Intelligent marketing method and system based on multi-stage prediction model fusion

Abstract

The invention discloses an intelligent marketing method and system based on multi-stage prediction model fusion. The method comprises the steps of obtaining user portrait feature data and historical marketing data of a target population to obtain initial data; performing prediction model training of three stages of connection, high intention and clicking on the initial data to obtain a connection model, a high intention model and a clicking model; generating a fusion model by combining the connection model, the high-intention model and the click model according to business requirements, and calculating a score corresponding to the target person according to the fusion model to generate a comprehensive score; and sorting the target population according to the comprehensive score, and selecting users whose scores meet requirements for resource delivery. By implementing the method provided by the invention, the overall marketing efficiency and accuracy can be realized, so that the problems of insufficient data utilization, label deviation and insufficient generalization ability of a single model in the prior art are solved.

Description

Technical Field

[0001] This invention relates to artificial intelligence, and more specifically to an intelligent marketing method and system based on the fusion of multi-stage prediction models. Background Technology

[0002] In the current field of AI-powered telemarketing or SMS marketing, businesses generally use a single predictive model to screen potential customers. These models are trained based on user profiles and behavioral data to determine which users are more likely to engage in specific conversion behaviors, such as clicking, registering, or placing an order. However, in practical applications, especially in industries such as B2C finance, insurance, and e-commerce, the process from contacting a customer to ultimately achieving conversion often involves multiple stages.

[0003] First, in the user data preparation phase, data sources are diverse, including internal CRM systems, active user databases, lead pools, and third-party data providers. The data covers user profiles, device usage, and historical outreach records. This raw data needs to be cleaned, labeled, and feature tables generated for subsequent analysis. Second, the multi-model screening phase aims to narrow down the target audience and improve ROI. This phase typically involves building and utilizing multiple predictive models, such as registration, order, and click prediction models, to estimate the probability of corresponding user behaviors. These models can be used independently for marketing strategy development or combined to rank the outputs, ultimately identifying the Top N audience segments or groups most likely to convert, for further processing by the AI ​​provider. Following this is the AI ​​voice outbound calling phase, where the provider makes calls using different types of network resources (such as landline numbers, virtual numbers, and real mobile numbers). Due to varying network quality, some lines are easily flagged as "harassment / fraud," while others may be blacklisted by operators, directly impacting reach and interception rates. Simultaneously, the same user may choose to block or flag calls from other companies due to frequent marketing campaigns. If the call is successfully connected, the AI ​​intent recognition stage begins. In this stage, AI determines the user's intent type based on factors such as keywords, speech rate, duration, emotion, and dialogue logic. This includes high intent, aversion or complaint, invalid / rejection, ambiguity / hesitation, and no clear intent or conversation turn. First, the user data preparation stage involves a wide range of data sources, including internal CRM systems, active user databases, lead pools, and third-party data providers. The data covers user profiles, device usage, and historical contact records. This raw data needs to be cleaned, labeled, and feature tables generated for subsequent analysis. Next, the multi-model screening stage aims to narrow down the target audience and improve ROI. This stage typically involves building and using multiple predictive models, such as registration prediction models, order prediction models, and click prediction models, to estimate the probability of corresponding user behaviors. These models can be used independently for marketing strategy development or combined to output ranking results, ultimately identifying the Top N audience groups or subgroups most likely to convert, for further processing by the AI ​​provider. Finally, the AI ​​voice outbound calling stage occurs, where the provider makes calls using different types of line resources (such as landline numbers, virtual numbers, and real mobile numbers). Due to varying line quality, some lines are easily flagged as "harassment / fraud," while others may be blacklisted by the operator, directly impacting the reach and blocking rates of the numbers. Simultaneously, the same user may choose to block or flag these calls due to frequent marketing campaigns from other companies. If the call is successfully connected, it enters the AI ​​intent recognition stage.At this stage, AI determines user intent based on factors such as keywords, speech rate, duration, emotion, and dialogue logic. Categories include high intent, aversion or complaint, invalid / rejection, vague / hesitant, and no clear intent or too few dialogue rounds. Based on this, it decides whether to send a follow-up SMS to the user. Finally, users judged as having high intent or partially vague intent receive an SMS containing product information and a short link. User click behavior is used not only as a key conversion metric for training the click model but also for evaluating overall marketing effectiveness. However, because SMS is only sent to users judged by AI as having high intent, the click model training data has an inherent bias, affecting the model's generalization ability and accuracy. However, since the text messages are only sent to users that the AI ​​judges to have high intent, this causes an inherent bias in the training data of the click model, affecting the model's generalization ability and accuracy.

[0004] In conclusion, while existing AI-powered telemarketing processes are complex and well-designed, they still face challenges in areas such as data path length, information loss, label bias, and the generalization ability of individual models. These issues limit the maximization of marketing efficiency, impacting the final business scale and ROI.

[0005] Therefore, it is necessary to design a new approach to achieve overall marketing efficiency and accuracy in order to address the problems of insufficient data utilization, label bias, and insufficient generalization ability of single models in existing technologies. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide an intelligent marketing method and system based on the fusion of multi-stage prediction models.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: an intelligent marketing method based on the fusion of multi-stage prediction models, comprising: Obtain user profile data and historical marketing data of the target audience to obtain initial data; The initial data is used to train prediction models for three stages: connection, high intent, and click, to obtain connection model, high intent model, and click model. Based on business needs, a fusion model is generated by combining the connection model, high intent model, and click model. The score corresponding to the target person is calculated based on the fusion model to generate a comprehensive score. The target audience is ranked based on the comprehensive score, and users whose scores meet the requirements are selected for resource allocation.

[0008] The further technical solution is as follows: after ranking the target audience based on the comprehensive score and selecting users whose scores meet the requirements for resource allocation, it also includes: Daily collection and analysis of actual marketing data feedback; use Isotonic Regression or Logistic Regression to calibrate the connection model, high intent model, and click model; and dynamically adjust the weights and parameters of the connection model, high intent model, and click model to continuously optimize their performance.

[0009] The further technical solution is as follows: Training a prediction model on the initial data for three stages—connection, high intent, and click—to obtain a connection model, a high intent model, and a click model, including: Based on the historical marketing data of the initial data, the call records with whether the call was successfully connected were selected and tagged. Combined with the corresponding user profile features as a sample set, the XGBoost algorithm was used to train the model to obtain the connection model. Based on the historical marketing data that has been connected according to the initial data, high-intent users are marked as positive samples and low-intent users are marked as negative samples according to the AI ​​voice recognition results, so as to use the XGBoost algorithm to train the model and obtain the high-intent model. Based on the initial data, for users with high intent to click on links in SMS messages, data showing clicks in SMS messages are marked as positive samples, and data showing no clicks in SMS messages are also marked as positive samples, forming a sample set to obtain the click model.

