Marketing analysis method, device, equipment and storage medium

Abstract

The application discloses a marketing analysis method, device and equipment and a storage medium. The method comprises the following steps: obtaining a historical business analysis report, performing quality evaluation on the historical business analysis report based on a preset screening rule, and determining a target sample set according to the quality evaluation result; performing supervised fine-tuning on a pre-trained large model according to the target sample set, obtaining a fine-tuned model; optimizing the fine-tuned model based on reinforcement learning, and obtaining a target analysis model; obtaining a plurality of business analysis objects to be analyzed and corresponding business data, inputting each business analysis object and the corresponding business data into the target analysis model, and outputting an analysis result corresponding to each business analysis object. Since the application performs data screening and model post-training, the target analysis model autonomously masters the parallel reasoning capability of multiple business analysis objects, and compared with the prior art, the application improves the accuracy and stability of the analysis result in marketing analysis.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a marketing analysis method, apparatus, device, and storage medium. Background Technology

[0002] Currently, workflow-based marketing intelligence systems are widely used in business scenarios such as customer acquisition and financial product marketing. These systems typically include an intent recognition module, an analysis point execution module, and a report generation module. Their core working principle is as follows: When a user query request is received, the intent recognition module first parses the query question to determine the list of target analysis points. Then, for each analysis point, the analysis point execution module retrieves the corresponding customer information data from the database and, combined with a pre-defined analysis approach (i.e., manually written prompts), calls a large model to generate an analysis report for that single analysis point. Finally, the report generation module summarizes and generates the final analysis report.

[0003] However, the aforementioned existing technical solutions still have significant limitations in practical applications, resulting in insufficient accuracy and stability of the analysis results. Specifically, the analysis approach relied upon by the analysis point execution module is usually preset manually, and its token length is too long or its description is ambiguous, which makes large models prone to attention distraction or logical deviation during the reasoning process, resulting in large fluctuations in the quality of the generated analysis results.

[0004] Therefore, there is an urgent need for a marketing analysis method that can improve the accuracy and stability of the analysis results in marketing analysis. Summary of the Invention

[0005] The main objective of this invention is to provide a marketing analysis method, apparatus, device, and storage medium, which aims to solve the technical problem of insufficient accuracy and stability of analysis results in existing marketing analysis.

[0006] To achieve the above objectives, the present invention provides a marketing analysis method, the method comprising the following steps: Obtain historical business analysis reports, conduct quality assessments on the historical business analysis reports based on preset screening rules, and determine the target sample set based on the quality assessment results; The pre-trained large model is supervised fine-tuned based on the target sample set to obtain the fine-tuned model. The fine-tuned model is optimized based on reinforcement learning to obtain the target analysis model; Obtain multiple business analysis objects to be analyzed and their corresponding business data, input each business analysis object and its corresponding business data into the target analysis model, and output the analysis results corresponding to each business analysis object.

[0007] Optionally, the step of obtaining historical business analysis reports, performing quality assessments on the historical business analysis reports based on preset screening rules, and determining the target sample set based on the quality assessment results includes: Obtain user interaction behavior feedback data and rule verification mark data corresponding to each historical business analysis report. The user interaction behavior feedback data includes a first type of feedback data representing positive evaluation and a second type of feedback data representing negative evaluation. Using the historical business analysis reports corresponding to the first type of feedback data as positive samples and the historical business analysis reports corresponding to the second type of feedback data and the rule verification mark data as negative samples, a preset classification model is trained to obtain a quality assessment model. The quality assessment model is used to assess the quality of each of the historical business analysis reports. Based on the quality assessment results, analysis reports that meet the preset threshold conditions are selected from the historical business analysis reports and determined as high-quality analysis reports. The input and output data of the high-quality analysis report are concatenated to generate the target sample set.

[0008] Optionally, the step of concatenating the input and output data of the high-quality analysis report to generate the target sample set includes: Extract the identifiers of multiple analysis points and the corresponding original business data contained in the high-quality analysis report as sample input features; Extract the logical derivation process and analysis conclusion corresponding to each analysis point identifier in the high-quality analysis report as sample output features; The sample input features and sample output features are structured and encapsulated according to the order of analysis points to generate the target sample set.

[0009] Optionally, the step of optimizing the fine-tuned model based on reinforcement learning to obtain the target analysis model includes: The sample input data from the target sample set is input into the fine-tuned model, and multiple candidate output results are generated based on a preset strategy. Each candidate output result is evaluated using a preset reward function to obtain a corresponding reward value. The preset reward function includes a format standardization reward function and a content accuracy reward function. Based on the reward value corresponding to each of the candidate output results, the relative advantage value corresponding to each of the candidate output results is determined, and the policy gradient is determined based on the relative advantage value. The parameters of the fine-tuned model are updated using a preset optimizer based on the policy gradient to obtain the target analysis model.

[0010] Optionally, the step of evaluating each candidate output result using a preset reward function to obtain the corresponding reward value includes: The matching degree between the structure and format of the candidate output result and the preset output template is evaluated using the format normalization reward function to obtain the format reward value; The semantic similarity between the candidate output and the reference output is evaluated using the content accuracy reward function to obtain a similarity reward value; The quality assessment model is used to predict the quality level of the candidate output results to obtain a quality reward value; The format reward value, similarity reward value, and quality reward value are weighted and fused based on preset weight coefficients to obtain the reward value corresponding to each candidate output result.

[0011] Optionally, the step of supervising the fine-tuning of the pre-trained large model based on the target sample set to obtain the fine-tuned model includes: The input sequence length and output sequence length of each sample in the target sample set are determined, and the maximum sequence length parameter of the pre-trained large model is determined based on a preset quantile statistical strategy. Each sample in the target sample set is padded or truncated according to the maximum sequence length parameter to obtain standardized training data; Based on the low-rank adaptive fine-tuning algorithm, the parameters of the pre-trained large model are fine-tuned using the standardized training data to obtain the fine-tuned model.

[0012] Optionally, the step of acquiring multiple business analysis objects to be analyzed and their corresponding business data, inputting each business analysis object and its corresponding business data into the target analysis model, and outputting the analysis results corresponding to each business analysis object includes: Receive user query requests and parse multiple business analysis objects and customer identification information to be analyzed from the user query requests; Based on the customer identification information, retrieve business data corresponding to each of the aforementioned business analysis objects from the business database; Based on the order of the business analysis objects, the identification information of each business analysis object and the corresponding business data are structurally concatenated to generate multi-object analysis input data; The multi-object analysis input data is input into the target analysis model in a single step, and the analysis results corresponding to each of the business analysis objects are output.

[0013] Furthermore, to achieve the above objectives, the present invention also proposes a marketing analysis device, the device comprising: The data filtering module is used to obtain historical business analysis reports, perform quality assessment on the historical business analysis reports based on preset filtering rules, and determine the target sample set based on the quality assessment results. The model fine-tuning module is used to perform supervised fine-tuning of the pre-trained large model based on the target sample set to obtain the fine-tuned model. The reinforcement learning module is used to optimize the fine-tuned model based on reinforcement learning to obtain the target analysis model; The results output module is used to acquire multiple business analysis objects to be analyzed and their corresponding business data, input each business analysis object and its corresponding business data into the target analysis model, and output the analysis results corresponding to each business analysis object.

[0014] Furthermore, to achieve the above objectives, the present invention also proposes a marketing analysis device, the device comprising: a memory, a processor, and a marketing analysis program stored in the memory and executable on the processor, the marketing analysis program being configured to implement the steps of the marketing analysis method as described above.

[0015] In addition, to achieve the above objectives, the present invention also proposes a storage medium storing a marketing analysis program, which, when executed by a processor, implements the steps of the marketing analysis method as described above.

