Sales demand analysis method and system based on AI and portrait

Abstract

The invention discloses a sales demand analysis method and system based on AI and portraits. The method comprises the steps of collecting multi-source heterogeneous data such as basic information, transaction records and online behaviors of customers; according to the method, a self-developed large model of the Chinese human life research and development center is adopted, implicit requirements in a customer behavior sequence are deeply mined through an attention mechanism, a hierarchical clustering algorithm is used for generating a tree-shaped clustering graph based on a Ward connection method, and high-value features with variances larger than 0.8 and absolute values of correlation coefficients larger than 0.5 are screened out in combination with a rule engine; it is ensured that the extracted features have statistical significance and are directly associated with business targets, invalid feature redundancy is avoided, the efficiency and accuracy of subsequent portrait construction are improved, product recommendation priorities are determined through a decision tree algorithm, and a Q-learning framework is combined to optimize communication verbal skills and cross sales strategies, so that the sales strategies meet the current demands of customers, and the customer experience is improved. Potential requirements are guided, and the conversion rate and the customer satisfaction degree are both improved.

Description

Technical Field

[0001] This invention relates to the field of sales analysis technology, specifically to a sales demand analysis method and system based on AI and user profiles. Background Technology

[0002] Traditional sales demand analysis relies on human experience, resulting in low efficiency, high subjectivity, and insufficient data utilization. Existing technologies often use static tagging systems to generate customer profiles, making it difficult to dynamically capture changes in customer behavior; demand prediction models are mostly based on linear regression or shallow machine learning, which have limited ability to model complex nonlinear relationships. This invention integrates AI technology with dynamic profiling to automate the entire process from multi-source data integration to intelligent strategy generation, significantly improving sales conversion rates and customer satisfaction. Summary of the Invention

[0003] To address the aforementioned technical issues, this technical solution provides a sales demand analysis method and system based on AI and user profiles, resolving the problems mentioned in the background section.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows: In a first aspect of the invention, a sales demand analysis method based on AI and user profiles is provided, comprising: Collect heterogeneous data from multiple sources, such as basic customer information, transaction records, and online behavior. Through data cleaning, deduplication, and standardization, form a structured dataset and establish a data association mapping table to achieve cross-dimensional data fusion. We use a pre-trained large language model to semantically encode structured datasets, extract customer behavior feature vectors through an attention mechanism, and optimize the feature space distribution using contrastive learning. Based on the semantic encoding results, the common characteristics of customer groups are identified through hierarchical clustering algorithm, and customer feature dimensions are selected by combining with the rule engine. Based on the extracted customer characteristic dimensions, a multi-dimensional profile system is constructed, including a customer preference matrix, a risk assessment model, and a purchase cycle. The Kalman filter algorithm is used to achieve dynamic updates of the profile. A demand prediction model is built based on generative adversarial networks. The model is input with customer profile feature vectors and predicted in the future insurance type preferences and coverage adjustment trends of customers through adversarial training between the generator and the discriminator. By integrating decision tree algorithms and reinforcement learning frameworks, a sales strategy guide is generated based on the prediction results of the demand forecasting model, which includes communication script templates, product recommendation priorities, and cross-selling strategies. Establish a closed-loop feedback mechanism for sales personnel, collect data on the effectiveness of strategy execution, and optimize model parameters through gradient descent algorithm to achieve continuous iterative improvement in model accuracy.

[0005] Preferably, the collection of multi-source heterogeneous data, such as basic customer information, transaction records, and online behavior, specifically includes: Retrieve basic information such as customer name, age, and contact information from the CRM system via a RESTful API interface; Front-end tracking technology is used to collect online behavior data such as user page browsing paths and click heatmaps. We obtain customers' historical transaction records from third-party platforms through compliant data crawlers, and irreversibly de-identify sensitive information such as ID card numbers and bank card numbers using the SHA-256 hash algorithm.

