A customer intention rating method and device based on cross-domain attention fusion, and a storage medium

Abstract

The application relates to the field of artificial intelligence and data mining technology, and particularly discloses a customer intention rating method and device based on cross-domain attention fusion and a storage medium, wherein the customer intention rating method based on cross-domain attention fusion comprises the following steps: collecting public domain data and private domain data in a whole life cycle before sales; pre-processing and feature extraction are performed on the public domain data and the private domain data to construct a public domain feature set and a private domain feature set; through a cross-domain attention fusion mechanism, the public domain feature set is taken as a query vector Q, the private domain feature set is taken as a key vector K and a value vector V, cross-domain attention weights are calculated, a fusion feature set is generated, and customer intention rating is performed based on the fusion feature set. The cross-domain attention fusion mechanism is introduced to improve the accuracy and reliability of customer intention rating.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and data mining technology, and in particular to a customer intention rating method, apparatus and storage medium based on cross-domain attention fusion. Background Technology

[0002] In sales management across various industries, accurately assessing the intent of potential customers is crucial for precise allocation of sales resources and improved conversion rates. Taking the automotive sales industry as an example, the process from initial lead generation to final sale typically involves multiple stages, including lead generation, store follow-up, in-store negotiation, and test drive experience—a process known as the pre-sales lifecycle. Throughout this process, sales teams can obtain relevant customer data from multiple channels. Depending on the data source, customer data can generally be divided into two main categories: public domain data and private domain data. Public domain data refers to information left by users through public channels, such as user contact information left by users on third-party automotive information platforms, social media, search engines, and other public channels, as well as user profile data and advertising feedback data provided by third-party data platforms. This type of data has a wide coverage and can reflect users' public interests and initial intentions, but its granularity is relatively coarse and its correlation with specific sales scenarios is weak. Private domain data refers to user behavior data collected by enterprises within their own channels, such as follow-up records in store CRM systems, private domain outreach behavior data (such as WeChat official account interactions, mini-program visits), and historical transaction records. This type of data is directly related to sales operations, has high information accuracy, but its coverage is relatively limited and depends on the depth of existing interactions between customers and enterprises.

[0003] In existing technologies, for example, Chinese patent CN111144938A discloses a method and system for sales lead rating applicable to the automotive industry. This method achieves lead classification by statically scoring online leads. While it introduces external data as a supplement, it merely integrates external data with online leads into simple data packets for model input, without deeply considering the inherent relationships between different types of data. Another example is Chinese patent CN116308756A, which discloses a method, device, equipment, medium, and product for determining target customers. This method fuses probabilities output from scenario models and general models separately. However, the core of this method lies in the probability integration at the model output layer; its input still processes data from different sources separately, without establishing a cross-domain data interaction mechanism at the feature level. Essentially, it remains a shallow fusion after independent processing. It is evident that existing technologies generally suffer from the following shortcomings: public and private domain data are simply concatenated or processed independently before being input into the model, without considering the deep relationships between them, resulting in low accuracy in customer intention rating. Summary of the Invention

[0004] The present invention aims to solve the technical problems existing in the prior art and provides a customer intention rating method, device and storage medium based on cross-domain attention fusion, which improves the accuracy and reliability of customer intention rating by introducing a cross-domain attention fusion mechanism.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solutions:

[0006] The customer intent rating method based on cross-domain attention fusion described in this invention includes:

[0007] Collect public and private domain data throughout the entire pre-sales lifecycle;

[0008] The public domain data and the private domain data are preprocessed and feature extracted to construct a public domain feature set and a private domain feature set;

[0009] By using a cross-domain attention fusion mechanism, the cross-domain attention weights are calculated using the public domain feature set as the query vector Q, the private domain feature set as the key vector K and the value vector V, and a fused feature set is generated.

[0010] Customer intent rating is performed based on the fused feature set.

