Artificial intelligence sales performance incentive system based on causal inference and Uplift modeling

Abstract

The invention relates to an AI-driven performance excitation system based on causal inference and promotion modeling technology. The technical field of the invention comprises data science and artificial intelligence, and aims to solve the technical problem that a traditional excitation system is difficult to accurately evaluate the excitation effect. According to the method, the actual influence of different incentive measures on sales performance is determined by using the causal inference model, and then the response of an individual to different incentive intensities is predicted by using the promotion modeling technology, so that the incentive strategy is optimized. The system can be widely applied to various sales and marketing scenes so as to improve the performance management effect of an enterprise and the accuracy of an incentive scheme.

Description

TECHNICAL FIELD

[0001] The artificial intelligence sales incentive system based on causal inference and uplift modeling relates to the technical field of artificial intelligence performance incentive systems, aiming to improve the efficiency and accuracy of sales performance management. In particular, this system combines causal inference models and uplift modeling techniques to explore the impact of different incentive measures on sales, thereby achieving personalized and efficient incentive management. BACKGROUND

[0002] In current enterprise management, how to effectively motivate sales personnel to improve sales performance is a key challenge. Traditional methods mainly rely on statistical analysis and empirical rules, but these methods often lack the precise ability to adjust individually. Causal inference and uplift modeling, as emerging technologies, can identify factors that truly affect sales performance by analyzing historical data and experimental effects, and accurately predict the effects of different incentive strategies.

[0003] The current industrial sales incentive method has the advantage of being able to make fine adjustments according to the different characteristics and working environment of individual sales personnel, thereby improving the effectiveness of incentive measures. However, traditional methods often fail to fully utilize the potential of advanced data analysis and modeling, especially in handling large-scale data and complex variable relationships.

[0004] Causal inference, as an important topic in interdisciplinary research, has been widely applied in statistics, computer science, public policy, economics and other fields. In particular, in the absence of randomized controlled trials, estimating causal effects from observational data is an attractive research direction. With the rapid development of machine learning technology, various causal effect methods have emerged, greatly enriching the toolbox of causal inference. In the industry, causal inference is widely used in advertising, recommendation systems, medical health and other fields. These methods not only help enterprises more accurately evaluate and optimize marketing activity effects, but also provide decision-makers with precise data-based recommendations, thereby improving the scientific nature and efficiency of decision-making.

[0005] Uplift modeling is an important method for modeling the heterogeneity of treatment effects based on individual characteristics. Unlike traditional average effect models, uplift models can identify the response differences of different groups after receiving a specific treatment, thereby more accurately predicting and optimizing personalized marketing strategies. In academia and industry, uplift models have been widely applied in personalized medicine, marketing and social science research, providing decision-makers with effective decision support tools.

[0006] Combining causal inference and uplift modeling can make up for the shortcomings of traditional sales incentive methods, and improve the personalization and accuracy of incentive strategies. Causal inference helps identify factors that truly affect sales performance, while uplift modeling can predict the differences in the effects of different incentive strategies at the individual level. This combination not only helps enterprises optimize the performance management of sales teams, but also provides more targeted incentives for individual salespeople, maximizing sales efficiency. By combining the advantages of causal inference and uplift modeling, enterprises can provide more intelligent and data-driven sales management solutions to effectively cope with market competition and business challenges. SUMMARY

[0007] The summary is divided into two parts, one is the combination of causal discovery and uplift model, and the other is the implementation of the specific system.

[0008] 1. First, introduce the combination of optimized causal inference modeling and uplift model:

[0009] In this invention, the combination of causal discovery modeling and uplift model is to optimize the incentive strategy of sales personnel through advanced data science methods. The following expands this part in detail:

[0010] 1.1 Input data:

[0011] Feature data of each salesperson Among them represents the jth feature of salesperson i. These features include not only historical sales data and personal background information, but also the following aspects:

[0012] - Historical sales data: including sales, transaction frequency, average amount per transaction, etc.

[0013] - Contains salesperson's sales incentives and sales.

[0014] - Sales behavior characteristics: such as sales cycle length, average amount per transaction, sales category preference, etc.

[0015] - Personal background information: including age, gender, education background, work experience, etc.

[0016] The relevance between sales personnel is represented by the graph G=(V,E), where V is the set of sales personnel, and E is the edge set connecting different sales personnel. This graph reflects the social relationships, cooperation history and information transmission path within the sales team. For example, an edge (i,j)∈E can represent the cooperation relationship or information exchange between sales personnel i and j.

[0017] Prior correlation graph G{prior} Directional assumptions that play a key role in guiding model training and effectiveness evaluation. The construction of prior correlation graphs is based on the following sources:

[0018] - Domain knowledge G {domain} : Based on industry characteristics and product characteristics, industry experts use their understanding and experience of sales business to establish prior relationships.