[0010] The further technical solution is as follows: A fusion model is generated based on business needs, combining the connection model, high-intent model, and click model; and the score corresponding to the target person is calculated based on the fusion model to generate a comprehensive score, including: Choose the optimization direction based on business needs; Users who connect and click according to the optimized direction are labeled as positive samples, and users who only connect but do not click are labeled as negative samples, thus forming labels and collecting user profile feature data to obtain the input set; Obtain the scores of the input set in the engagement model, high intent model, and click model; Logistic regression was used to train a fusion model by combining the ratings from the engagement model, high intent model, and click model as new features with the labels. An input set is constructed for the target audience and fed into the connection model, high intent model, and click model to determine the corresponding scores. The comprehensive score is then determined through the fusion model.

[0011] The further technical solution is as follows: the comprehensive score includes a linear combination of the weighted connection model, high intention model and click model scores through the sigmoid function, and the output conversion probability is in the range of [0, 1].

[0012] Its further technical solution is as follows: The daily collection and analysis of actual marketing data feedback, the use of Isotonic Regression or Logistic Regression to calibrate the connection model, high-intent model, and click model, and the dynamic adjustment of the weights and parameters of the connection model, high-intent model, and click model to continuously optimize their performance, including: Daily collection and analysis of connection rate, high intent rate, and click-through rate to initially assess the effectiveness of the model; Using data from the last three days, the scores of the engagement model, high intent model, and click model were calibrated using Isotonic Regression or Logistic Regression, and the weights of high-performing strategies were dynamically adjusted. The AUC and KS metrics were used to evaluate the distinguishing ability of the connection model, high intent model and click model, and the AUC and KS metrics were calculated by collecting a sample of the conversion audience. Calculate click-through rates among users to evaluate the accuracy of the engagement model, high intent model, and click model in high-potential market segments; The fusion model is optimized based on the data from the first three days, including adjusting weights, introducing new features, or improving the algorithm.

[0013] This invention also provides an intelligent marketing system based on the fusion of multi-stage prediction models, comprising: The data acquisition unit is used to acquire user profile feature data and historical marketing data of the target audience to obtain initial data; The training unit is used to train prediction models for three stages—connection, high intent, and click—on the initial data to obtain connection model, high intent model, and click model. The comprehensive scoring unit is used to generate a fusion model by combining the connection model, high intention model and click model according to business needs, and to calculate the score corresponding to the target person according to the fusion model to generate a comprehensive score. The sorting and delivery unit is used to sort the target audience based on the comprehensive score and select users whose scores meet the requirements for resource delivery.

[0014] Its further technical solutions include: The optimization unit is used to collect and analyze actual marketing data feedback daily, calibrate the connection model, high intent model, and click model using Isotonic Regression or Logistic Regression, and dynamically adjust the weights and parameters of the connection model, high intent model, and click model to continuously optimize their performance.

[0015] The present invention also provides a computer device, the computer device including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the above-described method.

[0016] The present invention also provides a storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0017] The advantages of this invention compared to existing technologies are as follows: This invention integrates user profile feature data of the target audience and historical marketing data as initial data, and sequentially trains targeted models for three key marketing stages: connection, high intent, and click, forming connection, high intent, and click models that respectively reflect the conversion probability of each stage. Then, based on specific business needs, these single-stage models are flexibly combined to generate a fusion model, which is used to calculate a comprehensive score for each target user. This score is used to rank and filter the target audience, prioritizing users with high scores for resource allocation. This process not only fully explores and utilizes the value of multi-source data and effectively corrects the bias problems that may be caused by single tags, but also significantly improves the generalization ability and predictive accuracy of the overall model through a multi-model fusion strategy. This ensures high conversion rates and ROI while improving marketing efficiency, overcoming the problems of insufficient data utilization, tag bias, and insufficient generalization ability of single models in existing technologies.

[0018] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A schematic diagram illustrating an application scenario of the intelligent marketing method based on multi-stage prediction model fusion provided in this embodiment of the invention; Figure 2 A flowchart illustrating the intelligent marketing method based on multi-stage prediction model fusion provided in an embodiment of the present invention; Figure 3 A schematic diagram of a sub-process of the intelligent marketing method based on multi-stage prediction model fusion provided in an embodiment of the present invention; Figure 4 A schematic diagram of a sub-process of the intelligent marketing method based on multi-stage prediction model fusion provided in an embodiment of the present invention; Figure 5 A flowchart illustrating an intelligent marketing method based on multi-stage prediction model fusion, provided as another embodiment of the present invention; Figure 6 A schematic block diagram of an intelligent marketing system based on multi-stage prediction model fusion provided in an embodiment of the present invention; Figure 7 A schematic block diagram of an intelligent marketing system based on multi-stage prediction model fusion, provided as another embodiment of the present invention; Figure 8 A schematic block diagram of a computer device provided for an embodiment of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0023] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0024] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0025] Please see Figure 1 and Figure 2 , Figure 1This is a schematic diagram illustrating an application scenario of the intelligent marketing method based on multi-stage prediction model fusion provided in an embodiment of the present invention. Figure 2 This is a schematic flowchart illustrating the intelligent marketing method based on multi-stage predictive model fusion provided in this invention. The method is applied to a server that interacts with terminals, integrating user profile feature data and historical marketing data. It uses the XGBoost algorithm to train predictive models for three stages: connection, high intent, and click. Based on business needs, these models are fused to generate a comprehensive score for precise targeting of specific demographics and resource allocation. Daily actual marketing feedback data is collected, and Isotonic Regression or Logistic Regression is used to calibrate the scores of each model and dynamically adjust their weights. Simultaneously, metrics such as AUC (Area Under the Curve) and Kolmogorov-Smirnov test are used to evaluate model performance, ensuring the accuracy and effectiveness of the models across different market segments. This method effectively solves the problems of insufficient data utilization, label bias, and insufficient generalization ability of single models in existing technologies, significantly improving overall marketing efficiency and accuracy.

[0026] Figure 2 This is a flowchart illustrating the intelligent marketing method based on multi-stage prediction model fusion provided in an embodiment of the present invention. Figure 2 As shown, the method includes the following steps S110 to S140.

[0027] S110. Obtain user profile feature data and historical marketing data of the target audience to obtain initial data.

[0028] In this embodiment, the initial data refers to detailed information about the target audience extracted from a database or data warehouse. This information includes, but is not limited to, user profile feature data and historical marketing data. Specifically: User profile feature data: This part of the data includes each user's static attributes and dynamic behaviors. Static attributes refer to relatively fixed information such as gender, age, and geographic location; while dynamic behaviors include activity levels (e.g., login frequency, usage duration), historical purchase records, and responses to different marketing campaigns. This data helps to depict a user's basic profile and preferences, providing rich input features for subsequent model training.

[0029] Historical Marketing Data: This section covers data from various marketing campaigns conducted targeting this demographic over a past period, including but not limited to call records (whether calls were connected), AI voice recognition results (high intent / low intent), SMS sending and click data, etc. Analyzing this data allows us to understand user performance in past marketing campaigns, assess which factors may have contributed to successful conversions, and apply these lessons to future campaigns.