[0011] Furthermore, the public domain data and the private domain data are preprocessed, specifically by filling missing numerical values ​​with 0, filling missing categorical values ​​with the string "NULL", performing equal-frequency binning on continuous value features, and performing one-hot encoding on categorical features.

[0012] Alternatively, feature extraction can be performed on the public domain data and the private domain data, specifically by deriving new features based on business rules.

[0013] Furthermore, after feature extraction, feature filtering is also required. Specifically, the information gain of each feature with respect to the target variable or the feature importance based on LightGBM is calculated, and features are filtered according to their importance ranking, removing redundant features with importance below a preset threshold.

[0014] Furthermore, the cross-domain attention fusion mechanism includes: mapping the public domain feature set to the query vector Q through a fully connected layer, mapping the private domain feature set to the key vector K and value vector V through two independent fully connected layers respectively, and calculating the cross-domain attention weights using a formula.

[0015]

[0016] Where d k The dimension is K; the attention output is concatenated with the original public domain features and private domain features to obtain the fused feature set.

[0017] Furthermore, after generating the fused feature set, the entire pre-sales lifecycle is divided into multiple stages, and a rating sub-model is constructed for each stage. The fused feature set is used to train the sub-models for each stage, and the comprehensive rating score is calculated through the stage transition probability matrix P.

[0018] Furthermore, the pre-sales lifecycle is divided into: unissued stage, unassigned stage, uncalled stage, follow-up stage, overdue follow-up stage, long-term no follow-up stage, and churn stage;

[0019] The rating sub-model is a binary classification sub-model based on LightGBM;

[0020] The stage transition probability matrix P has a dimension of 7×7, where the element P(i,j) represents the probability of transitioning from stage i to stage j, which is obtained based on historical data statistics.

[0021] The formula for calculating the comprehensive rating score is: R=ΣP(i,j)×W j ×R j W j Let R be the weight coefficient for stage j. j denoted as the rating score of the sub-model in stage j.

[0022] Furthermore, based on the event-driven mechanism, user behavior events are monitored. When a preset trigger event is detected, the historical rating score is decayed by combining a time decay function and the current rating is recalculated.

[0023] The time decay function is: , of which S original The historical rating score is represented by λ, the decay coefficient is represented by Δt, and the time interval since the last rating is represented by Δt. When Δt exceeds the preset threshold, the rating score for the corresponding stage will be automatically reduced by one level.

[0024] Furthermore, a multi-task portrait tag generation network is used to synchronously output a portrait tag set based on the fused feature set.

[0025] The device for customer intent rating based on cross-domain attention fusion according to the present invention includes:

[0026] The data acquisition module is used to collect public and private domain data throughout the entire pre-sales lifecycle.

[0027] The feature extraction module is used to preprocess and extract features from the public domain data and the private domain data;

[0028] The cross-domain attention fusion module is used to calculate cross-domain attention weights using the public domain feature set as the query vector Q and the private domain feature set as the key-value vector K and value vector V.

[0029] The multi-stage rating module is used to build sub-models for each stage and calculate the comprehensive rating score through the stage transition probability matrix P.

[0030] The event-driven update module is used to listen for behavioral events and trigger a recalculation of rating scores.

[0031] The image tag generation module is used to output image tag sets through a multi-task image tag generation network.

[0032] The present invention provides a storage medium storing a program that, when executed by a processor, implements the aforementioned customer intention rating method based on cross-domain attention fusion.

[0033] The customer intent rating method, device, and storage medium based on cross-domain attention fusion described in this invention, through a cross-domain attention fusion mechanism, uses a public domain feature set as the query vector and a private domain feature set as the key vector and value vector to calculate cross-domain attention weights. This enables adaptive learning of the deep correlation between public and private domain features, avoiding information redundancy and loss of cross-domain interaction information caused by simple concatenation of the two types of data. It achieves a leap from shallow concatenation to deep fusion of public and private domain data, thereby improving the accuracy and reliability of customer intent rating. Attached Figure Description

[0034] The above and other objects, features, and advantages of the invention will become clearer through a more detailed description of the preferred embodiments illustrated in the accompanying drawings. The same reference numerals denote the same parts throughout the drawings, and the drawings are not intentionally drawn to scale with actual dimensions; the focus is on illustrating the gist of the invention.

