3 months ago
We predict how a tweet will perform 🤯 How is that even possible? Let me tell you:
You're doubtful, right? But yes, it works. How can we be so sure? We secretly ran it for weeks on Tweet Hunter users for every new tweet published and here is how it performed 👇
We analyzed 23,472 tweets published and looked at how they "relatively" perform compared to the prediction. Meaning how those performed compared to the user's median number of engagements. Results are interesting: [img:E5asCCjqQ]
This is the most advanced AI we've ever built. Let's reveal a little bit of what is inside. (⚠️ it will get technical)
For months, we tried many approaches and they all failed. We were trying to estimate the number of likes and it was off... Probably because of the high variability and how extreme values can be.
So we drastically changed the approach and we now try to answer a simple question: is the tweet going to outperform. We don't need to know the # of likes, just if it's gonna be good or not.
What's crazy interesting is by doing this, we observe a strong correlation between the outperforming probability and the number of engagements received. So we do end up predicting the number of engagements even if we optimized the algo differently.
To do that, we used NLP state-of-the-art algorithms: - BERT from Google to get a relevant numeric representation of each tweet - a Non-Linear classifier over the BERT representation to obtain a prediction
We trained our model with tens of millions of tweets. Results were decent but not crazy. A tweet cannot be good no matter the audience. It's the audience-tweet fit that is the most important.
So we added an extra "per user" training layer. By doing this, the algo predicts if the tweet is going to generate more engagement than your personal average (actually median)
Side note, you may wonder what the algo is looking at With BERT, we found evidence that the algo takes into account: - emotions triggered - topics & keywords - consistency with previous tweets - general writing style