Data recommendation processing method and device based on dynamic calibration and related equipment

Abstract

The invention discloses a data recommendation processing method and device based on dynamic calibration and related equipment, relates to the technical field of data recommendation processing, and aims to overcome the defect of insufficient response capability of a static model to event dynamic evolution by combining a multi-dimensional influence factor tensor and performing customer portrait analysis on customer behavior data to obtain behavior characteristics. And dynamic coupling of market emotion and recommendation is realized. According to the multi-dimensional influence factor tensor, a compliance constraint vector obtained in advance and the relation logic tree after cognitive deviation correction, personalized recommendation information conforming to a preset rule is generated, it is ensured that output content conforms to compliance requirements, customer cognitive deviation is synchronously corrected, collaborative optimization of dynamic strategy suggestion generation and interpretable output is ensured, and user experience is improved. And personalized intelligent decision suggestions are formed, the deep response capability with personalized demands of clients in the recommendation information generation process is improved, and the deviation of the generated recommendation information is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data recommendation processing, more particularly to a data recommendation processing method and device based on dynamic calibration and related equipment. BACKGROUND

[0002] The current intelligent investment advisor system in the field of wealth management mainly relies on static rule engines and historical data backtesting models. After generating customer risk ratings through pre-set questionnaires, the system matches standardized investment portfolios to generate recommendation information such as investment suggestions and product recommendations.

[0003] In the existing process of generating recommendation information, collaborative filtering algorithms are usually used to recommend financial products. In the report generation stage of recommending financial products, most systems use template filling technology to convert investment suggestions into fixed format text, lacking the ability to respond in depth to market dynamics and customer individual needs. Moreover, the customer risk profile is static, and cannot deeply mine potential recommendation intentions and cognitive biases, resulting in deviations in the generated recommendation information.

[0004] Therefore, how to improve the depth of response to customer individual needs in the process of generating recommendation information and reduce the deviation of the generated recommendation information is a problem that needs to be solved by the present application. SUMMARY

[0005] Therefore, the present application discloses a data recommendation processing method and device based on dynamic calibration and related equipment, aiming to improve the depth of response to customer individual needs in the process of generating recommendation information and reduce the deviation of the generated recommendation information.

[0006] To achieve the above-mentioned purpose, the disclosed technical solution is as follows:

[0007] The first aspect of the present application discloses a data recommendation processing method based on dynamic calibration, which comprises:

[0008] acquiring heterogeneous data and customer behavior data;

[0009] performing event recognition on the heterogeneous data to obtain a multi-dimensional influence factor tensor;

[0010] performing multi-modal portrait construction on the customer behavior data to obtain behavior characteristics, and projecting the behavior characteristics to a preset decision space for hidden variable space mapping to obtain a relationship logic tree;

[0011] real-time monitoring whether there is a behavior risk point in the relationship logic tree;

[0012] If the relationship logic tree has a behavior risk point, correct the cognitive bias of the relationship logic tree;

[0013] According to the multi-dimensional influence factor tensor, the pre-acquired compliance constraint vector, and the relationship logic tree corrected for cognitive bias, personalized recommendation information conforming to a preset rule is generated.

[0014] Preferably, the event recognition on the heterogeneous data to obtain the multi-dimensional influence factor tensor comprises:

[0015] An embedding vector of an event type is extracted from the heterogeneous data by a pre-constructed multi-scale feature extractor;

[0016] An event influence weight matrix is calculated by an event impact intensity quantification model, the embedding vector of the event type, and a Granger causality test of market volatility;

[0017] According to the event influence weight matrix, a multi-dimensional influence factor tensor is generated.

[0018] Preferably, the multi-modal portrait construction on the customer behavior data to obtain behavior characteristics, and the projection of the behavior characteristics to a preset decision space for hidden variable space mapping to obtain a relationship logic tree, comprises:

[0019] Semantic clues are extracted from the customer behavior data by a dialogue state tracker, and behavior analysis is performed on the semantic clues by a behavior analysis engine to obtain behavior characteristics;

[0020] The behavior characteristics are projected to a preset decision space;

[0021] In the preset decision space, the behavior characteristics are matched with an asset logic framework library by attention to obtain a relationship logic tree.

[0022] Preferably, if the relationship logic tree has a behavior risk point, the relationship logic tree is corrected for cognitive bias, comprising:

[0023] If the relationship logic tree has a risk point, the type of the risk point is determined by a pre-constructed dynamic risk entropy model;

[0024] If the type of the risk point is loss aversion bias, an adjustment factor is injected into a decision model by a preset calibration formula to complete the cognitive bias correction of the relationship logic tree;

[0025] If the type of the risk point is confirmation bias, an opposite viewpoint generator is started;

[0026] By the opposite viewpoint generator, counterfactual scenario simulation, and counterfactual data, negative data is generated to complete the cognitive bias correction of the investment logic tree.

[0027] Preferably, the personalized recommendation information meeting the preset rules is generated according to the multi-dimensional influence factor tensor, the pre-acquired compliance constraint vector and the relationship logic tree corrected for cognitive bias, and the personalized recommendation information meeting the preset rules comprises:

[0028] A prefix vector is generated through the multi-dimensional influence factor tensor, the pre-acquired compliance constraint vector and the relationship logic tree corrected for cognitive bias.

[0029] The weight distribution of the attention mechanism in the generative model is adjusted through the prefix vector, so as to guide the generation of personalized recommendation information meeting the customer portrait and compliance requirements in terms of language style and terminology.

[0030] Preferably, the process of obtaining the compliance constraint vector comprises:

[0031] In the policy layer of the hierarchical control architecture, the model is trained in a preset combination space based on a reinforcement learning mode to calculate an optimal solution.

[0032] In the semantic layer of the hierarchical control architecture, real-time verification is performed through a preset knowledge graph, and when a potential violation expression is detected in the process of real-time verification, a corrector is triggered to reconstruct a sentence.

[0033] In the expression layer of the hierarchical control architecture, an adversarial training generative network is deployed, and through the adversarial training generative network and a discriminator of a pre-trained compliance consultant report library, ambiguous expressions and cognitive bias inducements are filtered.