[0019] - Historical data analysis G {historical} : Analyze past sales data to reveal the impact of different incentive strategies on sales performance trends, and establish relevant assumptions based on market response.

[0020] - Expert judgment G {expert} : Combine the experience of sales management experts and market trend forecasts to provide substantive support for the model's assumptions.

[0021] 1.2 Model

[0022] Step 1: Graph conversion and feature weighting

[0023] Use the graph neural network node2vec and GCN technology to convert the salesperson correlation graph G into a node embedding matrix, obtaining causal weights, and each node represents a salesperson's feature vector. These node embedding vectors capture the social network structure within the sales team and the relative weights of individual sales capabilities.

[0024] Feature weighting and fusion M2: Weight and fuse the feature vector of each salesperson with the learned causal weights to obtain the feature matrix M2. This step is shown as follows:

[0025] M1 = X • Causal Weights

[0026] Where Causal Weights is the weight learned from the prior correlation graph and causal inference model.

[0027] Step 2: Feature learning and causal modeling

[0028] Sales incentive correlation graph G incentive learning:

[0029] Learn the relationship between sales incentives and other features using GNN based on the prior correlation graph G prior and G. This includes:

[0030] - Establish a prior correlation graph G prior , describing the causal relationship between sales incentives and other salesperson features.

[0031] - Use GNN model to learn sales incentive correlation graph Gincentive , to reveal the impact of different incentive strategies on sales performance, resulting in M2. Sales-related graph G revenue Learning:

[0032] Similarly, using prior-related graph G prior and GNN to learn the relationship between sales and salesperson features. This includes:

[0033] - Establishing a prior-related graph G prior , describing the complex relationship between sales and salesperson features.

[0034] - Using GNN model, learning sales-related graph G revenue to understand the contribution of individual salesperson features to sales, resulting in M3.

[0035] Step 3: KDE image generation and CNN learning

[0036] In this step, we pair sales incentives X i and sales X j , using kernel density estimation KDE to generate density images. Then, using convolutional neural network (CNN) to learn these images, extracting feature vectors M2. The specific steps are as follows:

[0037] - KDE image generation: for each pair X i and X j , use KDE to generate its density image.

[0038] - CNN learning: use CNN to learn these density images, extract image features, and obtain feature vectors M4.

[0039] Step 4: Uplift Model modeling and optimization

[0040] Use causal inference models to calculate the Uplift value of each salesperson, which reflects the actual impact of incentive strategies on sales performance.

[0041]

[0042] where, and are the Uplift values of the intervention group and the control group, respectively.

[0043] Node embedding and GCN learning M3: through node embedding and graph convolution network (GCN), learn the final node embedding vector M3 of each salesperson, combining sales incentive and sales-related information.

[0044] Step 5: AUUC optimization

[0045] AUUC optimization: By optimizing the AUUC (Area Under the Uplift Curve), the most suitable incentive value for each salesperson is determined, thereby optimizing the implementation effect of the incentive strategy.

[0046]

[0047] |T|: The number of samples in the intervention group.

[0048] ∑: Summation symbol, indicating the summation of each sample in the intervention group.

[0049] i∈T: Indicates that i is in the intervention group T.

[0050] -y 1i : The response value (e.g., sales) of sample i after intervention.

[0051] -[t i =1]: Indicator function, 1 if sample t is in the intervention group; otherwise 0.

[0052] |C|: The number of samples in the control group.

[0053] j∈C: Indicates that j is in the control group C.

[0054] -y 0j : The response value (e.g., sales) of sample j without intervention.

[0055] -[t j =0]: Indicator function, 1 if sample j is in the control group; otherwise 0.

[0056] Finally, we determine the most suitable incentive value for each salesperson by optimizing the AUUC. AUUC is an important indicator for evaluating the effectiveness of different incentive strategies, and the optimized incentive strategy can significantly improve the performance of salespeople and the overall sales effect of the team.

[0057] 1.3 Summary

[0058] The invention combines advanced graph neural network technology and causal modeling methods, optimizing the incentive strategy for salespeople through deep learning and feature weighting. This innovative method not only improves the accuracy and efficiency of the incentive strategy, but also provides a more effective sales growth strategy for the sales team.

[0059] 2. System implementation

[0060] This system not only provides an intuitive interface for managers and salespeople, but also brings significant competitive advantage and economic benefits to the enterprise by combining advanced data analysis techniques and artificial intelligence algorithms. Managers can use real-time data and model analysis results provided by the system to accurately develop and dynamically adjust personalized incentive strategies. These strategies are not only based on sales targets and cost control, but also take into account the individual characteristics and performance of each salesperson, thereby improving the overall performance of the sales team.

