The Fifth Elephant 2019

Gathering of 1000+ practitioners from the data ecosystem

Participate

Improving product discovery via Hierarchical Recommendations!

Submitted by Neha Kumari (@neha-kumari) on Friday, 5 April 2019

Session type: Lecture Session type: Full talk of 40 mins

View proposal in schedule

Abstract

A recommendation engine’s primary goal is to surface personalised & relevant content to the user, content which satisfies explicit intent as well as serendipitous content that would otherwise be invisible. E-commerce categories such as Lifestyle, have a lot of flux, the trends last for a short time window and have their demand distributed across an extensive selection. In such cases, recommending product collections can be a better idea instead of individual products.

In this talk I will be talking about the recommendation system at Flipkart, our journey towards recommending collection and how it improved the product discovery and helped in solving cold start problem. I will also cover the relevance algorithm used which is a hybrid of collaborative and content-based recommendation, and how this is achieved at scale, where we have catalogue consisting of ~400M products and this will be ever growing in this ecommerce world.

The Why ?
1. To solve cold start problem and to improve product discovery which aids to coverage
2. Increasing basket size and improving engagement post purchase
3. Cross category discovery and search result augmentation

Outline

  • Introduction - a generic e-commerce Recommender System
  • Problem in a traditional Recommender System
  • Our journey towards recommending collections
    • Flipkart’s Taxonomy
    • Computing relevance via Cosine Similarity
    • Multiple approaches
    • Improvement and results
  • Conclusion

Speaker bio

Neha is a Senior Software Engineer with Recommendation team at Flipkart. She has worked on building scalable systems for product recommendations and personalisation. In the past she has also worked on Natural Language Processing. She is interested in building robust data processing pipelines at scale, and applying Machine Learning to solve challenging problems . She has graduated from IIT BHU. While not working on official projects, she involves herself in technical writing and blogging. She also contributes to the open source world by answering technical questions.

Links

Slides

https://docs.google.com/presentation/d/1YhDqTCLOPcxgn1wkVSuTq6iqPiRdfBwJt0L4a9QhGu4/edit?usp=sharing

Comments

  • Anwesha Sarkar (@anweshaalt) Reviewer 3 months ago

    Thank you for submitting the proposal. Submit your slides and preview video by 20th April (latest) it helps us to close the review process.

  • Zainab Bawa (@zainabbawa) Reviewer 2 months ago

    The slides are not publicly accessible, Neha. Change permission settings. Also, upload the two-min preview video explaining what this talk is about and why participants should attend?

    • Neha Kumari (@neha-kumari) Proposer 2 months ago

      I have updated the slides to be publicily accessible. Will update the video in a while.

      • Neha Kumari (@neha-kumari) Proposer 2 months ago

        I have added the video link as well.

  • Zainab Bawa (@zainabbawa) Reviewer 2 months ago

    We have evaluated the slides and have the following feedback:

    1. The problem has to be stated as a general problem statement rather than a Flipkart problem statement. This is because companies which don’t have Flipkart’s problems should be able to identify with the problem per se, rather than the company. For example, is the problem that there is a long-tail problem in the discovery of content about lifestyle products?
    2. The slides dive straight into the solution without explaining why this approach was considered for the solution? Which other approaches were considered? How was the evaluation done before finalizing that the problem will be solved in this manner?
    3. The key takeaways are not obvious because there are no details – such as before-after scenarios, data from implementation.
    4. Also, the slides don’t show what has changed after implementing this solution, and the challenges of using this solution at scale?

    The slides need to be much more detailed, incorporating the above feedback. Upload revised slides by or before 21 May in order to close the decision on your proposal.

  • Anwesha Sarkar (@anweshaalt) Reviewer a month ago

    Thank you for the submission we will get back to you soon.

  • Abhishek Balaji (@booleanbalaji) Reviewer a month ago

    Rehearsal scheduled

  • Neha Kumari (@neha-kumari) Proposer a month ago

    @Abhishek : I have made changes in my slide as per the feedback of the rehearsal, please review.

    • Abhishek Balaji (@booleanbalaji) Reviewer 27 days ago

      Thanks, will get it reviewed and share an update.

  • Abhishek Balaji (@booleanbalaji) Reviewer 25 days ago

    Hi Neha,

    Thanks for updating your content. Here’s the feedback from the reviews:

    1) There was not much improvement as expected from the feedback and reviews shared on the rehearsal.

    2) The talk would work much better as a crisp talk for 20 mins, where the key takeaway of the presentation can be arrived at faster.

    As next steps, I’m moving your proposal under waitlisted. This means that while your talk would not be relevant to the entire breadth of audience at The Fifth Elephant, a smaller audience could definitely benefit from hearing about this.

    We’re also putting together a Birds of a Feather session on recommendation systems and similarity matching. I’ll make sure to check with you on your participation for this at The Fifth Elephant.

Login with Twitter or Google to leave a comment