[0006] Preferably, the semantic encoding of the structured dataset specifically includes: The pre-trained large language model adopts the self-developed large model of China Life Research Center. The self-developed large model of China Life Research Center contains a 12-layer Transformer encoder with 768 hidden units per layer, a multi-head self-attention mechanism with 12 heads, and a total of 110M parameters. The model was fine-tuned based on a corpus specifically for the financial field. The corpus contains three types of data: insurance product terms and conditions text, customer consultation dialogue records, and financial news summaries. The fine-tuning process adopts a two-stage training strategy: the first stage is pre-training with a word-level masking language model task, and the second stage is learning the logical relationship of customer behavior sequences with a sentence-to-next-sentence prediction task. In the word-level prediction task, 15% of the words in the input text are randomly masked, of which 80% are replaced with [MASK] tags, 10% are replaced with random words, and 10% are kept as original words. The model recovers the original semantics by predicting the masked words. The loss function uses cross-entropy to calculate the difference between the predicted words and the real words. We construct a sentence pair dataset of customer behavior sequences, where positive examples are consecutive customer behavior sequence pairs and negative examples are randomly spliced ​​non-consecutive behavior sequence pairs. The model learns to determine whether a sentence pair is a consecutive sequence through a binary classification task, thereby improving its ability to understand the logical relationships of customer behavior. During the feature extraction stage, 12 self-attention heads are set up for parallel computation. Each attention head independently learns the dependencies of different dimensions in the customer behavior sequence. The output of each Transformer layer is standardized through layer normalization to stabilize the training process. Residual connections are used to add the original input to the layer normalized output to avoid the gradient vanishing problem and ensure effective training of deep networks.

[0007] Preferably, the step of identifying common characteristics of customer groups through hierarchical clustering algorithm and filtering customer characteristic dimensions by rule engine specifically includes: Each customer behavior feature vector is treated as an independent cluster, and the Euclidean distance matrix between clusters is calculated. The Euclidean distance is defined as follows: ; in, and respectively, feature vectors , In the eigenvalues ​​of dimension; The Ward join criterion is used to calculate the variance increment after cluster merging. The two clusters that minimize the total inter-cluster distance after merging are selected for merging. The formula for the Ward criterion is: ; in, and Clusters , The number of samples; A tree-like clustering graph is constructed through a recursive merging process. Each merging step records the corresponding distance threshold and merging path, forming a visual hierarchical structure. Feature filtering is performed using a two-dimensional threshold condition. Specifically, the variance of each feature dimension is calculated, and feature dimensions with a variance greater than 0.8 are retained. The absolute value of the Pearson correlation coefficient between the feature dimension and the target variable is calculated, and feature dimensions with an absolute value greater than 0.5 are retained. The customer characteristic dimensions are filtered out using the two-dimensional threshold conditions set above.

[0008] Preferably, in the multi-dimensional profiling system, the customer preference matrix uses cosine similarity to calculate the customer's preference score for different insurance products, with the specific formula as follows: ; Where A and B are the customer behavior vector and the product feature vector, respectively; In the multidimensional profiling system, the risk assessment model combines customer credit scores and historical compensation records to construct a Logistic regression model. The characteristic variables include income stability index, past compensation amount, and health questionnaire score. In the multi-dimensional profiling system, the purchase cycle is based on a time series prediction model established by a bidirectional LSTM network, with 256 hidden units set, and the ReLU activation function is used to capture the time dependence of customer purchase behavior.

[0009] Preferably, the demand prediction model based on generative adversarial networks, which inputs customer profile feature vectors and predicts future customer insurance type preferences and coverage adjustment trends through adversarial training between the generator and discriminator, specifically includes: It adopts a Transformer dual encoder-decoder structure, with 6 layers of encoder and 6 layers of decoder, each layer embedding an 8-head self-attention mechanism; the input is a random noise vector or a customer's historical behavior feature vector, and the long-distance dependencies in the sequence are captured through multi-head self-attention, and the output is the predicted customer's future behavior sequence; A convolutional neural network structure is adopted, with 3 convolutional layers, each with a kernel size of 3×3 and a stride of 1, to extract local features. Then, 2 fully connected layers are followed with 128 and 64 hidden units respectively to integrate global features. The LeakyReLU activation function with a negative slope of 0.2 is used to replace the traditional ReLU, avoiding the gradient vanishing problem and enhancing the ability to capture subtle feature differences. Using Wasserstein distance as the core of the loss function, the discriminator maximizes the difference between the real data distribution and the generated data distribution, and the generator minimizes this difference. A gradient penalty term is introduced to constrain the gradient norm of the discriminator to be close to 1, so as to prevent the generator from being unable to learn an effective distribution due to the discriminator being too powerful. An alternating training strategy is adopted: first, the generator is fixed and the discriminator is trained 5 times to fully learn the differences in data distribution; then, the discriminator is fixed and the generator is trained once to generate samples that are closer to the real distribution. By dynamically adjusting the ratio of training times, the optimization process of the two is balanced, ensuring that the generator gradually approaches the real data distribution.

[0010] Preferably, the decision tree algorithm uses a CART classification tree, the feature selection criterion is the Gini coefficient, overfitting is prevented through post-pruning, the reinforcement learning framework uses the Q-learning algorithm, the state space is defined as the customer profile feature vector, the action space is defined as the sales strategy combination, and the reward function is designed as follows: ; in , , ε is the weighting coefficient, which is used to balance exploration and utilization through an ε-greedy strategy. The initial ε is set to 0.9, and it decreases by 0.1 every 100 iterations.