[0030] First, based on a predetermined time window (e.g., the next N days), relevant data about the target population is extracted from the database. This step typically involves interacting with multiple data sources to ensure that all necessary information is collected.

[0031] Next, the extracted data is cleaned to handle issues such as missing values, duplicate records, and outliers. High-quality data is fundamental to ensuring the accuracy of the model, making this step crucial.

[0032] Based on business needs, feature engineering operations, such as feature selection and feature construction, are performed on the cleaned data. The goal is to extract the most valuable information for the prediction task and transform it into a format suitable for model use.

[0033] Finally, the compiled data is divided into training and testing sets. The training set is used to build models for each stage (connection model, high intent model, click model), while the testing set is used to evaluate model performance and ensure the effectiveness and reliability of the models.

[0034] Through the steps described above, the system obtains comprehensive and accurate initial data, providing a solid foundation for the subsequent multi-stage model training and fusion. This process not only enhances the model's understanding of real-world situations but also lays the cornerstone for ultimately achieving efficient and precise marketing strategies.

[0035] S120. Train the prediction models for the initial data in three stages: connection, high intent, and click, to obtain the connection model, high intent model, and click model.

[0036] In this embodiment, the call connection model is used to predict whether a user will answer the phone. By analyzing the user's call history (whether the call was successfully connected) and relevant user profile features (such as gender, age, activity level, etc.), a binary classification model is trained using the XGBoost algorithm to output a probability score for each user answering the call. A label of 1 indicates that the call was answered, and a label of 0 indicates that the call was not answered.

[0037] The high-intent model refers to using AI speech recognition technology to determine the level of interest shown by users during a call based on connected data, and then classifying users into high-intent (positive samples, label=1) and low-intent (negative samples, label=0). Similarly, the XGBoost algorithm is used to train the model to predict the probability that a user will become a high-intent customer after the call ends.

[0038] The click-through model targets high-intent users who have already sent and hung up an SMS message. It labels samples based on whether they clicked a link in the message (clicked link=1, not clicked link=0). Using these labels and corresponding user profile features as input, the XGBoost algorithm is applied again to train the model, aiming to predict the likelihood of a user clicking a link in an SMS message.

[0039] In one embodiment, please refer to Figure 3 The above step S120 may include steps S121 to S123.

[0040] S121. Based on the historical marketing data of the initial data, select the call records containing whether the call was successfully connected, label them, and use the corresponding user profile features as a sample set to train the model using the XGBoost algorithm to obtain the connection model.

[0041] In this embodiment, based on the historical marketing data of the initial data, data with call records indicating whether the call was successfully connected are filtered out and labeled (connected label=1, unconnected label=0), and combined with the corresponding user profile features as a sample set.

[0042] The XGBoost algorithm was used to train the sample set to obtain a connection model that can predict the likelihood of a user answering a call.

[0043] S122. Based on the historical marketing data that has been connected according to the initial data, high-intent users are marked as positive samples and low-intent users are marked as negative samples according to the AI ​​voice recognition results, so as to train the model using the XGBoost algorithm to obtain a high-intent model.

[0044] In this embodiment, based on the AI ​​voice recognition results, users with high intent are marked as positive samples (A-class label=1) and users with low intent are marked as negative samples (B-class and C-class label=0) from the historical marketing data that has been connected in the initial data.

[0045] Using these labeled data and their user profile features as input, the XGBoost algorithm is used to train a model, resulting in a high-intention model for evaluating the degree of high intent after a user makes a call.

[0046] S123. Based on the initial data, for users with high intent to click on links in SMS messages, data showing clicks in SMS messages are marked as positive samples, and data showing no clicks in SMS messages are marked as positive samples, to form a sample set and obtain the click model.

[0047] In this embodiment, for high-intent users who have already sent SMS messages in the initial data, data that have clicked on links in SMS messages are marked as positive samples (label=1), and data that have not clicked on links in SMS messages are marked as negative samples (label=0), forming a sample set.

[0048] The XGBoost algorithm was applied to train this sample set, and a click model was finally obtained to predict the probability of a user clicking on a link in a text message.

[0049] Through model training in these three stages, the system can provide accurate predictive capabilities for different marketing stages, thereby improving the overall conversion rate. Furthermore, these models are updated monthly to ensure they reflect the latest market dynamics and changes in user behavior, enhancing their adaptability and accuracy. This multi-stage modeling approach not only improves the model's predictive power but also achieves more efficient resource allocation and a higher ROI through the fine-grained division and optimization of each stage.

[0050] S130. Based on business needs, combine the connection model, high intention model, and click model to generate a fusion model, and calculate the score corresponding to the target personnel based on the fusion model to generate a comprehensive score.

[0051] In this embodiment, the fusion model refers to combining the prediction results of the connection model, the high intent model, and the click model through logistic regression or other fusion algorithms (such as Stacking) to form a final conversion probability score. This score not only considers the user's likelihood of answering the call and the degree of high intent, but also the likelihood of them clicking the SMS link, thus providing a more comprehensive and accurate assessment of the user's conversion potential.

[0052] The overall score is a numerical value between 0 and 1 generated for each user using the aforementioned fusion model. It represents the likelihood of the user completing the entire marketing process (from answering the phone call to clicking the link in the SMS message). This score can be used for ranking, helping to select the user groups most likely to convert for targeted resource allocation. The overall score consists of a linear combination of the weighted scores from the connection model, high-intent model, and click model using a sigmoid function, outputting a conversion probability in the range [0, 1].

[0053] In one embodiment, please refer to Figure 4 The above-mentioned step S130 may include steps S131 to S135.

[0054] S131. Select the optimization direction based on business needs.

[0055] In this embodiment, the optimization direction is based on determining the specific metrics that need to be optimized according to the current business objectives. For example, when improving ROI (Return on Investment), the focus might be on optimizing the click-through rate (CTR); if the goal is to expand the potential customer base, the focus might be on the click-to-call rate (CTR). This step determines the specific strategy for how to label samples and build the fusion model.

[0056] S132. According to the optimized direction, users who connect and click are positive samples, and users who only connect but do not click are negative samples, forming labels, and collecting user profile feature data to obtain the input set.

[0057] In this embodiment, the input set refers to the set of user samples that meet the selected optimization criteria. For example, if the goal is to optimize the click-through / connection ratio, users who "connected and clicked" will be treated as positive samples (label=1), while users who "connected but did not click" will be treated as negative samples (label=0). Simultaneously, detailed user profile data (such as gender, age, activity level, etc.) is collected as the input feature set for training the fusion model.

[0058] S133. Obtain the scores of the input set in the engagement model, high intent model, and click model.

[0059] In this embodiment, for each user in the input set, a pre-trained connection model, high intent model, and click model are used to predict their score (score_a, score_b, score_c). These scores represent the user's performance probability at each stage.

[0060] S134. Use logistic regression to train the fusion model by combining the ratings from the access model, high intent model, and click model as new features with the labels.