[0034] The compliance constraint vector is generated through the optimal solution, the corrector reconstructed sentence and the discriminator.

[0035] Preferably, after the personalized recommendation information meeting the preset rules is generated, the method further comprises:

[0036] Control operations are performed in the process of decoding the personalized recommendation information, wherein the control operations at least include a structured prompt engineering, a real-time compliance sentinel and an explanatory enhancer.

[0037] The second aspect of the application discloses a data recommendation processing device based on dynamic calibration, and the device comprises:

[0038] An acquisition unit is configured to acquire heterogeneous data and customer behavior data.

[0039] An identification unit is configured to perform event identification on the heterogeneous data to obtain a multi-dimensional influence factor tensor.

[0040] A mapping unit is configured to perform multi-modal portrait construction on the customer behavior data to obtain behavior characteristics, and project the behavior characteristics to a preset decision space for hidden variable space mapping to obtain a relationship logic tree.

[0041] A real-time monitoring unit is configured to monitor whether there is a behavior risk point in the relationship logic tree in real time.

[0042] A deviation correction unit is configured to correct the cognitive deviation of the relationship logic tree if there is a behavior risk point in the relationship logic tree.

[0043] A generation unit is configured to generate personalized recommendation information in accordance with the multi-dimensional influence factor tensor, the pre-acquired compliance constraint vector, and the relationship logic tree after the cognitive deviation correction.

[0044] The third aspect of the present application discloses a storage medium including stored instructions, wherein the instructions, when executed, control a device in which the storage medium is located to perform the data recommendation processing method based on dynamic calibration as described in any one of the first aspect.

[0045] The fourth aspect of the present application discloses an electronic device including a memory and one or more instructions, wherein the one or more instructions are stored in the memory and configured to be executed by one or more processors to perform the data recommendation processing method based on dynamic calibration as described in any one of the first aspect.

[0046] According to the above technical solutions, the present application discloses a data recommendation processing method, device and related equipment based on dynamic calibration, acquires heterogeneous data and customer behavior data, performs event recognition on the heterogeneous data to obtain a multi-dimensional influence factor tensor, performs customer portrait analysis on the customer behavior data to obtain behavior characteristics, projects the behavior characteristics to a preset decision space for attention matching to obtain a relationship logic tree, monitors whether there is a behavior risk point in the relationship logic tree in real time, corrects the cognitive deviation of the relationship logic tree if there is a behavior risk point in the relationship logic tree, and generates personalized recommendation information in accordance with the multi-dimensional influence factor tensor, the pre-acquired compliance constraint vector, and the relationship logic tree after the cognitive deviation correction.

[0047] The beneficial effects of the embodiments of the present application are that the multi-dimensional influence factor tensor is obtained by combining event recognition on heterogeneous data, and the behavior characteristics are obtained by customer portrait analysis on customer behavior data, so as to overcome the insufficient response capability of static models to dynamic evolution of events, realize dynamic coupling of market sentiment and recommendation and investment strategy, project the behavior characteristics to a preset decision space to obtain a relationship logic tree by attention matching, so as to break through the limitation of template report generation, and build a personalized logic reasoning chain through semantic understanding. And according to the multi-dimensional influence factor tensor, the pre-acquired compliance constraint vector and the relationship logic tree corrected by cognitive bias, the personalized recommendation information conforming to the preset rules is generated, so as to ensure that the output content conforms to the compliance requirements, correct the customer's cognitive bias synchronously, ensure the collaborative optimization of dynamic strategy suggestion generation and explainability output, form personalized intelligent decision suggestions, improve the deep response capability to customer's personalized needs in the process of generating recommendation information, and reduce the deviation of the generated recommendation information. BRIEF DESCRIPTION OF DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of the provided drawings.

[0049] Figure 1 A flowchart of a data recommendation processing method based on dynamic calibration disclosed by an embodiment of the present application is shown in the figure.

[0050] Figure 2 A flowchart of another data recommendation processing method based on dynamic calibration disclosed by an embodiment of the present application is shown in the figure.

[0051] Figure 3 A structure diagram of a data recommendation processing device based on dynamic calibration disclosed by an embodiment of the present application is shown in the figure.

[0052] Figure 4 A structure diagram of an electronic device disclosed by an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0053] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0054] In this application, the terms "comprise", "contain", or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such a process, method, article or device. Without more limitations, the element defined by the statement "comprises a" does not exclude the presence of additional identical elements in the process, method, article or device comprising the element.

[0055] As can be known from the background art, in the existing process of generating recommendation information, a collaborative filtering algorithm is usually used to recommend financial products. In the report generation link of recommending financial products, most systems use template filling technology to convert investment suggestions into fixed format text, lacking the ability to respond in depth to market dynamic changes and personalized needs of customers. And the customer risk portrait presents a static feature, which cannot deeply mine potential recommendation intentions and cognitive biases, resulting in deviation in the generated recommendation information.

[0056] To solve the above problems, the application discloses a data recommendation processing method, device and related equipment based on dynamic calibration, which combines event recognition of heterogeneous data to obtain a multi-dimensional influence factor tensor, and customer portrait analysis of customer behavior data to obtain behavior characteristics, so as to overcome the insufficient response ability of static models to event dynamic evolution, realize the dynamic coupling of market sentiment and recommendation and investment strategy. Project the behavior characteristics into a preset decision space to obtain a relationship logic tree through attention matching, so as to break through the limitation of template report generation, and construct a personalized logical reasoning chain through semantic understanding. And according to the multi-dimensional influence factor tensor, the pre-acquired compliance constraint vector and the relationship logic tree corrected for cognitive bias, generate personalized recommendation information conforming to the preset rules, ensure that the output content meets the compliance requirements, correct the customer's cognitive bias synchronously, ensure the collaborative optimization of dynamic strategy suggestion generation and explainability output, form personalized intelligent decision suggestions, improve the depth response ability to personalized needs of customers in the process of generating recommendation information, and reduce the deviation of the generated recommendation information. The specific implementation mode is specifically explained through the following embodiments.