[0011] Preferably, the salesperson feedback data is collected in real time via a mobile app, specifically including customer conversion rate after strategy execution, objection handling time, and customer satisfaction score. The model parameters are optimized using the Adam optimizer, with β1=0.9, β2=0.999, and ε=1e-8. The learning rate is initialized to 0.001, and a cosine annealing strategy is used to adjust the parameters every 20 rounds with a decay rate of 0.1. The gradient is calculated using the backpropagation algorithm, and gradient clipping is applied to prevent gradient explosion.

[0012] Preferably, the method further includes a real-time monitoring step: Establish a data quality monitoring system and use a sliding window mechanism to statistically analyze data integrity indicators for the past 7 days. Specific indicators include data missing rate, duplication rate, and outlier ratio. When the data missing rate exceeds 5%, a level 3 warning is triggered, and the associated data completion module is automatically invoked to fill in the missing values. The filling strategies include mean filling, hot card filling, and KNN-based prediction filling. When the data duplication rate exceeds 10%, a deduplication process is initiated, which identifies duplicate records by comparing hash fingerprints and retains the latest version.

[0013] In a second aspect of the invention, a sales demand analysis system based on AI and user profiles is also provided, comprising: The data collection module is used to collect heterogeneous data from multiple sources, such as customer basic information, transaction records, and online behavior. Through data cleaning, deduplication, and standardization, a structured dataset is formed, and a data association mapping table is established to achieve cross-dimensional data fusion. The extraction module is used to perform semantic encoding on the structured dataset using a pre-trained large language model, extract customer behavior feature vectors through an attention mechanism, and optimize the feature space distribution using contrastive learning. The filtering module is used to identify common characteristics of customer groups based on semantic encoding results through hierarchical clustering algorithm, and to filter out customer feature dimensions in combination with rule engine. The profile building module is used to construct a multi-dimensional profile system, including a customer preference matrix, a risk assessment model, and a purchase cycle, based on the extracted customer characteristic dimensions, and to use the Kalman filter algorithm to achieve dynamic profile updates. The demand forecasting module is used to build a demand forecasting model based on a generative adversarial network. It takes a customer profile feature vector as input and predicts the customer's future insurance type preference and the trend of the sum insured adjustment through adversarial training between the generator and the discriminator. The sales strategy module integrates decision tree algorithms and reinforcement learning frameworks to generate a sales strategy guide that includes communication script templates, product recommendation priorities, and cross-selling strategies based on the prediction results of the demand forecasting model. The sales feedback module is used to establish a closed-loop feedback mechanism for sales staff, collect data on the effectiveness of strategy execution, optimize model parameters through gradient descent algorithm, and achieve continuous iterative improvement of model accuracy.

[0014] Compared with existing technologies, this invention provides a sales demand analysis method and system based on AI and user profiles, which has the following beneficial effects: This invention establishes a data association mapping table to achieve deep integration of multi-source heterogeneous data such as customer basic information, transaction records, and online behavior, effectively breaking down the "data silos" between traditional systems. It employs a pre-trained large language model combined with contrastive learning, not only deeply mining implicit needs in customer behavior sequences (such as identifying unexpressed protection gaps) through an attention mechanism, but also using a hierarchical clustering algorithm based on the Ward join method to generate a tree-like clustering graph. Combined with a rule engine, it selects high-value features with variance > 0.8 and absolute correlation coefficient > 0.5, ensuring that the extracted features are both statistically significant and directly related to business objectives, avoiding redundant and ineffective features, and improving the efficiency and accuracy of subsequent profile construction. Furthermore, the constructed customer preference matrix, risk assessment model, and purchase cycle prediction module achieve dynamic profile updates through a Kalman filter algorithm—for example, when a customer adds a critical illness insurance inquiry, the system can adjust in real time. Its risk preference score and recommended coverage value ensure that the profile always reflects the latest customer behavior, solving the problem of lag in traditional static labeling systems and improving the timeliness of marketing strategies. By using a decision tree algorithm to determine product recommendation priorities and combining it with a Q-learning framework to optimize communication techniques and cross-selling strategies, the sales strategy not only meets the customer's current needs (e.g., recommending protection-type products to high-risk customers) but also guides potential needs (e.g., recommending education insurance to young families), achieving a dual improvement in conversion rate and customer satisfaction. In summary, this invention, through a closed-loop design of "data fusion - feature extraction - profile construction - demand prediction - strategy generation - feedback optimization," achieves three major leaps: from passive response to proactive prediction, from static labels to dynamic profiles, and from a single model to continuous evolution. This significantly improves sales efficiency and decision-making accuracy, and is applicable to precision marketing scenarios in multiple industries such as insurance, finance, and retail. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the method flow of S101-S107 in this invention; Figure 2 This is a schematic diagram of the method flow for S201-S203 in this invention; Figure 3 This is a schematic diagram of the method flow for S301-S305 in this invention; Figure 4 This is a schematic diagram of the method flow for S401-S405 in this invention; Figure 5 This is a schematic diagram of the method flow for S501-S504 in this invention. Detailed Implementation

