Inside the Tech is a blog series that goes hand-in-hand with our Tech Talks Podcast. Here, we dive further into key technical challenges we are tackling and share the unique approaches we are taking to do so. In this edition of Inside the Tech, we spoke with Senior Engineering Manager Michelle Gong to learn more about how the Personalization team’s work is helping Roblox users find experiences they’ll love.
What technical challenges are you solving for?
Our team – Personalization, which is in the Growth group – is responsible for providing our users with personalized and relevant recommendations. We want to empower people to find content they’ll love, to foster long-term engagement on Roblox, and to connect experiences with the people that are right for them.
Today, we have 66 million daily active users, but that number is increasing about 20% every year, and that means more and more data is coming in. So, a big technical challenge is maintaining real-time responsiveness and making sure personalized recommendations don’t require long waits, all without increasing serving costs. In fact, that’s one of the reasons why we completely rebuilt our backend infrastructure last year.
As we grow, we’re asking ourselves how we can improve the user experience without the need for a lot of additional compute power. We think machine learning could be part of the answer, but we’ve seen that ML solutions can use more compute resources — which raises costs — as the data models get bigger. That’s not scalable for us, so we’re working to improve real-time search and ranking without incurring those additional costs.
What are some of the innovative solutions we’re building to address these technical challenges?
We’re building a recommender system to help people discover the content that’s most relevant to them quickly. To do that, we’re learning how to apply the most advanced ML technologies to the problem. For example, we’ve incorporated self-supervised learning, advanced architectures and techniques from large language models (LLMs), and counterfactual evaluation in these systems.
There are many advanced pretrained LLMs, but we can’t use them directly because they incur high serving costs. Instead, we’re training our own models using techniques often employed to build LLMs. One example is sequence modeling, since both language and Roblox user play history are sequences. We want to understand which part of a user’s play history can predict their current and future interests and preferences. This model helps us do that.
At the same time, self-supervised representation learning is now being widely used in computer vision and natural language understanding, and we’re applying this technique to our recommendation systems.
What are the key learnings from doing this technical work?
Roblox’s goal is to connect a billion users, and to do that, we need to identify solutions that balance utility and cost. When we do this effectively, we’re able to invest more in our community.
For example, we decided to invest in our own data centers, and that bet is paying off. The biggest thing we learned is that when we have the resources and ability to do something ourselves, it’s more efficient to create something purpose-built than to pay for third-party technology. By building our platforms and our models from the ground up, we’re able to pursue innovative solutions that are optimized for our business and our resource constraints and requirements.
Which Roblox value do you think best aligns with how you and your team tackle technical challenges?
Respect the community. We care deeply about our creators and our developers. Their opinions really matter. We take developer feedback very seriously. I spend a lot of time answering developer questions directly in partnership with our Developer Relations Team. Taking the time to understand their feedback, and see how we can improve our platform for them, has helped us make sure we’re also focusing on the right things.
I’d also say take the long view. I joined Roblox because I really believe in Dave’s vision of taking the long view. In fact, in our day-to-day work, we avoid building short-term hacky solutions. Instead, we emphasize building principled, reliable, and scalable solutions because we are building for the future.
What excites you most about where Roblox and your team is headed?
We have so many unique challenges. Building recommender systems as a two-sided marketplace and for long-term user retention, is a huge opportunity for growth. But we’re also thinking about things like visual understanding and text understanding for use cases like recommendations, search, trust-and-safety, etc.
Also, we’re structured in a way that we can move really fast and be very efficient. Every team member is extremely driven and excited about the challenges we have. If this sounds like something you’re interested in, we’ve got a spot for you.
The post Inside the Tech – Solving for Personalization on Roblox appeared first on Roblox Blog.