[0057] It should be noted that the data recommendation processing method, device and related equipment based on dynamic calibration provided by the application can be used in the technical fields of data recommendation processing, intelligent investment, large model, etc. The above is only an example and does not limit the application of the data recommendation processing method, device and related equipment based on dynamic calibration provided by the application.

[0058] It can be understood that before using the technical solutions disclosed in the embodiments of the present application, the type of personal information involved in the present application, the use range, the use scenario, etc. should be informed to the user and the authorization of the user should be obtained according to relevant laws and regulations.

[0059] It can be understood that the data involved in the technical solutions (including but not limited to the data itself, the acquisition or use of the data) should comply with the requirements of relevant laws and regulations and relevant provisions.

[0060] Reference Figure 1 As shown in the figure, the data recommendation processing method based on dynamic calibration disclosed in the embodiments of the present application mainly includes the following steps:

[0061] S101: Acquire heterogeneous data and customer behavior data.

[0062] The heterogeneous data is the external data source, and the heterogeneous data includes but is not limited to market data, financial events, market sentiment, etc.

[0063] The market data includes but is not limited to exchange market information, etc.; the financial events are news, financial report release, etc., which are structured events extracted by the BERT-Fin event extraction model; the market sentiment includes social media sentiment analysis index, news public opinion heat map, etc.

[0064] The input layer of the heterogeneous data fusion pipeline of the event dynamic response module accesses the above asynchronous data in real time.

[0065] S102: Perform event identification on the heterogeneous data to obtain a multi-dimensional influence factor tensor (i.e. event factor tensor).

[0066] In S102, the event identification is performed on the heterogeneous data by the processing layer of the event dynamic response module, and the multi-dimensional influence factor tensor is output by the output layer of the event dynamic response module.

[0067] The process of obtaining the multi-dimensional influence factor tensor is specifically shown in A1-A3.

[0068] A1: Extract an embedding vector of an event type from the heterogeneous data by a pre-constructed multi-scale feature extractor.

[0069] In the processing layer of the event dynamic response module, the embedding vector of the event type is extracted from the heterogeneous data by the multi-scale feature extractor.

[0070] The multi-scale feature extractor is constructed by using a temporal convolutional network (TCN), and the dilated convolution coefficients of the multi-scale feature extractor are dynamically adjusted according to the financial event decay law.

[0071] Event types include but are not limited to market sentiment types, financial event types and market sentiment types.

[0072] A2: Calculate the event impact weight matrix through the event impact intensity quantification model, the embedding vector of the event type and the Granger causality test of the market volatility.

[0073] In A2, in the processing layer of the event dynamic response module, the event impact weight matrix is calculated through the event impact intensity quantification model, the embedding vector of the event type and the Granger causality test of the market volatility.

[0074] A3: Generate a multi-dimensional impact factor tensor according to the event impact weight matrix.

[0075] In the output layer of the event dynamic response module, a multi-dimensional impact factor tensor (time x asset category x impact dimension) is generated according to the event impact weight matrix, which dynamically triggers the strategy parameter optimizer.

[0076] The strategy parameter optimizer is configured to receive the output of the multi-dimensional impact factor tensor and dynamically adjust the specific parameters of the investment strategy based on the output to optimize asset allocation.

[0077] S103: Perform multi-modal portrait construction on the customer behavior data to obtain behavior characteristics, and project the behavior characteristics to a preset decision space for hidden variable space mapping to obtain a relationship logic tree.

[0078] In S103, the multi-modal portrait construction and hidden variable space mapping of the customer behavior data are performed by the intent deep understanding module, and finally a relationship logic tree is obtained. The relationship logic tree can be a product recommendation logic tree, a customer exclusive investment logic tree, etc.

[0079] The relationship logic tree includes a root node and a plurality of branch nodes. The root node is a risk tolerance target. The branch nodes can be used to associate asset allocation rules.

[0080] The process of obtaining the relationship logic tree is shown as B1-B3.

[0081] B1: Extract semantic clues from the customer behavior data through the dialogue state tracker, and perform behavior analysis on the semantic clues through the behavior analysis engine to obtain behavior characteristics.

[0082] The dialogue state tracker is based on a long text conversation memory network (such as Transformer-XL), which updates the intent in real time through semantic clues (such as avoiding inflation risk, product recommendation, etc.) in customer questions.

[0083] Real-time updating of intent through semantic clues in customer questions (such as avoiding inflation risk, product recommendations, etc.) enables the system to dynamically capture customer's immediate needs (such as "avoiding inflation risk" triggering anti-inflation asset allocation logic), avoiding the lag of relying on historical profiling and improving the real-time relevance of recommendations.

[0084] Behavioral analysis of semantic clues through the behavior analysis engine, resulting in transaction records, which are extracted by the LSTM-Attention network to focus on important customer behaviors, i.e., behavioral characteristics. Behavioral characteristics include but are not limited to frequent portfolio adjustment coefficients / black swan event reaction delays, etc.

[0085] B2: Projecting behavioral characteristics into a pre-set decision space.

[0086] The pre-set decision space is a low-dimensional decision space. The low-dimensional decision space refers to a situation where the number of features of the data is relatively small. Since in the low-dimensional decision space, the distance between data points is relatively close, the data is more compact, and the computational complexity is relatively low.

[0087] B3: In the pre-set decision space, the behavioral characteristics are matched with the asset logic framework library through attention, and the relationship logic tree is obtained.

[0088] Through attention matching, customer complex characteristics are compressed into key decision dimensions (such as risk appetite / liquidity demand), and joint regulation asset framework is accurately related.

[0089] The output of the relationship logic tree is used to generate an interpretable personalized allocation path (for example: root node "moderate risk"→branch "gold allocation ≤10%+ treasury bond weight ≥30%"), ensuring that the strategy meets the customer's essential goals.

[0090] S104: Real-time monitoring of behavioral risk points in the relationship logic tree.

[0091] In S104, through the cognitive bias correction module and the dynamic risk entropy model, real-time monitoring of behavioral risk points in the relationship logic tree is performed in a virtual stress test environment.

