Decision influence and proactive sale support in a chain of convenience stores

This paper presents a work in progress and initial design of a recommender system (RS) for active sale support within a large network of brick and mortar (or convenience) stores. There have been two datasets of historical transactional data provided for the pilot experiments. Each store consists of two kinds of shops; i.e., retail and cafeteria. Although these datasets contain various information about transactions, at our first experiment, they contain just a few information leading to customer identification and thus neither collaborative filtering nor content based techniques can be applied. Therefore, item co-occurrence approach and Naïve Bayes principle are chosen in order to build initial recommendation models with first promising results. Furthermore, discussions and solutions related to many real problems such as data sparsity, embedding of available features into recommendation models, benefits of item categorization, offline evaluation of proposed approaches over historical data, scalability and future personalization are presented in the work. Provided datasets are from real production and have larger sizes and required pre-processing and data transformation for efficient data manipulation and analysis. Various statistics and characteristics of transactional data are provided for practical view when working with similar kind of data, which can be interesting and useful for readers with similar research interests.