Let’s face it, we’ll never figure out the constant pattern and consistency of the almighty Instagram Algorithm (almost as powerful as YouTube Algorithm itself), since they use machine-learning for their system, all of which are improved by and rely on data.
So, really, how is Algorithm from 2019 to 2020? The more often we engage with a content or a content creator, then their content will appear in our uppermost feed. Instagram don’t universally favour photos over videos. Your feed is always customized for you, so it depends on how we each use our Instagram. If we’re watching a lot of those dance TikTok-like videos, we’re definitely going to get more of those videos in our feed. There was a ‘trick’ circulating around that liking or commenting on a post in the first 30 minutes would raise our post in the rank feed. Now, this can totally be a true/valid thing of course, but it’s all very short-term benefit, why? Because when the machine learning system knows and figures out that someone is trying to game the system, it would then presume “Wow, a new pattern. I have to improvise my Algorithm” so it’s not easily tricked anymore.
So then, what changed? Why is the Instagram Algorithm no longer time-based? Everyday lots of people post videos and photos on Instagram, and, well, obviously not all of them will get in our feed, right? Imagine having to see everything posted in the last 5 minutes and it’s just a barrage of body slimming vitamins or random pictures of people reposted all over everywhere. That’s why Instagram decided to make a ‘recommendation engine’ which tackles the ginormous amount of videos and photos uploaded every second. And to make sure this ‘recommendation engine’ works and flows smoothly, Instagram has to break apart the domains from each user with using IGQL, a.k.a. Instagram Graph Query Language. And from this IGQL, Instagram catalogues user data habit, whether it’s in the short-term or the long term.
But how exactly do they break apart the domains, or their posts from each user? Well, using this term called ‘Candidate Generation’, Instagram will use the users who’ve been interacting with each other before (liking each other’s posts, DM-ing each other, saving each other’s posts) so these users will know each other’s interest, despite not even really having that same interest for the both of them there. Then these users will be the ‘seed’, or samples to help other users find their interests. This of course doesn’t mean that they share an interest that fits with the Community Guidelines, which poses a problem with Instagram, which is why Instagram make their own system, known as, you guessed it, ‘Ranking Candidates’!
In ranking candidates, there are 3 process:
1. First stage, or more colloquially known as distillation model (think of it like going down through the pipeline tubes, like, literally), where Instagram picks 500 relevant posts from millions of ‘Candidates’, and then pick the highest-quality 150 posts from the 500.
2. Second stage, where Instagram picks 50 highest-quality posts from that same 150.
3. Final stage, where Instagram picks 25 highest-quality and most relevant posts from that same 50, of which they will then finally be the candidate for the first page of Instagram Explore.
Now, from that system alone, I’m sure some of you have questions that probably goes somewhere along like these:
1st Question: How does Instagram decide what’s relevant and what’s not for each stage?
Well, Instagram does that by predicting what you’re gonna do if given a posting (whether video or photo) of each of those ‘candidates’, whether it be a positive action such as a like or a save, or even negative action like pressing the ‘See Fewer Posts Like This’ button.
Follow-up 1st Question: But, dude, how can Instagram even make that prediction? Isn’t it all kinda random?
Yup, yup, and yup! All these predictions are taken from the cataloguing of every user’s data habit (short term and long term). And then these datas are given to the machine learning Algorithm, which will make the pattern from these datas for prediction process, or better known as Multilayer Perceptron. Cool term, huh? Almost like a name of a Transformer~
Fun fact: Even the engineers of Instagram’s machine learning Algorithm has no clue how it makes the pattern. Can you say Skynet?
2nd Question: So, does that mean we’re never gonna see contents that aren’t related to our interests?
Well, in order for a new interest to be able to get in alongside the other interests that are already there, the system put inside Explore has to be balanced in some way. And from there, Instagram created this new Heuristic Rule in their machine learning model, where they balance out the diversity of content given. How exactly? By downgrading the rank of a post that comes from the same author, or ‘seed’, better known as penalty factor. Very football-y.
And that there is why we’ll rarely see, almost borderline never see multiple posts in our Explore IG Feed by the same user.
Okay, so how do I outsmart the algorithm then?
Well, here are some quick tips for ya:
- Photo quality, but make it better…and A U T H E N T I C.
- Publish those stories consistently. Always.
- Three hours since you post? Respond to every comment made!
- First hour since you post? Engage with your followers!
1. Powered by AI: Instagram’s Explore recommender system | by Ivan Medvedev | Instagram Engineering (instagram-engineering.com)
2. Core Modeling at Instagram. At Instagram we have many Machine… | by Thomas Bredillet | Instagram Engineering (instagram-engineering.com)
3. Efficient tuning of online systems using Bayesian optimization — Facebook Research (fb.com)
Callasyah Erwinanda (Copywriter at Apiary Coworking Space)
Ridho Gani (Data Analyst at Apiary Coworking Space)