Essays on next best action in digital marketing using reinforcement learning

Guide(s)

Kumar, U Dinesh

Department

Decision Sciences

Area

Decision Sciences

University

Indian Institute of Management Bangalore

Place

Bangalore

Publication Date

3-31-2021

Year Awarded

March 2021

Year Completed

March 2021

Year Registered

June 2015

Abstract

Companies make numerous sequential marketing decisions to optimise varied economic and expansion goals. With deep penetration and growing acceptance of digital channels, many companies use digital interventions to accomplish these goals. The use of digital channels is further reinvigorated by the ease of data collection at various touch points, inexpensive storage of that data and improved analytical know-how to derive insights from this data. Companies realise that digital data can be used to design and fine-tune marketing strategies and determine the optimal marketing mix. This has drawn researchers and practitioners to take keen interest in determining the ‘best’ marketing action for every customer at different touch-points through the purchase journey. This is known as Next-Best Action (NBA) modelling, which forms the basis for this thesis. From the methodological standpoint, the thesis utilises reinforcement learning technique of Multi Armed Bandits (MAB) in Essay 1 and Essay 2 and Bagging and Boosting meta-algorithms in Essay 3 for designing the models. In this dissertation, we study three independent digital marketing utilisations of customer- level data based on appropriate NBA-based analytical models. The first essay designs an attributebased MAB model to determine the next best content action for a B2B customer. The results show that action features such as placement and format, individual user-level features such as last touch channel and firmographic features such as size of the customer firm have a significant impact on the choice of marketing action. The second essay utilises NBA modelling to predict forgetfulness, particularly in the online grocery use-case. The proposed model is a two-stage personalisation mechanism, which uses the likelihood of purchase of an item and a Bayesian bandit model to determine the list of forgotten items as the next- best recommendations.

Pagination

185p.

Copyright

Indian Institute of Management Bangalore

Document Type

Dissertation

DAC Chairperson

Kumar, U Dinesh

DAC Members

Venkatagiri, Shankar; Jonnalagedda, Sreelata

Type of Degree

Ph.D.

Relation

DIS-IIMB-FPM-P21-19

This document is currently not available here.

Request Access

Share

COinS