Facebook x DSSG: Attributing Metric Variations with Statistical Models

Agenda:
7:00pm - 7.20pm: Speaker presents
7:20pm - end: Q&A

To protect users, Facebook hires human reviewers to enforce content guidelines when machine learning systems are less confident in their decisions. Sauvik's team examines human review performance via multiple top-line metrics. One such metric is related to efficiency i.e. how fast a human reviewer is in reviewing content. These metrics may vary over time.

The core question: why is variance observed in these metrics? A non-trivial dip in such a metric could potentially indicate lower productivity and that poses risk. In other words, from a data perspective--what factors can be attribute to these metric movements to so as to 'fix' them individually?

Sauvik's talk will delve into a recipe for a good model that reliably informs the business on what to focus on. These include constituent steps like: data cleaning, feature selection, model diagnostics, and model evaluation, alongside alternative approaches.

Join Details:
https://us02web.zoom.us/j/88011410184?pwd=SkFRVzNmTk5HQTNwQVMzUkltamd5QT09
Meeting ID: 880 1141 0184
Passcode: ek@yhJNR3M

Speaker Bio:
Sauvik Dutta is a Data Science & Technical Analytics Lead for Facebook's Integrity, Product and Global Operations team in Singapore. To keep the Facebook ecosystem clean, Dutta works at the intersection of Product, Operations and Trust & Safety, all the while focusing on the integrity of the platform as an overarching theme. He is passionate about using data to solve complex business problems, and continues to be excited to continue to learn in this ever-evolving space.