[0035] Figure 1 This is a flowchart of a customer intent rating method based on cross-domain attention fusion in an embodiment of the present invention.

[0036] Figure 2 This is a schematic diagram of the cross-domain attention fusion module.

[0037] Figure 3 A schematic diagram of the structure of the network for generating multi-task profile labels. Detailed Implementation

[0038] To facilitate understanding of the present invention, a more comprehensive description of the present invention will be given below with reference to the accompanying drawings.

[0039] It should be noted that when a component is considered to be "connected" to another component, it can be directly connected to and integrated with the other component, or there may be an intervening component present. The terms "mounted," "one end," "the other end," and similar expressions used in this document are for illustrative purposes only.

[0040] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0041] In sales industries, such as automotive sales, effective customer lead management is crucial for improving sales conversion rates. Based on this, this invention specifically provides a customer intent rating method based on cross-domain attention fusion, see [link to relevant documentation]. Figure 1-3 This includes: collecting public and private domain data throughout the entire pre-sales lifecycle; "public domain data" can be information left by users through public channels, user profile data provided by third-party data platforms, and advertising feedback data, etc.; "private domain data" can be follow-up records, private domain outreach behavior data, and historical transaction records from the store's CRM system, etc.; preprocessing and feature extraction of the public and private domain data to construct public and private domain feature sets; calculating cross-domain attention weights using a cross-domain attention fusion mechanism, with the public domain feature set as the query vector Q, the private domain feature set as the key vector K, and the value vector V, to generate a fused feature set; and performing customer intention rating based on the fused feature set.

[0042] In a preferred embodiment, the public domain data and the private domain data are preprocessed, specifically by: filling missing numerical values ​​with 0, filling missing categorical values ​​with the string "NULL", performing equal-frequency binning on continuous value features, and performing one-hot encoding on categorical features; and extracting features from the public domain data and the private domain data, specifically by: constructing new features based on business rules, such as calculating statistical features such as the number of times a user browses in the past seven days and the number of times they leave leads in the past thirty days.

[0043] In a preferred embodiment, after feature extraction, feature filtering is also performed. Specifically, the information gain of each feature with respect to the target variable or the feature importance based on LightGBM is calculated, and the features are filtered according to the importance ranking, and redundant features with importance below a preset threshold are deleted.

[0044] In this embodiment, by preprocessing, extracting features, and filtering features from public and private domain data, the training efficiency and generalization ability of the model can be improved.

[0045] In a preferred embodiment, the cross-domain attention fusion mechanism includes: mapping the public domain feature set to the query vector Q through a fully connected layer, mapping the private domain feature set to the key vector K and the value vector V through two independent fully connected layers respectively, and calculating the cross-domain attention weights using a formula.

[0046]

[0047] The softmax function is a normalization exponential function used to transform the input vector into a probability distribution, d. k Let K be the dimension, K T Let K be the transpose of the key vector. The above method can scale the dot product to prevent gradient vanishing. The attention output is concatenated with the original public and private features to obtain the fused feature set. Through the cross-domain attention fusion mechanism, the model can adaptively learn the deep relationships between public and private features, avoiding information redundancy and loss of association caused by traditional simple concatenation methods.