[0016] This invention discloses a method for acquiring historical business analysis reports, evaluating the quality of these reports based on preset screening rules, determining a target sample set based on the evaluation results, supervising the fine-tuning of a pre-trained large model using the target sample set to obtain a fine-tuned model, optimizing the fine-tuned model using reinforcement learning to obtain a target analysis model, acquiring multiple business analysis objects to be analyzed and their corresponding business data, inputting each business analysis object and its corresponding business data into the target analysis model, and outputting the analysis results for each business analysis object. Because this invention determines the target sample set by evaluating the quality of historical business analysis reports, employs a two-stage training strategy combining supervised fine-tuning and reinforcement learning to obtain the target analysis model, and enables the processing of multiple business analysis objects in a single call, compared to existing technologies, this invention improves the accuracy and stability of analysis results in marketing analysis. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the first embodiment of the marketing analysis method of the present invention; Figure 2 This is a flowchart illustrating the second embodiment of the marketing analysis method of the present invention; Figure 3 This is a flowchart illustrating the third embodiment of the marketing analysis method of the present invention; Figure 4 This is a structural block diagram of the first embodiment of the marketing analysis device of the present invention; Figure 5 This is a schematic diagram of the marketing analysis device of the hardware operating environment involved in the embodiments of the present invention.

[0018] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0020] This invention provides a marketing analysis method, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the marketing analysis method of the present invention.

[0021] In this embodiment, the marketing analysis method includes steps S10 to S40: Step S10: Obtain historical business analysis reports, conduct quality assessments on the historical business analysis reports based on preset screening rules, and determine the target sample set based on the quality assessment results.

[0022] It should be noted that the executing entity in this embodiment can be a computer server device with data processing, network communication, and program execution functions applied in a financial marketing analysis scenario, such as a server, tablet computer, or personal computer, or an electronic device capable of performing the above functions (such as a marketing analysis device). The following uses a system containing a marketing analysis device (hereinafter referred to as the system) as an example to illustrate this embodiment and the following embodiments.

[0023] It should be understood that the aforementioned historical business analysis reports may refer to various business analysis reports that have been generated and stored by the system during its past operation, such as corporate customer acquisition analysis reports, financial product marketing analysis reports, and deposit business expansion analysis reports in the context of bank corporate marketing.

[0024] It is understood that the aforementioned preset screening rules may refer to the pre-set screening conditions or standards used to evaluate the quality of historical business analysis reports. These screening rules are used to select high-quality reports from a large number of historical business analysis reports, such as screening conditions set based on user interaction behavior feedback data, screening conditions based on manual rule marking, or a combination thereof.

[0025] It should be explained that the above quality assessment results may refer to the quantitative or graded evaluation results obtained after evaluating the quality of each historical business analysis report, such as score values ​​or grade labels.

[0026] It should be understood that the aforementioned target sample set may refer to a collection of high-quality historical business analysis reports that have been determined after quality assessment and screening and are used for subsequent model training. This collection includes the input and output data corresponding to each selected historical business analysis report.

[0027] In its implementation, when constructing the target sample set, the system first retrieves its stored historical business analysis reports. These reports originate from data generated and saved during previous business analysis tasks. Next, the system performs a quality assessment on each retrieved historical business analysis report according to preset filtering rules. The assessment process uses quality evaluation indicators defined in the preset filtering rules to quantitatively score or grade each report, resulting in a quality assessment result for each report. Then, based on the quality assessment results of each historical business analysis report, the system selects reports that meet the quality thresholds set by the preset filtering rules from all historical business analysis reports. These selected reports are then designated as the target sample set for subsequent model training.

[0028] To facilitate understanding, the following examples illustrate the concept, but do not impose specific limitations on this embodiment. For instance, in a corporate banking marketing scenario, the system retrieves one thousand customer acquisition analysis reports generated within the past month as historical business analysis reports. The system evaluates the quality of each analysis report according to preset screening rules. These rules include a scoring standard based on data from customer managers' likes and dislikes, as well as manually marked bad sample data. The system calculates a quality score for each report based on this scoring standard. Subsequently, the system filters out reports with quality scores higher than a preset threshold, for example, selecting the top 20% of reports by quality score, and defines these reports as the target sample set.

[0029] Step S20: Perform supervised fine-tuning of the pre-trained large model based on the target sample set to obtain the fine-tuned model.

[0030] Understandably, the aforementioned pre-trained large models can refer to large neural network models that have been pre-trained on massive amounts of general data and possess basic language understanding and generation capabilities, such as the Qwen series models and the LLaMA series models.

[0031] It should be understood that the aforementioned supervised fine-tuning can refer to further parameter adjustments to a pre-trained large model using a labeled target sample set, so as to improve the model's performance in specific business domains (such as bank marketing analysis scenarios). During the supervised fine-tuning process, the model will learn the mapping relationship between input data and output data in the target sample set.

[0032] Furthermore, the aforementioned fine-tuned model can refer to a model obtained by supervising the fine-tuning of a pre-trained large model, which has stronger analytical capabilities on a specific task.

[0033] In its implementation, after determining the target sample set, the system uses it as training data for supervised fine-tuning. The system takes the input data from each high-quality analysis report in the target sample set as model input and the corresponding output data as the supervision label for the model's expected output. Next, the system uses a pre-trained large model as the initial model and trains it in a supervised manner using the input data from the target sample set and the corresponding supervision label. During training, the system calculates the difference between the model's predicted output and the supervision label using the backpropagation algorithm and updates the parameters of the pre-trained large model based on this difference, gradually approximating the supervision label in the model's predictions on the target sample set. After multiple rounds of iterative training, the system obtains a fine-tuned model. This fine-tuned model inherits the general capabilities of the pre-trained large model and also possesses the ability to generate analysis results according to the analysis logic in the target sample set under specific business analysis scenarios.

[0034] To facilitate understanding, the following examples illustrate the concept, but do not impose specific limitations on this embodiment. For instance, in a corporate banking marketing scenario, the system described above has obtained a target sample set containing 500 high-quality corporate customer acquisition analysis reports through quality assessment and screening. The system acquires a Qwen3-14B pre-trained large model as the initial model. The system uses the input data of each report in the target sample set (e.g., multi-analysis point data containing basic corporate information and financial data) as model input, and the output data corresponding to each report (e.g., a complete analysis report containing the analysis ideas and conclusions of each analysis point) as supervision labels. The system uses these 500 samples to perform supervised fine-tuning training on the Qwen3-14B pre-trained large model. After parameter updates, a fine-tuned model is obtained. When this fine-tuned model sees new corporate data input, it can generate corresponding analysis results according to the analysis logic learned from the target sample set.

[0035] To improve the efficiency and stability of model fine-tuning, the step of supervising the fine-tuning of the pre-trained large model based on the target sample set to obtain the fine-tuned model may include: determining the input sequence length and output sequence length of each sample in the target sample set; determining the maximum sequence length parameter of the pre-trained large model based on a preset quantile statistical strategy; performing sequence padding or truncation on each sample in the target sample set according to the maximum sequence length parameter to obtain standardized training data; and fine-tuning the parameters of the pre-trained large model using the standardized training data based on a low-rank adaptive fine-tuning algorithm to obtain the fine-tuned model.

[0036] It should be explained that the above input sequence length may refer to the number of tokens contained in the input data of each sample in the target sample set after word segmentation. For example, an analysis input containing basic corporate information and financial data may contain a certain number of tokens after being processed by the word segmenter of a large model.

[0037] The output sequence length mentioned above can refer to the number of tokens contained in the output data of each sample in the target sample set after word segmentation. For example, the number of tokens contained in an output text containing multi-analysis point analysis conclusions after word segmentation.

[0038] It should be noted that the aforementioned preset quantile statistical strategy can refer to a statistical analysis method that sorts the sum of the input sequence length and the output sequence length of all samples and takes the value corresponding to a preset percentile as a reference value. For example, the 95th percentile means that the total length of 95% of all samples is less than or equal to this value.

