The Fifth Elephant 2017

On data engineering and application of ML in diverse domains

Wait, I can explain this! (ML models explaining their predictions)

Submitted by Ramprakash R (@ramprakashr) on Monday, 22 May 2017

Section: Crisp talk for data engineering track Technical level: Intermediate Status: Confirmed & Scheduled


Today ML/AI is being used in mission critical applications. However, it is still difficult for a human being to trust a black-boxy ML algorithm. Wouldn’t it be cool if an algorithm could also explain why it had predicted a particular result and thereby strengthen it’s voice? That’s what exactly this talk is all about. Would walk you through how we implemented a model explainer for ZOHO’s ML suite and the story of pushing it into production.


  1. Motivation behind the problem statement
  2. Local & Interpretable explanations
  3. Overall system design
  4. Case study - Our churn predictor
  5. Questions

Speaker bio

Own the Machine Learning / Deep Learning product stack at Zoho Corporation. Have made high impact full stack ML/DL releases, reaching over a million users and have scaled and tweaked the platform accordingly!



Preview video


  • Abhishek Balaji (@booleanbalaji) 3 years ago

    Hi Ramprakash,

    Please upload a two-min preview video explaining what the talk is about and what the key takeaway is for participants. We need this information by 29 May to evaluate your proposal.

    • Ramprakash R (@ramprakashr) Proposer 2 years ago

      Hi Abhishek,

      Have added the same. Thank you.

  • Amit Kapoor (@amitkaps) 2 years ago

    Hi Ramprakash - Would love to see more detail on the talk here. Are you covering just LIME or more techniques here. It would be great to see a broader coverage of making ML models more interpretable e.g.

    • Ramprakash R (@ramprakashr) Proposer 2 years ago

      Hi Amit,

      As mentioned in the rehearsal, this is not about LIME. This is about taking our inspirations from LIME and writing a production ready model explainer, the pitfalls we faced in the process and the story of pushing this to production.

      So answering your questions,
      1. Covering LIME ? NO!
      2. Broader coverage? Definitely yes!

      Here is the link to the repo :

Login to leave a comment