Development of Classification of Customer Complaints Using Deep Learning

In recent years, the automotive industry has been making efforts to develop vehicles that satisfy customers’ emotions rather than malfunctions by improving the durability of vehicles. The durability and reliability of vehicles sold in the U.S. can be determined through the VDS (Vehicle Dependability Study) published by JD Power. The VDS is index which is the number of complaints per 100 units released by J.D. POWER in every year. It investigates customers who have used it for 3 years after purchasing a new car and consists of 177 specific problems grouped into 8 categories such as PT, ACEN, FCD, Exterior. The VDS-4 has been strengthened since the introduction of the new evaluation system VDS-5 in 2015. In order to improve the VDS index, it is important to gather various customer complaints such as internet data, warranty data, Enprecis data and clarify the problem and cause. Enprecis data is survey of customer complaints by on-line in terms of VDS. In the case of warranty and Enpreics data, it is easy to analyze because it is already categorized, but internet data is difficult to classify because it is unstructured data collected randomly at various internet sites, and the amount of data is big. In this paper, we developed classification technology for internet data using deep learning method such as TF-IDF and Word2Vec. This technology automatically classifies 8 categories and extract keywords of each category even if you don’t read the articles.