[0039] Understandably, the maximum sequence length parameter mentioned above can refer to the maximum number of tokens that a pre-trained large model can process during training. Any tokens exceeding this length will be truncated, for example, 8192 tokens.

[0040] Furthermore, the aforementioned sequence padding can refer to the operation of adding padding markers to the end of the sequence to make the sequence length reach the maximum sequence length when the sequence length of the sample is less than the maximum sequence length. The aforementioned truncation can refer to the operation of removing the portion exceeding the length from the beginning or end of the sequence when the sequence length of the sample is greater than the maximum sequence length.

[0041] Accordingly, the standardized training data mentioned above can refer to a dataset with a uniform sequence length obtained after sequence padding or truncation, which can be directly used for batch training of large models.

[0042] It should be noted that the aforementioned low-rank adaptive fine-tuning algorithm can refer to a parameter-efficient model fine-tuning method. This method involves adding a low-rank decomposition matrix next to the weight matrix of a pre-trained large model, and freezing the original pre-trained weights by updating only the parameters of the low-rank matrix during training. For example, the LoRA (Low-Rank Adaptation) algorithm can achieve effective model adaptation while maintaining low computational resource consumption.

[0043] In its implementation, after obtaining the target sample set, the system reads all samples in the target sample set and calculates the length of the input sequence corresponding to the input data and the length of the output sequence corresponding to the output data for each sample.

[0044] Next, the system adds the input sequence length to the output sequence length for each sample to obtain the total sequence length for each sample. The system then sorts the total sequence lengths of all samples and, based on a preset quantile statistical strategy, determines the total sequence length value corresponding to a preset percentile. This value is then used as the maximum sequence length parameter for the pre-trained large model.

[0045] Subsequently, the system performs length standardization on each sample in the target sample set according to the determined maximum sequence length parameter: if the total sequence length of a sample is less than the maximum sequence length parameter, the system adds padding to the end of the sample's sequence to make its total length reach the value corresponding to the maximum sequence length parameter; if the total sequence length of a sample is greater than the maximum sequence length parameter, the system cuts off the excess portion from the end of the sample's sequence according to a preset truncation rule, making the total length of the truncated sample equal to the value corresponding to the maximum sequence length parameter; if the total sequence length of a sample is equal to the maximum sequence length parameter, the system does not pad or truncate the sample. After performing the above processing on all samples, the system obtains standardized training data.

[0046] Subsequently, the system uses a low-rank adaptive fine-tuning algorithm to fine-tune the parameters of a pre-trained large model using standardized training data: the system adds a low-rank decomposition matrix next to the weight matrix of the pre-trained large model, keeps the original pre-trained weights unchanged during training, and only updates the parameters of the low-rank decomposition matrix. Through backpropagation and iterative updates by the optimizer, the fine-tuned model is finally obtained.

[0047] For example, in a corporate banking marketing scenario, the target sample set obtained by the above system contains one thousand high-quality analysis reports. The system performs word segmentation on the input and output data of each report, calculates the input sequence length and output sequence length, and adds them together to obtain the total sequence length of each report. The preset quantile statistical strategy is set to the 95th quantile. The system sorts the total sequence lengths of the one thousand reports from shortest to longest, taking the total sequence length of the 950th sample as a reference value. Assuming this value is 8000, the system determines the maximum sequence length parameter to be 8192 (usually an integer power of 2 and not less than this reference value). Using 8192 as the maximum sequence length parameter, the system performs sequence padding or truncation on each report, padding samples with a total sequence length less than 8192 to 8192, and truncating samples with a total sequence length exceeding 8192 to 8192, thus obtaining standardized training data. Subsequently, the system employs the LoRA low-rank adaptive fine-tuning algorithm, adding a low-rank decomposition matrix with a rank of 8 next to the weight matrix of the pre-trained large model Qwen3-14B. The parameters of the low-rank matrix are updated using standardized training data. After two rounds of training iterations, the fine-tuned model is obtained.

[0048] It should be understood that by adaptively determining the maximum sequence length through quantile statistics strategy, and in conjunction with sequence standardization and low-rank adaptive fine-tuning, dynamic adaptation between data length distribution and model processing capabilities is achieved. This reduces memory usage and computational overhead while ensuring information integrity, significantly improving the efficiency and stability of model fine-tuning.

[0049] Step S30: Optimize the fine-tuned model based on reinforcement learning to obtain the target analysis model.

[0050] Understandably, reinforcement learning can refer to a machine learning training paradigm in which an agent obtains reward or punishment signals by interacting with the environment and adjusts its own behavioral strategy based on these signals to maximize the cumulative reward. In this embodiment, it is specifically manifested as the process of using a fine-tuned model as the initial strategy and iteratively optimizing the model parameters by generating output results and receiving reward feedback.

[0051] Accordingly, the aforementioned target analysis model can refer to the final model obtained after fine-tuning the model and optimizing it through reinforcement learning. This model has higher analytical accuracy, standardized output format, and parallel processing capability for multiple analysis points in business analysis tasks.

[0052] In its implementation, after obtaining the fine-tuned model, the system uses it as the initial policy model for reinforcement learning training. The system inputs the business analysis task's data as state input to the fine-tuned model, which then generates corresponding analysis results as action outputs. The system evaluates the analysis results generated by the fine-tuned model according to a pre-defined reward evaluation mechanism, calculating a reward value to measure the quality of the generated results. Based on the reward value, the system calculates the policy gradient and uses this gradient to iteratively update the parameters of the fine-tuned model, making the model more likely to produce outputs with higher reward values ​​in subsequent generation processes. After multiple rounds of reinforcement learning iterations, the system determines the model with updated parameters as the target analysis model.

[0053] To facilitate understanding, the following explanation uses examples, but does not impose specific limitations on this embodiment. For example, in a corporate banking marketing scenario, the system has already obtained a supervised fine-tuned Qwen3-14B-SFT model. The system uses this fine-tuned model as the initial policy model for reinforcement learning and optimizes it using the GRPO reinforcement learning algorithm. The system inputs the input data from the target sample set into the fine-tuned model, which generates multiple candidate analysis conclusions. The system calculates a reward value for each candidate output based on the format reward function and the content reward function based on the quality assessment model, calculates the policy gradient based on the reward value, and updates the model parameters through the optimizer. After one round of reinforcement learning iteration training, the system determines the trained model as the target analysis model.

[0054] Step S40: Obtain multiple business analysis objects to be analyzed and their corresponding business data, input each business analysis object and its corresponding business data into the target analysis model, and output the analysis results corresponding to each business analysis object.

[0055] It should be explained that the above-mentioned business analysis objects can refer to specific dimensions or items that require specific business analysis. For example, in the scenario of corporate banking marketing, business analysis objects can include different analysis points such as basic enterprise information analysis, financial status analysis, upstream and downstream industrial chain analysis, and potential product demand analysis.

[0056] Accordingly, the aforementioned business data may refer to the original data information corresponding to each business analysis object. For example, for the analysis object of basic enterprise information, the corresponding business data may include the enterprise name, registered address, years of establishment, registered capital, etc.; for the analysis object of financial status, the corresponding business data may include financial indicators such as operating revenue, net profit, and debt-to-equity ratio.

[0057] It is understandable that the above analysis results may refer to the conclusive output generated by the target analysis model after reasoning and analyzing each business analysis object, such as the conclusion on the business operation stability generated for the basic information analysis object, or the debt repayment ability assessment conclusion generated for the financial status analysis object.

