TL;DR
This paper introduces an intelligent method, Improved Cost-Sensitive Learning Support Vector Machine (ICS-SVM), to efficiently identify "lead users" in customer collaborative product innovation (CCPI). By considering the varying costs of misclassifying users, this approach aims to improve the accuracy and effectiveness of identifying these valuable contributors.
What Problem Does This Paper Address?
Customer Collaborative Product Innovation (CCPI) relies heavily on "lead users" – individuals with advanced needs and product knowledge who can drive novel product development. However, accurately identifying these lead users within a large user base is challenging. Traditional methods are often subjective, time-consuming, and struggle with imbalanced datasets where lead users are a minority. This paper addresses the need for a more efficient, objective, and accurate method to pinpoint these crucial innovators.
Methods
The core of this paper is the development of an Improved Cost-Sensitive Learning Support Vector Machine (ICS-SVM). Traditional Support Vector Machines (SVMs) treat all misclassifications equally. This paper recognizes that misclassifying a lead user as an ordinary user incurs a much higher "opportunity cost" than the reverse. To address this, they integrate cost-sensitive learning into the SVM framework. This means assigning different penalties for different types of misclassifications.
The method involves:
- Feature Engineering: Identifying and quantifying key characteristics of lead users across three categories: user general attributes (e.g., age, income), user activity attributes (e.g., recency, frequency of online participation), and user knowledge attributes (e.g., R&D knowledge, functional requirements understanding).
- ICS-SVM Model: Adapting the standard SVM by incorporating misclassification costs into the objective function. This allows the model to prioritize minimizing the cost of misclassifying lead users.
- PSO Optimization: Using a Particle Swarm Optimization (PSO) algorithm to tune the parameters of the ICS-SVM model, including the kernel function parameters and the penalty factor, to further enhance performance.
Datasets
The study utilizes a dataset from a real-world Chinese company focused on intelligent electronic products. This dataset comprises 96 users, with 13 identified as lead users and 83 as ordinary users. The data includes the quantified factors for lead user identification and the ground truth classification for each user. This dataset was randomly split into training (80 users) and testing (16 users) sets.
Key Insights and Ideas
- The Asymmetry of Misclassification Costs: The most significant insight is the explicit recognition and mathematical modeling of the unequal costs associated with misclassifying lead users versus ordinary users. This is a crucial step beyond standard classification.
- Comprehensive Feature Set: The paper proposes a well-defined set of 21 factors spanning general, activity, and knowledge attributes, providing a robust basis for lead user identification.
- Intelligent Methods for User Innovation: The study reinforces the trend of leveraging machine learning for complex user-centric tasks, moving away from purely manual or subjective methods.
- PSO for Parameter Tuning: Employing PSO for optimizing the ICS-SVM parameters demonstrates a practical approach to fine-tuning complex models for better performance.
Key Results
The ICS-SVM method demonstrated strong performance in identifying lead users. In a comparison against a standard BP neural network and a regular RS-SVM, the ICS-SVM achieved:
- Overall Correct Prediction Rate (Z): 93.75% (same as RS-SVM, better than BP).
- Weighted Overall Correct Prediction Rate (WZ): 97.56% (significantly higher than BP and RS-SVM).
- First Type Error Rate (FZ1): 0% (perfectly identifying lead users as lead users).
- Second Type Error Rate (FZ2): 9.09% (identifying ordinary users as lead users).
- Average Cost (EC): 6.25% (substantially lower than BP and RS-SVM), reflecting the effective minimization of misclassification costs.
Limitations and Future Work
- Limitations: The paper acknowledges that the misclassification costs were set based on expert suggestions and the ratio of lead users to ordinary users. While this is a practical approach, these costs might vary across different industries and contexts. The study was conducted on a single company's data, limiting its generalizability.
- Future Work: The authors propose extending this research to cross-industry comparisons to enhance universality. They also plan to investigate the ethical aspects of lead user identification, considering procedural and distributed fairness. Furthermore, they aim to apply the ICS-SVM method in big data contexts to test and improve its scalability and universality.
Summary updated