While this may not be the best direct answer to your question, I believe Zhou et al. 2016 and Pidsley et al. 2016 can give you possible reasons for filtering probes on the EPIC array.

Other than that, highly overlapping probes may both have cross-hybridization potential. However, I would generally prefer to focus on regions with multiple (preferably, at least 4 probes) within a CpG Island. So, if the probes don't overlap (but are both in the same promoter for a gene), I think that can give you some additional confidence. Also, unless the cross-hybridization is a huge problem, using methylation thresholds (like average beta > 0.7 in one group and average beta < 0.3 in the other group) may help. Those are the sort of things that I have implemented in the COHCAP package.

Check this paper: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4909830/