[0058] In its implementation, after obtaining the target analysis model, upon receiving a user's business analysis request, the system retrieves multiple business analysis objects to be analyzed and the corresponding business data for each object. The system then organizes the identification information of each business analysis object and its corresponding business data according to a preset data organization structure, forming input data that conforms to the input format requirements of the target analysis model. Next, the system inputs the organized input data into the target analysis model for inference and calculation. After processing the input data, the target analysis model generates corresponding analysis results for each business analysis object and outputs the analysis results for all business analysis objects in a structured format. The system receives the analysis results output by the target analysis model and provides these results to the user as the final analysis conclusion.

[0059] It should be added that, after the step of obtaining the target analysis model, the process may further include: deploying the target analysis model to the analysis point execution module of the workflow to replace the original base model in the analysis point execution module; when a user query request containing multiple business analysis objects is received, the target analysis model is called through the analysis point execution module, and the target analysis model generates the analysis results corresponding to each of the business analysis objects at once.

[0060] It should be understood that by deploying the post-trained target analysis model into the workflow to replace the original base model, when receiving multi-object analysis requests, all analysis results can be generated in a single call, eliminating the need to call external large model interfaces separately for each analysis point. This effectively reduces the number of interface calls and token consumption, avoids rate limiting risks, significantly improves system processing efficiency, and reduces response latency.

[0061] To improve the overall system throughput and processing efficiency, the step of acquiring multiple business analysis objects to be analyzed and their corresponding business data, inputting each business analysis object and its corresponding business data into the target analysis model, and outputting the analysis results corresponding to each business analysis object may include: receiving a user query request; parsing the multiple business analysis objects to be analyzed and customer identification information from the user query request; obtaining business data corresponding to each business analysis object from the business database based on the customer identification information; structurally concatenating the identification information and corresponding business data of each business analysis object according to the order of the business analysis objects to generate multi-object analysis input data; and inputting the multi-object analysis input data into the target analysis model in a single step, and outputting the analysis results corresponding to each business analysis object.

[0062] It is understandable that the aforementioned user query request may refer to a data instruction initiated by a user to the system that includes business analysis requirements. This request usually includes the specific object identifier that the user wants to analyze and the target customer identifier to be analyzed. For example, a query request submitted by a bank account manager through the system interface to conduct customer acquisition analysis for a specific target company.

[0063] The aforementioned customer identification information may refer to identification data used to uniquely identify target customers, such as customer number, unified social credit code, or enterprise registration number. The system retrieves original business data related to the customer from the database based on this identification.

[0064] It should be noted that the aforementioned business database can refer to a storage system that is internal or externally connected to the system and is used to store customer business-related data. This database contains data records of multiple dimensions, such as basic customer information, financial information, and transaction information.

[0065] It should be explained that the order of the business analysis objects mentioned above can refer to the arrangement order of multiple business analysis objects in the input data. This arrangement order is used to determine the order of the business data and analysis results corresponding to each business analysis object in the structured and concatenated text sequence.

[0066] It should be added that the order of the above business analysis objects can be determined by the order of analysis objects specified in the user's query request, or by the system's preset arrangement rules for analysis objects, such as arranging them according to the business logic dependencies of the analysis objects, according to the priority of the analysis objects, or according to the preset number of the analysis objects.

[0067] Understandably, in a marketing analytics scenario, the order of business analytics objects determines the order in which the content of each analytics object is read when the target analytics model receives multi-object analytics input data, and also determines the order in which the conclusions corresponding to each analytics object are arranged when the target analytics model outputs analytics results.

[0068] The aforementioned multi-object analysis input data can refer to a complete data structure generated by combining the identification information and corresponding business data of multiple business analysis objects in a specific format, which can be directly input into the target analysis model. This data structure contains all the analysis basis information required for all the objects to be analyzed.

[0069] It is understandable that the above-mentioned structured splicing can refer to combining and arranging the identification information and corresponding business data of each business analysis object with specific delimiters or hierarchical structures according to preset format rules to form a standardized data string that meets the input format requirements of the target analysis model.

[0070] In its implementation, the system receives user query requests, which contain multiple business analysis objects the user wishes to analyze, as well as customer identification information used to identify target customers. The system then parses the received query request, extracting all the business analysis objects and customer identification information to be analyzed.

[0071] Next, the system retrieves the original business data corresponding to the customer identification information obtained from the parsing, and extracts the business data corresponding to each business analysis object from the original business data obtained.

[0072] Subsequently, the system, based on the order of the business analysis objects, structurally splices the identification information of each business analysis object and the corresponding business data according to preset splicing rules, combining the identification information and business data of all business analysis objects into a complete set of multi-object analysis input data.

[0073] Finally, the system inputs the generated multi-object analysis input data into the target analysis model in a single call. The target analysis model processes the multi-object analysis input data uniformly, generates corresponding analysis results for each business analysis object, and outputs them.

[0074] To facilitate understanding, the following example illustrates the concept, but does not impose specific limitations on this embodiment. For instance, in a corporate banking marketing scenario, a bank account manager submits a customer acquisition analysis query request for a target company through the system interface. This request includes three business analysis objects: "Basic Enterprise Information Analysis," "Financial Health Analysis," and "Industry Competitiveness Analysis," along with the target company's Unified Social Credit Code as customer identification information. Upon receiving the query request, the system parses out the three business analysis objects and the Unified Social Credit Code. Based on the Unified Social Credit Code, the system retrieves the company's original business data from the business database, obtaining the company's business registration information as the business data corresponding to "Basic Enterprise Information Analysis," the financial statement data for the past three years as the business data corresponding to "Financial Health Analysis," and the industry ranking report data as the business data corresponding to "Industry Competitiveness Analysis." Following the order of the three business analysis objects, the system structurally combines the identification information of the three objects and their corresponding business data according to a preset template to generate a multi-object analysis input data. The system inputs this multi-object analysis data into the target analysis model via a single API call. The target analysis model then outputs analysis conclusions for "enterprise basic information analysis," "financial health analysis," and "industry competitiveness analysis" in one go.

[0075] It should be understood that by structurally concatenating the identification information and corresponding business data of multiple business analysis objects, multi-object analysis input data is generated and input into the target analysis model in a single step. This achieves parallel inference of multiple analysis points, significantly reduces the number of model interface calls, avoids triggering rate limiting mechanisms, reduces end-to-end response latency, and improves the overall throughput and processing efficiency of the system.

[0076] This embodiment discloses a method for acquiring historical business analysis reports, evaluating the quality of these reports based on preset screening rules, determining a target sample set based on the evaluation results, performing supervised fine-tuning of a pre-trained large model using the target sample set to obtain a fine-tuned model, optimizing the fine-tuned model using reinforcement learning to obtain a target analysis model, acquiring multiple business analysis objects to be analyzed and their corresponding business data, inputting each business analysis object and its corresponding business data into the target analysis model, and outputting analysis results for each business analysis object. Because this embodiment determines the target sample set by evaluating the quality of historical business analysis reports, employs a two-stage training strategy combining supervised fine-tuning and reinforcement learning to obtain the target analysis model, and achieves the ability to process multiple business analysis objects in a single call, this embodiment improves the accuracy and stability of analysis results in marketing analysis compared to existing technologies.

[0077] refer to Figure 2 , Figure 2 This is a flowchart illustrating the second embodiment of the marketing analysis method of the present invention.

[0078] Based on the first embodiment described above, in this embodiment, step S10 includes steps S101 to S104: Step S101: Obtain user interaction behavior feedback data and rule verification mark data corresponding to each historical business analysis report. The user interaction behavior feedback data includes a first type of feedback data representing positive evaluation and a second type of feedback data representing negative evaluation.

[0079] It should be noted that the aforementioned user interaction feedback data can refer to evaluative data generated by users during their interaction with historical business analysis reports. This data reflects the user's level of recognition or satisfaction with the report's quality, such as recorded data generated by users liking, disliking, rating, or marking the report as useful.