[0092] A virtual stress test environment is constructed through a dynamic strategy backtest simulator, which injects historical extreme scenarios and random event shocks (such as Monte Carlo simulation), as well as robustness evaluation indicators such as the Sharpe ratio decay rate and maximum asset drawdown of the strategy under a 99% confidence interval, to construct a virtual stress test environment.

[0093] Sharpe ratio decay rate: a strategy excess return stability indicator in stress testing (such as a 47% decay from a Sharpe ratio of 1.5 to 0.8 under a 99% confidence interval).

[0094] Maximum drawdown: the maximum loss in net asset value from peak to trough in extreme scenarios (e.g. -20%).

[0095] Dynamic risk entropy model, i.e. risk preference map. The dynamic risk entropy model can be constructed by questionnaire results (preliminary static portrait) and flow data.

[0096] The preliminary static portrait refers to the relatively fixed risk preference and financial condition assessment results obtained through the standard questionnaire filled out by the customer.

[0097] Flow data, i.e. behavior data flow or dynamic interaction data. In addition to fund flow data (such as deposit, withdrawal, transfer frequency and size), it also includes but is not limited to the following types: transaction behavior data flow (order placement frequency, transaction execution time, holding period, buy / sell point selection), interactive behavior data flow (page dwell time on APP or client, number of queries for certain news or research reports), market response data flow (customer login APP activity, asset query behavior when facing market rally, market crash, or specific event news such as Fed rate hike).

[0098] The types of risk points include but are not limited to commodity recommendation risk points and financial risk points.

[0099] S105: If there is a behavior risk point in the relationship logic tree, correct the cognitive bias of the relationship logic tree.

[0100] The process of correcting the cognitive bias of the relationship logic tree is shown as C1-C4.

[0101] C1: If there is a risk point in the relationship logic tree, determine the type of risk point through the pre-constructed dynamic risk entropy model.

[0102] In C1, if there is a risk point in the relationship logic tree, the relationship logic tree risk point can be compared with the risk preference map through the dynamic risk entropy model to determine the type of risk point.

[0103] C2: If the type of risk point is loss aversion bias, inject an adjustment factor into the decision model through a pre-set calibration formula to complete the cognitive bias correction of the relationship logic tree.

[0104] In C2, if the type of risk point is loss aversion bias, inject an adjustment factor into the decision model through a pre-set calibration formula, such as "probability" and "value". The purpose of injecting the adjustment factor is to balance the irrational fear of loss and make the risk assessment closer to the "real" psychological weight.

[0105] Among them, the pre-set calibration formula includes but is not limited to the mental calibration formula of prospect theory (CPT).

[0106] Event factor injection timing can be dynamically adjusted (static preposition or real-time update).

[0107] C3: If the type of risk point is confirmation bias, start the opposite view generator.

[0108] C4: Generate counterfactual data through the opposite view generator, counterfactual scenario simulation, and counterfactual data to complete the cognitive bias correction of the investment logic tree.

[0109] For the convenience of understanding the process of real-time monitoring of whether there is a behavioral risk point in the relationship logic tree, this paper takes an example to illustrate:

[0110] For example, real-time monitoring of behavioral finance risk points in the relationship logic tree (logic chain); where the logic chain constructs the GAT asset causal diagram, which can be replaced by causal discovery algorithms (such as PC algorithm, data-driven causal discovery) or knowledge graph reasoning (supporting structured multi-hop reasoning);

[0111] If the finance risk point is loss aversion bias, inject a probability-weighted adjustment factor (CPT theory-driven);

[0112] If the finance risk point is confirmation bias, activate the opposite view generator (based on counterfactual data augmentation);

[0113] Monitor the thinking chain of investment decisions in a timely manner, and detect psychological traps (loss aversion, confirmation bias, etc.) through behavioral finance knowledge;

[0114] When loss aversion is found, immediately use the "psychological calibration formula" of prospect theory (CPT). Inject adjustment factors to "probability" and "value" in the decision-making model, the purpose is to balance your irrational fear of loss, and make risk assessment closer to the "real" psychological weight;

[0115] When confirmation bias is found, immediately start the opposite view generator, based on counterfactual scenario simulation and data, to create convincing "counterfactual evidence". The purpose is to break the information cocoon house and ensure that the system considers the neglected risks and counterpoints.

[0116] If the customer over-configures AI stocks (confirmation bias), the system will automatically generate counterfactual evidence:

[0117] Data simulation: inject historical technology stock bubble burst data (such as -78% of Nasdaq index in 2000);

[0118] Opposite view: output the quantitative conclusion "current AI plate margin rate has reached nearly 98% of the past ten years, and the probability of excess return is less than 15%, please invest carefully".

[0119] S106: generating personalized recommendation information conforming to the preset rules according to the multi-dimensional influence factor tensor, the pre-acquired compliance constraint vector and the relationship logic tree after the correction of the cognitive bias.

[0120] In S106, a prefix vector is generated through the multi-dimensional influence factor tensor, the compliance constraint vector and the relationship logic tree after the correction of the cognitive bias, and the weight distribution of the attention mechanism in the generation model is adjusted through the prefix vector to guide the generation of personalized recommendation information conforming to the customer portrait and compliance requirements in terms of language style and terminology.

[0121] The hierarchical control architecture is adopted to run through the whole process of generating the compliance constraint vector for the text. The process of obtaining the compliance constraint vector is shown as D1-D4.

[0122] D1: In the policy layer of the hierarchical control architecture, a reinforcement learning model is trained to explore the optimal solution in a preset combination space.

[0123] The preset combination space includes but is not limited to a commodity recommendation combination space and an investment portfolio space.

[0124] For example, in the policy layer, a reinforcement learning model is trained to explore the optimal solution in an investment portfolio space, and the reward function design needs to be double-constrained by risk value and maximum drawdown to ensure that the output strategy does not break the preset risk boundary.

[0125] D2: In the semantic layer of the hierarchical control architecture, real-time verification is performed through a preset knowledge graph.

[0126] D3: In the verification process, when a potential violation expression is detected, a modifier is triggered to reconstruct the sentence.

