Sales recommender system

Abstract

The salespeople of a local frozen foods supplier have the problem of determining which products will be required by the customers they will be visiting on a particular day so that they can load their van accordingly. Their decisions are based on past experience and what they think will sell that day. We propose a Sales Recommender System which can predict which items will sell based on past sales data of similar days and determine if the predictions are a good recommendation, that is, if the total products which are predicted to be sold constitute a good day of sales. In this dissertation, we investigate whether such a system is possible using the sales data of the frozen foods supplier. Traditional recommender systems make recommendations based on either the prior selection of items by a user or on the selections made by similar users. This is not possible in our case due to the type of problem and data that we have. Although the data contains all information about the transactions, we lack other information such as the size of the customer. We, therefore, take a novel approach. In our data, we include environmental factors, such as each day's temperature and the week number, to investigate if and how these affect sales. Moreover, we add a predictive component to the system which employs a statistical predictive modelling technique. We also describe a class of NP-hard problems known as the Knapsack Problems (KP). We draw parallels between them and our problem by modelling our problem as a variant of KP. There are heuristics, which use dynamic programming techniques and run in pseudo-polynomial time, to solve KP. However, there are a number of differences between the two, so we do not apply any of these solutions. The predictive model that the system builds from the given data set is known as a generalized linear model ( GLM). The system uses it to predict the quantities of each product which should be sold to a particular customer on a particular day. It then gives a recommendation about the predictions made. To evaluate the system we divided the data into a training set to build models and a testing set to evaluate the predictions and recommendations that the system gives for these models. We describe the results given for a total of four different models. The two models with a Poisson distribution were discarded since their prediction are less than the actual quantities sold. However, the other models predict that the salespeople could have sold more than they did. The results give the predicted quantities for all products which should have been sold and show that almost all predictions make good recommendations. We, thus, conclude that the developed Sales Recommender System could be of benefit to the supplier's salespeople to help them load the correct quantities of each item depending on the day and the customers they will be visiting. Indeed, the project was a success as we developed the system successfully and showed that it can be of use.