[0016] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0017] Example 1 Please refer to Figure 1 As shown, in a first aspect of the present invention, a sales demand analysis method based on AI and user profiles is provided, comprising: S101. Collect heterogeneous data from multiple sources, such as basic customer information, transaction records, and online behavior. Through data cleaning, deduplication, and standardization, form a structured dataset and establish a data association mapping table to achieve cross-dimensional data fusion. S102. Use a pre-trained large language model to perform semantic encoding on the structured dataset, extract customer behavior feature vectors through an attention mechanism, and use contrastive learning to optimize the feature space distribution. S103. Based on the semantic encoding results, the common characteristics of the customer group are identified through hierarchical clustering algorithm, and the customer characteristic dimensions are selected by combining the rule engine. S104. Based on the extracted customer characteristic dimensions, construct a multi-dimensional profile system including a customer preference matrix, risk assessment model, and purchase cycle, and use the Kalman filter algorithm to achieve dynamic profile updates. S105. Construct a demand prediction model based on generative adversarial networks. Input customer profile feature vectors and predict customers' future insurance type preferences and coverage adjustment trends through adversarial training between the generator and discriminator. S106. Integrating decision tree algorithm and reinforcement learning framework, a sales strategy guide including communication script templates, product recommendation priorities and cross-selling strategies is generated based on the prediction results of the demand prediction model. S107. Establish a closed-loop feedback mechanism for sales staff, collect data on the effectiveness of strategy execution, optimize model parameters through gradient descent algorithm, and achieve continuous iterative improvement of model accuracy.

[0018] Those skilled in the art will understand that this invention, by establishing a data association mapping table, achieves deep integration of multi-source heterogeneous data such as customer basic information, transaction records, and online behavior, effectively breaking down the "data silos" between traditional systems. Employing a pre-trained large language model combined with contrastive learning, it not only deeply mines implicit needs in customer behavior sequences (such as identifying unexpressed protection gaps) through an attention mechanism, but also uses a hierarchical clustering algorithm based on the Ward join method to generate a tree-like clustering graph. Combined with a rule engine, it filters out high-value features with variance > 0.8 and absolute correlation coefficient > 0.5, ensuring that the extracted features are both statistically significant and directly related to business objectives, avoiding redundant and ineffective features, and improving the efficiency and accuracy of subsequent profile construction. Furthermore, the constructed customer preference matrix, risk assessment model, and purchase cycle prediction module achieve dynamic profile updates through a Kalman filter algorithm—for example, when a customer adds a critical illness insurance inquiry, the system can adjust their risk preference score and recommended coverage value in real time, ensuring that the profile always reflects the customer's latest behavioral status, solving the lag problem of traditional static labeling systems, and improving the timeliness of marketing strategies.

[0019] Please refer to Figure 2 As shown, the system collects heterogeneous data from multiple sources, including basic customer information, transaction records, and online behavior. Specifically, this includes: S201. Obtain basic information such as customer name, age, and contact information from the CRM system through the RESTful API interface; S202. Use front-end tracking technology to collect online behavior data such as user page browsing paths and click heatmaps; S203. Obtain customer historical transaction records from third-party platforms through compliant data crawlers, and use the SHA-256 hash algorithm to irreversibly de-identify sensitive information such as ID card numbers and bank card numbers.