[0048] In a preferred embodiment, after generating the fused feature set, the pre-sales lifecycle is further divided into multiple stages. A rating sub-model is constructed for each stage, and the fused feature set is used to train each stage's sub-model. A comprehensive rating score is then calculated using the stage transition probability matrix P. Specifically, the pre-sales lifecycle can be divided into: undistributed stage, unassigned stage, uncalled stage, follow-up stage, overdue follow-up stage, long-term lack of follow-up stage, and churn stage. A rating sub-model is independently constructed for each stage. The rating sub-model is a binary classification sub-model based on LightGBM and is trained using the fused feature set of the corresponding stage. The stage transition probability matrix P has a dimension of 7×7, where element P(i,j) represents the probability of transitioning from stage i to stage j, which is obtained based on historical data statistics. The comprehensive rating score is calculated as: R=ΣP(i,j)×W j ×R j W j Let R be the weight coefficient for stage j. j Let R be the rating score of the sub-model in stage j. Based on the numerical range of the comprehensive rating score R, the rating results can be divided into five levels: S, A, B, C, and D. Among them, level S (R≥0.85) indicates an extremely high conversion probability, level A (0.70≤R<0.85) indicates a high conversion probability, level B (0.55≤R<0.70) indicates a medium conversion probability, level C (0.40≤R<0.55) indicates a low conversion probability, and level D (R<0.40) indicates a low conversion probability.

[0049] In this embodiment, the aforementioned two independent fully connected layers can be divided into a first fully connected layer and a second fully connected layer. The first fully connected layer inputs public domain cue features and uses a formula... To obtain the query vector Q, where X pub As a public domain clue characteristic, W Q This is a linear transformation matrix that maps public domain features to the query matrix; the second fully connected layer takes private domain behavioral features as input and uses the formula... and formula To obtain the key vector K and value vector V, which are used for public domain query retrieval and weighting, where X pri As a characteristic of private domain behavior, W K W is a linear transformation matrix that maps private domain features to the key matrix. V This is a linear transformation matrix that maps private domain features to a value matrix. Additionally, public domain clue features can be third-party automotive information, social platforms, search engines, etc., while private domain behavioral features can be store check-in, test drive records, price quotes, financial applications, etc. Two fully connected layers are used, where the query vector Q is responsible for asking questions from the public domain, and the value vector V and key vector K are responsible for answering and providing information from the private domain. These two fully connected layers are connected by a learnable W... Q W K W V Align heterogeneous features to the same dimension, It can use public domain intent to dynamically filter private domain behavior information, thereby generating cross-domain fusion features.

[0050] In a preferred embodiment, user behavior events are monitored based on an event-driven mechanism. When a preset trigger event is detected, the current rating is recalculated after the historical rating score is decayed using a time decay function. The event-driven mechanism monitors the user behavior event stream through a message queue (such as Kafka). The preset trigger events may include user browsing of the vehicle model page, user information submission, store follow-up feedback, and user check-in at the store. When a trigger event is detected, the corresponding stage of the rating recalculation process is immediately triggered to achieve real-time dynamic updates of the rating.

[0051] In this embodiment, the time decay function is: , of which S original The rating score is the historical rating score, λ is the decay coefficient (e.g., a value between 0.01 and 0.05), and Δt is the time interval since the last rating, in days. When Δt exceeds a preset threshold (e.g., 30 days), the rating score for the corresponding period is automatically reduced by one level. This method ensures the timeliness of the rating; customer ratings that have not seen new interactions for a long time naturally decay, preventing expired ratings from misleading sales resource allocation.