[0061] In this embodiment, the three model scores (score_a, score_b, score_c) obtained in step S133 are used as a new feature set (new_features), combined with the previously defined labels, and a new fusion model is trained using the logistic regression method. This fusion model can output a comprehensive score (score_d), reflecting the user's conversion potential throughout the entire marketing process.

[0062] S135. Construct an input set for the target audience and input it into the connection model, high intent model, and click model to determine the corresponding score, and then determine the comprehensive score through the fusion model.

[0063] In this embodiment, for the entire target audience awaiting marketing, steps S132 to S134 are repeated: first, an input set is constructed; then, the inputs are fed into three basic models to obtain their respective scores; finally, these scores are used as new features to input into the fusion model to calculate the comprehensive score (score_d) for each user. Based on this comprehensive score, the Top N users can be selected for priority resource allocation to maximize the expected conversion rate and ROI.

[0064] This multi-stage modeling and fusion mechanism not only improves the predictive power of the model, but also ensures the robustness and adaptability of the system through dynamic correction and real-time evaluation. In particular, it can effectively improve the return on investment per unit of reach cost when resources are limited.

[0065] S140. Sort the target group according to the comprehensive score, and select users whose scores meet the requirements for resource allocation.

[0066] In this embodiment, based on the comprehensive score (score_d) of the target audience calculated by the previously constructed fusion model, these users are ranked, and the top N users with the highest scores are selected for resource allocation. This process ensures that limited resources are used most effectively, maximizing the expected conversion rate and ROI.

[0067] For each target user to be marketed to, a comprehensive score (score_d) between 0 and 1 is calculated using a fusion model based on their scores in the connection model, high intent model, and click model (score_a, score_b, score_c). This score reflects the likelihood that the user will complete the entire marketing process (from answering the phone call to finally clicking the link in the SMS message).

[0068] All target users are ranked in descending order of their overall rating. This step forms the basis for subsequent selections, placing users with higher conversion potential at the top of the list.

[0069] Based on current business needs and available resources, determine a specific threshold or decide to select the top N users. For example, when resources are very limited, only the top 10% of users with the highest scores may be selected; while when resources are relatively abundant, the criteria can be relaxed appropriately, increasing the number of selected users.

[0070] Given the changing market environment and user behavior, this threshold or the selection of Top N is not fixed. The selection criteria can be dynamically adjusted based on recent performance data (such as actual connection rate, high intent rate, click-through rate, etc.) to adapt to the latest market demands and changes.

[0071] Targeted marketing campaigns, such as AI-powered telemarketing and sending hang-up text messages, are executed for the selected high-scoring user groups. Because these users have been identified as having high conversion potential, this method can significantly improve marketing effectiveness and return on investment (ROI) compared to randomly selecting users for marketing.

[0072] Real-time feedback and optimization: After each campaign, collect and analyze actual marketing results (including but not limited to metrics such as connection rate, high intent rate, and click-through rate) to further optimize future marketing strategies. For example, if a specific group performs significantly better than expected, it can receive more attention and resources in subsequent campaigns.

[0073] Step S140 not only effectively filters the target audience but also maximizes the effectiveness of marketing campaigns even with limited resources. This method, based on multi-stage modeling and integrated scoring mechanisms, is particularly suitable for scenarios requiring refined operations and efficient resource allocation, and performs exceptionally well in complex marketing processes such as AI voice + SMS joint marketing.

[0074] In this embodiment, TOP N refers to the top few, and the corresponding number can be determined according to the actual situation.

[0075] The method in this embodiment aims to improve the final conversion rate by constructing a multi-stage model (including an engagement model, a high-intent model, and a click model) and performing fusion modeling to output a comprehensive score that better reflects the overall marketing conversion process. Specific objectives include: improving the model's end-to-end predictive capability, achieving integrated modeling from engagement to conversion; correcting end-point label bias through predictive scoring at each stage; prioritizing users most likely to convert when resources are limited by utilizing the fusion model's ranking capabilities; designing reasonable model fusion methods (such as logistic regression weighting, stacking, etc.) to adapt to flexible adjustments in campaign strategies; and enabling self-checking and correction of model performance by combining historical samples and recent performance.

[0076] First, multiple models are trained. The connection model is trained using the "whether the call was successfully connected" status from the call logs of the past N days as labels, where connected calls are labeled 1 and unconnected calls are labeled 0. The high-intent model trains a high-intent prediction model by assigning positive labels to high-intent users in the connected data of the past N days using AI speech recognition; for example, category A is labeled 1, while categories B and C are labeled 0. The click model uses the successfully sent SMS data after high intent users in the past N days, labeling "clicked short link" as a positive sample (1) and unclicked links (0). Each model uses a unified user profile feature system, including gender, age, activity level, geographic location, and historical behavior. These three basic models are trained monthly to maintain model freshness.

[0077] Specifically, taking call records from the past month as an example, users whose calls were connected are labeled 1, and those whose calls were not connected are labeled 0. Features include the user's profile and behavior (e.g., gender, age, activity level, geographic location, historical behavior). Using these features as input and the labels as tags, an XGBoost model is trained to obtain the connection model. For users who made and connected calls in the past month, AI voice recognition identifies high-intent users (Category A) as labeled 1, and low-intent users (Categories B and C) as labeled 0. Features again include the user's profile and behavior. Using these features as input and the labels as tags, an XGBoost model is trained to obtain the high-intent model. For users who were identified as having high intent by AI voice recognition in the past month and successfully received a hang-up SMS, users who clicked the short link in the SMS are labeled 1, and those who did not click the short link are labeled 0. Features include the user's profile and behavior. Using these features as input and the labels as tags, an XGBoost model is trained to obtain the click model.

[0078] Secondly, model fusion is performed. The user's scores across the three models are obtained. a score b score c ), where score a Corresponding connection model, score b Corresponding to the high intent model, score cCorresponding click model. The modeling metrics are determined based on current business needs. For example, if a high ROI is required, the click / connection metric needs to be optimized; if the click audience needs to be expanded, the click / call metric needs to be optimized. Taking the optimization of the click / connection metric as an example, using data from the past n days, "connected and clicked" samples are labeled as positive samples (1), and "connected but not clicked" samples are labeled as negative samples (0). A fusion model D is trained using logistic regression (or other fusion algorithms). The fusion score is output. d = LR(score a score b score c Taking a high ROI requirement as an example, we need to optimize the click / connection metrics. We obtain connection data from the past 3 days. From this data, users who connected and clicked the short link are labeled as 1, and users who connected but did not click are labeled as 0 (including users who connected but did not receive a hang-up SMS, and users who connected and received a hang-up SMS but did not click). We obtain the user profiles and behaviors (such as gender, age, activity level, geolocation, historical behavior, etc.) for the current date as features, and input them into the connection model, high intent model, and click model respectively, thus outputting scores for the three models, denoted as scores. a score b score c The score obtained in the previous step a score b score c As new features (new_features), logistic regression (LR) is used to model new_features and labels to obtain a fusion model. Features are then obtained from the full target audience using the same method, and input into the engagement model, high-intent model, and click model to obtain the score. a score b score c The new features, `new_features`, are input into the fusion model to obtain `score_d`. Taking logistic regression as an example, the structure of the fusion model is: `scored = σ(w)`. T ⋅new_features+b), where σ is the sigmoid function, with an output range of [0, 1]. Its advantage is strong interpretability; the weighting of each model's score in the conversion probability prediction is clearly visible. Specifically, ; Finally, audience selection is performed. In each campaign, the top N users are selected based on their score_d for resource allocation (AI telemarketing).