[0127] In the semantic layer, a knowledge graph (such as a financial regulation knowledge graph) is used to integrate regulatory rules to implement real-time verification, and when a potential violation expression is detected, a BERT-based modifier is triggered to reconstruct the sentence.

[0128] D4: In the expression layer of the hierarchical control architecture, an adversarial training generation network is deployed, and through the adversarial training generation network and the discriminator of the pre-trained compliance advisor report library, ambiguous expressions and cognitive bias inducements are filtered to complete the process of obtaining the compliance constraint vector using the hierarchical control architecture.

[0129] In the expression layer, an adversarial training generation network (GAN) is deployed, and a discriminator pre-trained on a compliance advisor report library is used to filter ambiguous expressions and cognitive bias inducements, such as automatically replacing violation promises such as "absolute returns" with compliance terms such as "historical backtracking returns".

[0130] The expression of the prefix vector is shown in formula (1).

[0131] Prefix vector = [event factor tensor] ⊕ [relation logic tree after cognitive bias correction]

[0132] ⊕ [compliance constraint vector] (1)

[0133] For example, the generative large model synthesis engine and the Prefix-Tuning efficient fine-tuning technology can be used to optimize the generation of personalized recommendation information and model reports.

[0134] In the relation logic tree after cognitive bias correction, the logic proof chain is constructed, and the prefix vector is generated through the multi-dimensional influence factor tensor, the compliance constraint vector, and the logic proof chain.

[0135] Logic proof chain construction:

[0136] Through the graph attention network (GAT), the causal reasoning path between asset allocation nodes is generated (such as "Fed rate hike → US Treasury yield rise → gold allocation weight down 5%").

[0137] Through the fine-tuned model, precise control is generated to generate personalized recommendation information:

[0138] For example, the prefix vector composed of events, customers, and compliance is used to adjust the weight distribution of the attention mechanism inside the large model, guiding the system to focus on knowledge and patterns highly related to the current event, the customer, and compliance requirements, and suppressing irrelevant or potentially illegal generated content, using language style and terminology that meets customer profiles and compliance requirements. Ensure that the final report content is relevant, highly personalized, and naturally compliant.

[0139] Specific Prefix-Tuning control generation examples are as follows:

[0140] Prefix vector [Fed rate hike event factor] ⊕ [customer low-risk logic tree] ⊕ [prohibition of "high yield" compliance constraints] guides the model:

[0141] 1. Attention redistribution: suppress "technology growth stock" recommendations and strengthen "high-rated bond" generation weights;

[0142] 2. Output term replacement: automatically replace "high yield" with "historical volatility lower than benchmark".

[0143] Build a trusted logic (such as GAT reasoning chain):

[0144] Treat assets as interconnected graph network nodes, use graph attention network (GAT) to simulate the dynamic transmission process of market events among assets, automatically mine the most influential and most relevant causal relationship chains, and convert these chains into clear, verifiable natural language reasoning paths, embedded in the report as core arguments.

[0145] After generating the personalized recommendation information conforming to the preset rules, the personalized recommendation information is decoded, and a control operation is performed in the process of decoding the personalized recommendation information, wherein the control operation at least includes a structured prompt project, a real-time compliance sentinel, and an explanatory enhancer.

[0146] Structured prompt project:

[0147] Decompose the investment suggestion into a chain template (precondition → argument → operation suggestion).

[0148] Real-time compliance sentinel:

[0149] The output token is matched with the subgraph of the regulatory knowledge graph (red line knowledge base), and the illegal word is automatically replaced with the recorded term (such as "guaranteed" → "principal guarantee type").

[0150] Explainable enhancer:

[0151] Insert attribution markers (▲ data-driven ● customer preference ■ compliance requirements) at key decision points to form machine-readable logical anchors. For example, generate the suggestion: "add government bonds (▲ 10-year yield rises to 3.5% ● customer risk aversion level A ■ comply with stable product regulatory guidelines)", where: the ▲ mark indicates the data-driven basis; the ● mark indicates the customer preference; and the ■ mark indicates the compliance requirement.

[0152] In order to facilitate the understanding of the process of data recommendation processing based on dynamic calibration, combined with Figure 2 Example:

[0153] For example, heterogeneous data is obtained from external data sources, and customer behavior data is obtained from customer terminals; wherein the external data sources include but are not limited to market quotations, news, financial reports, etc.; the customer behavior data includes but is not limited to transaction behavior and questionnaire feedback;

[0154] The heterogeneous data is sent to the event factor engine through the event extraction API interface, so as to identify events and quantify influences of the heterogeneous data through the event factor engine, and obtain a multi-dimensional influence factor tensor;

[0155] The customer behavior data is subjected to multi-modal portrait construction, i.e. customer portrait analysis, to obtain behavior characteristics, and the behavior characteristics are projected to a preset decision space for hidden variable space mapping to obtain a relationship logic tree; the customer portrait analysis at least includes behavior pattern extraction, risk entropy fusion, and dynamic risk graph construction (i.e. risk preference modeling); wherein the behavior pattern extraction and the risk preference modeling can be calculated in parallel and then fused;

[0156] Obtain a compliance constraint vector injected with compliance rules;

[0157] The relationship logic tree is monitored by the consultant intelligent agent in real time to determine whether there is a behavior risk point.

[0158] If the relationship logic tree has a behavior risk point, the relationship logic tree is corrected by a cognitive bias correction module; wherein the cognitive bias correction at least includes loss aversion adjustment and confirmation bias intervention;

[0159] The generative large model synthesis engine uses the GAT graph attention network after GAT causal reasoning and model fine-tuning (such as Prefix-Tuning), generates personalized investment consultant information (investment consultant information) according to the multi-dimensional influence factor tensor, the compliance constraint vector and the relationship logic tree after the cognitive bias correction, and conforms to the preset rules; wherein the model fine-tuning includes but is not limited to Prefix-Tuning, LoRA and other efficient fine-tuning methods.

[0160] The cognitive bias correction and report generation sequence of the present application can be interchangeable (first correction and then generation, or first generation of a preliminary draft and then targeted correction).

