] Step S206: Based on the intent vector, risk tolerance level, and preset cross-domain product knowledge graph of the current round, determine the product recommendation list for the current round.

[0041] Among them, the cross-domain product knowledge graph refers to the knowledge graph database built into the recommendation system for the structured association of various types of financial products. Its core is to build a relationship network of nodes and edges for products in different fields through the association dimensions such as product attributes, risk level, applicable scenarios, and return characteristics, so as to achieve rapid matching and retrieval across product types.

[0042] The product recommendation list refers to a collection of financial products that match the user's current needs and risk tolerance, obtained through processes such as intent tag matching and risk level filtering, based on the intent vector of the current round, the updated risk tolerance level, and the cross-domain product knowledge graph. It is the final recommendation result output by the system to the target user.

[0043] This application's embodiments acquire dialogue data from each round of the target user, encode the current round's user input data using a semantic coding model to obtain an intent vector, and calculate the drift degree compared to the previous round's intent vector. This accurately captures the dynamic changes in user needs and intents across multiple rounds of consultation, avoiding mismatches between recommended products and actual needs due to demand drift. When the intent drift degree is less than a threshold, scenario-related features are extracted as dynamic context features. Combined with the user's emotion score and initial risk benchmark value, the risk tolerance level is updated, enabling timely adjustment of risk boundaries based on dynamic information in the dialogue, reducing compliance risks. Simultaneously, based on the current round's intent vector, the updated risk tolerance level, and cross-domain product knowledge graphs, a product recommendation list is determined, enabling real-time customized filtering across business lines. This overcomes the limitations of traditional recommendation logic, such as low efficiency and the inability to recommend only preset solutions, meeting diverse user needs and improving the accuracy of product recommendations.

[0044] In some optional implementations of this embodiment, step 203, calculating the intent drift between the intent vector of the current round and the intent vector of the previous round, specifically includes the following steps: Calculate the similarity between the intent vector of the current round and the intent vector of the previous round; obtain the preset adjustment parameters, perform difference processing on the adjustment parameters and the similarity, and obtain the intent drift degree.

[0045] Similarity refers to an indicator used to quantify the semantic correlation between the intent vector of the current round and the intent vector of the previous round. Its value ranges from [0,1]. The closer the value is to 1, the more similar the user needs and intentions corresponding to the two intent vectors are; the closer the value is to 0, the greater the difference in intent and the more significant the change in user needs. This indicator is calculated using vector similarity algorithms, commonly including cosine similarity, Euclidean distance, and Pearson correlation coefficient.

[0046] The adjustment parameter refers to a fixed value pre-set based on historical dialogue data, business scenario requirements, and the accuracy requirements for calculating intent drift. It is a key calibration parameter for converting similarity into intent drift. The difference processing refers to the process of using preset mathematical operation rules to perform numerical calculations on the adjustment parameter and similarity, inversely converting the similarity index into intent drift. Its core logic is "intent drift = adjustment parameter - similarity." This operation achieves an inverse mapping relationship: "higher similarity, lower drift; lower similarity, higher drift," making the calculation result more closely match the semantic definition of "drift" (i.e., the magnitude of intent change).

[0047] In one example, the intelligent recommendation system uses the cosine similarity algorithm to calculate the similarity of intent vectors, with a preset adjustment parameter of 1.0 and a drift threshold of 0.3. In the first round of dialogue, the target user, Li, inputs "I want to invest in bank wealth management products with an annualized return of around 4%", and the system obtains the intent vector V1=[0.3,0.4,0.2,0.1] through a semantic encoding model. In the second round of dialogue, Li inputs "I want to invest in 3-year bank wealth management products with an annualized return of around 4%", and the system encodes the intent vector V2=[0.32,0.41,0.22,0.08]. The system calls the cosine similarity algorithm to calculate the similarity between V1 and V2, obtaining a similarity of 0.97 through vector dot product and modulus calculation. The system then obtains the preset adjustment parameter of 1.0 and performs a difference operation between 1.0 and 0.97, i.e., 1.0-0.97=0.03, resulting in an intent drift of 0.03. Since the drift is less than the preset threshold of 0.3, the system determines that the user's intent has not changed significantly and continues to extract the "3-year period" scenario-related features to support risk level updates and product recommendations.

[0048] This application's embodiments calculate the similarity between the intent vector of the current round and the intent vector of the previous round, and introduce preset adjustment parameters for difference processing to obtain the intent drift degree. This method can quantify the degree of change in user demand intent across multiple rounds of dialogue. Compared to traditional recommendation methods based solely on static profiles, it can more accurately capture dynamic changes in user demand intent, such as hesitation, regret, or drift, providing a reliable basis for subsequent accurate recommendations based on actual user needs and effectively avoiding the problem of mismatch between recommended products and actual user needs.

