Login
Talks & Seminars
Title: Demand Forecasting Using Trees, Neural Nets, and Probabilistic LSTMs
Dr. Srayanta Mukherjee, Flipkart
Date & Time: July 26, 2017 11:00
Venue: Conference Room, C Block, 01st Floor, Department of Computer Science and Engineering, Kanwal Rekhi (KReSIT) Building
Abstract:
In an e-commerce operations management framework, coordination of the various arms of the supply chain via meticulous planning is critical to the health of the supply chain. In the collaborative planning process, unless accurate forecasts are available at an SKU level, the buying and replenishment of SKUs is chaotic, leading to lost opportunity in the form of Out of Stocks, poor inventory health and poor regional utilization, thus crippling the supply-chain process. While at lower granularities, simple time-series or ARIMA models are often sufficient for a good projection, at the SKU level the only recourse is to implement complex non-parametric machine learning approaches. The challenges that come up are many including sparsity, ambiguity and unavailability of data, inherent ad hoc ness of existing processes and infrastructure to name a few. Recently, we developed two forecasting methodologies, a rule-based decision tree model and a deep neural network algorithm. The decision tree approach combines ensembling and boosting approaches while the neural network is a probabilistic LSTM aided by a causal layer driven by a feed-forward algorithm. Both methods perform admirably, with the neural network outperforming the decision tree model. The performance is consistent across time, event and business as usual weeks, as well as over various business units.
Speaker Profile:
My background is mostly in Computational Biophysics working on protein structure prediction, protein-protein interactions and brownian dynamics simulations. I did my PhD in the University of Kansas, though most of it was actually spent at the University of Michigan. Following my PhD, I did 2 postdocs, 1 at the Univ. of Kansas, Med Center working on single-particle reconstruction of Cryo-electron microscopy images and the second at the Georgia working on stochastic simulations of large molecular machines. After a short stint at a pharmaceutical company, Ipsen, in Boston, I returned to India just over a year back. Have been working at Flipkart since May 2016, and lead the Planning-research team of Flipkart. Our team mostly works on demand forecasting of existing and new products, automated design of inventory selection and assortment planning for Flipkart.
List of Talks

Webmail

Username:
Password:
Faculty CSE IT
Forgot Password
    [+] Sitemap     Feedback