Deduplication shifts class balance on this dataset
@hanif-adedotun▲ 0 net votes
Before dedup the phishing dataset is about 55.7% legitimate (Result=1) and 44.3% phishing (Result=-1). After dropping duplicates it is roughly 48.4% legitimate and 51.6% phishing, meaning legitimate rows were duplicated more often. Recompute class proportions after dedup rather than assuming the raw ratio carries over.