[0049] In some optional implementations, step 205, obtaining the target user's sentiment score in the current round, specifically includes the following steps: Using a pre-established sentiment analysis model, sentiment tendency analysis is performed on each text segment in the current round of dialogue data to obtain an initial sentiment value for each text segment. Based on the initial sentiment value, the dialogue data of the current round is comprehensively calculated to determine the overall sentiment score of the current round. If the overall sentiment score is greater than or equal to a preset score threshold, the overall sentiment score is determined as the sentiment score of the current round. If the overall sentiment score is less than the score threshold, secondary sentiment feature extraction is performed on the dialogue data of the current round to obtain supplementary sentiment features. By supplementing the sentiment features, the overall sentiment score is adjusted to obtain the sentiment score of the current round.

[0050] The sentiment analysis model refers to an AI model pre-trained on dialogue corpora, possessing the ability to recognize and quantify natural language emotions. Its core function is to analyze the emotional tendency in text and convert it into numerical indicators. Sentiment tendency analysis refers to the process by which the sentiment analysis model performs semantic analysis on each text segment of the current round of dialogue data, determining its emotional category (positive, negative, neutral) and initially quantifying it. This process breaks down the semantic structure of the text segment, extracts features such as emotional keywords and modal particles, and combines them with the context to determine the intensity of the emotional tendency, avoiding misjudgments caused by isolated analysis. The initial sentiment value refers to the numerical result output by the sentiment analysis model after performing sentiment tendency analysis on each text segment, used to quantify the emotional intensity of that segment. The value can range from -1 to 1. The closer the value is to 1, the more positive the emotion; the closer it is to -1, the more negative the emotion; 0 represents neutral emotion. The absolute value reflects the intensity of the emotion.

[0051] The comprehensive calculation refers to the process of integrating the initial sentiment value of each text segment into the overall sentiment level of the current round of dialogue data using a pre-set mathematical algorithm (such as weighted summation, averaging, and weighting by length). The overall sentiment score, obtained through comprehensive calculation, is a numerical indicator representing the overall sentiment state of the current round of dialogue data. Its value range is consistent with the initial sentiment value and is the core basis for determining whether secondary sentiment feature extraction is needed. The score threshold is a pre-set critical value based on the needs of the financial business scenario, the statistical characteristics of historical sentiment data, and the accuracy requirements of risk assessment. It is used to determine whether the overall sentiment score is within the range requiring further optimization.

[0052] Secondary emotion feature extraction refers to the process where, when the overall emotion score is below a threshold, the system employs a more refined feature extraction algorithm to deeply mine hidden information behind emotions from the dialogue data. Supplementary emotion features refer to the structured feature information obtained through secondary emotion feature extraction, used to explain the background of negative emotions and refine emotional attributes, including dimensions such as emotion attribution, emotion targeting, and emotion reversibility.

[0053] In one example, the sentiment analysis model employs a BERT-based fine-tuned model with a preset score threshold of -0.5. Secondary sentiment feature extraction focuses on three core dimensions: "sentiment attribution, scope of influence, and reversibility." The target user, Li, is currently engaged in a dialogue where the message is: "The market has been falling recently, and I've lost a lot of money on my previously invested funds. I'm particularly worried that I won't be able to recover my remaining principal in the short term." The system first semantically splits this dialogue into three text segments: "The market has been falling recently," "I've lost a lot of money on my previously invested funds, and I'm particularly worried that I won't be able to recover my remaining principal in the short term," and "I feel like I won't be able to recover my principal in the short term." Through sentiment analysis using the sentiment analysis model, combined with training experience with sentiment corpora in financial scenarios, the initial sentiment values ​​for the three segments are -0.5, -0.8, and -0.6, respectively. The system assigns weights to text fragments based on their semantic importance (with higher weights for core concerns, at 0.2, 0.6, and 0.2 respectively), and calculates the overall sentiment score as follows: (-0.5 × 0.2) + (-0.8 × 0.6) + (-0.6 × 0.2) = -0.1 - 0.48 - 0.12 = -0.7. This score is below the preset threshold of -0.5, triggering a secondary sentiment feature extraction process. The system uses sentiment attribution algorithms and contextual analysis to deeply mine and supplement sentiment features from the dialogue data: "Sentiment Attribution: Short-term fund losses", "Scope of Impact: Only currently held funds", "Reversibility: High (the user did not deny the long-term investment logic)", "Sentiment Intensity: Temporary anxiety". Based on these supplementary features, the system activates a preset adjustment rule: "The adjustment coefficient for temporary negative emotions caused by short-term market fluctuations is +0.4." This corrects the overall emotion score, resulting in an emotion score of -0.7 + 0.4 = -0.3 for the current round. This approach retains the user's current anxiety while eliminating the impact of short-term fluctuations on misjudgments of long-term risk preferences, thus better reflecting the user's true emotional state.

[0054] This application first uses a pre-established sentiment analysis model to analyze the sentiment tendency of each text segment in the current round of dialogue data, obtaining an initial sentiment value and comprehensively calculating the overall sentiment score, which can initially grasp the user's emotional state. When the overall sentiment score is lower than a preset threshold, further sentiment features are extracted to obtain supplementary sentiment features, and the overall sentiment score is adjusted accordingly. This multi-level and refined sentiment analysis method can more accurately capture the user's true emotions in the dialogue and avoid the errors of single analysis.