[0061] Salespeople can use the system to understand their individual performance and team goals in real time, making it clearer to understand their contribution and growth space in the team. The system uses operational optimization algorithms to intelligently output the incentive method for each business based on sales data and participant input. These methods not only include completion rate and incentive coefficient in detail, but also can be dynamically adjusted according to real-time conditions to ensure the timeliness and effectiveness of the incentive strategy.

[0062] In summary, this comprehensive system not only improves the scientificity and efficiency of management decision-making, but also effectively promotes the collaboration and performance improvement of the sales team. By combining advanced data analysis and artificial intelligence technology, enterprises can better adapt to market changes and achieve business growth and long-term competitive advantage. BRIEF DESCRIPTION OF DRAWINGS

[0063] Figure 1 Causal Inference Modeling and Uplift Model Combination

[0064] From left to right, the matrix first contains the feature data of the salesperson Horizontal represents a salesperson, and vertical each column represents a feature. First, we will use node2vec and GCN technology to convert the salesperson association graph G into node embedding matrix to get causal weights. Combine the feature vector of each salesperson and Causal Weights to get M1. Use the prior correlation graph G prior and GNN to learn the relationship between sales incentives, sales and salesperson characteristics. Establish the prior correlation graph G prior , describing the causal relationship between sales incentives and other salesperson characteristics, use GNN model to learn the sales incentive correlation graph G incentive , to reveal the impact of different incentive strategies on sales performance, get M2; Similarly, establish the prior correlation graph G prior , describing the complex relationship between sales and salesperson characteristics, use GNN model to learn the sales correlation graph G revenue , to understand the contribution of individual salesperson characteristics to sales, get M3. At the same time, we also combine sales incentives X i and sales X jPairing, using kernel density estimation KDE to generate density images. Then, using convolutional neural network (CNN) to learn these images, extract feature vectors M4. Through the learning of max pooling and mlp, finally, combine the three vectors to get the sales incentive alignment sales uplift of each salesperson. By optimizing AUUC (Area Under the Uplift Curve), determine the most suitable incentive value for each salesperson, so as to optimize the implementation effect of incentive strategy.

[0065] Figure 2 System implementation (management perspective)

[0066] By inputting sales targets, cost controls, and participant numbers, the system can combine causal inference modeling, UpliftModel, and operational optimization algorithms to generate incentive policies for overall sales products. Management personnel can dynamically adjust incentive strategies based on the optimization suggestions provided by the system. This not only includes the achievement of global targets, but also enables the formulation of specific incentive measures for different products and sales regions, thereby enhancing overall performance.

[0067] Figure 3 System implementation (salesperson perspective)

[0068] Based on the historical performance and incentive elasticity of individual salespersons, the system can generate personalized incentive schemes. These schemes are derived from causal inference modeling, Uplift Model, and operational optimization results, ensuring that each salesperson receives targeted support. The schemes include target sales, recommended sales strategies, and incentive requirements for meeting targets, allowing salespersons to clearly understand their tasks and maximize their performance potential. DETAILED DESCRIPTION

[0069] The specific implementation of the present invention includes the following key steps, covering data collection and preprocessing, causal discovery modeling, Uplift modeling analysis, system implementation and optimization, as well as continuous iterative optimization and automated processing.

[0070] Step 1: Data collection and preprocessing

[0071] Collect a large amount of historical sales data and individual feature data from the sales data system, and perform preprocessing and cleaning to ensure data quality and integrity.

[0072] Data includes:

[0073] - Historical sales data: sales, transaction frequency, average amount per transaction, etc.

[0074] - Sales behavior characteristics: sales cycle length, sales category preference, etc.

[0075] Person background information: age, gender, education background, work experience, etc.

[0076] Step 2: Causal discovery modeling

[0077] Using pre-processed data, apply neural network architectures such as CNN towers and GNN towers to learn the distribution characteristics and structural rules between variables, respectively, to identify effective causal relationships.

[0078] Step 3: Uplift modeling analysis

[0079] According to the results of causal discovery, select appropriate intervention variables and use uplift modeling technology to analyze the impact of different incentive measures on sales performance, and generate personalized incentive programs.

[0080] Step 4: System implementation and optimization

[0081] Design and implement the software platform of the artificial intelligence performance incentive system, provide intuitive user interface and function modules, support the operation needs of managers and sales personnel, and continuously optimize the performance and functions of the system.

[0082] The present invention provides a new sales performance management solution for enterprises by combining advanced causal inference and uplift modeling technology, which can effectively improve the performance and work efficiency of the sales team, and realize more efficient incentive management and resource allocation.