[0079] A phased modeling structure based on multi-stage labels; using multiple model scores for fusion instead of a single model scoring system; the fused model outputs a unified score, suitable for multi-strategy audience selection and delivery optimization; supports dynamic fine-tuning of the fused model score using recent click data, call data, etc. (e.g., using logistic regression or Isotonic Regression correction); the scoring ranking has stability and monotonicity, suitable for reliable evaluation in strongly biased delivery scenarios.

[0080] The method in this embodiment constructs a link modeling system for the three stages of AI voice connection - high intent judgment - SMS click, trains three sub-models respectively, and outputs the final conversion probability score through weighted fusion (such as logistic regression, stacking), effectively capturing the sequential dependency features of the passive behavior link and making up for the modeling deficiencies of traditional click prediction models in this type of scenario.

[0081] By combining real delivery feedback results (such as clicks and connections) from the past T days with historical model scores, unsupervised methods (such as equal-frequency binning + click-rate mapping, Isotonic Regression, etc.) are used to dynamically correct the model output, thereby improving the model's adaptability to resource channel fluctuations and sample drift.

[0082] An automated modeling and update mechanism of "monthly model update + daily calibration" is established. The system can evaluate the changes in model performance in real time based on the actual campaign results on T+1, and trigger alarms and model retraining when the results are abnormal, so as to ensure the long-term stable operation of the system under complex marketing processes.

[0083] In a delivery environment where only the Top N samples can be observed, an evaluation mechanism is proposed based on the decreasing (monotonic) click-through rate of the model score interval to solve the problem that the traditional full-sample AUC / KS becomes invalid due to the lack of sample acquisition.

[0084] Applied to AI voice + SMS combined marketing scenarios, this model can significantly improve ROI per unit reach cost. Actual deployment verification shows that this integrated model significantly improves click / connection metrics and ROI, achieving ROI of 2-13 at high-scoring stages. Notably, when network resources are limited, the audience selected by this model has a higher ROI compared to conventional tasks.

[0085] Furthermore, the method in this embodiment is built based on a large-scale historical campaign sample and takes into account multiple variables in the marketing chain (such as routes, suppliers, user characteristics, etc.), and has strong robustness, especially in resource-constrained scenarios.

[0086] The aforementioned intelligent marketing method based on multi-stage predictive model fusion integrates user profile feature data of the target audience and historical marketing data as initial data. It then sequentially trains targeted models for three key marketing stages: engagement, high intent, and click, forming engagement, high intent, and click models that respectively reflect the conversion probability of each stage. These single-stage models are then flexibly combined according to specific business needs to generate a fusion model. This model is used to calculate a comprehensive score for each target user, which is then used to rank and filter the target audience, prioritizing users with higher scores for resource allocation. This process not only fully leverages the value of multi-source data and effectively corrects the bias that may arise from single-label approaches, but also significantly improves the generalization ability and predictive accuracy of the overall model through a multi-model fusion strategy. This ensures high conversion rates and ROI while improving marketing efficiency, overcoming the problems of insufficient data utilization, label bias, and insufficient generalization ability of single models in existing technologies.

[0087] Figure 5 This is a flowchart illustrating an intelligent marketing method based on the fusion of multi-stage prediction models, provided in another embodiment of the present invention. Figure 5 As shown, the intelligent marketing method based on multi-stage prediction model fusion in this embodiment includes steps S210-S250. Steps S210-S240 are similar to steps S110-S140 in the above embodiment and will not be described again here. The following details the additional step S250 in this embodiment.

[0088] S250. Collect and analyze actual marketing data feedback daily, use Isotonic Regression or Logistic Regression to calibrate the connection model, high intent model, and click model, and dynamically adjust the weights and parameters of the connection model, high intent model, and click model to continuously optimize their performance.

[0089] In one embodiment, step S250 described above may include steps S251 to S255.

[0090] S251. Collect and analyze connection rate, high intent rate and click-through rate daily to initially assess the effectiveness of the corresponding model.

[0091] In this embodiment, statistical analysis is performed on the connection rate, high intent rate, and click-through rate based on the daily actual delivery data feedback. This analysis serves as the basis for initially determining the effectiveness of the currently used connection model, high intent model, and click-through model.

[0092] S252. Using data from the last three days, calibrate the scores of the engagement model, high intent model, and click model using Isotonic Regression or Logistic Regression, and dynamically adjust the weights of high-performing strategies.

[0093] In this embodiment, the output scores of each model are calibrated using actual deployment data from the past three days, and methods such as Isotonic Regression or Logistic Regression are used to correct model prediction biases.

[0094] Based on the recent performance of different strategies, the weights of each model in the fusion model are dynamically adjusted to ensure that the fusion model can more accurately reflect the actual situation.

[0095] S253. Use AUC and KS metrics to evaluate the distinguishing ability of the connection model, high intent model and click model, and collect conversion population samples to calculate AUC and KS metrics.

[0096] In this embodiment, AUC (Area Under Curve) and KS (Kolmogorov-Smirnov statistic) are important indicators for measuring the ability of a binary classification model to distinguish between positive and negative samples.

[0097] Collect a sample of converted users and use these metrics to evaluate the performance of the engagement model, high intent model, and click model, and identify areas for improvement.

[0098] S254. Calculate the click-through rate among users to evaluate the accuracy of the connection model, high intent model, and click model in high-potential market segments.

[0099] In this embodiment, the click-through rate is calculated for a specific market segment or high-potential user group, and the accuracy of the model's predictions for this user group is evaluated.

[0100] By comparing the difference between the actual click-through rate and the model's predicted value, the model parameters or feature selection can be further optimized to improve the prediction accuracy for specific groups of people.

[0101] S255. Optimize the fusion model based on the data from the previous three days, including adjusting weights, introducing new features, or improving the algorithm.

[0102] In this embodiment, the fusion model is re-examined and optimized based on the data from the previous three days. This may involve adjusting the weights of individual models, adding new feature variables, or trying different algorithm combinations.

[0103] This process aims to continuously iterate and update the model to better adapt to market changes and improve overall marketing efficiency and return on investment (ROI).

[0104] Through this series of steps, the system can perform daily maintenance and optimization of the connection model, high intent model, and click model, ensuring that they are always in optimal condition and providing strong support for subsequent resource deployment.

[0105] The implementation of this invention also includes a crucial step—online correction and evaluation. This step aims to further optimize the model's performance and effectiveness through real-time feedback and dynamic adjustments.

