The Fifth Elephant 2020 edition

80 Million products across 80+ categories is what Flipkart’s Search enables discovery for. And, in a user’s journey of discovering products, she is shown with autosuggest suggestions to choose from while typing a query. These suggestions don’t just help users in choosing a well formed query with minimal typing effort, there is more to it.

This talk briefly touches upon the opportunities that decorating these suggestions brings to us.
After setting the context of how product popularity has led to a never ending loop in the system leading to this bias, I’ll be walking the audience through our journey of solving the problem of less sought categories not visible on autosuggest due to it.

We’ll start discussing our journey with an implementation that randomly chooses store decorations and the unexpected learnings that it gave us. Further ahead, we’ll look at the possible rewards that are relevant to autosuggest and the observations from our first reward based decoration selection algorithm which pretty much solves for the bias but misses to make its mark on the constraints that the problem poses. We’ll look at how looking at rewards as distributions gave further improvement but affected our metrics for quite some time initially. Introducing priors helped us with reducing the initial adjustment period and also showed interesting patterns around the impact of priors on overall convergence. We’ll close the discussion with the learnings at each step in our journey and the future work.

Outline

Problem Background

1. Autosuggest in search
2. Role of decoration as a two-way communication channel with the user
3. What is the bias that we are trying to solve for and why is it there in the first place (with illustrations)
4. Problem Definition : Goals & Constraints
5. Issues with the existing reward (continuing the same illustration)

Journey of solving for it

1. Explore Exploit as a solution

2. First step towards solving : Random Selection

• User Experience View
• Observations
• Merits of starting with random exploration

3. Moving towards performance reward based exploration

• Choice of reward and its pros and cons
• Our way of implementing a performance based exploration algorithm
• Convergence Illustrations
• Movement in overall store visibility landscape
• There was still scope for improvement, so what next?

4. Need to account for regret along with reward

• Visualising store decorations as Beta distributions. sampling on them for decoration selection
• Convergence improvement
• Movement in overall store visibility landscape
• Observations : Slower convergence

6. Solve for faster convergence

• How priors affect movement in arms and the gains observed from having priors in sampling

7. Future Work

Note : All sections will include illustrations, metrics movement and changing suggestions for the aforementioned example

Speaker bio

Pranjal is a Software Development Engineer with Flipkart Search

Links

• www.linkedin.com/in/pranjalsanjanwala

Slides

https://docs.google.com/presentation/d/1Ek27iIBoFB23-h_YVpkYpd50qaKx-Lb4PIhISeHRJPg/edit?usp=sharing