6 Simple Steps On How Trell Uses Machine Learning to Iterate Quickly and Efficiently
Machine learning has helped Trell scale up rapidly. Right now, we are catering to the interests of 80 million+ people with millions of videos posts coming in everyday. Though we deployed manual tagging of videos in the initial days, 4 years later, advanced algorithms do this job for us. Our users are happy too, with access to more personalised content in real-time.
The many advantages of ML are well-known. But using ML efficiently, in a way that helps the organization grow swiftly yet steadily, is the real challenge. Here is our journey that helped identify 6 simple steps for efficient iteration that made us massive:
- Understanding the problem:
In 2016, we had a team that manually mapped content to users. We would pick up content and place it in the user feed when the daily influx of videos was manageable. A year later, the volume of content increased and so we introduced NSFW checks and basic personalization. For instance, if a user liked 10 videos, with 3 being recipes and 7 being travel-related, the next time they got on the app, we would show 1 recipe and 2 travel videos. This helped us increase user engagement by 20%.
2. Setting the right expectations:
Good things take time. When user personalisation was identified as what we had to solve for, we didn’t get overwhelmed. Instead, we focused on the problems that were easiest to solve but yield the maximum result. In 2019, we achieved a major breakthrough with install-source based biases. Through this, if a user had logged on to the Trell website, we knew their preferences before they downloaded the app. This helped us personalise our users’ feed in real-time once they were on the app even for the first session, when there is still not enough information about the user’s activity on the app. The result? An increase in user engagement by 35%.
3. Using shelf-models:
“Not Invented Here” or NIH is usually condemned by the tech-industry as it discourages the use of technical resources that are not made “in-house”. We didn’t start with ML from the get-go. We used a correlation study to prove our hypothesis and then moved on to using AWS Personalize with our solutions.
4. Don’t over engineer solutions:
Instead of striving for perfection in the first go, we introduce more complexity and robustness through constant product improvement and innovation. While we always had content tagging — first we started with object tagging and then moved on to introducing speech detection and language analysis — small but significant additions helped our algorithms immensely. This helped increase user time spent on the app by 45%.
5. Implementing incremental learnings:
As an ML-based model needs constant tweaks and adjustments, focus on iteration and implementation more than getting it perfectly right the first time. Your learnings through successes and failures will help scale the model and get better results. Today, our users spend 800% more time on the app than they ever did before.
6. Taking data-driven decisions:
Without data, there is no ML. After starting off with category-level data, we enriched it with details like user demographics, collaborative filtering, object detection in videos, audio analysis. Currently, we are using Natural Language Processing (NLP) to first translate content in regional languages to English, and then using algorithms to identify key aspects of our videos content for more stream-lined content tagging.
In the next couple of months, we will be able to learn more about the videos shared including embeddings and activity. We will also use explicit user input in the form of like/dislikes, reporting videos and user type affinity and continue iterating quickly and efficiently.
Hope this was helpful for everyone who wants to use machine learning for good results and growth!