Prediction of consumer preferences by artificial neural networks method: An application in the retail sector

In view of the rapidly developing world economy, increasing global cooperation, easy access to information by consumers thanks to technological developments and constantly changing consumer wishes, data mining in the retail sector and marketing has become more today. In order to produce products or services that can meet the expectations of consumers by understanding their demands, the data obtained from many channels such as shopping records in retail stores, actions on social media, online shopping records, blog posts, online complaint channels are processed and valuable information is obtained and this information is taken into consideration during production. Store managers also benefit from this information in decision-making processes. This study aims to put forward a behavioral model by examining retail store customers' past product or service purchase behaviors and predict the products or services that customers can purchase in the future with the help of this model. Via the formation of prediction model, consumers with similar behavior were grouped with clustering analysis. The study consists of five chapters. The first four chapters are comprised of a comprehensive literature search and the last chapter presents a field research. In the literature section, studies on the concept of consumer, consumer purchase behaviors, retailing concept, consumer behaviors in retail sector and data mining process are examined to set a theoretical background. In the fifth part of the study, the application was carried out using quantitative research techniques such as association rules, clustering and prediction model. After the data was processed, clustering analysis was performed and the consumers with similar behavior were grouped. In addition, Market Basket Analysis was performed with the help of Apriori Algorithm and association rules were established to determine the relationship between the products purchased. The established association rules were transformed into an prediction model with the help of artificial neural networks, and finally, a comparison was made by Logistic Regression Analysis to measure the effectiveness of artificial neural networks. Basic analyzes (frequency distribution, data conversion, transpose, etc.), clustering analysis and logistic regression analysis were performed with the help of SPSS 22 and Visual Basic. The association rule and prediction model were established with the help of MATLAB program. In this study, quantitative research method was used. The population of the study consists of the consumers aged over 18 who shop from a retail store in Kütahya. The data was collected from the 2016, 2017 and 2019 shopping records of the consumers who are in the database of the retail shop. In the study, 26.543 shopping records which belong to the 489 costumers were used. The shopping consists of 45.650 purchased products, 40 product groups, and 3.982 product records. At the end of this study, 103 association rules were created, and 26 association rules which have over 90% confidence value and 4 demographic variables were used as input parameter in the prediction model. As there were 26 association rules to be predicted in total, the output layer was determined as 26. The prediction was conducted with the artificial neural networks and the logistic regression methods, and the results were compared. While the artificial neural networks had 99, 9621% correct prediction result with the maximum of 15 iterations, the logistic regression method had 89,830% correct prediction result with the maximum of 20 iterations. It turns out from the analysis results that the artificial neural networks method gave higher level of correct prediction results when compared to logistic regression method. In the cluster analysis, 10 clusters were obtained. Since the number of the product groups were high, the quality of clustering were low. The distribution of data into clusters is balanced, and the differences between the clusters has high level of importance. In the clusters which were composed of five variables (gender, age, the month of the shopping, the time of the shopping, and the product groups), the level of importance of the product groups within a cluster was obtained as 86%, while the other variables had 100% level of importance.