sharkzero

@argentHQ @Starknet @abdelstark It's good to be chosen for gas free swap week in @avnu_fi . Thanks guys.

gm We are please to introduce the machine learning model that will be used for @theblocktheory_ mint The purpose of this model is to create a list of addresses less likely to be bots but also familiar with Solana NFT trading Theses addresses are then ranked and the higher the rank, the higher the likelihood of having a winning mint, it’s not a luck based raffle system but a merit based raffle system The data We used @flipsidecrypto to extract all the Solana addresses that interacted with secondary marketplaces starting 01/01/2023 till 25/02/2024 Besides volume and number of trades, we added more features to feed the machine learning algorithm : - Nb_days_since_first_trade: Calculates the number of days between the first and last trade recorded, it shows how active the wallet is - Avg_nb_minutes_between_trades: frequency of trades allow to better identify bots, we could have false positives with sweeps - Avg_nb_hours_between_hold_trade: this calculate how long the NFT was held before it was sold, the sql query can still be improve on this part The final dataset after cleaning contains 326 023 addresses that traded Solana NFTs The machine learning model We used K-Means Clustering to identify different cluster of addresses and pinpoint bot activity, after filtering out addresses based on these clusters Kmeans clustering “K-Means Clustering divides data into clusters where each point belongs to the cluster with the nearest mean, updating cluster centers based on member points. This process iterates until the clusters are stable and well-separated.” After testing a range from 3 to 10 clusters, 7 clusters provided the most insightful classification Cluster 0 and 1 groups the most regular traders, but we only selected cluster 1 as it has more active traders with higher volume traded Ranking We initially tried using Gradient boosting but it gave weird results, so we went for a more manual approach with a filter filtering of low activity addresses(70k addresses from the Kmeans cluster) and gave a weight for a few normalized features Depending on how much, how often, how long and if an address buys more thans sells NFTs, their ranking increase How are these addresses related to the mint? The mint will have a limit of 5 attempts per address, it fully public there is no whitelist restriction If 1000 addresses out of these 43k address mint, the higher the rank the higher the chance of minting the NFT If only 500 of those try to mint, and some attempted more than once, it will run a second loop In total there will be 5 loops over these address, and if there is still spots left, it goes to public where it’s luck based We think this is a better approach than a pure raffle based system as it favors ppl that try more attempts and have more funds, and it rewards active and long term holders This is the first iteration of the list, it will be improved and also it will be open to other data scientist Wen mint? This week, doing a few more testing on the onchain program and we will announce a date All the dataset and script are open source and on github, link on second tweet