Network Quality Algorithms
Inauthentic and coordinated behaviors are rampant on social media. This is especially true with larger influencers that brands often pay exorbitant amounts of money to in order to access their community. This called for the development of several algorithms that could assess the quality of these influencers, and in particular help to locate high quality micro-influencers that were more affordable, more approachable, and pound-for-pound drove more results for brands.
Coordination Score
While many were trying to determine the “botty-ness” of any one account, I took the route of identifying clusters of accounts that appeared to be coordinating their activity (i.e. owned by the same person), for example by following the same account or engaging with the same post.
The most obvious clue of coordination was large numbers of followers with similar creation timestamps. I developed a score that analyzed this using simple standard deviations-- if there were a suspiciously large number of accounts (>3.5σ) that all created their accounts on the same day, each account in that cluster had their Coordination Score incremented. These suspicious accounts were then able to be tracked across their other following & engagement actions and subtracted from legitimate results.
Network Influence Score
If two accounts both have 100,000 followers, whose are better? Generally, higher quality accounts have a diverse and distributed audience, across a few metrics:
-
Followers with account ages from the start of Twitter to now.
-
Followers of different account sizes.
-
Followers with different posting activity.
-
Followers with different follower/following ratios
The strategy I developed to estimate this involves XY scatter plotting the accounts across these different metrics and analyzing the variance and distribution of followers across the metrics. More distribution across these metrics yields a higher score.
Follower Quality Score
The first score I developed involved a simple letter grade of the bulk follower quality. The simplest means to determine this is to count a few key metrics across the followers:
-
Percentage of 0-10 follower accounts
-
Number of verified accounts
-
Average followers per account
This yielded a score that provided a quick litmus test of the account and provides a simple pass/fail during influencer evaluation
Results
Our first client using these scores is a well-known web3 bridging platform having launched an underperforming v2 and came to us to help them jumpstart their growth. Using the above algorithms, we worked with them to select and onboard high scoring micro-influencers which drove their platform usage by a staggering 13x within one month.