[0106] Specifically, the system collects feedback data the day after each campaign, using key metrics such as actual connection rate, high intent rate, and click-through rate. This data directly reflects the model's performance in real-world applications. To further improve the model's accuracy and adaptability, the system can use data from the past three days to calibrate the model score using methods such as Isotonic Regression or Logistic Regression. If certain strategies or models perform well, the system can also dynamically adjust their weights to better contribute to the calculation of the final score_d.

[0107] To comprehensively evaluate the model's performance, the system uses a variety of metrics, including AUC, KS, TopN click-through rate, and cost-per-unit conversion rate. These metrics measure the model's performance and effectiveness from different perspectives. For example, AUC and KS can assess the model's overall discrimination ability, while TopN click-through rate directly reflects the model's effectiveness in filtering high-value users. Cost-per-unit conversion rate focuses on the model's economic benefits in actual marketing campaigns.

[0108] In practice, the first step is to evaluate the campaign's effectiveness. By calculating the connection rate, high intent rate, and click-through rate of the day's campaign data, the system can determine the difference between the target audience selected by the fusion model and the overall market average. These metrics not only reflect the model's selection effectiveness but also provide a preliminary assessment of its validity. The specific calculation formulas are as follows: Connection rate = number of connected calls divided by number of calls made; High intent rate = number of high-intent Category A users identified by AI divided by the number of connected calls; Click-through rate = number of people who clicked the short link divided by the number of people who received the hang-up SMS.

[0109] Next, the system uses metrics such as AUC, KS, and TopN click-through rate to evaluate the model's performance for the day. After each day's marketing campaign ends, the system collects converted users as a sample, labeling users who clicked the short link as label=1 and other users as label=0. Then, it obtains the fusion model score_d for this batch of users through the previous steps, and uses score_d and label as input to calculate AUC and KS values ​​to evaluate the model's performance.

[0110] To calculate the Top N click-through rate, the system sorts users based on their score_d, selects the Top N users with the highest scores, and calculates their click-through rates. This metric directly reflects the model's effectiveness in selecting high-value users.

[0111] Finally, to further optimize the model, the system will update the fusion model using data from the past three days. The specific method can be found in the training process of the fusion model mentioned earlier. By introducing new data and feedback, the model is continuously optimized and adjusted to ensure optimal performance in a constantly changing market environment.

[0112] Figure 6 This is a schematic block diagram of an intelligent marketing system 300 based on multi-stage prediction model fusion provided in an embodiment of the present invention. Figure 6 As shown, corresponding to the above-described intelligent marketing method based on multi-stage prediction model fusion, the present invention also provides an intelligent marketing system 300 based on multi-stage prediction model fusion. This intelligent marketing system 300 includes a unit for executing the above-described intelligent marketing method based on multi-stage prediction model fusion, and the system can be configured in a server. Specifically, please refer to... Figure 6 The intelligent marketing system 300 based on the fusion of multi-stage prediction models includes a data acquisition unit 301, a training unit 302, a comprehensive scoring determination unit 303, and a ranking and delivery unit 304.

[0113] The data acquisition unit 301 is used to acquire user profile feature data and historical marketing data of the target audience to obtain initial data; the training unit 302 is used to train the prediction models of the initial data in three stages: connection, high intent, and click, to obtain the connection model, high intent model, and click model; the comprehensive score determination unit 303 is used to generate a fusion model based on business needs by combining the connection model, high intent model, and click model, and calculate the score corresponding to the target personnel based on the fusion model to generate a comprehensive score; the ranking and delivery unit 304 is used to rank the target audience according to the comprehensive score and select users whose scores meet the requirements for resource delivery.

[0114] In one embodiment, the training unit 302 includes: The connection model training subunit is used to filter out data from the historical marketing data of the initial data that shows whether the call was successfully connected, and label it. Combined with the corresponding user profile features as a sample set, the XGBoost algorithm is used to train the model to obtain the connection model. The high intent model training subunit is used to mark high intent users as positive samples and low intent users as negative samples based on the historical marketing data that has been connected based on the initial data, according to the AI ​​voice recognition results, so as to train the model using the XGBoost algorithm to obtain the high intent model. The click model training subunit is used to mark data with clicked links in SMS messages as positive samples and data with unclicked links in SMS messages as positive samples based on the initial data of high-intent users who have already sent SMS messages, in order to form a sample set and obtain the click model.

[0115] In one embodiment, the comprehensive scoring determination unit 303 includes: The selection subunit is used to select the optimization direction based on business needs; the input set construction subunit is used to label users who connect and click as positive samples and users who only connect but do not click as negative samples according to the optimization direction, forming labels and collecting user profile feature data to obtain the input set; the rating acquisition subunit is used to obtain the rating of the input set in the connection model, high intent model and click model; the fusion training subunit is used to use logistic regression to use the ratings in the connection model, high intent model and click model as new features and combine them with the labels for training to obtain the fusion model; the synthesis subunit is used to construct an input set for the target audience and input it into the connection model, high intent model and click model to determine the corresponding rating, and then use the fusion model to determine the comprehensive rating.

[0116] Figure 7 This is a schematic block diagram of an intelligent marketing system 300 based on multi-stage prediction model fusion, provided in another embodiment of the present invention. Figure 7 As shown, the intelligent marketing system 300 based on multi-stage prediction model fusion in this embodiment is based on the above embodiment with the addition of optimization unit 305.

[0117] The optimization unit 305 is used to collect and analyze actual marketing data feedback daily, calibrate the connection model, high intent model, and click model using IsotonicRegression or Logistic Regression, and dynamically adjust the weights and parameters of the connection model, high intent model, and click model to continuously optimize their performance.

[0118] In one embodiment, the optimization unit 305 includes: The system comprises the following sub-units: a data collection sub-unit, which collects and analyzes daily connection rate, high intent rate, and click-through rate to initially assess the effectiveness of the corresponding models; an adjustment sub-unit, which uses data from the past three days to calibrate the scores of the connection model, high intent model, and click-through model using Isotonic Regression or Logistic Regression, and dynamically adjusts the weights of high-performing strategies; a metrics collection sub-unit, which uses AUC and KS metrics to evaluate the discriminative power of the connection model, high intent model, and click-through model, and collects conversion user samples to calculate AUC and KS metrics; a calculation sub-unit, which calculates the click-through rate among users to evaluate the accuracy of the connection model, high intent model, and click-through model in high-potential market segments; and an optimization sub-unit, which optimizes the fusion model based on the data from the previous three days, including adjusting weights, introducing new features, or improving the algorithm.

[0119] It should be noted that those skilled in the art can clearly understand that the specific implementation process of the intelligent marketing system 300 based on the fusion of multi-stage prediction models and its various units can be found in the corresponding descriptions in the aforementioned method embodiments. For the sake of convenience and brevity, these details will not be repeated here.

[0120] The aforementioned intelligent marketing system 300 based on the fusion of multi-stage prediction models can be implemented as a computer program, which can, for example... Figure 8 It runs on the computer device shown.