[0055] In some optional implementations, step S205, based on the target user's initial risk baseline value, dynamic context features, and sentiment score, updates the target user's risk tolerance level, specifically including the following steps: The initial risk baseline value for the target user is obtained, which is determined through a pre-set questionnaire. Based on the pre-set emotion weights, the emotion scores are weighted to obtain the emotion-weighted result. Based on the pre-set scenario weight coefficients, the dynamic context features are weighted to obtain the scenario-weighted result. The initial risk baseline value, the emotion-weighted result, and the scenario-weighted result are summed to obtain the target risk level value. Based on the target risk level value, the risk tolerance level of the target user is updated.

[0056] Among them, the emotion weight refers to the weight coefficient set in advance based on business risk assessment rules, historical emotion impact data, and compliance requirements, used to quantify the degree of influence of emotion score on risk tolerance level. The emotion weighted result refers to the numerical result obtained by multiplying the current round's emotion score by the preset emotion weight, which is the final quantitative impact value of emotion factors on risk tolerance level.

[0057] Among them, the scenario weight coefficient refers to the preset weight parameters for different scenario dimensions (such as funding source, investment period, debt status, etc.) in the dynamic context features, which are used to distinguish the priority of the impact of different scenarios on the risk tolerance level. The scenario weighted result refers to the comprehensive numerical result obtained by multiplying the quantified values ​​of each scenario dimension in the dynamic context features with the corresponding scenario weight coefficients and then summing them.

[0058] Among them, the target risk level value refers to the numerical indicator obtained by summing the initial risk benchmark value, the emotion-weighted result, and the scenario-weighted result, and is the core basis for updating the risk tolerance level.

[0059] In one example, target user Zhao obtained an initial risk baseline score of 65 (balanced) through a pre-set questionnaire. The current round's emotion score is -0.6 (negative anxiety), with a pre-set negative emotion weight of 0.3. The weighted emotion calculation yields: -0.6 × 0.3 × 10 (score scaling factor) = -1.8 points. The dynamic context features are "just bought a house and needs to make monthly mortgage payments" (quantified value -0.3) and "no other debts" (quantified value 0.1), with corresponding scenario weight coefficients of 0.4 and 0.2 respectively. The scenario weighted result is: (-0.3 × 0.4 + 0.1 × 0.2) × 10 = -1.0 points. Summing the initial baseline score, emotion, and scenario weighted results yields the target risk level score: 65 - 1.8 - 1.0 = 62.2 points. Based on this value, the system fine-tunes Zhao's risk tolerance level from balanced to "stable with a slight balance," ensuring a precise match between the risk level and the current emotion and scenario state.

[0060] This application first determines the initial risk baseline value for the target user through a pre-set questionnaire, laying the foundation for risk assessment. Then, based on pre-set emotion weights, the emotion score is weighted, and based on pre-set scenario weight coefficients, the dynamic context features are weighted, highlighting the different degrees of influence of emotion and dialogue scenario in risk assessment. The target risk level value is obtained by summing the initial risk baseline value with the weighted results of emotion and scenario, comprehensively considering multiple factors. This multi-dimensional, weighted fusion calculation method can more accurately and comprehensively update the user's risk tolerance level based on dynamic information in the user's dialogue, effectively avoiding the problem that traditional static models cannot adjust risk boundaries in a timely manner, and better adhering to the appropriateness principle of product recommendations.

[0061] In some optional implementations, step 206, based on the intent vector of the current round, risk tolerance level, and preset cross-domain product knowledge graph, determines the product recommendation list for the current round, specifically including the following steps: The intent vector of the current round is matched and filtered with the intent tags of the preset cross-domain product knowledge graph to obtain a set of products associated with the intent; based on the risk tolerance level, the products in the product set are subjected to risk filtering to obtain a set of candidate products after risk filtering; the product utility index of each product in the candidate product set is obtained; based on the product utility index, each product in the candidate product set is sorted to determine the product recommendation list of the current round.

[0062] Among them, matching and filtering refers to the process of semantically comparing the intent vector of the current round with the intent tags in the cross-domain product knowledge graph using a preset algorithm, and filtering out tags and corresponding products that highly match the user's intent. The product set refers to the set of all financial products related to the user's current intent obtained from the cross-domain product knowledge graph after intent tag matching and filtering, which is the initial candidate pool for product recommendations.

[0063] Risk filtering refers to the process of verifying the risk compliance of products in the initial product set based on the user's updated risk tolerance level, and removing products whose risk levels exceed the user's tolerance range. The candidate product set refers to the set of products whose risk levels match the user's risk tolerance level after risk filtering, and is the direct source of the final product recommendation list. Product utility metrics refer to core evaluation indicators used to quantify the degree to which products meet user needs, and are a key basis for product ranking, including but not limited to dimensions such as expected rate of return, product duration, liquidity, cost, and suitability for specific scenarios.