[0083] Detailed implementation steps

[0084] 1. Data collection and preprocessing

[0085] Automatically collect a large amount of historical sales data and individual characteristic data from the sales data system, and perform preprocessing and cleaning to ensure data quality and integrity. This step includes data deduplication, outlier processing, data standardization, etc.

[0086] 2. Causal discovery modeling

[0087] Using pre-processed data, apply neural network architectures such as CNN towers and GNN towers to learn the distribution characteristics and structural rules between variables, respectively, to identify effective causal relationships. The specific modeling process uses the Pytorch framework for training, and is performed once a month to ensure that the model can adapt to the latest data and market changes.

[0088] 3. Uplift modeling analysis

[0089] Based on the results of causal discovery, appropriate intervention variables are selected, and uplift modeling techniques are used to analyze the impact of different incentives on sales performance, generating personalized incentive programs. Uplift models are also trained and optimized using the Pytorch framework.

[0090] 4. System implementation and optimization

[0091] The software platform for designing and implementing the artificial intelligence performance incentive system provides an intuitive user interface and functional modules to support the operational needs of managers and sales personnel, and continuously optimizes the performance and functionality of the system.

[0092] 5. Continuous iterative optimization and automated processing

[0093] After the system is implemented, MLOps (Machine Learning Operations) technology is used for continuous iterative optimization and automated processing, mainly including the following aspects:

[0094] - Continuous Integration and Continuous Deployment (CI / CD):

[0095] - Automated model building, testing, and deployment.

[0096] - After monthly training is completed, the model is automatically deployed to the production environment.

[0097] - Model monitoring and management:

[0098] - Real-time monitoring of model performance to detect data drift and changes in model effectiveness.

[0099] - Set threshold, when AUUC reaches 0.7 and above, consider the model reasonable and can be deployed to production environment.

[0100] - Automated data pipeline:

[0101] - Achieve automated data collection, preprocessing, and storage to ensure data timeliness and quality.

[0102] - Data pipeline includes data extraction, data cleaning, feature engineering, etc. to ensure data can be used for model training and optimization in a timely manner.

[0103] - Model feedback and update:

[0104] - Based on actual usage and feedback, continuously adjust and optimize the model.

[0105] - Train new data every month to maintain the accuracy and real-time performance of the model.

Claims

1. An artificial intelligence sales performance incentive system, characterized by: ● Combining causal inference models and uplift modeling techniques to optimize salesperson incentive strategies. ● Using graph neural networks and GCN technology to convert salesperson association graphs into node embedding matrices, and learning relationships between sales incentives and sales volume and salesperson characteristics based on prior correlation graphs and GNN models. ● Using kernel density estimation (KDE) to generate density images, and learning and extracting feature vectors through convolutional neural networks (CNN) for calculating the uplift value of each salesperson. ● Determining the most suitable incentive value for each salesperson by optimizing AUUC (Area Under the Uplift Curve) to improve the overall sales performance of the sales team.

2. The above-mentioned artificial intelligence sales performance incentive system, further characterized by: ● The system provides real-time data and model analysis results to help managers dynamically adjust personalized incentive strategies, taking into account individual characteristics and performance of salespeople, thereby optimizing the performance of the sales team.

3. A method for optimizing salesperson incentive strategies, comprising the following steps: ● Based on the characteristics of salespeople and sales behavior characteristics, construct salesperson association graphs and prior correlation graphs. ● Use graph neural networks and GCN technology to convert association graphs into node embedding matrices, and learn causal inference models and uplift models. ● According to the density image generated by KDE, extract feature vectors through CNN to calculate the uplift value of each salesperson. ● Determine the most suitable personalized incentive value by optimizing AUUC to optimize the implementation effect of salesperson incentive strategies.

4. A sales performance incentive system based on advanced data analysis techniques and artificial intelligence algorithms, comprising: ● The operation interface provides real-time data and model analysis results to support managers in developing and adjusting personalized incentive strategies. ● The system uses operational optimization algorithms to intelligently output the incentive method for each business based on sales data, dynamically adjusts the implementation effect of the incentive strategy, and improves the collaboration and performance improvement of the sales team.

5. A computer-readable storage medium for optimizing salesperson performance management, storing instructions that can cause a computer to perform the following steps: ● Receive the characteristics of salespeople and sales behavior characteristics. ● Construct salesperson association graphs and prior correlation graphs. ● Use graph neural networks and GCN technology to convert association graphs into node embedding matrices, and learn causal inference models and uplift models. ● According to the density image generated by KDE, extract feature vectors through CNN to calculate the uplift value of each salesperson. ● Determine the most suitable personalized incentive value by optimizing AUUC to optimize the performance management strategy of salespeople.

Images