S4RAH.com

My account growth has been unprecedented this week, averaging 103,100 views per day, with almost a million impressions in one day last week.

I decided to take some time to dive into the Twitter Blog and read up on how they decide to rank tweets. This turned out to be a rather in depth dive into the overall workings of frameworks, datasets, graphs, and algorithms. I’ve learned that the Twitter framework is comprised of multiple frameworks, algorithms, software, and hardware that come together to form the feed we experience. There are overall frameworks that apply to all accounts, and then there are frameworks comprised of algorithms that handle specifics scenarios, such as in network and out of network ranking.

RealGraph and GraphJet

In-network ranking has a everything to do with RealGraph. RealGraph is a framework developed at Twitter to assess the strength of connections between users based on their interactions. It’s a directed and weighted graph that predicts the likelihood of future interactions between users. This framework utilizes historical data and machine learning models like logistic regression to assign weights to edges in the graph. These weights represent the probability of future interactions between specific users on Twitter. This is the framework that large accounts want to focus on, to sustain user base which will boost impressions.

Out of network ranking relies on GraphJet. Graphjet began operating in 2014, initially addressing the cold start problem for new users. Currently it powers various services, including injecting relevant tweets into users timelines. On the hardware configuration with Intel Xeon processors, GraphJet sustains one million edge insertions per second during cold start and drops to typical engagement rates of tens of thousands per second. In fact, a single GraphJet server supports up to 500 recommendation requests per second, with latency profiles (p50 = 19ms [meaning 50% of the recommendation requests were processed within 19ms], p90 = 27ms, and p99 = 33ms) for subgraph SALSA (mmmmm…. salsa) algorithm usage. The Subgraph SALSA is the algorithm smaller accounts will want to focus on if growth is their goal.

So what is the Subgraph SALSA algorithm and how does it work? 

The Subgraph SALSA algorithm creates recommendations by looking at a smaller group of connections within a bigger network. It’s like focusing on friends of friends rather than everyone.

Key points about this algorithm:

• It finds a smaller group of connections based on a starting set, like a user’s close friends. Then, it shares scores among these connections in a way that shows which tweets are more popular within this group.
• It divides scores equally among these connections, considering how many connections each person has. It does this back and forth between people and tweets. This result shows a list of tweets based on how popular they are within this smaller group but it sacrifices knowing about some indirect connections to make things faster. Depending on what you need, this tradeoff can be important.

In short, the Subgraph SALSA algorithm gives recommendations by focusing on a smaller group, making things faster but possibly missing some less obvious connections between people and tweets.

So what does all this mean anyway, and why should you care? 

If you don’t care about growing on Twitter, then you shouldn’t care. Go right on your merry way. This stuff is not keeping me awake at night either. But I’ve learned more about algorithms reading through these docs than I have through all 3 of my bootcamps. And really, if some of those sweet sweet Elon bucks are temping to you, here is my main takeaway:

1. If the algorithm is repeatedly showing you tweets from a user that means that that user has an audience that is well suited to following you also. Interacting with their posts by way of commenting and quote tweeting is a way to draw their audience to your timeline.
2. Unfortunately, influencers matter, if you want to grow. This doesn’t mean you need to obsess, but their audience is the fastest way to organic growth. If you had a bucket of Oysters to sell, you would go to a city, not to the middle of the forest where nobody is. This is the same theory. Go to where the people already are.
3. Although quality over quantity is true, quantity is key for new accounts. The algorithm will assess every new piece of content you put up there, and you just never know what might resonate with your audience, which then may resonate with their audience, and so on, and so forth. *You* may think you’re sharing a lot. You’re audience might never see most of your posts, however. The algorithm requires your content to grow.

Please note this advice is a simplified overview of a very complex framework. This advice is also targeted to accounts under one million, which I am pretty sure is every single person in my network. If this advice resonates with you, I’d love it if you share. Thanks for reading!