[0064] In one example, the target user Zhang's current intent vector is "low risk + 3-year term + retirement purpose," and the cross-domain product knowledge graph's intent tag set includes tags such as "low risk" and "retirement purpose." The system matches and filters the intent vector with the tags, resulting in a product set containing "XX 3-year stable financial management," "XX retirement annuity insurance," and "XX aggressive fund." Zhang's updated risk tolerance level is stable (58 points). The system performs risk filtering on the product set, removing "XX aggressive fund" (risk level 85 points), resulting in a candidate product set. The utility indicators of the candidate products are obtained and weighted: "XX 3-year stable financial management" has a comprehensive utility score of 8.2, and "XX retirement annuity insurance" has 7.8. The products are sorted in descending order of utility score to determine the product recommendation list, prioritizing the display of "XX 3-year stable financial management" and then "XX retirement annuity insurance," which aligns with the intent and meets compliance requirements.

[0065] This application's embodiments filter out a set of related products by matching the current round's intent vector with intent tags from a cross-domain product knowledge graph. This fully leverages the rich information of cross-business line knowledge graphs, breaking down business boundaries and enabling initial product screening across business lines to meet diverse user needs. Risk filtering of the product set is performed based on risk tolerance levels, strictly adhering to the "appropriateness principle" of product recommendations to ensure that the recommended product risk matches the user's risk tolerance, reducing compliance risks. Product utility metrics are obtained and sorted to determine the recommendation list. By comprehensively considering various product utility factors, the recommendation results are more closely aligned with the user's actual needs, improving recommendation accuracy and effectively solving the problems of inefficiency and inability to meet diverse needs in traditional recommendation logic.

[0066] In some optional implementations, the step "matching and filtering the intent vector of the current round with the intent tags of a preset cross-domain product knowledge graph to obtain a set of products associated with the intent" specifically includes the following steps: The intent tag set is extracted from a pre-defined cross-domain product knowledge graph. The intent tag set contains multiple tag information related to products. The intent vector of the current round is matched and calculated with each intent tag in the intent tag set to obtain the relevance score between each intent tag and the intent vector of the current round. Based on the relevance score, each intent tag is sorted and filtered to determine multiple target intent tags. The multiple target intent tags are matched with the corresponding nodes in the cross-domain product knowledge graph to obtain the corresponding product set.

[0067] The intent tag set refers to the combination of tags extracted from the cross-domain product knowledge graph, used to represent the core adaptation intent of various products. It serves as a bridge connecting user intent vectors and products. Tag information refers to the structured descriptive information contained in each tag in the intent tag set, including tag name, corresponding product attributes, adaptation scenario description, associated product node ID, etc., which is the core data for establishing the association between tags and products. Matching calculation refers to the process of using a vector similarity algorithm (such as cosine similarity) to calculate the degree of semantic association between the intent vector of the current round and the vector corresponding to each tag in the intent tag set. This process pre-encodes the tags into vectors with the same dimension as the intent vector, and measures the degree of matching by calculating the similarity between vectors. The relevance score refers to the quantitative indicator obtained through matching calculation, used to represent the degree of semantic association between the intent vector and each intent tag. The value range can be [0,1], and the closer the value is to 1, the higher the relevance and the more closely the intent and tag match. This score is the core basis for tag sorting and filtering.

[0068] Among them, the target intent tags refer to the top N (N being a preset number) intent tags with the highest relevance after being sorted according to their relevance scores. These are the key tags for accurately matching the user's core intent. Corresponding nodes refer to the product nodes in the cross-domain product knowledge graph that are associated with the target intent tags. Each node stores complete structured information about a single product (such as product name, risk level, return, term, etc.) and is the core data unit of the knowledge graph. Association matching refers to the process of querying the selected target intent tags with product nodes in the cross-domain product knowledge graph, obtaining the corresponding product information through the preset mapping relationship between tags and nodes.

[0069] In one example, the pre-defined cross-domain product knowledge graph's intent tag set includes tags such as "low risk," "3-year term," "for retirement purposes," and "short-term financial management," each coded as a vector. The target user, Li, has an intent vector for the current round encoded as [0.2, 0.5, 0.3, 0]. The system performs cosine similarity matching between this vector and each tag vector, obtaining relevance scores: "for retirement purposes" 0.91, "low risk" 0.87, "3-year term" 0.83, and "short-term financial management" 0.25. The top three tags are selected in descending order of relevance scores, identifying "for retirement purposes," "low risk," and "3-year term" as the target intent tags. These three tags are then matched with corresponding nodes in the knowledge graph to obtain products such as "XX Bank 3-year retirement financial management" and "XX Life Insurance retirement annuity insurance (3-year term)," all associated with these three tags, forming a product set associated with the intent.

[0070] This application's embodiments extract a set of intent tags containing multiple product-related tag information from a pre-defined cross-domain product knowledge graph, laying the foundation for accurate matching of user intents. A relevance score is calculated by matching the current round's intent vector with each intent tag, quantifying the degree of fit between the user intent and product tags, and accurately capturing user needs. Multiple target intent tags are determined by sorting and filtering according to the relevance score, allowing the selection of tags that best match the user's current intent and avoiding interference from irrelevant information. The target intent tags are then matched with corresponding nodes in the knowledge graph to obtain a product set. Based on the user's real-time intent, relevant products can be quickly located from a large and complex product matrix, enabling real-time customized filtering across business lines and effectively solving the problem of existing systems struggling to customize product filtering across business lines in real time.

