Customer satisfaction assessment using neural networks modelling.

Dan, Mihaela ; Vasilache, Simona ; Dima, Mihaela Alina 等

1. INTRODUCTION

The works of Schlenker (1980) and Schlenker and Weigold (1990) stress, in terms of social psychology, the importance of impression management. In their turn, Grayson and Schulman (2000) argued that services marketing is a combination of the qualities of the service provider and of the favorable outcomes of the situation, whose interaction makes a good impression of the service, in the eyes of the customer. As far as the service is concerned, the impression of the customer is based on some elements to which he/ she is sensitive. In this context, we may ask which are the dominant criteria, which have to be taken into account?

In order to deal with these problems, neural networks approaches were lately proposed, in customer satisfaction and loyalty studies (Clerfeuille et al., 2008). Comparative studies dedicated to the robustness of neural network approaches as compared with other types of quantitative approaches were conducted (Poubanne et al., 2006), and they proved the ability of the neural networks to deal with non-linear transformations, due to their redundant architecture, which simulates the associative structure of the brain.

We will use this approach, of the artificial neural networks, to assess the customer satisfaction of a sample of informed clients, veterinarians, in respect to the services offered by a veterinary pharmacy.

2. METHODOLOGY

We used a sample of 200 veterinarians, out of which 150 were previous customers of the considered pharmacy, and 50 are prospective customers. They were asked to rate, on a Likert scale from 1 to 5 (1 being the least satisfactory, 5 being the most satisfactory), the quality of the service they received, based on three parameters: cost, delivery, accessibility. We have also recorded their age, their location (urban/ rural), their professional status (senior or junior professional). The ones who were already customers of the considered pharmacy were asked on their intention to re-buy, whether it was present or absent.

We trained and tested a neural network, which we used in order to classify the prospective customers, in those who will re-buy, based on their satisfaction, and those who would not re-buy.

The main limitations of the research arise from the size of the sample, which was not thought, in this stage, as being representative for the entire population of veterinarians, country-wide, or for the population of veterinary pharmacies. Also, the three parameters we have selected sketch a model which is reductive, fit for statistical analysis purposes, and thus, some of the complexity of the real-life choices is inevitably lost.

Further research may properly address these limitations by enlarging the sample, according to criteria of statistical representativeness, and by opting for a stratified sample, by either years of practice, or practice in the public/ private sector, or average yearly earnings. The model can be further developed by including adjacent criteria of choice, in such a way that it adequately captures most of the real-life situation.

3. RESULTS

The results focus on the model which was developed by training a neural network to classify customers (i.e., veterinarians) in returning and not returning, and on discussing the two types of estimation errors and their effects for the customer relationships strategy of the respective veterinary pharmacy.

The model summary for the neural network is presented in Table 1:

The model is very accurate in predicting the re-buying intention of the customers, thus being useful in accurately predicting the buying intention of the prospective customers.

The classification of the customers is presented in Table 2.

All the people who wouldn't buy are classified correctly, which means that the model is better suited for classifying non-repeated buyers than repeated buyers. 98.8% of the training cases were classified correctly. In the holdout sample, the overall percent of the cases classified correctly is 93, which means that the model is accurate in approximately nine cases of ten.

The sensitivity and specificity of the model are shown in Figure 1.

[FIGURE 1 OMITTED]

The area under the curve, .982, indicates that, for a randomly selected repeated buyer and a randomly selected non-repeated buyer, there is a .982 probability that the probability to buy, predicted by the model, will be greater for the buyer than for the non-buyer, which is very accurate.

The predicted by observed chart is presented in Figure 2.

[FIGURE 2 OMITTED]

The symbols above the 0.5 mark stay for correct predictions, while the symbols below this mark are incorrect predictions. The model is better at predicting the customers who will buy again, only few cases being below the 0.5 limit.

The pharmacy should consider which is the cost of classifying a re-buying customer as a customer who is going to buy again (Type I error), and which is the cost of classifying a customer who is not going to buy again as a customer who is going to come back (Type II error).

If the pharmacy is interested in reducing Type I error, that is, keeping only those customers who are going to come again, then it can give up the top 40% of its clients, which include approximately 90% of the clients who won't buy again. Still, this would mean lowering the amount of clients with about half, which may not be a sound option for the pharmacy. If, on the contrary, it wants to increase its client base, but also reduce the probability that they won't buy again, it has to give up the first 20% customers, which contain approximately 40% of the non-buyers. This measure, consistent with better customer loyalty programs, will ensure a higher rate of re-buy.

4. CONCLUSIONS

The neural network model we developed is a decision-aiding tool for a veterinary pharmacy which wants to filter its coming back clients, and to reduce the number of the clients who won't buy again its services.

We have trained the neural network model taking the decision to buy again as the dependent variable, and the characteristics of the service, quality, fast delivery and accessibility as factors.

The age of the customers, in this particular case veterinarians, their clinics' localization, in the urban or rural regions, and their professional status, that of a senior or junior medical professional were considered as covariates.

The accuracy of the model is outstandingly good, as it predicts correctly about nine cases out of ten. Based on this model, the pharmacy can decide which clients it wants to keep, and which are the measures to be taken in order to increase the loyalty of the remaining, non-repeated buying customers.

5. REFERENCES

Schlenker B.R. (1980). Impression management: The self-concept, social identity, and interpersonal relations. Monterey: Brooks/Cole, pp 75-85

Schlenker, B.R., Weigold, M.F. (1990). Self-consciousness and self-presentation: Being autonomous versus appearing autonomous. Journal of Personality and Social Psychology 59, pp 820-828

Grayson, K., Shulman D. (2000). Impression Management in Services Marketing, Handbook of Services Marketing (Dawn Iacobucci and Teresa Swartz eds.), Thousand Oaks, CA: Sage, pp 51-67

Clerfeuille, F., Poubanne, Y., Vakrilova, M., Petrova, G. (2008). Evaluation of consumer satisfaction using the tetra-class model. Research in Social and Administrative Pharmacy, vol. 4, issue 3, pp 258 - 271

Poubanne, Y., Clerfeuille, F., Chandon, J.L. (2006). Variation within service categories and customer satisfaction: A segment-based approach using the tetra-class model, Journal of Targeting, Measurement and Analysis for Marketing, vol. 15, issue 1, pp 30-46

Tab. 1. Model summary

Training Cross Entropy Error 5.652

 Percent Incorrect 1.2%
 Predictions

 Stopping Rule Used Maximum number
 of epochs(100)
 exceeded

 Training Time 00:00:00.157

Holdout Percent Incorrect 7.0%
 Predictions

Dependent Variable: buy

Tab. 2. Classification of the cases

Sample Observed Predicted

 0 1 Percent
 Correct

Training 0 44 1 97.8%

 1 0 39 100.0%

 Overall 52.4% 47.6% 98.8%
 Percent

Holdout 0 42 3 93.3%

 1 3 38 92.7%

 Overall 52.3% 47.7% 93.0%
 Percent

Dependent Variable: buy