The point of all this is that a single transaction that connects two addresses together is not necessarily enough to link two businesses together, absent additional evidence.
The paper addresses this issue:
Although someone would use the CoinJoin method [9] to combine UTXOs from multiple senders into a single transaction to make it more challenging to determine the relationship between input and output addresses, we detect this method has not been adopted by the exchange so far.
Also, a classification model that is accurate 96% of the time (it is unclear how you are measuring accuracy) has very high accuracy. My first reaction to that high of claimed accuracy is that you might have
data leakage. I can't point to the source without looking at your specific steps to train your model, which understandably may not be something you want to share.
I suppose they are presenting a model more than a software. So far, the model seems to me to be solid up to the extent that a good heuristic-based data mining model could be. The implementation is not open and it is not good news, so the results presented are highly suspicious.
For example, consider a conspiracy theory to be true: A shady exchange (such as Bittrex) with very low liquidity and a high incentive to put itself in the top 10 list and faking high volumes of trade, as a part of its scam, hires a team of technical writers and they publish an acceptable analysis model and faking privately generated results in favor of the exchange.