[0071] In some optional implementations, after comparing the intended drift degree with a preset drift degree threshold in step 204, the following steps are further included: If the intention drift is greater than or equal to the preset drift threshold, a request confirmation message is generated and sent to the target user; the response message returned by the target user based on the request confirmation message is received; features are extracted from the dialogue data, request confirmation message and response message of the current dialogue round to obtain scene fusion features, and the scene fusion features are used as dynamic context features.

[0072] Among them, the demand confirmation information is a structured interactive message automatically generated by the system when the intent drift is greater than or equal to a preset threshold. Its core function is to verify the user's true intent and avoid deviations in subsequent scenario feature extraction and risk level calculation due to misjudgment of intent. Its content design needs to accurately locate the key differences in intent changes, use clear and guiding natural language, and provide users with options for clear responses, such as confirmation, denial, or supplementary explanation, to adapt to the interaction logic of multi-turn dialogues.

[0073] The response information, received by the target user after receiving the request confirmation information from the system, is returned through the user terminal such as a mobile app or web-based customer service interface. It serves as crucial supplementary data for the system to determine the user's true intent. Scene fusion features are a structured feature set obtained by the system after comprehensively extracting features from the original dialogue data, request confirmation information, and user response information for the current round. When intent drift exceeds the limit, these features replace the original scene-related features as dynamic contextual features. Its core characteristic lies in "integrating multi-source interactive data," encompassing not only the features of the user's initial request but also the features of the user's supplementary requests during intent confirmation, effectively avoiding the problem of missing scene information caused by single data sources.

[0074] In one example, the target user, Chen, initially inquired about "a one-year investment product with an annualized return of around 3%." The system encoded the user's input data using a semantic coding model, obtaining intent vector V1. In the current round, Chen input, "I've heard that retirement products are good, and I'd like to learn more about them." The system again processed this information, obtaining intent vector V2. The intent drift between V1 and V2 was then calculated, resulting in 0.45, while the preset drift threshold was 0.3. Clearly, the intent drift was greater than the threshold. The system then generated a request confirmation message: "You previously inquired about one-year investment products, and now you've mentioned retirement products. Would you like to switch to inquiries about retirement-related products (such as retirement investment or retirement annuity insurance)? If you have other needs, please provide further details." This message was sent to Chen's app client. After seeing the message, Chen replied, "I do want to learn about retirement-related products, but I prefer products with lower risk and flexible withdrawal options, as I may have unexpected cash needs later." After receiving the reply, the system performs comprehensive feature extraction on the dialogue data of the current round ("I heard that elderly care products are good, I would like to learn more about them"), the demand confirmation information, and the reply information. Finally, it obtains scenario fusion features such as "demand direction: elderly care related products", "risk preference: low risk", "liquidity demand: flexible withdrawal", and "demand adjustment: excluding high-risk, fixed-term elderly care products". These scenario fusion features are used as dynamic context features to provide data support for subsequent updates to Mr. Chen's risk tolerance level.

[0075] In this embodiment, when an intent drift is detected to be greater than or equal to a preset drift threshold, a demand confirmation message is generated and sent to the target user. This allows for timely detection of significant changes in the user's intent, proactively engaging with the user for confirmation, and preventing subsequent recommendation errors due to system misjudgment of user intent. Receiving the user's response based on the demand confirmation message provides clear feedback on the user's changed needs. Extracting scenario fusion features from the current dialogue round's data, demand confirmation message, and response message as dynamic context features comprehensively integrates multiple aspects of information, accurately capturing key elements of the user's needs during dynamic changes, thereby more accurately understanding the user's real-time needs.

[0076] It should be emphasized that, in order to further ensure the privacy and security of the aforementioned dialogue data, intent vectors, intent drift, scene association features, and product recommendation lists, these data can also be stored in a blockchain node.

[0077] The blockchain referred to in this application is a novel application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms. Essentially, a blockchain is a decentralized database, a chain of data blocks linked together using cryptographic methods. Each data block contains information about a batch of network transactions, used to verify the validity of the information (anti-counterfeiting) and generate the next block. A blockchain can include an underlying blockchain platform, a platform product service layer, and an application service layer.

[0078] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence (AI) refers to the theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0079] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by instructing related hardware through computer-readable instructions. These computer-readable instructions can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods. The aforementioned storage medium can be a non-volatile storage medium such as a magnetic disk, optical disk, or read-only memory (ROM), or random access memory (RAM).