[0020] Please refer to Figure 3 As shown, semantic encoding of structured datasets specifically includes: S301. The pre-trained large language model adopts the self-developed large model of China Life R&D Center. The self-developed large model of China Life R&D Center contains 12 layers of Transformer encoder, each layer has 768 hidden units, 12 heads of multi-head self-attention mechanism, and a total of 110M parameters. S302. Model fine-tuning is carried out based on a financial field-specific corpus. The corpus contains three types of data: insurance product terms and conditions text, customer consultation dialogue records, and financial news summaries. The fine-tuning process adopts a two-stage training strategy: the first stage is pre-training with a word-level mask language model task, and the second stage is learning the logical relationship of customer behavior sequence with a sentence-to-next-sentence prediction task. S303. In the word-level prediction task, 15% of the words in the input text are randomly masked, of which 80% are replaced with [MASK] tags, 10% are replaced with random words, and 10% are kept as original words. The model recovers the original semantics by predicting the masked words. The loss function uses cross-entropy to calculate the difference between the predicted words and the real words. S304. Construct a sentence pair dataset of customer behavior sequences, where positive examples are consecutive customer behavior sequence pairs and negative examples are randomly spliced ​​non-consecutive behavior sequence pairs. The model learns to determine whether a sentence pair is a consecutive sequence through a binary classification task, thereby improving its ability to understand the logical relationship of customer behavior. S305. In the feature extraction stage, 12 self-attention heads are set up for parallel computation. Each attention head independently learns the dependencies of different dimensions in the customer behavior sequence. The output of each Transformer layer is standardized through layer normalization to stabilize the training process. Residual connections are used to add the original input to the layer normalized output to avoid the gradient vanishing problem and ensure the effective training of deep networks.

[0021] Please refer to Figure 4 As shown, hierarchical clustering algorithms are used to identify common characteristics of customer groups, and a rule engine is used to filter out customer characteristic dimensions, specifically including: S401. Treat each customer behavior feature vector as an independent cluster, and calculate the Euclidean distance matrix between clusters. The Euclidean distance is defined as: ; in, and respectively, feature vectors , In the eigenvalues ​​of dimension; S402. Calculate the variance increment after cluster merging using the Ward's join criterion, and select the two clusters that minimize the total inter-cluster distance after merging for merging. The Ward's criterion formula is: ; in, and Clusters , The number of samples; S403. Construct a tree-like clustering graph through a recursive merging process. Record the corresponding distance threshold and merging path for each merging step to form a visual hierarchical structure. S404. Set two-dimensional threshold conditions for feature filtering. Specifically, calculate the variance value of each feature dimension and retain feature dimensions with a variance greater than 0.8. Also, calculate the absolute value of the Pearson correlation coefficient between the feature dimension and the target variable and retain feature dimensions with an absolute value greater than 0.5. S405. Filter out customer feature dimensions using the dual-dimensional threshold conditions set above.

[0022] In the multi-dimensional profiling system, the customer preference matrix uses cosine similarity to calculate customer preference scores for different insurance products. The specific formula is as follows: ; Where A and B are the customer behavior vector and the product feature vector, respectively; In the multidimensional profiling system, the risk assessment model combines customer credit scores and historical compensation records to construct a Logistic regression model. The characteristic variables include income stability index, past compensation amount, and health questionnaire score. In the multi-dimensional profiling system, the purchase cycle is based on a time series prediction model established by a bidirectional LSTM network, with 256 hidden units set, and the ReLU activation function is used to capture the time dependence of customer purchase behavior.

[0023] Please refer to Figure 5 As shown, a demand prediction model is constructed based on a generative adversarial network. Taking a customer profile feature vector as input, the model predicts future customer insurance type preferences and coverage adjustment trends through adversarial training between the generator and discriminator. Specifically, this includes: S501 adopts a Transformer dual encoder-decoder structure, with 6 layers of encoder and 6 layers of decoder, each layer embedding an 8-head self-attention mechanism; the input is a random noise vector or a customer's historical behavior feature vector, and the long-distance dependencies in the sequence are captured through multi-head self-attention, and the output is the predicted customer's future behavior sequence. S502 adopts a convolutional neural network structure with 3 convolutional layers, each with a kernel size of 3×3 and a stride of 1 to extract local features. It is followed by 2 fully connected layers with 128 and 64 hidden units respectively for global feature integration. The LeakyReLU activation function with a negative slope of 0.2 is used to replace the traditional ReLU, avoiding the gradient vanishing problem and enhancing the ability to capture subtle feature differences. S503. Using Wasserstein distance as the core of the loss function, the discriminator maximizes the difference between the real data distribution and the generated data distribution, and the generator minimizes this difference. A gradient penalty term is introduced to constrain the gradient norm of the discriminator to be close to 1, so as to prevent the generator from being unable to learn an effective distribution due to the discriminator being too powerful. S504. An alternating training strategy is adopted: first, the generator is fixed and the discriminator is trained 5 times to fully learn the differences in data distribution; then, the discriminator is fixed and the generator is trained once to generate samples that are closer to the real distribution. By dynamically adjusting the ratio of training times, the optimization process of the two is balanced to ensure that the generator gradually approaches the real data distribution.

