Editor’s note: Since launching its Spaces feature, Twitter has demonstrated that hearing people’s voices can bring conversations on Twitter to life in a completely new way. Next, it aimed to make it easier for customers to join and listen to live conversations they personally care about. In this blog, we learn how the Twitter Spaces Engineering team is bringing this vision to life with AutoML, powering a new ML heuristic which serves personalized recommendations to Twitter customers. The authors would like to thank Chuan Lu, Joe Balistreri, Chen-Rui Chou, Pablo Jablonski, Alberto Parrella, Pradip Thachile and Sam Lee from Twitter, as well as Helin Wang from Google, for contributions to this blog.

Since Twitter introduced Spaces in 2020 to enable live audio conversations on its platform, the Twitter Spaces Engineering team has been continually testing, building, and updating this feature in the open. Today, anyone can join, listen, and speak in a Space on Twitter, and the feature’s popularity has taken off. But this success also poses a challenge: with millions of people creating and joining Spaces at any time, how can they find the Spaces to engage with while they’re happening? Taking this as an opportunity to further improve the experience of its customers, Twitter has turned to machine learning (ML) and cloud technology for answers.

“ML fits into the natural progression of Twitter consumer and revenue product building, especially for a product feature such as Spaces,” explains Diem Nguyen, Senior Machine Learning Engineer and Data Scientist at Twitter. “We launched Spaces with a base-line algorithm using the ‘most popular’ heuristic which assumes that if a Space is popular, there’s a good chance you’d like it too. But our aim is to leverage ML to surface the most interesting and relevant Spaces to a particular Twitter customer, making it easier for them to find and join the conversations they personally care about. This is a complex functionality that Google Cloud ML capabilities help us to enable.”

Setting the stage for building new features with limited ML resources

While looking for the right tools to power this vision, Nguyen and her team started evaluating in December 2021 whether the Vertex AI platform and AutoML in particular could solve challenges observed when they first started building Spaces. These included a lack of dedicated ML resources to build and deploy the product feature, and the need to work on a multi-cloud environment.

“We had three key questions in mind during our assessment,” Nguyen explains. “Can we realistically deploy the AutoML model off-platform? Once deployed, can it solve for the request load that we get from the service we’re serving (in this case, the Spaces tab)? And finally, can we develop and maintain such a solution without a dedicated team of ML experts for this project?” The answer to all three questions was yes.

Positive answers motivated the Spaces Engineering team to take the solution to production in February 2022. “We started using AutoML Tables to train high-accuracy models with minimal ML expertise or effort, alleviating our resource constraint,” says Nguyen of the results. “Soon AutoML also stood out for its high performance and for supporting easy deployment beyond the Google Cloud Platform, making it ideal for this project hosted in a multi-cloud environment.”

Increasing customer engagement at speed with accurate ML predictions

With a classification model in place to predict the probability of user engagement in a particular Space, Twitter now aims to optimize its model with aggregated data around Twitter features that can help it better understand customer preferences. For example, if a customer has historically engaged with a particular topic and a new Space matches that topic, the ML model increases the score of that Space being served to that user on the Spaces tab.

Because Spaces are live audio conversations, the Spaces tab needs to be ranked to customers in near real time so they don’t miss out. With this in mind, Twitter’s model currently performs 900 queries per second on the Spaces tab, and evaluates 50,000 candidates per second. Meanwhile, 99% of these requests are faster than 100 milliseconds, and 90% of requests are faster than 50 milliseconds.

To measure the success of this project, Nguyen’s team conducted A/B experiments around key customer engagement metrics–A stands for the ‘most popular’ heuristic previously in production, and B is the new AutoML model which seeks to personalize Spaces recommendations to the interests of individual Twitter users. Three months into the project, the numbers were encouraging. “After deploying our AutoML Tables solution we saw an increase of 1.96% in Spaces daily active customers, which is one of our key metrics. We also noticed an increase of 1.99% in Spaces join in rates, and an increase of 8.42% in user clicks to explore a Space,” Nguyen shares. “These are positive signals that users are now engaging more with the Spaces tab service on the Twitter app, which is exactly what we set out to do with this project.”

Powering new use cases with hands-off ML frameworks

With this first solution running in production to improve the performance of the Spaces tab, Nguyen starts to ask how else it might support the experience of Twitter users moving forward. “The Spaces tab is a small surface on the Twitter app. With our current ML solution we’re some distance away from serving our home tab traffic, which is where a lot of our traffic happens and therefore would involve a much bigger-scale operation. Getting there will take some work but we’re evaluating the possibility of optimizing our model performance for this in collaboration with Google Cloud,” says Nguyen.