[0080] Further reference Figure 3 As a response to the above Figure 2 To implement the method shown, this application provides an embodiment of a product recommendation device, which is similar to... Figure 2 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0081] like Figure 3 As shown, the product recommendation device 400 of this embodiment includes: an acquisition module 401, an encoding module 402, a calculation module 403, a comparison module 404, an update module 405, and a determination module 406. Wherein: The acquisition module 401 is used to acquire the dialogue data of each round of dialogue of the target user, including the user input data; The encoding module 402 is used to encode the user input data of the current round to obtain the intent vector of the current round; The calculation module 403 is used to calculate the intention drift degree between the intention vector of the current round and the intention vector of the previous round. The comparison module 404 is used to compare the intent drift degree with the preset drift degree threshold. If the intent drift degree is less than the drift degree threshold, the current round of dialogue data is used to extract features to obtain scene association features, and the scene association features are used as dynamic context features. The update module 405 is used to obtain the target user's sentiment score in the current round, and update the target user's risk tolerance level based on the target user's initial risk benchmark, dynamic context features and sentiment score; The determination module 406 is used to determine the product recommendation list for the current round based on the intent vector of the current round, the risk tolerance level, and the preset cross-domain product knowledge graph.

[0082] This application's embodiments acquire dialogue data from each round of the target user, encode the current round's user input data using a semantic coding model to obtain an intent vector, and calculate the drift degree compared to the previous round's intent vector. This accurately captures the dynamic changes in user needs and intents across multiple rounds of consultation, avoiding mismatches between recommended products and actual needs due to demand drift. When the intent drift degree is less than a threshold, scenario-related features are extracted as dynamic context features. Combined with the user's emotion score and initial risk benchmark value, the risk tolerance level is updated, enabling timely adjustment of risk boundaries based on dynamic information in the dialogue, reducing compliance risks. Simultaneously, based on the current round's intent vector, the updated risk tolerance level, and cross-domain product knowledge graphs, a product recommendation list is determined, enabling real-time customized filtering across business lines. This overcomes the limitations of traditional recommendation logic, such as low efficiency and the inability to recommend only preset solutions, meeting diverse user needs and improving the accuracy of product recommendations.

[0083] In one embodiment, the calculation module 403 includes: The similarity calculation submodule is used to calculate the similarity between the intent vector of the current round and the intent vector of the previous round. The difference processing submodule is used to obtain preset adjustment parameters, perform difference processing on the adjustment parameters and similarity, and obtain the intention drift degree.

[0084] In one embodiment, the update module 405 includes: The analysis submodule is used to perform sentiment analysis on each text segment in the current round of dialogue data using a pre-established sentiment analysis model, and obtain the initial sentiment value corresponding to each text segment. The comprehensive calculation submodule is used to perform comprehensive calculations on the dialogue data of the current round based on the initial emotion value, and determine the overall emotion score of the current round. The determination submodule is used to determine the overall emotion score as the emotion score for the current round if the overall emotion score is greater than or equal to a preset score threshold. The extraction submodule is used to perform secondary emotional feature extraction on the dialogue data of the current round if the overall emotion score is less than the score threshold, so as to obtain supplementary emotional features of the dialogue data. The adjustment submodule is used to adjust the overall emotion score by supplementing emotion features, so as to obtain the emotion score for the current round.

[0085] In one embodiment, the update module 405 includes: The first acquisition submodule is used to acquire the initial risk benchmark value of the target user. The initial risk benchmark value is determined by a preset questionnaire. The first weighted calculation submodule is used to perform weighted calculation on the emotion score based on the preset emotion weights to obtain the emotion weighted result; The second weighted calculation submodule is used to perform weighted calculation on dynamic context features based on preset scene weight coefficients to obtain scene weighted results; The third weighted calculation submodule is used to sum the initial risk benchmark value, the sentiment weighted result, and the scenario weighted result to obtain the target risk level value; The update submodule is used to update the risk tolerance level of the target user based on the target risk level value.

[0086] In one embodiment, the determining module 406 includes: The matching and filtering submodule is used to match and filter the intent vector of the current round with the intent tags of the preset cross-domain product knowledge graph to obtain a set of products associated with the intent. The filtering submodule is used to perform risk filtering on products in the product set based on risk tolerance level, so as to obtain a set of candidate products after risk filtering. The second acquisition submodule is used to acquire the product utility index of each product in the candidate product set; The sorting submodule is used to sort each product in the candidate product set based on product utility metrics to determine the product recommendation list for the current round.

[0087] In one embodiment, the matching and filtering submodule is further configured to extract an intent tag set from a preset cross-domain product knowledge graph, the intent tag set containing multiple tag information associated with products; perform matching calculations on the intent vector of the current round and each intent tag in the intent tag set to obtain the relevance score between each intent tag and the intent vector of the current round; sort and filter each intent tag according to the relevance score to determine multiple target intent tags; and perform association matching between the multiple target intent tags and corresponding nodes in the cross-domain product knowledge graph to obtain the corresponding product set.

[0088] In one embodiment, the product recommendation device 400 further includes: The generation module is used to generate a request confirmation message and send the request confirmation message to the target user if the intention drift is greater than or equal to a preset drift threshold. The receiving module is used to receive the response information returned by the target user based on the request confirmation information; The extraction module is used to extract features from the dialogue data, request confirmation information and response information of the current dialogue round to obtain scene fusion features, which are then used as dynamic context features.

