Shoya Yoshida

Machine Learning @ Instagram/Threads - Personalization & Recommender Systems | Columbia CS | 🇯🇵

[Shoya Yoshida @decisiontree7]

Happy to announce that, a couple weeks ago, I started a new job as a Machine Learning Engineer at Instagram, specifically working on Threads! 👀

そして日本語で呟く用のアカウントをスレッズで初めました。よければあっちで繋がってください!　https://t.co/EecV0Jbpf8

[Shoya Yoshida @decisiontree7]

@invinc4u Thank you for your interest!! I'm not quite sure what the policy is around that so let me find out!

[Shoya Yoshida @decisiontree7]

I had the pleasure of giving a lecture about recommender systems at Columbia's Applied ML class! Hopefully students got some clarity on what I think is a criminally under-discussed topic in college education despite its prevalence in the industry.

[Shoya Yoshida @decisiontree7]

@g3shandar Thank you Gayatri 🥺

[Shoya Yoshida @decisiontree7]

Talk title: "FitNExT: Leveraging Transformers with Context-Augmented Start Tokens to Generate Recommendations for New Users in Connected Fitness at Peloton"

[Shoya Yoshida @decisiontree7]

Very excited to be in Singapore for #Recsys2023! Come check out my presentation on how we tackled the user cold-start problem at Peloton at the BehavRec workshop, 9/19 9:10 - 9:30 SGT!

[Shoya Yoshida @decisiontree7]

@shanegJP 「日本人はアイディアとかものづくりの文化が強い」に共感しました。変な例ですが日本ではボカロが発達して音楽が作れる人が増えたようにChatGPTとの親和性も比較的高いと思います。

[Shoya Yoshida @decisiontree7]

My blog article on revamping the Peloton homescreen with personalized rows is now up - please check it out! https://t.co/FyOuSf9eke

[Shoya Yoshida @decisiontree7]

Got to meet some cool folks at the NYC MLOps Meetup hosted at the Spotify office today!

[Shoya Yoshida @decisiontree7]

Finally, the online performance did reveal that the model trained with resurfacing awareness did significantly better than without!

[Shoya Yoshida @decisiontree7]

#RecSys2022 Highlight #3 - Industry Keynote of Context-Aware Recommender system (CARS) Workshop - "Contextual Product Recommendation at Wayfair" by Jeffrey Mei. Presentation Slides: https://t.co/0oj1ITTQya

[Shoya Yoshida @decisiontree7]

This also begs the question - why not just always have the item that the user ends up buying as the positive training data? Well, they did try that, and it so turned out that the users didn't like that because users want to explore a little first (which makes sense)!