“As a product-led company, we focus on continually improving the customer experience and we want to iterate faster to get to that point. AutoML brings that value to our product teams because it is so hands-off. You don’t need to write any model code in order to reap the benefits from this machine learning framework; AutoML automatically experiments with many different model architectures and comes up with a state-of-the-art model that addresses your needs. So while it is not a one-size-fits-all solution, it is a great solution with the potential to power many more Twitter use cases,” she concludes.

Who would have known that today technology would enable us with the ability to use machine learning to track vessel activity, and make pattern inferences to help address IUU (illegal, unreported, and unregulated) fishing activities. What’s even more noteworthy is that we now have the computing power to share this information publicly in order to enable fair and sustainable use of our ocean. 

An amazing group of humans at the nonprofit Global Fishing Watch took on this massive big data challenge and succeeded. You can immediately access their dynamic map on their website globalfishingwatch.org/map that is bringing greater transparency to fishing activity and supporting the creation and management of marine protected areas throughout the world.

In our second episode of our People and Planet AI series we were inspired by their ML solution to this challenge, and we built a short video and sample with all the relevant code you need to get started with building a basic time-series classification model in Google Cloud, and visualize it in an interactive  map. 

Architecture

These are the components used to build a model for this sample:

• Global Fishing Watch GitHub: where we got the data
• Apache Beam: (open source library) runs on Dataflow. 
• Dataflow: (Google’s data processing service) creates 2 datasets; 1 for training a model and the other to evaluate its results.
• TensorflowKeras: (high level API library) used to define a machine learning model, which we then train in Vertex AI.
• Vertex AI: (a platform to build, deploy, and scale ML models) we train and output the model.

Pricing and steps

The total cost to run this solution was less than $5. 

There are seven steps we went through with their approximate time and cost:

Why do we use a time series classification model? 

Vessels in the ocean are constantly moving, which creates distinctive patterns from a satellite view.

 We can train a model to recognize the shapes of a vessel’s trajectory. Large vessels are required to use the automatic identification system, or AIS. The GPS-like transponders  regularly broadcast a vessel’s maritime mobile service identity, or MMSI, and other critical information to nearby ships, as well as to terrestrial and satellite receivers. While AIS is designed to prevent collisions and boost overall safety at sea, it has turned out to be an invaluable system for monitoring vessels and detecting suspicious fishing behavior globally.

One tricky part is that the MMSI data location signal (which includes a timestamp, latitude, longitude, distance from port, and more) is not emitted at regular intervals. AIS broadcast frequency changes with vessel speed (faster at higher speeds), and not all AIS messages that are broadcast are received – terrestrial receivers require line-of-sight, satellites must be overhead, and high vessel density can cause signal interference. For example, AIS messages might be received frequently as a vessel leaves the docks and operates near shore, then less frequently as they move further offshore until satellite reception improves.  This is challenging for a machine learning model to interpret. There are too many gaps in the data, which makes it hard to predict.

A way to solve this is to normalize the data and generate fixed-sized hourly windows. Then the model can predict if the vessel is fishing or not fishing for each hour.

It could be hard to know if a ship is fishing or not by just looking at its current position, speed, and direction. So we look at the data from the past as well, looking at the future could also be an option if we don’t need to do real time predictions. For this sample, it seemed reasonable to look 24 hours into the past to make a prediction. This means we need at least 25 hours of data to make a prediction for a single hour (24 hours in the past + 1 current hour). But we could predict longer time sequences as well. In general, to get hourly predictions, we need (n+24) hours of data.

Options to deploy and access the model

For this sample specifically we used Cloud Run to host the model as a web app so that other apps can call it to make predictions on an ongoing basis; this is our favorite in terms of pricing if you need to access your model from the internet over an extended period of time (charged per prediction request). You can also host it directly from Vertex AI where you trained and built the model, just note there is an hourly cost for using those VMs even if they are idle. If you do not need to access the model over the internet, you can make predictions locally or download the model onto a microcontroller if you have an IoT sensor strategy.

Want to go deeper?

If you found this project interesting and would like to dive deeper either into the specifics of the thought process behind each step of this solution or even run through the code in your own project (or test project); we invite you to check out our interactive sample hosted on Colab, which is a free Jupyter notebook.  It serves as a guide with all the steps to run the sample, including visualizing the predictions on a dynamically moving map using an open source Python library called Folium. 

There’s no prior experience required! Just click “open in Colab” which is linked at the bottom of GitHub.

You will need a Google Cloud Platform project. If you do not have a Google Cloud project you can create one with the free $300 Google Cloud credit, you just need to ensure you set up billing, and later delete the project after testing the desired sample.

🌏🌎🌍 We hope to inspire you to build other beautiful climate-related solutions.