Analyzing imbalanced online consumer review data in product design using geometric semantic genetic programming. (October 2021)

• Record Type: Journal Article
• Authors: Chan, Kit Yan; Kwong, C.K.; Jiang, Huimin
• Is Part Of: Engineering applications of artificial intelligence. Volume 105(2021)
• Publication Date: 2021-10
• Subjects: New product development -- Social media -- Online customer reviews -- Imbalanced data mining -- Multi-objective optimization -- Genetic programming; Engineering -- Data processing -- Periodicals; Artificial intelligence -- Periodicals; Expert systems (Computer science) -- Periodicals; Ingénierie -- Informatique -- Périodiques; Intelligence artificielle -- Périodiques; Systèmes experts (Informatique) -- Périodiques; Artificial intelligence; Engineering -- Data processing; Expert systems (Computer science); Periodicals; 620.00285
• DOI: 10.1016/j.engappai.2021.104442
• Languages: English
• ISSNs: 0952-1976

Abstract

Abstract: To develop a successful product, understanding the relationship between customer satisfaction (CS) and design attributes of a new product is essential. Nowadays IoT technologies are used to collect online review data from social media. More representative CS models are developed using online review data. However, online review data is imbalanced, since popular products receive more online consumer reviews and unpopular products receive less. When imbalanced data is used, CS models learn the characteristics of majority data while rarely learning minority data. Misleading analysis for product development is made since the CS model is biased to popular products. This paper proposes an approach to generate nondominated CS models which learn equally to imbalanced data from popular and unpopular products. A multi-objective optimization problem is formulated to learn equally in imbalanced data. This problem is proposed to be solved by the geometric semantic genetic programming (GSGP); a Pareto set of nondominated CS models is generated by the GSGP. Product designers select the most preferred models in the Pareto set. The preferred nondominated CS model attempts to tradeoff unpopular and popular products, to determine optimal design attributes and maximize the CS. The case study shows that the proposed GSGP is able to generate CS models with more accurate CS predictions compared to the commonly used methods. The proposed GSGP also generates a Pareto set of nondominated CS models which equally learn consumer reviews for those dryers. Based on the Pareto set, the design team selects the most preferred CS model. Highlights: Customer satisfaction (CS) models are developed by online review data. Data is imbalanced. Unpopular products have fewer; Popular products have more. Multi-objective genetic programming is used to tackle non-balanced data. Product designers select the most preferred model among all nondominance CS models. Better prediction can be achieved by the proposed nondominance CS models.