People you may know:
how to find your old friends

Where it all started

Many of you have probably come across a “People You May Know” section while scrolling through your feed on Facebook, Instagram, or any other social network. This is a commonly used feature across social platforms that suggests other users you might want to follow or add as friends. Its primary purpose is to help grow your social graph and ultimately, your engagement.

In fact, increasing user activity and growing connections are at the heart of almost every social network. Social activities motivate people to return to the platform and stay engaged with the product.

Before the rise of TikTok and other content-oriented platforms, social networks like Facebook, and the one I worked for were fundamentally built around growing the friend graph as the core infrastructure for communication and interaction.

That’s why this growth driver was always considered essential to the overall product strategy.

The challenges we had

New connections within the social network have a significant impact on our daily active users — a fact once again confirmed by our data science team.
We realised that we hadn’t seen any growth in new connections, which in turn affected key business metrics such as daily active users (DAU) and engagement indicators like stickiness (DAU/MAU):

This raised an important challenge: why aren’t users connecting with each other? Are we failing to surface relevant people for them to connect with?

Discovery Stage

After several weeks of investigation (discovery stage), this root cause led the team to identify a few issues that could have been affecting the current situation.

We grouped them into 3 key categories:

First, some of our ML pipelines were misconfigured, which meant that certain data wasn’t reaching the recommendation systems and therefore wasn’t being taken into account
Second, some potential relevant features were entirely missing, and we worked with the feature engineering team to develop new indicators and signals to improve the quality of recommendations
Third, our recommender system updates were not optimal and were happening with delays. Ideally, we wanted the system to respond almost instantly to changes in user behavior or new inputs. For example, when a user adds a new contact to their phonebook, the recommendation logic should adapt in near real-time

Action #1

Revise all data pipelines and optimise them in order to guarantee that we store everything properly and use all available information to perform recommendations

Action #2

Invest in feature engineering, or more accurate, in re-engineering, because we noticed some low-effort actions can be taken to significantly improve the overall performance

Action #3

Add more sources of information which could be useful in RecSys

Action #4

Optimise infrastructure to maintain more data sources, faster real-time updates of recommendations, because this was crucial for use-cases when people just join social networks and we had almost zero-information about their preferences and connections

Achievements

After several months of working on this project, we were able to stop the decline in user activity that had started a few months earlier. Following the launch of a new experiment focused on delivering more relevant recommendations, we saw a significant impact user-created connections increased by over 20% MoM
For me as a leader, it was a great personal achievement where I discovered the power of RecSys and influence and impact they may give. Of course, if they got cooked properly. This was one of the most successful outcomes we had achieved as a team, and it came after integrating an additional data source that provided richer signals about user relationships
From that point on, the service began to grow actively, as we continued to improve both the recommendation models and the user experience. We also expanded our data sources and optimized specific interaction flows across the service

Lessons Learned

Over the project development, I learnt many critical things as a manager working with ML:

The importance of monitoring and observability of ML pipelines
Regular revision of all available data sources, which can drive your ML product
Flexible and scalable infrastructure, which open the doors for faster experiments and iterations