Enabling the smart basket for big basket

Abstract

BigBasket.com is India’s first comprehensive online grocery store. They currently have operations in the 7 Indian cities of Bangalore, Mumbai, Delhi, Mysore, Chennai, Hyderabad and Pune. The objective of the project was to address the following business problem: ? To improve the accuracy of predictions of the Smart basket recommendatory system A team under the guidance of Prof. Dinesh Kumar had already undertaken this project and applied multiple association algorithms such as ranking analysis, Jaccardian coefficient analysis and Apriori analysis. After reviewing their model and considering different approaches, the technique used for improving accuracy of predictions during the course of this project were Logistic Regression and Clustering Analysis. The association techniques used to recommend a smart basket do not take into account the frequency of purchase behaviour of customers. A customer who purchases a certain item would not purchase the same item again until he has consumed his supplies of the same item. To take into account this frequency impact, this project has used logistic regression models to predict whether a customer would buy a particular item based on his past purchase behaviour. The variables used for prediction include average frequency of purchase of an item and the time since last purchase. Hence the frequency impact is taken into account. 90 logistic regression models were built to predict the purchase decision of customers for 90 different categories. The results show that the accuracy defined as the percentage of items correctly predicted in a customer’s basket is pretty high with 100% accuracy for 46% of the orders and between 80%-100% for 33% of the orders. This was followed by the clustering analysis which was used to find out categories which are frequently purchased together. The K-means clustering approach was used and value of k was varied from 5 to 35. 9 suitable clusters were identified using k value of 30 and outlier clusters were ignored. The next steps include updating the analysis using more recent data. Also, the logistic regression analysis predicts high number of categories in a basket for certain products. The categories need to be ranked for each basket and only the top ranking categories should be suggested to the customer. Evolved techniques such as Markov chain optimization may be used to further refine the results