[0024] The decision tree algorithm uses a CART classification tree, with the Gini coefficient as the feature selection criterion. Overfitting is prevented through post-pruning. The reinforcement learning framework employs the Q-learning algorithm. The state space is defined as the customer profile feature vector, the action space as the sales strategy combination, and the reward function is designed as follows: ; in , , ε is the weighting coefficient, which is used to balance exploration and utilization through an ε-greedy strategy. The initial ε is set to 0.9, and it decreases by 0.1 every 100 iterations.

[0025] Salesperson feedback data is collected in real time via a mobile app, including customer conversion rate after strategy implementation, objection handling time, and customer satisfaction rating. Model parameter optimization uses the Adam optimizer with β1=0.9, β2=0.999, and ε=1e-8. The learning rate is initialized to 0.001, and a cosine annealing strategy is used to adjust the parameters every 20 rounds with a decay rate of 0.1. The gradient is calculated using the backpropagation algorithm, and gradient clipping is applied to prevent gradient explosion.

[0026] The method also includes a real-time monitoring step: Establish a data quality monitoring system and use a sliding window mechanism to statistically analyze data integrity indicators for the past 7 days. Specific indicators include data missing rate, duplication rate, and outlier ratio. When the data missing rate exceeds 5%, a level 3 warning is triggered, and the associated data completion module is automatically invoked to fill in the missing values. The filling strategies include mean filling, hot card filling, and KNN-based prediction filling. When the data duplication rate exceeds 10%, a deduplication process is initiated, which identifies duplicate records by comparing hash fingerprints and retains the latest version.

[0027] In a second aspect of the invention, a sales demand analysis system based on AI and user profiles is also provided, comprising: The data collection module is used to collect heterogeneous data from multiple sources, such as basic customer information, transaction records, and online behavior. Through data cleaning, deduplication, and standardization, it forms a structured dataset and establishes a data association mapping table to achieve cross-dimensional data fusion. The extraction module uses a pre-trained large language model to semantically encode the structured dataset, extracts customer behavior feature vectors through an attention mechanism, and optimizes the feature space distribution using contrastive learning. The filtering module is used to identify common characteristics of customer groups based on semantic encoding results through hierarchical clustering algorithms, and to filter out customer feature dimensions in combination with the rule engine. The profile building module is used to construct a multi-dimensional profile system based on the extracted customer feature dimensions, including a customer preference matrix, risk assessment model, and purchase cycle, and uses the Kalman filter algorithm to achieve dynamic profile updates. The demand forecasting module is used to build a demand forecasting model based on a generative adversarial network. It takes a customer profile feature vector as input and predicts the customer's future insurance type preferences and coverage adjustment trends through adversarial training between the generator and the discriminator. The sales strategy module integrates decision tree algorithms and reinforcement learning frameworks to generate a sales strategy guide that includes communication script templates, product recommendation priorities, and cross-selling strategies based on the prediction results of the demand forecasting model. The sales feedback module is used to establish a closed-loop feedback mechanism for sales staff, collect data on the effectiveness of strategy execution, and optimize model parameters through the gradient descent algorithm to achieve continuous iterative improvement in model accuracy.

[0028] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A sales demand analysis method based on AI and user profiles, characterized in that: include: Collect heterogeneous data from multiple sources, such as basic customer information, transaction records, and online behavior. Through data cleaning, deduplication, and standardization, form a structured dataset and establish a data association mapping table to achieve cross-dimensional data fusion. We use a pre-trained large language model to semantically encode structured datasets, extract customer behavior feature vectors through an attention mechanism, and optimize the feature space distribution using contrastive learning. Based on the semantic encoding results, the common characteristics of customer groups are identified through hierarchical clustering algorithm, and customer feature dimensions are selected by combining with the rule engine. Based on the extracted customer characteristic dimensions, a multi-dimensional profile system is constructed, including a customer preference matrix, a risk assessment model, and a purchase cycle. The Kalman filter algorithm is used to achieve dynamic updates of the profile. A demand prediction model is built based on generative adversarial networks. The model is input with customer profile feature vectors and predicted in the future insurance type preferences and coverage adjustment trends of customers through adversarial training between the generator and the discriminator. By integrating decision tree algorithms and reinforcement learning frameworks, a sales strategy guide is generated based on the prediction results of the demand forecasting model, which includes communication script templates, product recommendation priorities, and cross-selling strategies. Establish a closed-loop feedback mechanism for sales personnel, collect data on the effectiveness of strategy execution, and optimize model parameters through gradient descent algorithm to achieve continuous iterative improvement in model accuracy.

