Access Type

Open Access Dissertation

Date of Award

January 2012

Degree Type

Dissertation

Degree Name

Ph.D.

Department

Industrial and Manufacturing Engineering

First Advisor

Ratna B. Chinnam

Abstract

Cross-selling and product bundling are prevalent strategies in the retail sector. Instead of static bundling offers, i.e. giving the same offer to everyone, personalized dynamic cross-selling generates targeted bundle offers and can help maximize revenues and profits. In resolving the two basic problems of dynamic cross-selling, which involves selecting the right complementary products and optimizing the discount, the issue of computational complexity becomes central as the customer base and length of the product list grows. Traditional recommender systems are built upon simple collaborative filtering techniques, which exploit the informational cues gained from users in the form of product ratings and rating differences across users. The retail setting differs in that there are only records of transactions (in period X, customer Y purchased product Z). Instead of a range of explicit rating scores, transactions form binary datasets; 1-purchased and 0-not-purchased. This makes it a one-class collaborative filtering (OCCF) problem. Notwithstanding the existence of wider application domains of such an OCCF problem, very little work has been done in the retail setting. This research addresses this gap by developing an effective framework for dynamic cross-selling for online retailing.

In the first part of the research, we propose an effective yet intuitive approach to integrate temporal information regarding a product's lifecycle (i.e., the non-stationary nature of the sales history) in the form of a weight component into latent-factor-based OCCF models, improving the quality of personalized product recommendations. To improve the scalability of large product catalogs with transaction sparsity typical in online retailing, the approach relies on product catalog hierarchy and segments (rather than individual SKUs) for collaborative filtering. In the second part of the work, we propose effective bundle discount policies, which estimate a specific customer's interest in potential cross-selling products (identified using the proposed OCCF methods) and calibrate the discount to strike an effective balance between the probability of the offer acceptance and the size of the discount. We also developed a highly effective simulation platform for generation of e-retailer transactions under various settings and test and validate the proposed methods.

To the best of our knowledge, this is the first study to address the topic of real-time personalized dynamic cross-selling with discounting. The proposed techniques are applicable to cross-selling, up-selling, and personalized and targeted selling within the e-retail business domain. Through extensive analysis of various market scenario setups, we also provide a number of managerial insights on the performance of cross-selling strategies.

Download

Share

COinS