[0080] The first type of feedback data mentioned above can refer to data representing positive evaluations in user interaction behavior feedback data, such as the behavior record generated by users clicking the "like" button on the report, or the record of users giving the report a high score.

[0081] The second type of feedback data mentioned above can refer to data representing negative evaluations in user interaction behavior feedback data, such as the behavior records generated by users clicking the "dislike" button on the report, or the records of users giving low scores to the report.

[0082] It should be explained that the aforementioned rule verification tag data can refer to the tag data generated by the system or administrators after automatically or manually reviewing historical business analysis reports according to preset verification rules. This tag data is used to identify whether the report has quality defects such as content errors, logical defects, or format problems. For example, a "bad sample" tag is added after the preset rule engine detects that the report has data anomalies or logical contradictions.

[0083] Step S102: Using the historical business analysis report corresponding to the first type of feedback data as a positive sample, and the historical business analysis report corresponding to the second type of feedback data and the rule verification mark data as a negative sample, train the preset classification model to obtain a quality assessment model.

[0084] It should be understood that the aforementioned positive samples can refer to sample data labeled as positive categories during model training. In this step, it specifically refers to historical business analysis reports that are considered to be of good quality based on positive user interaction behavior, and are used as positive examples when training the classification model.

[0085] The aforementioned negative samples can refer to sample data labeled as negative categories during model training. In this step, it specifically refers to historical business analysis reports that are deemed to be of poor quality based on negative user interaction behavior or rule verification failures, and are used as negative examples when training the classification model.

[0086] It should be explained that the aforementioned preset classification model can refer to a pre-selected machine learning model used to perform binary classification tasks. This model can divide samples into one of two categories based on the input feature data, such as the BERT model, the TextCNN model, etc. In this embodiment, it is specifically used to perform binary classification judgment on the quality of historical business analysis reports.

[0087] Accordingly, the aforementioned quality assessment model can refer to a model obtained after training a pre-defined classification model, which has the ability to score the quality of historical business analysis reports or predict their quality level. This model can receive an analysis report as input and output the quality evaluation result of the report.

[0088] Step S103: The quality of each of the historical business analysis reports is evaluated using the quality assessment model. Based on the quality assessment results, analysis reports that meet the preset threshold conditions are selected from the historical business analysis reports and determined as high-quality analysis reports.

[0089] Understandably, the aforementioned quality assessment can refer to the process of using a trained quality assessment model to score or classify each historical business analysis report. This process outputs a quantitative assessment score or probability value to characterize the quality level of the report.

[0090] It should be understood that the aforementioned preset threshold conditions may refer to pre-set judgment criteria for screening high-quality reports. This criterion may be a score threshold. Reports whose quality assessment results meet this threshold (e.g., greater than or equal to a certain score) are considered high-quality reports, while reports that do not meet this threshold are excluded.

[0091] It should be explained that the aforementioned high-quality analysis reports may refer to historical business analysis reports whose evaluation results meet the preset threshold conditions after being evaluated by the quality assessment model. These reports are considered to have high content quality, logical accuracy, and practical value, and are suitable as target samples for subsequent model training.

[0092] Step S104: Concatenate the input and output data of the high-quality analysis report to generate the target sample set.

[0093] It should be understood that the input data mentioned above can refer to the raw data content used as model input in a high-quality analysis report. For example, in a bank marketing analysis scenario, the input data can include raw business data from multiple dimensions such as basic customer information, financial indicators, and transaction records.

[0094] The aforementioned output data can refer to the analytical conclusions that serve as the model output in a high-quality analysis report, such as the textual content generated after analyzing the input data, including the analytical approach, reasoning process, and final analytical conclusions.

[0095] In its implementation, the system acquires all historical business analysis reports and simultaneously acquires user interaction behavior feedback data and rule verification mark data corresponding to each historical business analysis report. The user interaction behavior feedback data includes a first type of feedback data representing positive evaluations and a second type of feedback data representing negative evaluations.

[0096] Next, the system uses the historical business analysis reports corresponding to the first type of feedback data as positive samples, and the historical business analysis reports corresponding to the second type of feedback data and the historical business analysis reports corresponding to the rule verification marker data as negative samples, combining the positive and negative samples into a training dataset. The system then uses this training dataset to train a pre-defined classification model, updates the parameters of the model using a backpropagation algorithm, and obtains a quality assessment model after training.

[0097] Subsequently, the system inputs all historical business analysis reports one by one into the quality assessment model. The quality assessment model then evaluates the quality of each historical business analysis report and outputs the corresponding quality assessment results. Based on the quality assessment results of each historical business analysis report, the system filters out historical business analysis reports that meet preset threshold conditions and identifies the filtered reports as high-quality analysis reports.

[0098] Subsequently, the system concatenates the input and output data of the high-quality analysis reports separately, merging the input data of all high-quality analysis reports into a unified input part and merging the output data of all high-quality analysis reports into a unified output part. The input and output parts together constitute the target sample set.

[0099] To improve the quality of subsequent model training, the step of concatenating the input and output data of the high-quality analysis report to generate a target sample set may include: extracting multiple analysis point identifiers and corresponding original business data contained in the high-quality analysis report as sample input features; extracting the logical derivation process and analysis conclusions corresponding to each analysis point identifier in the high-quality analysis report as sample output features; and structurally encapsulating the sample input features and the sample output features according to the analysis point order to generate the target sample set.

[0100] Understandably, the aforementioned analysis point identifier can refer to the name or code used to uniquely identify an analysis point or analysis dimension, such as the name of an analysis point like "Enterprise Basic Information Analysis", "Financial Health Analysis", or "Industry Competitiveness Analysis", which is used to distinguish different analysis dimensions in the report.

[0101] The aforementioned raw business data can refer to the direct business information corresponding to each analysis point without model reasoning or processing. For example, for the "financial health analysis" analysis point, the corresponding raw business data can include raw indicator data such as operating revenue, net profit, and debt-to-equity ratio.

[0102] It should be explained that the above logical deduction process can refer to the intermediate reasoning steps and thought processes that the model performs before generating analytical conclusions. For example, before arriving at the conclusion that "the company has strong debt repayment ability", the model goes through the reasoning chain of "the debt-to-asset ratio is lower than the industry average and the current ratio is greater than 2, therefore the debt repayment ability is strong".

[0103] It is understandable that the above-mentioned structured encapsulation can refer to organizing and packaging sample input features and sample output features according to a preset data organization structure to form a data object with a unified format and hierarchical structure, so that the encapsulated data can be directly used for supervised fine-tuning training of the model.

[0104] In its implementation, after determining a high-quality analysis report, the system performs data extraction on each report. From each report, the system extracts all analysis point identifiers and the corresponding original business data, using these as sample input features.

[0105] Next, the system extracts the logical derivation process corresponding to each analysis point identifier and the analysis conclusion for that analysis point from the same high-quality analysis report, and uses all extracted logical derivation processes and analysis conclusions as sample output features. Following the order of the analysis points, the system arranges the analysis point identifiers and corresponding original business data in the extracted sample input features according to the order of the analysis points, and arranges the logical derivation processes and analysis conclusions corresponding to each analysis point identifier in the sample output features according to the same order of analysis points.

[0106] Subsequently, the system encapsulates the arranged sample input and output features according to a preset structured format, forming a complete data sample containing both input and output parts. The system performs the above extraction, sorting, and encapsulation operations on each high-quality analysis report, and then aggregates the complete data samples generated from all high-quality analysis reports to form the target sample set.