[0089] To address the aforementioned technical problems, embodiments of this application also provide a computer device. Please refer to [link / reference needed]. Figure 4 , Figure 4 This is a basic structural block diagram of the computer device in this embodiment.

[0090] Computer device 6 includes a memory 61, a processor 62, and a network interface 63 that are interconnected via a system bus. It should be noted that only computer device 6 with memory 61, processor 62, and network interface 63 is shown in the figure; however, it should be understood that it is not required to implement all the components shown, and more or fewer components can be implemented alternatively. Those skilled in the art will understand that the computer device described herein is a device capable of automatically performing numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), embedded devices, etc.

[0091] Computer devices can include desktop computers, laptops, handheld computers, and cloud servers. These devices allow for human-computer interaction with users through keyboards, mice, remote controls, touchpads, or voice-activated devices.

[0092] The memory 61 includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 61 may be an internal storage unit of the computer device 6, such as the hard disk or memory of the computer device 6. In other embodiments, the memory 61 may also be an external storage device of the computer device 6, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device 6. Of course, the memory 61 may also include both the internal storage unit and the external storage device of the computer device 6. In this embodiment, the memory 61 is typically used to store the operating system and various application software installed on the computer device 6, such as computer-readable instructions for product recommendation methods. In addition, the memory 61 may also be used to temporarily store various types of data that have been output or will be output.

[0093] In some embodiments, processor 62 may be a central processing unit (CPU), controller, microcontroller, microprocessor, or other data processing chip. This processor 62 is typically used to control the overall operation of the computer device 6. In this embodiment, processor 62 is used to execute computer-readable instructions stored in memory 61 or to process data, such as computer-readable instructions for executing product recommendation methods.

[0094] The network interface 63 may include a wireless network interface or a wired network interface, which is typically used to establish a communication connection between the computer device 6 and other electronic devices.

[0095] This application's embodiments acquire dialogue data from each round of the target user, encode the current round's user input data using a semantic coding model to obtain an intent vector, and calculate the drift degree compared to the previous round's intent vector. This accurately captures the dynamic changes in user needs and intents across multiple rounds of consultation, avoiding mismatches between recommended products and actual needs due to demand drift. When the intent drift degree is less than a threshold, scenario-related features are extracted as dynamic context features. Combined with the user's emotion score and initial risk benchmark value, the risk tolerance level is updated, enabling timely adjustment of risk boundaries based on dynamic information in the dialogue, reducing compliance risks. Simultaneously, based on the current round's intent vector, the updated risk tolerance level, and cross-domain product knowledge graphs, a product recommendation list is determined, enabling real-time customized filtering across business lines. This overcomes the limitations of traditional recommendation logic, such as low efficiency and the inability to recommend only preset solutions, meeting diverse user needs and improving the accuracy of product recommendations.

[0096] This application also provides another embodiment, namely, providing a computer-readable storage medium storing computer-readable instructions that can be executed by at least one processor to cause the at least one processor to perform the steps of the product recommendation method described above.

[0097] This application's embodiments acquire dialogue data from each round of the target user, encode the current round's user input data using a semantic coding model to obtain an intent vector, and calculate the drift degree compared to the previous round's intent vector. This accurately captures the dynamic changes in user needs and intents across multiple rounds of consultation, avoiding mismatches between recommended products and actual needs due to demand drift. When the intent drift degree is less than a threshold, scenario-related features are extracted as dynamic context features. Combined with the user's emotion score and initial risk benchmark value, the risk tolerance level is updated, enabling timely adjustment of risk boundaries based on dynamic information in the dialogue, reducing compliance risks. Simultaneously, based on the current round's intent vector, the updated risk tolerance level, and cross-domain product knowledge graphs, a product recommendation list is determined, enabling real-time customized filtering across business lines. This overcomes the limitations of traditional recommendation logic, such as low efficiency and the inability to recommend only preset solutions, meeting diverse user needs and improving the accuracy of product recommendations.

[0098] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods of the various embodiments of this application.

[0099] Obviously, the embodiments described above are only some embodiments of this application, not all embodiments. The accompanying drawings show preferred embodiments of this application, but do not limit the patent scope of this application. This application can be implemented in many different forms; rather, the purpose of providing these embodiments is to provide a more thorough and comprehensive understanding of the disclosure of this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing specific embodiments, or make equivalent substitutions for some of the technical features. Any equivalent structures made using the content of this application's specification and drawings, directly or indirectly applied to other related technical fields, are similarly within the scope of patent protection of this application.

[0100] The software tools or components not belonging to our company that appear in the embodiments of this application are merely examples and do not represent actual use.