[0121] Please see Figure 8 , Figure 8 This is a schematic block diagram of a computer device provided in an embodiment of this application. The computer device 500 can be a server, wherein the server can be a standalone server or a server cluster composed of multiple servers.

[0122] See Figure 8 The computer device 500 includes a processor 502, a memory, and a network interface 505 connected via a system bus 501. The memory may include a non-volatile storage medium 503 and internal memory 504.

[0123] The non-volatile storage medium 503 may store an operating system 5031 and a computer program 5032. The computer program 5032 includes program instructions that, when executed, cause the processor 502 to perform an intelligent marketing method based on the fusion of multi-stage predictive models.

[0124] The processor 502 provides computing and control capabilities to support the operation of the entire computer device 500.

[0125] The internal memory 504 provides an environment for the operation of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute an intelligent marketing method based on the fusion of multi-stage predictive models.

[0126] This network interface 505 is used for network communication with other devices. Those skilled in the art will understand that... Figure 8 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 500 to which the present application is applied. The specific computer device 500 may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0127] The processor 502 is used to run a computer program 5032 stored in the memory to perform the following steps: Acquire user profile feature data and historical marketing data of the target audience to obtain initial data; train prediction models for three stages—connection, high intent, and click—on the initial data to obtain connection model, high intent model, and click model; generate a fusion model by combining the connection model, high intent model, and click model according to business needs, and calculate the score corresponding to the target personnel based on the fusion model to generate a comprehensive score; rank the target audience according to the comprehensive score, and select users with scores that meet the requirements for resource allocation.

[0128] In one embodiment, after implementing the steps of sorting the target population based on the comprehensive score and selecting users with scores that meet the requirements for resource allocation, the processor 502 further implements the following steps: Daily collection and analysis of actual marketing data feedback; use Isotonic Regression or Logistic Regression to calibrate the connection model, high intent model, and click model; and dynamically adjust the weights and parameters of the connection model, high intent model, and click model to continuously optimize their performance.

[0129] In one embodiment, when the processor 502 trains the prediction model for the three stages of connection, high intent, and click on the initial data to obtain the connection model, high intent model, and click model, it specifically implements the following steps: Based on the historical marketing data of the initial data, call records showing whether a call was successfully connected are selected and tagged. These records are then used as a sample set, and the XGBoost algorithm is used to train a model to obtain a connection model. Based on the historical marketing data of connected calls from the initial data, high-intent users are marked as positive samples and low-intent users as negative samples according to AI voice recognition results. The XGBoost algorithm is then used to train a model to obtain a high-intent model. Based on the high-intent users who have already sent SMS messages from the initial data, data showing clicks on links in the SMS messages are marked as positive samples, and data showing no clicks on links in the SMS messages are also marked as positive samples, forming a sample set to obtain a click model.

[0130] In one embodiment, when the processor 502 generates a fusion model by combining the connection model, high intent model, and click model according to business requirements, and calculates the score corresponding to the target person based on the fusion model to generate a comprehensive score, the processor 502 specifically implements the following steps: Optimization directions are selected based on business needs; users who connect and click are labeled as positive samples, and users who only connect but do not click are labeled as negative samples, forming labels, and user profile feature data is collected to obtain an input set; the scores of the input set in the connection model, high intent model, and click model are obtained; logistic regression is used to train the connection model, high intent model, and click model as new features combined with the labels to obtain a fusion model; an input set is constructed for the target audience and input into the connection model, high intent model, and click model to determine the corresponding scores, and the comprehensive score is determined by the fusion model.

[0131] The comprehensive score includes a conversion probability in the range [0, 1] by linearly combining the weighted scores of the connection model, high intent model, and click model using the sigmoid function.

[0132] In one embodiment, when the processor 502 implements the steps of collecting and analyzing actual marketing data feedback daily, calibrating the connection model, high intent model, and click model using Isotonic Regression or Logistic Regression, and dynamically adjusting the weights and parameters of the connection model, high intent model, and click model to continuously optimize their performance, the processor 502 specifically implements the following steps: Daily collection and analysis of connection rate, high intent rate, and click-through rate (CTR) are used to initially assess the effectiveness of the corresponding models. The scores of the connection model, high intent model, and click-through model are calibrated using Isotonic Regression or Logistic Regression based on data from the past three days, and the weights of high-performing strategies are dynamically adjusted. The AUC and KS metrics are used to evaluate the discriminative power of the connection model, high intent model, and click-through model, and AUC and KS metrics are calculated by collecting samples from converted users. Click-through rate is calculated among users to evaluate the accuracy of the connection model, high intent model, and click-through model in high-potential market segments. The fusion model is optimized based on the data from the previous three days, including adjusting weights, introducing new features, or improving the algorithm.

[0133] It should be understood that in the embodiments of this application, the processor 502 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0134] It will be understood by those skilled in the art that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program includes program instructions and can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0135] Therefore, the present invention also provides a storage medium. This storage medium can be a computer-readable storage medium. The storage medium stores a computer program, wherein when executed by a processor, the computer program causes the processor to perform the following steps: Acquire user profile feature data and historical marketing data of the target audience to obtain initial data; train prediction models for three stages—connection, high intent, and click—on the initial data to obtain connection model, high intent model, and click model; generate a fusion model by combining the connection model, high intent model, and click model according to business needs, and calculate the score corresponding to the target personnel based on the fusion model to generate a comprehensive score; rank the target audience according to the comprehensive score, and select users with scores that meet the requirements for resource allocation.

[0136] In one embodiment, after executing the computer program to implement the steps of sorting the target population based on the comprehensive score and selecting users with scores that meet the requirements for resource allocation, the processor further implements the following steps: Daily collection and analysis of actual marketing data feedback; use Isotonic Regression or Logistic Regression to calibrate the connection model, high intent model, and click model; and dynamically adjust the weights and parameters of the connection model, high intent model, and click model to continuously optimize their performance.

[0137] In one embodiment, when the processor executes the computer program to train the prediction model for the three stages of connection, high intent, and click on the initial data to obtain the connection model, high intent model, and click model, the processor specifically implements the following steps: Based on the historical marketing data of the initial data, call records showing whether a call was successfully connected are selected and tagged. These records are then used as a sample set, and the XGBoost algorithm is used to train a model to obtain a connection model. Based on the historical marketing data of connected calls from the initial data, high-intent users are marked as positive samples and low-intent users as negative samples according to AI voice recognition results. The XGBoost algorithm is then used to train a model to obtain a high-intent model. Based on the high-intent users who have already sent SMS messages from the initial data, data showing clicks on links in the SMS messages are marked as positive samples, and data showing no clicks on links in the SMS messages are also marked as positive samples, forming a sample set to obtain a click model.