[0107] To facilitate understanding, the following examples illustrate the concepts, but do not impose specific limitations on this embodiment. For instance, in a corporate banking marketing scenario, a high-quality analysis report contains three analysis points, in the following order: "Basic Enterprise Information Analysis," "Financial Health Analysis," and "Industry Competitiveness Analysis." The system extracts the identifiers of the three analysis points and the corresponding raw business data from the report as sample input features. The raw business data for "Basic Enterprise Information Analysis" includes the company name, registration date, and business scope; the raw business data for "Financial Health Analysis" includes operating revenue and debt-to-asset ratio; and the raw business data for "Industry Competitiveness Analysis" includes market share ranking. The system extracts the logical derivation process and analysis conclusions corresponding to each of the three analysis points from the same report as sample output features. The logical derivation process for "Financial Health Analysis" is: "Operating revenue has increased for three consecutive years, and the debt-to-asset ratio is 45%, lower than the industry average of 60%, therefore profitability is stable and debt repayment risk is low," and the corresponding analysis conclusion is: "The company has a good financial health." The system arranges the identifiers of the three analysis points in the sample input features and their corresponding original business data in that order, following the sequence of "Enterprise Basic Information Analysis," "Financial Health Analysis," and "Industry Competitiveness Analysis." Similarly, it arranges the logical derivation processes and analysis conclusions of the three analysis points in the sample output features in the same order. The arranged input and output features are then packaged into a complete data sample. The system performs this process on all high-quality analysis reports, obtaining multiple complete data samples, which are then aggregated to generate the target sample set.

[0108] It should be understood that by extracting multiple analysis point identifiers and original business data as input features, and simultaneously extracting the logical derivation process and analysis conclusions as output features, the model learns the parallel analysis capability and autonomous reasoning logic of multiple business objects in supervised fine-tuning. The structured encapsulation according to the order of analysis points ensures the standardization and coherence of the input-output mapping relationship, effectively improving the model training quality.

[0109] This embodiment discloses a method for acquiring user interaction feedback data and rule verification marker data corresponding to each historical business analysis report. The user interaction feedback data includes a first type of feedback data representing positive evaluations and a second type of feedback data representing negative evaluations. Using the historical business analysis reports corresponding to the first type of feedback data as positive samples and the historical business analysis reports corresponding to the second type of feedback data and the rule verification marker data as negative samples, a preset classification model is trained to obtain a quality assessment model. The quality assessment model is then used to assess the quality of each historical business analysis report. Based on the quality assessment results, analysis reports that meet preset threshold conditions are selected from each historical business analysis report and identified as high-quality analysis reports. The input and output data of the high-quality analysis reports are then concatenated to generate a target sample set. Because this embodiment uses automated quality assessment by integrating real user interaction feedback and rule verification markers, it avoids the subjective bias and inefficiency of manual selection, accurately identifies high-quality analysis reports, and constructs a target sample set. This provides a high-quality and reliable data foundation for subsequent model fine-tuning, effectively improving the model training effect.

[0110] refer to Figure 3 , Figure 3 This is a flowchart illustrating the third embodiment of the marketing analysis method of the present invention.

[0111] Based on the above embodiments, in this embodiment, step S30 includes steps S301 to S304: Step S301: Input the sample input data from the target sample set into the fine-tuned model, and generate multiple candidate output results based on a preset strategy.

[0112] It should be explained that the aforementioned preset strategy can refer to a pre-defined control method for sampling and generating multiple candidate results from the output probability distribution of the fine-tuned model. This strategy is used to control the model to generate diverse output results during the inference process.

[0113] Furthermore, the aforementioned candidate output results can refer to multiple different output texts generated by the fine-tuned model for the same sample input data. Each output text is an expression of a possible analysis result of the input data, and there are differences in content or expression among the multiple candidate output results.

[0114] Step S302: Evaluate each candidate output result using a preset reward function to obtain the corresponding reward value. The preset reward function includes a format standardization reward function and a content accuracy reward function.

[0115] It is understandable that the aforementioned preset reward function can refer to a set of pre-defined mathematical functions or evaluation rules used to quantitatively evaluate candidate output results. This function is used to measure the quality of candidate output results generated by the model in multiple dimensions.

[0116] The aforementioned format-standard reward function can refer to an evaluation function in the preset reward function specifically used to assess whether the candidate output results conform to the preset output template requirements at the structural format level. For example, it can check whether the candidate output results contain the output structure of all analysis points, whether the identification format of each analysis point is correct, and whether the layout format of the analysis conclusions conforms to the predetermined specifications.

[0117] The aforementioned content accuracy reward function can refer to an evaluation function in the preset reward function specifically used to assess the consistency between the candidate output and the reference output at the semantic content level. For example, content accuracy can be assessed by calculating the text similarity between the candidate output and the reference output.

[0118] It should be noted that the above reward value can refer to the quantitative value output by the preset reward function after evaluating the candidate output results. This value is used to characterize the overall quality of the candidate output results in each evaluation dimension defined by the preset reward function.

[0119] Step S303: Based on the reward value corresponding to each candidate output result, determine the relative advantage value corresponding to each candidate output result, and determine the policy gradient based on the relative advantage value.

[0120] It should be explained that the aforementioned relative advantage value can refer to a relative quantitative value calculated based on the reward value corresponding to each candidate output result, used to measure the performance of each candidate output result relative to other candidate output results. This relative advantage value is obtained by subtracting the baseline or normalizing the reward value of each candidate output result. For example, the difference obtained by subtracting the average reward value of all candidate output results from the reward value of a certain candidate output result is the relative advantage value of that candidate output result.

[0121] Understandably, the aforementioned policy gradient can refer to the computational cost of the parameter update direction and magnitude of the target analysis model during reinforcement learning training. This policy gradient is calculated based on the relative advantage value of each candidate output result and the corresponding policy probability ratio, and is used to guide the update of model parameters, so that the model tends to produce output results with higher relative advantage values ​​in subsequent generation processes.

[0122] In its implementation, after obtaining the reward value for each candidate output, the system aggregates the reward values ​​for all candidate outputs. The system calculates a baseline reference value based on the reward values ​​of each candidate output, which can be the average of the reward values ​​of all candidate outputs or other statistical measures. The system compares the reward value of each candidate output with the baseline reference value, calculating the deviation of each candidate output's reward value from the baseline reference value, and uses the calculated deviation as the relative advantage value for each candidate output. Next, based on the relative advantage values ​​of all candidate outputs, and using reinforcement learning policy gradient calculation methods, the system calculates the policy gradient used to update the fine-tuned model parameters. Specifically, the system multiplies the relative advantage value of each candidate output by its corresponding policy probability ratio, aggregates the products of all candidate outputs, and determines the aggregated result as the policy gradient.

[0123] Step S304: Update the parameters of the fine-tuned model using a preset optimizer based on the policy gradient to obtain the target analysis model.

[0124] It should be explained that the aforementioned preset optimizer can refer to a pre-defined optimization algorithm tool used to update model parameters based on policy gradients during model training. This optimizer iteratively adjusts the model's weight parameters based on the calculated policy gradient direction and magnitude, combined with its built-in hyperparameters such as learning rate and momentum. For example, the preset optimizer could be the AdamW optimizer, which adjusts the learning rate of each parameter through adaptive moment estimation, enabling stable and efficient parameter updates during training.

[0125] It is understandable that the parameters of the fine-tuned model mentioned above can refer to numerical variables such as the trainable weight matrix and bias terms within the fine-tuned model. These parameters are continuously adjusted through iterative updates of the optimizer during training, so that the output performance of the model gradually optimizes in the desired direction.