[0052] In a preferred embodiment, a multi-task profile label generation network synchronously outputs a profile label set based on the fused feature set. The profile label set may include one or more of the following: replacement vehicle model prediction, purchase intention prediction, interest preference labels, and brand preference labels. In this embodiment, the multi-task profile label generation network adopts a shared-bottom structure, including a shared bottom layer network and four task towers. The shared bottom layer network consists of multiple fully connected layers, for example, three layers, each with 256 neurons and a ReLU activation function. It receives the fused feature set and extracts shared feature representations. The four task towers are the replacement vehicle model prediction tower, the purchase intention prediction tower, the interest preference classification tower, and the brand preference classification tower. The replacement vehicle model prediction tower can output multi-class classification, corresponding to each candidate vehicle model; the purchase intention prediction tower can output binary classification, high / low intention; the interest preference classification tower can output multi-label classification, such as price sensitivity, configuration focus, brand loyalty, etc.; and the brand preference classification tower can output multi-class classification, corresponding to each competing brand. In this embodiment, each task tower independently outputs its corresponding profile label, and the loss function is a weighted sum of the cross-entropy losses of each task. In this embodiment, after the profile tags are generated, they can be provided to store consultants and CRM systems for the purpose of developing personalized follow-up strategies and marketing plans. For example, for customers with a high car purchase inclination and who are interested in a certain competitor's brand, a comparative test drive activity can be promoted in a targeted manner.

[0053] This invention also provides an embodiment of a customer intention rating device based on cross-domain attention fusion, comprising: a data acquisition module for collecting public domain data and private domain data throughout the pre-sales lifecycle; a feature extraction module for preprocessing and extracting features from the public domain data and the private domain data; a cross-domain attention fusion module for calculating cross-domain attention weights using the public domain feature set as the query vector Q, the private domain feature set as the key-value vector K, and the value vector V; a multi-stage rating module for constructing sub-models for each stage and calculating a comprehensive rating score through a stage transition probability matrix P; an event-driven update module for monitoring behavioral events and triggering a recalculation of the rating score; and a profile tag generation module for generating a network-output profile tag set through multi-task profile tag generation.

[0054] In a preferred embodiment, the event-driven update module further includes a time decay unit, used to adjust the update time according to the time decay time. Historical rating scores are attenuated.

[0055] The present invention also provides an embodiment of a storage medium on which a program is stored, which, when executed by a processor, implements the above-described customer intention rating method based on cross-domain attention fusion.

[0056] The customer intent rating method, device, and storage medium based on cross-domain attention fusion described in this invention, through a cross-domain attention fusion mechanism, uses a public domain feature set as the query vector and a private domain feature set as the key vector and value vector to calculate cross-domain attention weights. This enables adaptive learning of the deep correlation between public and private domain features, avoiding information redundancy and loss of cross-domain interaction information caused by simple concatenation of the two types of data. It achieves a leap from shallow concatenation to deep fusion of public and private domain data, thereby improving the accuracy and reliability of customer intent rating.

[0057] This invention incorporates a cross-domain attention fusion mechanism to introduce the deep correlation between public and private domain features into customer intention rating; it achieves dynamic tracking throughout the entire lifecycle through a multi-stage rating model and stage transition probability matrix; it enables real-time dynamic updates of the rating through an event-driven mechanism and time decay function; and it generates a multi-dimensional profile, including replacement vehicle model prediction, purchase tendency prediction, interest preference tags, and brand interest tags, while outputting the rating score through a multi-task profile tag generation network. This overall technical solution effectively solves the problems of static scoring, shallow feature fusion, and single output in existing technologies, significantly improving the accuracy and timeliness of customer intention rating, and providing technical support for the precise allocation of sales resources and the formulation of personalized marketing strategies.

[0058] In this application, unless otherwise expressly specified and limited, "above" or "below" the second feature can mean that the first feature is in direct contact with the second feature, or that the first feature is in indirect contact with the second feature through an intermediate medium. Furthermore, "above," "on top of," and "over" the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0059] In the description of this specification, the references to terms such as "preferred embodiment," "another embodiment," "other embodiment," or "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0060] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A customer intent rating method based on cross-domain attention fusion, characterized in that, include: Collect public and private domain data throughout the entire pre-sales lifecycle; The public domain data and the private domain data are preprocessed and feature extracted to construct a public domain feature set and a private domain feature set; By using a cross-domain attention fusion mechanism, the cross-domain attention weights are calculated using the public domain feature set as the query vector Q, the private domain feature set as the key vector K and the value vector V, and a fused feature set is generated. Customer intent rating is performed based on the fused feature set.

