PAW Business Expert Round 4: AI for Sales & Service
Thursday, June 17, 2021
- Expert Round: Extracting Insights with Cluster Inspection Toolkit (Annemarie Paul)
Extracting meaningful insights from clustering of a given dataset is hard and laboursome. We propose a toolkit containing three techniques to derive insights from arbitrary clustering solutions and / or low-cardinality classifications faster:
1. Cluster plotting: Feature inputs are usually high dimensional. To represent the cluster solution in a low dimensional (~plottable) space, Uniform Manifold Approximation and Projection (UMAP) is employed.
2. Cluster assignment: Which features and what interactions drive the clustering? The cluster assignment can be reverse-engineered in an interpretable manner using gradient boosted trees and shaply values.
3. Feature inspection: The contribution of a specific feature on the global clustering space can be approximated by combining a low dimensional UMAP projection with a kernel ridge regression.
The python-based toolkit will be shared with the audience.
- How Deep Learning Helped to Classify Large-Scale B2B Marketing, Sales, and Web Traffic (Rohit Kewalramani & Justin Chien)
The account engagement platform at 6sense involves analysing traffic flow through customers’ website, marketing automation, and customer relationship tools. Prospect accounts visit certain webpages and interact with marketing / sales teams – all influencing their buying decision. Our algorithm standardises data from these sources and uses NLPs context and word sequence to predict categories that are part of buying intent models. We leverage ONNX that boosts inference by 10x (i.e. 8.5M URLS/day).
- Expert Round: Predicting Customer Attrition during COVID-19 – (Evelina Stoikou & Daniel Hellwig) As the world adjusts to the post-pandemic environment, the nature of customer attrition is changing. The pandemic has shortened the attrition cycle, altered the underlying causes, and changed the profile of customers who call in to cancel their service. At the same time, historic data if often not applicable anymore. Predictive analytics can be used to better understand customer attrition by tracking geo-specific indicators, segmenting customer groups in new ways, and creating new features.