[0126] In its implementation, the system calculates the policy gradient and then inputs it into a preset optimizer. The system uses the preset optimizer to calculate the parameter update amount for each parameter based on the policy gradient. This update amount is obtained by multiplying the policy gradient by the learning rate built into the preset optimizer and then processing it through the optimizer's internal momentum adjustment mechanism. Next, the system applies the calculated parameter update amount to the current parameters of the fine-tuned model, subtracting or adding the corresponding update amount to the current parameter values ​​to obtain the updated parameter values. After completing a single parameter update, the system determines whether a preset training termination condition is met, such as reaching a preset number of training epochs or the model performance no longer improving. If the training termination condition is not met, the system continues to the next reinforcement learning iteration, i.e., re-executes the process of inputting sample input data from the target sample set into the updated model to generate candidate output results, calculate reward values, determine relative advantage values, and calculate policy gradients, and then updates the parameters again using the preset optimizer. If the training termination condition is met, the system stops training and identifies the model with completed parameter updates after all reinforcement learning iterations as the target analysis model.

[0127] To improve the stability of the reinforcement learning training process and the reliability of the target analysis model output, the step of evaluating each candidate output result using a preset reward function to obtain a corresponding reward value may include: evaluating the degree of matching between the structural format of the candidate output result and the preset output template using the format normalization reward function to obtain a format reward value; evaluating the semantic similarity between the candidate output result and the reference output result using the content accuracy reward function to obtain a similarity reward value; predicting the quality level of the candidate output result using the quality assessment model to obtain a quality reward value; and weighting and fusing the format reward value, the similarity reward value, and the quality reward value based on preset weight coefficients to obtain the reward value corresponding to each candidate output result.

[0128] It should be understood that by comprehensively evaluating candidate outputs through three dimensions—format standardization, content accuracy, and quality level—and generating the final reward value through a weighted fusion method, the one-sidedness of a single evaluation dimension is avoided. At the same time, the quality evaluation model is reused for quality prediction, ensuring that the evaluation criteria are consistent with the data screening stage, which significantly improves the stability of the reinforcement learning training process and the reliability of the target analysis model output.

[0129] It is understandable that the aforementioned preset output template may refer to a pre-defined standard format framework for standardizing the structure and format of candidate output results. This template specifies the order of each analysis point in the output results, the marking method of each analysis point identifier, the paragraph structure of the analysis conclusions, and the overall layout format requirements.

[0130] The aforementioned reference output results may refer to the standard output data corresponding to the current candidate output results in the target sample set, such as the logical derivation process and analysis conclusions contained in a high-quality analysis report, which serve as a benchmark for evaluating the accuracy of the candidate output results.

[0131] It should be noted that the semantic similarity mentioned above refers to the degree of similarity between the candidate output and the reference output at the semantic level, which is quantified by calculating the text overlap or vector distance between the two, such as by using indicators like BLEU score or ROUGE score.

[0132] The aforementioned quality level can refer to the quality category or quality score output by the quality assessment model after classifying and predicting the quality of candidate output results. This quality level is used to measure the degree of excellence or inferiority of candidate output results at the overall quality level.

[0133] Understandably, the aforementioned preset weight coefficients may refer to the pre-set weight allocation ratios used when weighting and fusing the format reward value, similarity reward value, and quality reward value. For example, the three weight coefficients may be 0.3, 0.3, and 0.4, respectively, which correspond to the proportions of the three reward values ​​in the final fusion result.

[0134] Accordingly, the aforementioned weighted fusion can refer to the calculation method of multiplying multiple reward values ​​by their respective preset weight coefficients and then summing them to obtain the comprehensive reward value.

[0135] This embodiment discloses a method for inputting sample input data from the target sample set into the fine-tuned model, generating multiple candidate output results based on a preset strategy; evaluating each candidate output result using a preset reward function to obtain a corresponding reward value, the preset reward function including a format conformity reward function and a content accuracy reward function; determining the relative advantage value corresponding to each candidate output result based on the reward value, and determining the strategy gradient based on the relative advantage value; and updating the parameters of the fine-tuned model using a preset optimizer according to the strategy gradient to obtain the target analysis model. Because this embodiment generates multiple candidate outputs through a preset sampling strategy and combines multi-dimensional reward function evaluation based on format conformity and content accuracy, the model autonomously explores the optimal output strategy in reinforcement learning, determines the strategy gradient using the relative advantage value, and optimizes the model parameters, guiding the model to balance structural conformity and content accuracy. This allows for autonomous model optimization without relying on manual analysis, improving the inference quality and output stability of the target analysis model compared to existing technologies.

[0136] Furthermore, embodiments of the present invention also propose a storage medium storing a marketing analysis program, which, when executed by a processor, implements the steps of the marketing analysis method described above.

[0137] Reference Figure 4 , Figure 4 This is a structural block diagram of the first embodiment of the marketing analysis device of the present invention.

[0138] like Figure 4 As shown, the marketing analysis device proposed in this embodiment of the invention includes: a data filtering module 501, a model fine-tuning module 502, a reinforcement learning module 503, and a result output module 504.

[0139] The data filtering module 501 is used to obtain historical business analysis reports, perform quality assessment on the historical business analysis reports based on preset filtering rules, and determine the target sample set based on the quality assessment results.

[0140] The model fine-tuning module 502 is used to perform supervised fine-tuning of the pre-trained large model based on the target sample set to obtain the fine-tuned model.

[0141] The reinforcement learning module 503 is used to optimize the fine-tuned model based on reinforcement learning to obtain a target analysis model.

[0142] The result output module 504 is used to acquire multiple business analysis objects to be analyzed and their corresponding business data, input each business analysis object and its corresponding business data into the target analysis model, and output the analysis results corresponding to each business analysis object.

[0143] The model fine-tuning module 502 is further configured to determine the input sequence length and output sequence length of each sample in the target sample set, determine the maximum sequence length parameter of the pre-trained large model based on a preset quantile statistical strategy, perform sequence padding or truncation on each sample in the target sample set according to the maximum sequence length parameter to obtain standardized training data, and fine-tune the parameters of the pre-trained large model using the standardized training data based on a low-rank adaptive fine-tuning algorithm to obtain the fine-tuned model.

[0144] The result output module 504 is further configured to receive user query requests, parse multiple business analysis objects and customer identification information to be analyzed from the user query requests; obtain business data corresponding to each of the business analysis objects from the business database based on the customer identification information; structurally concatenate the identification information of each business analysis object and the corresponding business data according to the order of the business analysis objects to generate multi-object analysis input data; input the multi-object analysis input data into the target analysis model in a single step, and output the analysis results corresponding to each of the business analysis objects.

[0145] This device embodiment discloses the following steps: acquiring historical business analysis reports; evaluating the quality of these reports based on preset screening rules; determining a target sample set based on the evaluation results; supervising the fine-tuning of a pre-trained large model using the target sample set to obtain a fine-tuned model; optimizing the fine-tuned model using reinforcement learning to obtain a target analysis model; acquiring multiple business analysis objects to be analyzed and their corresponding business data; inputting each business analysis object and its corresponding business data into the target analysis model; and outputting the analysis results for each business analysis object. Because this device embodiment determines the target sample set by evaluating the quality of historical business analysis reports, employs a two-stage training strategy combining supervised fine-tuning and reinforcement learning to obtain the target analysis model, and achieves the ability to process multiple business analysis objects in a single call, compared to existing technologies, this device embodiment improves the accuracy and stability of analysis results in marketing analysis.

[0146] Based on the first embodiment of the marketing analysis device of the present invention described above, a second embodiment of the marketing analysis device of the present invention is proposed.

[0147] In this embodiment, the data filtering module 501 is further configured to acquire user interaction behavior feedback data and rule verification mark data corresponding to each historical business analysis report. The user interaction behavior feedback data includes a first type of feedback data representing positive evaluation and a second type of feedback data representing negative evaluation. Using the historical business analysis report corresponding to the first type of feedback data as a positive sample and the historical business analysis report corresponding to the second type of feedback data and the rule verification mark data as a negative sample, a preset classification model is trained to obtain a quality assessment model. The quality assessment model is used to assess the quality of each historical business analysis report, and based on the quality assessment results, analysis reports that meet preset threshold conditions are selected from each historical business analysis report and determined as high-quality analysis reports. The input data and output data of the high-quality analysis reports are concatenated to generate a target sample set.