2. The sales demand analysis method based on AI and user profiles according to claim 1, characterized in that, The collection of multi-source heterogeneous data, including basic customer information, transaction records, and online behavior, specifically includes: Retrieve basic information such as customer name, age, and contact information from the CRM system via a RESTful API interface; Front-end tracking technology is used to collect online behavior data such as user page browsing paths and click heatmaps. We obtain customers' historical transaction records from third-party platforms through compliant data crawlers, and irreversibly de-identify sensitive information such as ID card numbers and bank card numbers using the SHA-256 hash algorithm.

3. The sales demand analysis method based on AI and user profiles according to claim 2, characterized in that, The semantic encoding of the structured dataset specifically includes: The pre-trained large language model adopts the self-developed large model of China Life Research Center. The self-developed large model of China Life Research Center contains a 12-layer Transformer encoder with 768 hidden units per layer, a multi-head self-attention mechanism with 12 heads, and a total of 110M parameters. The model was fine-tuned based on a corpus specifically for the financial field. The corpus contains three types of data: insurance product terms and conditions text, customer consultation dialogue records, and financial news summaries. The fine-tuning process adopts a two-stage training strategy: the first stage is pre-training with a word-level masking language model task, and the second stage is learning the logical relationship of customer behavior sequences with a sentence-to-next-sentence prediction task. In the word-level prediction task, 15% of the words in the input text are randomly masked, of which 80% are replaced with [MASK] tags, 10% are replaced with random words, and 10% are kept as original words. The model recovers the original semantics by predicting the masked words. The loss function uses cross-entropy to calculate the difference between the predicted words and the real words. We construct a sentence pair dataset of customer behavior sequences, where positive examples are consecutive customer behavior sequence pairs and negative examples are randomly spliced ​​non-consecutive behavior sequence pairs. The model learns to determine whether a sentence pair is a consecutive sequence through a binary classification task, thereby improving its ability to understand the logical relationships of customer behavior. During the feature extraction stage, 12 self-attention heads are set up for parallel computation. Each attention head independently learns the dependencies of different dimensions in the customer behavior sequence. The output of each Transformer layer is standardized through layer normalization to stabilize the training process. Residual connections are used to add the original input to the layer normalized output to avoid the gradient vanishing problem and ensure effective training of deep networks.

4. The sales demand analysis method based on AI and user profiles according to claim 3, characterized in that, The process of identifying common characteristics of customer groups through hierarchical clustering algorithms and combining this with a rule engine to filter customer characteristic dimensions specifically includes: Each customer behavior feature vector is treated as an independent cluster, and the Euclidean distance matrix between clusters is calculated. The Euclidean distance is defined as follows: ； in, and They are the feature vectors , In the eigenvalues ​​of dimension; The Ward join criterion is used to calculate the variance increment after cluster merging. The two clusters that minimize the total inter-cluster distance after merging are selected for merging. The formula for the Ward criterion is: ； in, and Clusters , The number of samples; A tree-like clustering graph is constructed through a recursive merging process. Each merging step records the corresponding distance threshold and merging path, forming a visual hierarchical structure. Feature filtering is performed using a two-dimensional threshold condition. Specifically, the variance of each feature dimension is calculated, and feature dimensions with a variance greater than 0.8 are retained. The absolute value of the Pearson correlation coefficient between the feature dimension and the target variable is calculated, and feature dimensions with an absolute value greater than 0.5 are retained. The customer characteristic dimensions are filtered out using the two-dimensional threshold conditions set above.

5. The sales demand analysis method based on AI and user profiles according to claim 4, characterized in that, In the multi-dimensional profiling system, the customer preference matrix uses cosine similarity to calculate the customer's preference score for different insurance products. The specific formula is as follows: ； Where A and B are the customer behavior vector and the product feature vector, respectively; In the multidimensional profiling system, the risk assessment model combines customer credit scores and historical compensation records to construct a Logistic regression model. The characteristic variables include income stability index, past compensation amount, and health questionnaire score. In the multi-dimensional profiling system, the purchase cycle is based on a time series prediction model established by a bidirectional LSTM network, with 256 hidden units set, and the ReLU activation function is used to capture the time dependence of customer purchase behavior.

