Amey Dharwadker

Engineering Manager, Machine Learning @Facebook. Building video recommender systems serving billions of users. Writing on Machine Learning & AI.

[Amey Dharwadker @ameydhar]

@thesamparr Interesting product, but a spell and grammar check would go a long way in building customer trust!

[Amey Dharwadker @ameydhar]

Fascinating read on @Google OKR discussion covering search and ads long vs short term metrics, financial goals for search and cannibalization across teams.

[Deedy @deedydas]

Shashi, one of the engg leads for Google Search, in a leaked internal email says "there's no way I'm signing up for a daily active users goal (OKR)".

The focus was always on user value, not growth at all costs.

There are SOME companies who would not think twice about it.

[Amey Dharwadker @ameydhar]

@karpathy How did you ensure that the model hadn’t seen these questions and answers in its training data already?

[Amey Dharwadker @ameydhar]

@icwsm Interested, would love to help!

[Amey Dharwadker @ameydhar]

CDR-Adapter tackles data sparsity and cold-start in cross-domain recsys, separating the model from the mapping function. It's a plug-and-play module for flexible knowledge transfer and efficient fine-tuning with minimal training costs.

Link: https://t.co/VuMibcsYiJ

[Amey Dharwadker @ameydhar]

ControlRec: Bridging the Semantic Gap between Language Model and Personalized Recommendation

Novel contrastive prompt learning framework that aligns user IDs and natural language features, employing Heterogeneous Feature Matching (HFM) and Instruction Contrastive Learning (ICL).

[Tony Russell-Rose @tonygrr]

Delighted to announce the winners of the 2023 BCS Search Industry Awards. Congratulations to
https://t.co/VG9ULoxl8w, Wikiframe Visual Graph, @ameydhar and @FlaxSearch.
Thanks also to our judges @molnaragnes, Andy Hind and Peter Cotroneo
https://t.co/xmzRdu4l2Q

[Amey Dharwadker @ameydhar]

Delighted to be a part of this insightful conversation with @jaygshah22, discussing the fascinating world of building and deploying billion-scale video recommender systems, tips on how to get started or transition into AI, etc. Do check out the podcast!

[Jay Shah, Ph.D. @jaygshah22]

My podcast with @ameydhar on building Video Recommendation systems at @Meta for @facebook-Watch (2 billion+ users), improving user experience and engagement using #MachineLearning & some tips for acing at ML roles in the industry.

https://t.co/QJpM6zLFay

[Amey Dharwadker @ameydhar]

@facebook @TheWebConf @FacebookWatch @banarasi_ghat Retweet the first tweet of this thread to spread the word and share these works with your network interested in industrial #RecommenderSystems and #MachineLearning for personalization!

https://t.co/cxOvNUyMhG

[Amey Dharwadker @ameydhar]

🚨 Exciting news! 🚨 Two papers from our @facebook Video Recommendations team have been accepted for publication at ACM @TheWebConf (WWW 2023)! Our #recsys work highlights how we're enhancing video watching experiences for users worldwide. 👇

[Amey Dharwadker @ameydhar]

🚨 Exciting news! 🚨 Two papers from our @facebook Video Recommendations team have been accepted for publication at ACM @TheWebConf (WWW 2023)! Our #recsys work highlights how we're enhancing video watching experiences for users worldwide. 👇

[Amey Dharwadker @ameydhar]

@facebook @TheWebConf Both these works have been successfully deployed in production on @FacebookWatch, a popular video discovery and sharing platform serving billions of users.

Huge shoutout to all my amazing collaborators! @banarasi_ghat

[Khushhall Chandra Mahajan @banarasi_ghat]

Our two papers have been accepted at @TheWebConf (WWW '23), showcasing outstanding work of my team at @facebook Video Recommendations in enhancing the user's watch experience.
https://t.co/XiDlEPohZP
https://t.co/nHPtMGdL46

#recsys #ml

[Amey Dharwadker @ameydhar]

@thealexbanks Nice stuff! Thanks for the summary 😊

[Amey Dharwadker @ameydhar]

Dynamically Expandable Graph Convolution for Streaming Recommendation

Proposes a novel algorithm for streaming recommendation, disentangling outdated short-term preferences from useful long-term preferences, capturing individual-level preference shifts.

https://t.co/nwqFaJjsoM

[Amey Dharwadker @ameydhar]

Debiased Contrastive Learning for Sequential Recommendation

Proposes a Debiased Contrastive learning paradigm for Recommendation (DCRec) that unifies sequential pattern encoding and collaborative relation modeling to tackle popularity bias.

Paper: https://t.co/9EZq9F3ndg

[Lenny Rachitsky @lennysan]

I don't say this often, but today's podcast episode with Keith Yandell is a must listen. Seriously, go listen to it.

Of the many lessons on leadership Keith shares from this time at @DoorDash and @Uber, one of my favorites is the power a simple doc like this can have on your org

[Amey Dharwadker @ameydhar]

@PeterDiamandis The ethical implications of AI creating seemingly "real" content raises questions about accountability and control.

[Amey Dharwadker @ameydhar]

Two-Stage Constrained Actor-Critic for Short Video Recommendation

Proposes a novel method to optimize short video recommendations, balancing both main goal (cumulative watch time) and auxiliary goals (interactions) using reinforcement learning.

https://t.co/TObqrRU0Dt