Use of neural networks in customer satisfaction index model

Abstract

The concept of customer satisfaction has attracted much attention in recent years. A key motivation for the growing emphasis on customer satisfaction is that higher satisfaction can lead to a stronger competitive position resulting in higher market share and profit, reduce price elasticity, lower business cost, reduce failure cost, and reduce the cost of attracting new customers. The purpose of customer satisfaction index (CSI) models is to measure the quality of the goods and services as experienced by the customers that consume them. The independent and uniform measurement characteristics of the CSI model provide a useful tool for tracking performance and systematic benchmarking over time. A major advantage of the measurement model is the use of generic questions, which are sufficiently flexible to be used across a wide variety of products and services. In this study, a new customer satisfaction index model is developed considering the previous CSI models such as American Customer Satisfaction Index and European Customer Satisfaction Index. The proposed model was applied for Turkish mobile phone sector since the competition in this industry results a dynamic product development and an increasing demand for that products. CSI models are designed as a structural equation model (SEM) which consists of well established theories and approaches in customer behaviour. The constructs of the CSI models are latent variables indirectly described by a block of measurement variables. The structural model of the proposed CSI consists of 6 latent variables with their 23 observable variables. The latent variables of the model are company image, customer expectations, perceived quality, perceived value, customer satisfaction and customer loyalty. A survey instrument, developed to measure the manifest variables, was conducted to 700 mobile phone users. Besides the model questions, some demographic questions (e.g. age, gender, education level etc.) are also included in the survey. The structural model of the present model is analyzed using variance based Partial Least Squares (PLS) method. The main concern of the PLS is related to the explanatory power of the path model along with the significance level of standardized regression weights. An iterative scheme of simple and/or multiple regressions contingent on the particular model is performed until a solution converges on a set of weights. The general applicability of a SEM model depends on the reliability and validity of the modelling results. Reliability and validity of the proposed CSI model were assessed by checking unidimensionalty of the blocks, individual item reliability, convergent validity and discriminant validity. All test results satisfy the crucial requirements for validity and reliability of structural model. In this study, a feed forward neural network model is proposed as an alternative to simple or multiple regression methods for the inner model estimation of the CSI. The use of artificial neural networks (NN) gained popularity in different fields, and some studies have demonstrated the superiority of NN over multiple regression. NN simulates human cognition by modelling the inherent parallelism of neural circuits in the brain using mathematical models of how the circuits function. However, the NN approach has been applied more recently to customer satisfaction and loyalty analysis. In this study there are 5 different inner models estimating 5 different endogenous latent variables of the CSI model. Each inner model was estimated using three-layer feed forward neural networks. A sigmoid function is used in the hidden layer, and a linear function in the output layer. The data was divided into two sets, 75 percent for training, and 25 percent for testing. Training is performed using the Levenberg Marquardt back propagation algorithm, and the weights are initialized using Nguyen Widrow algorithm. After the NN is trained it was evaluated for the test data. Root mean squared error and R square scores were used as performance criteria. The use of NN provides a powerful estimation for the inner models used in the CSI. The results of the CSI model as a whole can be a valuable guide for the managers in formulating competitive marketing strategies. Considering the results of the CSI model, the limited resources of the firms can be allocated for critical factors which have important impacts on satisfaction. In conclusion, the CSI model provides important information for the purchase decisions of the customers and lead to improvements in the quality of goods and services they consume.

Özet

Müşteri memnuniyet indeks modelleri son yıllarda birçok ülkede yaygın olarak uygulanmaktadır. Müşteri memnuniyet indekslerinin en büyük özelliği yapısındaki ölçüm faktörlerinin ürün ve hizmet sektöründe rahatlıkla kullanılabilmesine imkan sağlamasıdır. Bu sayede güvenilir bir ölçekle firmalar arasında, sektörler arasında ve ülke çapında karşılaştırmalı bir memnuniyet ölçümü mümkün olmaktadır. Müşteri memnuniyet indeks modelleri, müşteri memnuniyeti ile ilişkili gizli (latent) değişkenler ve bu gizli değişkenleri ölçen ölçüm değişkenleri arasındaki ilişkilerden oluşan yapısal eşitlik modelleridir. Bu çalışmada diğer ülkelerde kullanılan müşteri memnuniyet indeks modellerinden yola çıkarak ülkemiz şartlarında kullanılabilecek bir müşteri memnuniyet indeks modeli geliştirilip test edilmiştir. Model, 6 gizli değişken ve bunlara bağlı toplam 23 ölçüm değişkeninden oluşmaktadır. Modele uygun olarak hazırlanan anket formu kullanılarak, cep telefonu sektöründe 700 kullanıcı ile yüz-yüze anket uygulaması yapılmıştır. Müşteri memnuniyet indeks modelindeki gizli değişkenler ve ölçüm değişkenleri arasındaki ilişkiler kısmi en küçük kareler yöntemiyle tahmin edilmiştir. Müşteri memnuniyet indeks modelinin güvenilirlik ve geçerlilik test sonuçları modelin genel uygulanabilirliğini göstermektedir. Çalışmada ayrıca, gizli değişkenler arası ilişkilerin oluşturduğu iç modellerin tahmininde Yapay Sinir Ağları (YSA) metodu önerilmiştir. Yapısal modeller için tasarlanan YSA modeli 3 katmanlı ileri beslemeli ve geri yayılımlı bir modeldir. Gizli katmanda sigmoid transfer fonksiyonu, çıkış katmanında ise doğrusal transfer fonksiyonu kullanılmıştır. Bu şekilde 5 farklı iç modelin tahmini için yapay sinir ağları metodunun kullanımı modellerin açıklayıcılık gücünü artırmıştır.