Claims

1. A product recommendation method, characterized in that, Includes the following steps: Acquire dialogue data for each round of conversation of the target user, the dialogue data including user input data; The user input data for the current round is encoded to obtain the intent vector for the current round; Calculate the intent drift between the intent vector of the current round and the intent vector of the previous round; The intent drift is compared with a preset drift threshold. If the intent drift is less than the drift threshold, then feature extraction is performed on the dialogue data of the current round to obtain scene association features, and the scene association features are used as dynamic context features. Obtain the target user's sentiment score in the current round, and update the target user's risk tolerance level based on the target user's initial risk baseline value, the dynamic context features, and the sentiment score; Based on the intent vector of the current round, the risk tolerance level, and the preset cross-domain product knowledge graph, the product recommendation list for the current round is determined.

2. The method according to claim 1, characterized in that, The step of calculating the intent drift between the intent vector of the current round and the intent vector of the previous round specifically includes: Calculate the similarity between the intent vector of the current round and the intent vector of the previous round; Obtain preset adjustment parameters, perform difference processing on the adjustment parameters and the similarity to obtain the intention drift degree.

3. The method according to claim 1, characterized in that, The step of obtaining the target user's emotion score in the current round specifically includes: Using a pre-established sentiment analysis model, sentiment tendency analysis is performed on each text segment in the current round of dialogue data to obtain the initial sentiment value corresponding to each text segment; Based on the initial emotion value, the dialogue data of the current round is comprehensively calculated to determine the overall emotion score of the current round; If the overall emotion score is greater than or equal to a preset score threshold, then the overall emotion score is determined as the emotion score for the current round. If the overall emotion score is less than the score threshold, then a second emotion feature extraction is performed on the dialogue data of the current round to obtain supplementary emotion features of the dialogue data. By supplementing the emotional features, the overall emotional score is adjusted to obtain the emotional score for the current round.

4. The method according to claim 1, characterized in that, The step of updating the risk tolerance level of the target user based on the initial risk benchmark value, the dynamic context features, and the sentiment score specifically includes: Obtain the initial risk benchmark value of the target user, which is determined through a pre-set questionnaire; Based on preset emotion weights, the emotion scores are weighted and calculated to obtain the emotion weighted result; Based on preset scene weight coefficients, the dynamic context features are weighted to obtain the scene weighting result; The target risk level value is obtained by summing the initial risk benchmark value, the emotion-weighted result, and the scenario-weighted result. Based on the target risk level value, update the risk tolerance level of the target user.

5. The method according to claim 1, characterized in that, The step of determining the product recommendation list for the current round based on the intent vector of the current round, the risk tolerance level, and a preset cross-domain product knowledge graph specifically includes: The intent vector of the current round is matched and filtered with the intent tags of the preset cross-domain product knowledge graph to obtain a set of products associated with the intent; Based on the risk tolerance level, the products in the product set are subjected to risk filtering to obtain a risk-filtered candidate product set. Obtain the product utility index for each product in the candidate product set; Based on the product utility index, each product in the candidate product set is sorted to determine the product recommendation list for the current round.

6. The method according to claim 5, characterized in that, The step of matching and filtering the intent vector of the current round with the intent tags of a preset cross-domain product knowledge graph to obtain a set of products associated with the intent specifically includes: Extract intent tag set from a preset cross-domain product knowledge graph, wherein the intent tag set contains multiple tag information associated with the product; The intent vector of the current round is matched with each intent tag in the intent tag set to obtain the correlation score between each intent tag and the intent vector of the current round; Based on the relevance scores, the intent tags are sorted and filtered to determine multiple target intent tags; The multiple target intent tags are associated and matched with the corresponding nodes in the cross-domain product knowledge graph to obtain the corresponding product set.

7. The method according to claim 1, characterized in that, After the step of comparing the intended drift degree with a preset drift degree threshold, the method further includes: If the intention drift is greater than or equal to a preset drift threshold, a request confirmation message is generated and sent to the target user. Receive the response information returned by the target user based on the request confirmation information; Feature extraction is performed on the dialogue data of the current dialogue round, the request confirmation information, and the response information to obtain scene fusion features, which are then used as dynamic context features.

8. A product recommendation device, characterized in that, include: The acquisition module is used to acquire dialogue data for each round of dialogue of the target user, the dialogue data including user input data; The encoding module is used to encode the user input data in the current round to obtain the intent vector of the current round; The calculation module is used to calculate the intention drift degree between the intention vector of the current round and the intention vector of the previous round. The comparison module is used to compare the intent drift degree with a preset drift degree threshold. If the intent drift degree is less than the drift degree threshold, then the dialogue data of the current round is used to extract features to obtain scene association features, and the scene association features are used as dynamic context features. The update module is used to obtain the target user's sentiment score in the current round, and update the target user's risk tolerance level based on the target user's initial risk benchmark value, the dynamic context features, and the sentiment score; The determination module is used to determine the product recommendation list for the current round based on the intent vector of the current round, the risk tolerance level, and a preset cross-domain product knowledge graph.

9. A computer device, characterized in that, The device includes a memory and a processor, wherein the memory stores computer-readable instructions, and the processor executes the computer-readable instructions to implement the steps of the product recommendation method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-readable instructions that, when executed by a processor, implement the steps of the product recommendation method as described in any one of claims 1 to 7.

Images