[0161] Optionally, in the scene embodiment of the investment consultant test on the customer's trading behavior, the present application tests the customer's trading behavior by constructing a prototype product. The sample scene is explained as follows:

[0162] A middle-aged manufacturing investor, the traditional questionnaire has been determined as "cautious" risk preference for a long time. When a sudden policy adjustment of a certain emerging industry in China triggered a shock in the industry chain, his account showed typical stress behavior: during the continuous decline of the industry chain related stocks in which he had a large position, he frequently placed orders and canceled them, and at the same time, a large amount of funds were transferred into a money market fund account but no operation was performed. The system real-time recognizes double cognitive bias early warning: the loss aversion coefficient is significantly higher than the safety threshold, and there is a confirmation bias feature of excessive search for "policy favorable interpretation".

[0163] Innovative effect:

[0164] 1. Dynamic risk modeling breaks through the static limitation:

[0165] The traditional system relies on historical questionnaire results and still maintains the "hold and watch" suggestion. The present system captures the fund risk aversion behavior as the core risk signal (behavior data weight dominant) through multi-source real-time sensing:

[0166] Capture fund risk aversion behavior as the core risk signal (behavior data weight dominant);

[0167] Fusion of industry fluctuation event impact caused by policy adjustment (event factor reinforcement);

[0168] Realize the dynamic reevaluation of risk preference and trigger targeted adjustment instructions;

[0169] Expected effect: generate risk control suggestions in advance during the policy window period, which significantly improves the timeliness compared with traditional methods.

[0170] 2. Cognitive bias intervention realizes decision optimization:

[0171] For loss aversion, the system starts a cognitive restructuring mechanism: separate the loss-making assets to an independent evaluation module, and demonstrate the long-term impact of different decision paths through a probability model. For confirmation bias, automatically generate a "policy impact analysis framework" with multiple perspectives, and implant the objective evolution rules of similar historical events.

[0172] Expected effect: User decision-making efficiency is improved to several times that of traditional scenarios, and key suggestion execution rate achieves a magnitude breakthrough.

[0173] 3. Generative logic proof builds trusted decision-making:

[0174] The report generation engine establishes a "policy-industry chain-enterprise fundamentals" three-level transmission chain:

[0175] From macro policy adjustment to midstream supply and demand changes, replace vague recommendations with traceable data logic closed loop;

[0176] Expected effect: The generated recommendation report has complete decision-making causal links, and through compliance verification, it generates personalized recommendation information that meets the preset rules, ensuring that the output content meets the compliance requirements.

[0177] The three major technological breakthroughs expected to be demonstrated in this case:

[0178] Risk perception dynamic: behavioral data becomes the main source of risk judgment, breaking the static nature of questionnaires;

[0179] Cognitive intervention initiative: psychological mechanism modeling effectively reduces non-rational decision-making impulses;

[0180] Report generation logic: verifiable decision-making replaces templated output.

[0181] This application breaks through the static limitations of traditional intelligent investment advisors, achieving the following effects:

[0182] 1. Dynamic event response: Through TCN adaptive inflation convolution and event intensity quantification model, replace fixed time window analysis method, significantly reduce strategy adjustment delay.

[0183] 2. Intention depth analysis innovatively integrates behavioral pattern analysis and dynamic risk entropy modeling, improving customer intention recognition accuracy by at least 30% over existing levels.

[0184] 3. Compliance generation control uses a three-layer real-time verification architecture to suppress regulatory violations to at least 50% below the current level within milliseconds.

[0185] Cognitive bias intervention first algorithmizes the prospect theory (CPT), reducing customer non-rational decision-making by at least 40% through real-time injection of behavior correction factors.

[0186] Each module forms a closed loop verification in the dynamic strategy backtracking test simulator, ensuring that the final output asset allocation report has strategy traceability, market adaptability and regulatory penetration, fundamentally solving the industry pain point of coexistence of strategy rigidity and compliance risk in the wealth management field.

[0187] The present application aims to build a large model intelligent recommendation (adviser) system based on dynamic cognitive calibration, and the core solves the following technical problems:

[0188] (1) Overcome the insufficient response capability of static models to dynamic evolution of events such as financial events, and realize dynamic coupling of market sentiment and investment strategy;

[0189] (2) Break through the limitations of template report generation, and build individualized logical reasoning chain through semantic understanding;

[0190] (3) Ensure that the output content meets the compliance requirements, and correct the customer's cognitive bias simultaneously. Focus on solving the collaborative optimization of dynamic strategy suggestion generation and explainability output, forming intelligent decision suggestions different from simple product recommendations.

[0191] At the dynamic adaptation level, the present application improves the strategy adjustment response speed by at least 50% through the financial event time series modeling module; in the dimension of personalized service, the investment suggestion adoption rate is increased by at least 30% through the intent deep understanding model, and the cognitive bias correction module effectively reduces the frequency of irrational transactions by 30%; in the compliance control link, the special decoder compresses the regulatory clause violation rate to less than half of the existing level.

[0192] The beneficial effects of the embodiments of the present application are: combining the multi-dimensional influence factor tensor obtained by event recognition on heterogeneous data and the behavior characteristics obtained by customer portrait analysis on customer behavior data to overcome the insufficient response capability of static models to dynamic evolution of events, realize dynamic coupling of market sentiment and recommendation, and investment strategy. Project the behavior characteristics into the preset decision space to get the relationship logic tree, so as to break through the limitations of template report generation, and build individualized logical reasoning chain through semantic understanding. And according to the multi-dimensional influence factor tensor, the pre-acquired compliance constraint vector and the relationship logic tree after correction of cognitive bias, generate individualized recommendation information that meets the preset rules, ensure that the output content meets the compliance requirements, correct the customer's cognitive bias simultaneously, ensure the collaborative optimization of dynamic strategy suggestion generation and explainability output, form individualized intelligent decision suggestions, improve the deep response capability to customer's individualized needs in the process of generating recommendation information, and reduce the deviation of generated recommendation information.