[0148] The data filtering module 501 is further configured to extract multiple analysis point identifiers and corresponding original business data contained in the high-quality analysis report as sample input features; extract the logical derivation process and analysis conclusions corresponding to each analysis point identifier in the high-quality analysis report as sample output features; and encapsulate the sample input features and the sample output features in a structured manner according to the analysis point order to generate a target sample set.

[0149] Other embodiments or specific implementations of the marketing analysis device of the present invention can be found in the above-described method embodiments, and will not be repeated here.

[0150] This application provides a marketing analysis device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the marketing analysis method in Embodiment 1 above.

[0151] The following is for reference. Figure 5 The diagram illustrates a structural schematic of a marketing analytics device suitable for implementing embodiments of this application. The marketing analytics device in these embodiments may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 5 The marketing analytics device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0152] like Figure 5 As shown, the marketing analytics device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.) that can perform various appropriate actions and processes according to a program stored in a read-only memory 1002 or a program loaded from a storage device 1003 into a random access memory 1004. The random access memory 1004 also stores various programs and data required for the operation of the marketing analytics device. The processing unit 1001, the read-only memory 1002, and the random access memory 1004 are interconnected via a bus 1005. An input / output interface 1006 is also connected to the bus. Typically, the following systems can be connected to the input / output interface 1006: input devices 1007 including, for example, a touchscreen, touchpad, keyboard, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 1003 including, for example, magnetic tape, hard disk, etc.; and communication devices 1009. Communication device 1009 allows the marketing analytics device to communicate wirelessly or wiredly with other devices to exchange data. While the figure shows a marketing analytics device with various systems, it should be understood that implementing or having all of the systems shown is not required. More or fewer systems may be implemented alternatively.

[0153] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0154] The marketing analysis device provided in this application, employing the marketing analysis method described in the above embodiments, can solve the technical problem of insufficient accuracy and stability of analysis results in existing marketing analysis. Compared with the prior art, the beneficial effects of the marketing analysis device provided in this application are the same as those of the marketing analysis method provided in the above embodiments, and other technical features of the marketing analysis device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0155] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0156] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0157] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0158] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0159] 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 the present invention, 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 read-only memory / random access memory, 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 described in the various embodiments of the present invention.

[0160] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A marketing analysis method, characterized in that, The method includes: Obtain historical business analysis reports, conduct quality assessments on the historical business analysis reports based on preset screening rules, and determine the target sample set based on the quality assessment results; The pre-trained large model is supervised fine-tuned based on the target sample set to obtain the fine-tuned model. The fine-tuned model is optimized based on reinforcement learning to obtain the target analysis model; Obtain multiple business analysis objects to be analyzed and their corresponding business data, input each business analysis object and its corresponding business data into the target analysis model, and output the analysis results corresponding to each business analysis object.

2. The marketing analysis method as described in claim 1, characterized in that, The steps of obtaining historical business analysis reports, evaluating the quality of the historical business analysis reports based on preset screening rules, and determining the target sample set based on the quality evaluation results include: Obtain user interaction behavior feedback data and rule verification mark data corresponding to each historical business analysis report. The user interaction behavior feedback data includes a first type of feedback data representing positive evaluation and a second type of feedback data representing negative evaluation. Using the historical business analysis reports corresponding to the first type of feedback data as positive samples and the historical business analysis reports corresponding to the second type of feedback data and the rule verification mark data as negative samples, a preset classification model is trained to obtain a quality assessment model. The quality assessment model is used to assess the quality of each of the historical business analysis reports. Based on the quality assessment results, analysis reports that meet the preset threshold conditions are selected from the historical business analysis reports and determined as high-quality analysis reports. The input and output data of the high-quality analysis report are concatenated to generate the target sample set.

3. The marketing analysis method as described in claim 2, characterized in that, The step of concatenating the input and output data of the high-quality analysis report to generate the target sample set includes: Extract the identifiers of multiple analysis points and the corresponding original business data contained in the high-quality analysis report as sample input features; Extract the logical derivation process and analysis conclusion corresponding to each analysis point identifier in the high-quality analysis report as sample output features; The sample input features and sample output features are structured and encapsulated according to the order of analysis points to generate the target sample set.

4. The marketing analysis method as described in claim 2, characterized in that, The step of optimizing the fine-tuned model based on reinforcement learning to obtain the target analysis model includes: The sample input data from the target sample set is input into the fine-tuned model, and multiple candidate output results are generated based on a preset strategy. Each candidate output result is evaluated using a preset reward function to obtain a corresponding reward value. The preset reward function includes a format standardization reward function and a content accuracy reward function. Based on the reward value corresponding to each of the candidate output results, the relative advantage value corresponding to each of the candidate output results is determined, and the policy gradient is determined based on the relative advantage value. The parameters of the fine-tuned model are updated using a preset optimizer based on the policy gradient to obtain the target analysis model.

5. The marketing analysis method as described in claim 4, characterized in that, The step of evaluating each candidate output result using a preset reward function to obtain the corresponding reward value includes: The matching degree between the structure and format of the candidate output result and the preset output template is evaluated using the format normalization reward function to obtain the format reward value; The semantic similarity between the candidate output and the reference output is evaluated using the content accuracy reward function to obtain a similarity reward value; The quality assessment model is used to predict the quality level of the candidate output results to obtain a quality reward value; The format reward value, similarity reward value, and quality reward value are weighted and fused based on preset weight coefficients to obtain the reward value corresponding to each candidate output result.

6. The marketing analysis method as described in claim 1, characterized in that, The step of supervising the fine-tuning of the pre-trained large model based on the target sample set to obtain the fine-tuned model includes: Determine the input sequence length and output sequence length of each sample in the target sample set, and determine the maximum sequence length parameter of the pre-trained large model based on a preset quantile statistical strategy; Each sample in the target sample set is padded or truncated according to the maximum sequence length parameter to obtain standardized training data; Based on the low-rank adaptive fine-tuning algorithm, the parameters of the pre-trained large model are fine-tuned using the standardized training data to obtain the fine-tuned model.

7. The marketing analysis method as described in claim 1, characterized in that, The steps of acquiring multiple business analysis objects to be analyzed and their corresponding business data, inputting each business analysis object and its corresponding business data into the target analysis model, and outputting the analysis results corresponding to each business analysis object include: Receive user query requests and parse multiple business analysis objects and customer identification information to be analyzed from the user query requests; Based on the customer identification information, retrieve business data corresponding to each of the business analysis objects from the business database; Based on the order of the business analysis objects, the identification information of each business analysis object and the corresponding business data are structurally concatenated to generate multi-object analysis input data; The multi-object analysis input data is input into the target analysis model in a single step, and the analysis results corresponding to each of the business analysis objects are output.

8. A marketing analysis device, characterized in that, The device includes: The data filtering module is used to obtain historical business analysis reports, perform quality assessment on the historical business analysis reports based on preset filtering rules, and determine the target sample set based on the quality assessment results. The model fine-tuning module is used to perform supervised fine-tuning of the pre-trained large model based on the target sample set to obtain the fine-tuned model. The reinforcement learning module is used to optimize the fine-tuned model based on reinforcement learning to obtain the target analysis model; The results output module is used to acquire multiple business analysis objects to be analyzed and their corresponding business data, input each business analysis object and its corresponding business data into the target analysis model, and output the analysis results corresponding to each business analysis object.

9. A marketing analysis device, characterized in that, The device includes: a memory, a processor, and a marketing analytics program stored in the memory and executable on the processor, the marketing analytics program being configured to implement the steps of the marketing analytics method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium stores a marketing analysis program, which, when executed by a processor, implements the steps of the marketing analysis method as described in any one of claims 1 to 7.

Images