6. The sales demand analysis method based on AI and user profiles according to claim 5, characterized in that, The demand prediction model based on generative adversarial networks (GANs) takes a customer profile feature vector as input and predicts future customer insurance type preferences and coverage adjustment trends through adversarial training between the generator and discriminator. Specifically, it includes: It adopts a Transformer dual encoder-decoder structure, with 6 layers of encoder and 6 layers of decoder, each layer embedding an 8-head self-attention mechanism; the input is a random noise vector or a customer's historical behavior feature vector, and the long-distance dependencies in the sequence are captured through multi-head self-attention, and the output is the predicted customer's future behavior sequence; A convolutional neural network structure is adopted, with 3 convolutional layers, each with a kernel size of 3×3 and a stride of 1, to extract local features. Then, 2 fully connected layers are followed with 128 and 64 hidden units respectively to integrate global features. The LeakyReLU activation function with a negative slope of 0.2 is used to replace the traditional ReLU, avoiding the gradient vanishing problem and enhancing the ability to capture subtle feature differences. Using Wasserstein distance as the core of the loss function, the discriminator maximizes the difference between the real data distribution and the generated data distribution, and the generator minimizes this difference. A gradient penalty term is introduced to constrain the gradient norm of the discriminator to be close to 1, so as to prevent the generator from being unable to learn an effective distribution due to the discriminator being too powerful. An alternating training strategy is adopted: first, the generator is fixed and the discriminator is trained 5 times to fully learn the differences in data distribution; then, the discriminator is fixed and the generator is trained once to generate samples that are closer to the real distribution. By dynamically adjusting the ratio of training times, the optimization process of the two is balanced, ensuring that the generator gradually approaches the real data distribution.

7. The sales demand analysis method based on AI and user profiles according to claim 6, characterized in that, The decision tree algorithm uses a CART classification tree, with the Gini coefficient as the feature selection criterion. Overfitting is prevented through post-pruning. The reinforcement learning framework employs the Q-learning algorithm. The state space is defined as the customer profile feature vector, the action space as the sales strategy combination, and the reward function is designed as follows: ； in , , ε is the weighting coefficient, which is used to balance exploration and utilization through an ε-greedy strategy. The initial ε is set to 0.9, and it decreases by 0.1 every 100 iterations.

8. The sales demand analysis method based on AI and user profiles according to claim 7, characterized in that, The salesperson feedback data is collected in real time via a mobile app, specifically including customer conversion rate after strategy execution, objection handling time, and customer satisfaction score. The model parameters are optimized using the Adam optimizer, with β1=0.9, β2=0.999, and ε=1e-8. The learning rate is initialized to 0.001, and a cosine annealing strategy is used to adjust the parameters every 20 rounds with a decay rate of 0.

1. The gradient is calculated using the backpropagation algorithm, and gradient clipping is applied to prevent gradient explosion.

9. The sales demand analysis method based on AI and user profiles according to claim 8, characterized in that, The method also includes a real-time monitoring step: Establish a data quality monitoring system and use a sliding window mechanism to statistically analyze data integrity indicators for the past 7 days. Specific indicators include data missing rate, duplication rate, and outlier ratio. When the data missing rate exceeds 5%, a level 3 warning is triggered, and the associated data completion module is automatically invoked to fill in the missing values. The filling strategies include mean filling, hot card filling, and KNN-based prediction filling. When the data duplication rate exceeds 10%, a deduplication process is initiated, which identifies duplicate records by comparing hash fingerprints and retains the latest version.

10. A sales demand analysis system based on AI and user profiles, used to implement the sales demand analysis method based on AI and user profiles as described in any one of claims 1-9, characterized in that, include: The data collection module is used to collect heterogeneous data from multiple sources, such as customer basic information, transaction records, and online behavior. Through data cleaning, deduplication, and standardization, a structured dataset is formed, and a data association mapping table is established to achieve cross-dimensional data fusion. The extraction module is used to perform semantic encoding on the structured dataset using a pre-trained large language model, extract customer behavior feature vectors through an attention mechanism, and optimize the feature space distribution using contrastive learning. The filtering module is used to identify common characteristics of customer groups based on semantic encoding results through hierarchical clustering algorithm, and to filter out customer feature dimensions in combination with rule engine. The profile building module is used to construct a multi-dimensional profile system, including a customer preference matrix, a risk assessment model, and a purchase cycle, based on the extracted customer characteristic dimensions, and to use the Kalman filter algorithm to achieve dynamic profile updates. The demand forecasting module is used to build a demand forecasting model based on a generative adversarial network. It takes a customer profile feature vector as input and predicts the customer's future insurance type preference and the trend of the sum insured adjustment through adversarial training between the generator and the discriminator. The sales strategy module integrates decision tree algorithms and reinforcement learning frameworks to generate a sales strategy guide that includes communication script templates, product recommendation priorities, and cross-selling strategies based on the prediction results of the demand forecasting model. The sales feedback module is used to establish a closed-loop feedback mechanism for sales staff, collect data on the effectiveness of strategy execution, optimize model parameters through gradient descent algorithm, and achieve continuous iterative improvement of model accuracy.

Images