[0193] Based on the above embodiments Figure 1The disclosed data recommendation processing method based on dynamic calibration also discloses a data recommendation processing device based on dynamic calibration, as shown in the figure. Figure 3 The data recommendation processing device based on dynamic calibration includes

[0194] The acquisition unit 301 is configured to acquire heterogeneous data and customer behavior data.

[0195] The recognition unit 302 is configured to perform event recognition on the heterogeneous data to obtain a multi-dimensional influence factor tensor.

[0196] The mapping unit 303 is configured to perform multi-modal portrait construction on the customer behavior data to obtain behavior characteristics, and project the behavior characteristics to a preset decision space for implicit variable space mapping to obtain a relationship logic tree.

[0197] The real-time monitoring unit 304 is configured to monitor whether there is a behavior risk point in the relationship logic tree in real time.

[0198] The bias correction unit 305 is configured to correct the relationship logic tree for cognitive bias if there is a behavior risk point in the relationship logic tree.

[0199] The generation unit 306 is configured to generate personalized recommendation information conforming to a preset rule according to the multi-dimensional influence factor tensor, a pre-acquired compliance constraint vector, and the relationship logic tree corrected for cognitive bias.

[0200] Further, the recognition unit 302 includes

[0201] The extraction module is configured to extract an embedding vector of an event type from the heterogeneous data through a pre-constructed multi-scale feature extractor.

[0202] The calculation module is configured to calculate an event influence weight matrix through an event impact intensity quantification model, the embedding vector of the event type, and a Granger causality test of market volatility.

[0203] The first generation module is configured to generate the multi-dimensional influence factor tensor according to the event influence weight matrix.

[0204] Further, the mapping unit 303 includes

[0205] The extraction and analysis module is configured to extract semantic clues from the customer behavior data through a dialogue state tracker, and perform behavior analysis on the semantic clues through a behavior analysis engine to obtain behavior characteristics.

[0206] The projection module is configured to project the behavior characteristics to a preset decision space.

[0207] An attention matching module is configured to perform attention matching between the behavior characteristics and the asset logic framework library in a preset decision space to obtain a relationship logic tree.

[0208] Further, the bias correction unit 305 includes:

[0209] A determination module is configured to determine the type of the risk point by using a pre-constructed dynamic risk entropy value model if the relationship logic tree has a risk point.

[0210] An injection module is configured to inject an adjustment factor into the decision model by using a preset calibration formula to complete the cognitive bias correction of the relationship logic tree if the type of the risk point is loss aversion bias.

[0211] A starting module is configured to start an opposite viewpoint generator if the type of the risk point is confirmation bias.

[0212] A second generation module is configured to generate negative data by using the opposite viewpoint generator, the counterfactual scenario simulation, and the counterfactual data to complete the cognitive bias correction of the investment logic tree.

[0213] Further, the generation unit 306 includes:

[0214] A third generation module is configured to generate a prefix vector by using the multi-dimensional influence factor tensor, the pre-acquired compliance constraint vector, and the relationship logic tree after the cognitive bias correction.

[0215] A fourth generation module is configured to adjust the weight distribution of the attention mechanism in the generative model by using the prefix vector to guide the generation of personalized recommendation information in a language style and terms that meet the customer portrait and compliance requirements.

[0216] Further, the generation unit 307 for obtaining the compliance constraint vector includes:

[0217] A calculation module is configured to calculate an optimal solution in a preset combination space by using a model trained in a reinforcement learning manner in a policy layer of a hierarchical control architecture.

[0218] A triggering module is configured to perform real-time verification by using a preset knowledge graph in a semantic layer of the hierarchical control architecture, and to trigger a corrector to reconstruct a statement when detecting a potential violation expression in the process of real-time verification.

[0219] A fifth generation module is configured to deploy an adversarial training generation network in an expression layer of the hierarchical control architecture, and to filter ambiguous expressions and cognitive bias inducements by using the adversarial training generation network and a discriminator of a pre-trained compliance advisor report library.

[0220] A sixth generation module is configured to generate a compliance constraint vector by using the optimal solution, the corrector-reconstructed statement, and the discriminator.

[0221] Further, the data recommendation processing device based on dynamic calibration further comprises:

[0222] A control operation unit configured to perform control operations in the process of decoding the personalized recommendation information, wherein the control operations at least include a structured prompt engine, a real-time compliance sentinel, and an explanatory enhancer.

[0223] The application embodiment has the beneficial effects that: the multi-dimensional influence factor tensor is obtained by combining event recognition on heterogeneous data, and the behavior characteristics are obtained by customer portrait analysis on customer behavior data, so as to overcome the insufficient response capability of a static model to dynamic evolution of events, realize dynamic coupling of market sentiment and recommendation and investment strategy, project the behavior characteristics to a preset decision space to obtain a relationship logic tree, so as to break through the limitation of template report generation, build a personalized logic reasoning chain through semantic understanding, generate personalized recommendation information in accordance with the multi-dimensional influence factor tensor, the pre-acquired compliance constraint vector, and the relationship logic tree corrected in terms of cognitive bias, ensure that the output content meets the compliance requirements, correct the customer cognitive bias synchronously, ensure the collaborative optimization of dynamic strategy suggestion generation and explainability output, form personalized intelligent decision suggestions, improve the deep response capability to customer personalized needs in the process of generating recommendation information, and reduce the bias of the generated recommendation information.

[0224] The application embodiment further provides a storage medium, which comprises stored instructions, wherein when the instructions are executed, the device where the storage medium is located performs the data recommendation processing method based on dynamic calibration.

[0225] The application embodiment further provides an electronic device, a structural schematic diagram of which is shown in Figure 4 The electronic device specifically comprises a memory 401 and one or more than one instruction 402, wherein the one or more than one instruction 402 is stored in the memory 401 and is configured to execute the one or more than one instruction 402 executed by the one or more than one processor 403 to execute the data recommendation processing method based on dynamic calibration.

[0226] For each method embodiment described above, in order to simply describe, the method embodiments are all described as a series of action combinations, but those skilled in the art should know that the application is not limited to the action sequence described, because according to the application, some steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the specification all belong to preferred embodiments, and the actions and modules involved are not necessarily necessary for the application.