2. The customer intention rating method based on cross-domain attention fusion according to claim 1, characterized in that, The public domain data and the private domain data are preprocessed, specifically by filling missing numerical values ​​with 0, filling missing categorical values ​​with the string "NULL", performing equal-frequency binning on continuous value features, and performing one-hot encoding on categorical features. Alternatively, feature extraction can be performed on the public domain data and the private domain data, specifically by deriving new features based on business rules.

3. The customer intention rating method based on cross-domain attention fusion according to claim 2, characterized in that, After feature extraction, feature filtering is also required. Specifically, the information gain of each feature with respect to the target variable or the feature importance based on LightGBM is calculated, and features are filtered according to their importance, removing redundant features with an importance lower than a preset threshold.

4. The customer intention rating method based on cross-domain attention fusion according to claim 1, characterized in that, The cross-domain attention fusion mechanism includes: mapping the public domain feature set to the query vector Q through a fully connected layer, mapping the private domain feature set to the key vector K and value vector V through two independent fully connected layers respectively, and calculating the cross-domain attention weights using a formula. ， The softmax function is a normalization exponential function used to transform the input vector into a probability distribution, d. k Let K be the dimension, K T Let K be the transpose of the key vector; the attention output is concatenated with the original public and private features to obtain the fused feature set.

5. The customer intention rating method based on cross-domain attention fusion according to claim 1, characterized in that, After generating the fused feature set, the entire pre-sales lifecycle is divided into multiple stages. A rating sub-model is constructed for each stage, and the fused feature set is used to train the sub-models for each stage. The comprehensive rating score is then calculated using the stage transition probability matrix P.

6. The customer intent rating method based on cross-domain attention fusion according to claim 5, characterized in that, The pre-sales lifecycle is divided into: undistributed stage, unallocated stage, uncalled stage, follow-up stage, overdue follow-up stage, long-term lack of follow-up stage, and churn stage. The rating sub-model is a binary classification sub-model based on LightGBM; The stage transition probability matrix P has a dimension of 7×7, where the element P(i,j) represents the probability of transitioning from stage i to stage j, which is obtained based on historical data statistics. The formula for calculating the comprehensive rating score is: R=ΣP(i,j)×W j ×R j W j Let R be the weight coefficient for stage j. j denoted as the rating score of the sub-model in stage j.

7. The customer intention rating method based on cross-domain attention fusion according to claim 5, characterized in that, Based on the event-driven mechanism, user behavior events are monitored. When a preset trigger event is detected, the historical rating score is decayed by combining a time decay function and the current rating is recalculated. The time decay function is: , of which S original The historical rating score is represented by λ, the decay coefficient is represented by Δt, and the time interval since the last rating is represented by Δt. When Δt exceeds the preset threshold, the rating score for the corresponding stage will be automatically reduced by one level.

8. The customer intent rating method based on cross-domain attention fusion according to any one of claims 1-7, characterized in that, A multi-task portrait label generation network is used to synchronously output a portrait label set based on the fused feature set.

9. A device for customer intention rating based on cross-domain attention fusion, characterized in that, include: The data acquisition module is used to collect public and private domain data throughout the entire pre-sales lifecycle. The feature extraction module is used to preprocess and extract features from the public domain data and the private domain data; The cross-domain attention fusion module is used to calculate cross-domain attention weights using the public domain feature set as the query vector Q and the private domain feature set as the key-value vector K and value vector V. The multi-stage rating module is used to build sub-models for each stage and calculate the comprehensive rating score through the stage transition probability matrix P. The event-driven update module is used to listen for behavioral events and trigger a recalculation of rating scores. The image tag generation module is used to output image tag sets through a multi-task image tag generation network.

10. A storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements the customer intent rating method based on cross-domain attention fusion as described in any one of claims 1-8.

Images