<xmr-ack[m]> "github.com/MAGICGrants..." <- I dont know why I cant give you a 'tada' too :p

So, you say you can predict the real signer with 23% certainty (instead of 9% 1/11)? Is this number also valid for the 50 real txs? Nice job.

This right here is why Seraphis is important

> > <@ack-j:matrix.org> MAGICGrants/Monero-Fund #15#issuecomment-1145354923... (full message at libera.ems.host/_matrix/media/r0/do…209f33ec33f38f99c8344ab6099e5e7559a)

jeffro256: dangerousfreedom The 23% accuracy was using a preliminary dataset of transaction collected in February. I was able to achieve up to 27% out of sample testing accuracy with a gradient boosted classifier two months back. This dataset is known to have flaws with its collection method and that is why I am collecting a new one as we speak, which should be done by July 1st. The dataset and training of the models however was

done with an undersampled subset of the dataset, eliminating the guess newest heuristic. This also has not been tested against the ~50 mainnet transactions yet, hopefully I will have that done by this or next week. Also I agree, even though these are preliminary results they show the importance of increasing the anonymity set, and a sublinear solution like Seraphis is perfect. If the v15 hardfork didnt get pushed back so far I

would have also collected an additional dataset with size 16 ring signatures to quantify the added privacy benefits.

Cool :)