The problem Persona bot-wallet detection solves

Persona helps dapps acquire new users via marketing campaigns. Persona focuses on quality campaigns and provides the best ROI to the dapps. In the recent past, we see that most of these airdrop marketing campaigns are claimed by bots. So although there is a high engagement with the new user acquisition campaign, the dapps don't see any improvements to their daily active users. This is huge monetary loss as well.

For this hackathon, Persona built a bot detection model that would help Persona to weed out the bots from the campaign's target audience. This in turn leads to better click-through and conversion rates and provides better value to dapps from these user acquisition marketing campaigns.

Today most dapps are broken or abandoned due to lack of real users. To build sustainable dapps there is a need for real users. To understand dapp performance, we can perform bot analysis on the current users of a dapp to identify real users vs bots, this metric is more valuable than just daily active users.

Challenges we ran into

Challenges we ran into:

1. Data availability: We currently don't have bot labels to perform supervised learning. We overcame this problem by identifying bots and labeling them based on anomalies detected in the average time between two send transactions. This approach is based on our literature review - https://arxiv.org/abs/1810.01591

2. Data size: There are 186M wallets in the ethereum transaction table. Working with this huge data set with inactive wallets adds noise to the ML models. For the purpose of this hackathon and to create a proof of concept, we picked the active users of a popular dapp - Opensea.

3. Accuracy calculation: We don't have bot labels to compute the accuracy of our model. We applied transfer learning, that is, trained the model using data from November and predicted labels on data from October. These predicted labels were compared with the labels from the time series anomaly detection approach.