[0138] In one embodiment, when the processor executes the computer program to generate a fusion model based on business needs by combining the connection model, high intent model, and click model, and calculates the score corresponding to the target person based on the fusion model to generate a comprehensive score, the processor specifically implements the following steps: Optimization directions are selected based on business needs; users who connect and click are labeled as positive samples, and users who only connect but do not click are labeled as negative samples, forming labels, and user profile feature data is collected to obtain an input set; the scores of the input set in the connection model, high intent model, and click model are obtained; logistic regression is used to train the connection model, high intent model, and click model as new features combined with the labels to obtain a fusion model; an input set is constructed for the target audience and input into the connection model, high intent model, and click model to determine the corresponding scores, and the comprehensive score is determined by the fusion model.

[0139] The comprehensive score includes a conversion probability in the range [0, 1] by linearly combining the weighted scores of the connection model, high intent model, and click model using the sigmoid function.

[0140] In one embodiment, when the processor executes the computer program to collect and analyze actual marketing data feedback daily, calibrate the connection model, high intent model, and click model using Isotonic Regression or Logistic Regression, and dynamically adjust the weights and parameters of the connection model, high intent model, and click model to continuously optimize their performance, the processor specifically implements the following steps: Daily collection and analysis of connection rate, high intent rate, and click-through rate (CTR) are used to initially assess the effectiveness of the corresponding models. The scores of the connection model, high intent model, and click-through model are calibrated using Isotonic Regression or Logistic Regression based on data from the past three days, and the weights of high-performing strategies are dynamically adjusted. The AUC and KS metrics are used to evaluate the discriminative power of the connection model, high intent model, and click-through model, and AUC and KS metrics are calculated by collecting samples from converted users. Click-through rate is calculated among users to evaluate the accuracy of the connection model, high intent model, and click-through model in high-potential market segments. The fusion model is optimized based on the data from the previous three days, including adjusting weights, introducing new features, or improving the algorithm.

[0141] The storage medium can be any computer-readable storage medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0142] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0143] In the embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For example, the division of each unit is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0144] The steps in the method of this invention can be adjusted, merged, or reduced in order according to actual needs. The units in the system of this invention can be merged, divided, or reduced according to actual needs. Furthermore, the functional units in the various embodiments of this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0145] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0146] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An intelligent marketing method based on the fusion of multi-stage prediction models, characterized in that, include: Obtain user profile data and historical marketing data of the target audience to obtain initial data; The initial data is used to train prediction models for three stages: connection, high intent, and click, to obtain connection model, high intent model, and click model. Based on business needs, a fusion model is generated by combining the connection model, high intent model, and click model. The score corresponding to the target person is calculated based on the fusion model to generate a comprehensive score. The target audience is ranked based on the comprehensive score, and users whose scores meet the requirements are selected for resource allocation.

2. The intelligent marketing method based on multi-stage prediction model fusion according to claim 1, characterized in that, After ranking the target audience based on the comprehensive score and selecting users who meet the score requirements for resource allocation, the process further includes: Daily collection and analysis of actual marketing data feedback; use Isotonic Regression or Logistic Regression to calibrate the connection model, high intent model, and click model; and dynamically adjust the weights and parameters of the connection model, high intent model, and click model to continuously optimize their performance.

3. The intelligent marketing method based on multi-stage prediction model fusion according to claim 1, characterized in that, The training of prediction models for three stages—connection, high intent, and click—on the initial data to obtain connection, high intent, and click models includes: Based on the historical marketing data of the initial data, the call records with whether the call was successfully connected were selected and tagged. Combined with the corresponding user profile features as a sample set, the XGBoost algorithm was used to train the model to obtain the connection model. Based on the historical marketing data that has been connected according to the initial data, high-intent users are marked as positive samples and low-intent users are marked as negative samples according to the AI ​​voice recognition results, so as to use the XGBoost algorithm to train the model and obtain the high-intent model. Based on the initial data, for users with high intent to click on links in SMS messages, data showing clicks in SMS messages are marked as positive samples, and data showing no clicks in SMS messages are also marked as positive samples, forming a sample set to obtain the click model.

4. The intelligent marketing method based on multi-stage prediction model fusion according to claim 1, characterized in that, The process involves generating a fusion model based on business needs, combining the connection model, high-intent model, and click model, and calculating the score corresponding to the target personnel based on the fusion model to generate a comprehensive score, including: Choose the optimization direction based on business needs; Users who connect and click according to the optimized direction are labeled as positive samples, and users who only connect but do not click are labeled as negative samples, thus forming labels and collecting user profile feature data to obtain the input set; Obtain the scores of the input set in the engagement model, high intent model, and click model; Logistic regression was used to train a fusion model by combining the ratings from the engagement model, high intent model, and click model as new features with the labels. An input set is constructed for the target audience and fed into the connection model, high intent model, and click model to determine the corresponding scores. The comprehensive score is then determined through the fusion model.

5. The intelligent marketing method based on multi-stage prediction model fusion according to claim 4, characterized in that, The comprehensive score includes a linear combination of the weighted connection model, high intent model, and click model scores using the sigmoid function, outputting a conversion probability in the range [0, 1].

6. The intelligent marketing method based on multi-stage prediction model fusion according to claim 2, characterized in that, The daily collection and analysis of actual marketing data feedback, using Isotonic Regression or Logistic Regression to calibrate the connection model, high-intent model, and click model, and dynamically adjusting the weights and parameters of the connection model, high-intent model, and click model to continuously optimize their performance, including: Daily collection and analysis of connection rate, high intent rate, and click-through rate to initially assess the effectiveness of the model; Using data from the last three days, the scores of the engagement model, high intent model, and click model were calibrated using Isotonic Regression or Logistic Regression, and the weights of high-performing strategies were dynamically adjusted. The AUC and KS metrics were used to evaluate the distinguishing ability of the connection model, high intent model and click model, and the AUC and KS metrics were calculated by collecting a sample of the conversion audience. Calculate click-through rates among users to evaluate the accuracy of the engagement model, high intent model, and click model in high-potential market segments; The fusion model is optimized based on the data from the first three days, including adjusting weights, introducing new features, or improving the algorithm.

7. An intelligent marketing system based on the fusion of multi-stage prediction models, characterized in that, include: The data acquisition unit is used to acquire user profile feature data and historical marketing data of the target audience to obtain initial data; The training unit is used to train prediction models for three stages—connection, high intent, and click—on the initial data to obtain connection model, high intent model, and click model. The comprehensive scoring unit is used to generate a fusion model by combining the connection model, high intention model and click model according to business needs, and to calculate the score corresponding to the target person according to the fusion model to generate a comprehensive score. The sorting and delivery unit is used to sort the target audience based on the comprehensive score and select users whose scores meet the requirements for resource delivery.

8. The intelligent marketing system based on multi-stage prediction model fusion according to claim 7, characterized in that, Also includes: The optimization unit is used to collect and analyze actual marketing data feedback daily, calibrate the connection model, high intent model, and click model using Isotonic Regression or Logistic Regression, and dynamically adjust the weights and parameters of the connection model, high intent model, and click model to continuously optimize their performance.

9. A computer device, characterized in that, The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method as described in any one of claims 1 to 6.

10. A storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 6.

Images