[0227] It should be noted that each of the embodiments in the specification is described in progressive mode, and each embodiment focuses on the difference from other embodiments. The same and similar parts between embodiments can be referred to each other. For system embodiments, since they are basically similar to method embodiments, the description is relatively simple, and the relevant parts can be referred to the part of the method embodiment.

[0228] The steps in the method of each embodiment of the present application can be adjusted, combined and reduced in sequence according to actual needs.

[0229] Finally, it should be noted that in this paper, relationship terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations.

[0230] The above description of the disclosed embodiments enables a person skilled in the art to implement or use the present application. Various modifications to the embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

[0231] The above is only the preferred embodiment of the present application. It should be noted that for ordinary skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, which should be considered as the protection scope of the present application.

Claims

1. A data recommendation processing method based on dynamic calibration, characterized in that, The method comprises: acquiring heterogeneous data and customer behavior data; performing event recognition on the heterogeneous data to obtain a multi-dimensional influence factor tensor; performing multi-modal portrait construction on the customer behavior data to obtain behavior characteristics, and projecting the behavior characteristics to a preset decision space for implicit variable space mapping to obtain a relationship logic tree; real-time monitoring of whether there is a behavior risk point in the relationship logic tree; if the relationship logic tree has a behavior risk point, correcting the relationship logic tree for cognitive bias; generating personalized recommendation information conforming to a preset rule according to the multi-dimensional influence factor tensor, a pre-acquired compliance constraint vector and the relationship logic tree corrected for cognitive bias.

2. The method of claim 1, wherein, The event recognition on the heterogeneous data to obtain a multi-dimensional influence factor tensor comprises: extracting an event type embedding vector from the heterogeneous data through a pre-constructed multi-scale feature extractor; calculating an event influence weight matrix through an event impact intensity quantification model, the event type embedding vector and a Granger causality test of market volatility; generating a multi-dimensional influence factor tensor according to the event influence weight matrix.

3. The method of claim 1, wherein, The multi-modal portrait construction on the customer behavior data to obtain behavior characteristics, and projecting the behavior characteristics to a preset decision space for implicit variable space mapping to obtain a relationship logic tree comprises: extracting semantic clues from the customer behavior data through a dialogue state tracker, and performing behavior analysis on the semantic clues through a behavior analysis engine to obtain behavior characteristics; projecting the behavior characteristics to a preset decision space; in the preset decision space, performing attention matching between the behavior characteristics and an asset logic framework library to obtain a relationship logic tree.

4. The method of claim 1, wherein, If the relationship logic tree has a behavior risk point, correcting the relationship logic tree for cognitive bias comprises: if the relationship logic tree has a risk point, determining the type of the risk point through a pre-constructed dynamic risk entropy model; if the type of the risk point is loss aversion bias, injecting an adjustment factor into a decision model through a preset calibration formula to complete the cognitive bias correction of the relationship logic tree; if the type of the risk point is confirmation bias, starting an opposing view generator; generating negative data through the opposing view generator, counterfactual scenario simulation and counterfactual data to complete the cognitive bias correction of the investment logic tree.

5. The method of claim 1, wherein, The generation of personalized recommendation information conforming to a preset rule according to the multi-dimensional influence factor tensor, a pre-acquired compliance constraint vector and the relationship logic tree corrected for cognitive bias comprises: generating a prefix vector through the multi-dimensional influence factor tensor, the pre-acquired compliance constraint vector and the relationship logic tree corrected for cognitive bias; adjusting the weight distribution of the attention mechanism in the generative model through the prefix vector to guide the generation of personalized recommendation information conforming to the customer portrait and compliance requirements in terms of language style and terminology.

6. The method of claim 1, wherein, The process of obtaining a compliance constraint vector comprises: training a model in a preset combination space to calculate an optimal solution based on reinforcement learning in a policy layer of a hierarchical control architecture; In the semantic layer of the hierarchical control architecture, real-time verification is performed through a preset knowledge graph. In the process of real-time verification, when a potential violation expression is detected, a corrector is triggered to reconstruct the statement; In the expression layer of the hierarchical control architecture, an adversarial training generation network is deployed to filter ambiguous expressions and cognitive bias induction through the adversarial training generation network and a discriminator of a pre-trained compliance consultant report library; An optimal solution, the corrector reconstructed statement and the discriminator are used to generate a compliance constraint vector.

7. The method of claim 1, wherein, After generating personalized recommendation information that meets the preset rules, the method further includes: During decoding of the personalized recommendation information, a control operation is performed, wherein the control operation at least includes a structured prompt engineering, a real-time compliance sentinel and an explanatory enhancer.

8. A dynamic calibration-based data recommendation processing apparatus, characterized by comprising: The apparatus includes: An acquisition unit configured to acquire heterogeneous data and customer behavior data; An identification unit configured to perform event identification on the heterogeneous data to obtain a multi-dimensional influence factor tensor; A mapping unit configured to perform multi-modal portrait construction on the customer behavior data to obtain behavior characteristics, and project the behavior characteristics to a preset decision space for implicit variable space mapping to obtain a relationship logic tree; A real-time monitoring unit configured to monitor whether there is a behavior risk point in the relationship logic tree in real time; A bias correction unit configured to correct cognitive bias of the relationship logic tree if there is a behavior risk point in the relationship logic tree; A generation unit configured to generate personalized recommendation information that meets preset rules according to the multi-dimensional influence factor tensor, a pre-acquired compliance constraint vector and the relationship logic tree after cognitive bias correction.

9. A storage medium, characterized by The storage medium includes stored instructions, wherein when the instructions are executed, the device where the storage medium is located performs the data recommendation processing method based on dynamic calibration according to any one of claims 1 to 7.

10. An electronic device, comprising: The apparatus includes a memory, and one or more instructions, wherein the one or more instructions are stored in the memory and are configured to be executed by one or more processors to perform the data recommendation processing method based on dynamic calibration according to any one of claims 1 to 7.

Images