Our conversation today is with Luke Guerdan, PhD student at Carnegie Mellon’s Human Computer Interaction Institute. Luke’s work examines the safety and validity of data-driven algorithms deployed in high-stakes decision-making settings. During our discussion, Luke shares key insights from his recent research on the topic of data label correctness. We unpack the suprisingly numerous, and often subtle ways that incorrect or inappropriate data labels can undermine machine learning initiatives in the real world.

Links: 
 - https://lukeguerdan.com
 - https://twitter.com/lukeguerdan

Timestamps:
(00:00:00) Introduction
(00:02:03) Excitement over imperfect labels in data science.
(00:04:44) Racially biased algorithm failed to identify medical need.
(00:06:49) Identifying challenges in evaluating predictive models.
(00:10:11) Importance of assessing issue and model performance.
(00:15:51) Customer success interventions impact customer retention predictions.
(00:18:26) Difficulty quantifying differences between customer satisfaction and prediction. Importance of understanding intervention impacts on outcomes.
(00:25:57) Organizational issues in data science projects. People and data literacy matter.
(00:30:54) Considered customer cohorts, varied predictions, human dynamics.
(00:34:49) Avoid complexity, bring assumptions to light.
(00:38:04) Issues in political and organizational communication hinder production. Lack of tools to address data problems.
(00:41:04) Subscribe, rate, contact us for feedback.

Tune in for more insights and do not forget to rate or review on your favorite platform!