New Math Method Catches Fake 'Groups' Hiding in Your Data
Standard tools may invent teams that never actually worked together
When researchers study who works with whom — students in group projects, users in online forums, tags on a post — they feed that information into software that hunts for patterns and "holes" (gaps in the structure that can reveal something meaningful, like a missing bridge between two clusters). The standard method builds its picture from pairs: if A overlaps with B, B with C, and A with C, it concludes that all three formed a group. Researchers call that a "clique" — everyone connected to everyone.
The authors, Koichi Yasutake, Wakana Tsuji and Hitoshi Inoue, tested this against a method that only counts sets of things genuinely seen together. They call it the "Dowker complex," named after a mathematician. On four Stack Exchange forums, 46% to 91% of the assumed three-way groups were "phantom" — they never actually occurred together. Worse, the standard method erased between 522 and 1,970 real patterns across those datasets. In one popular teaching-analytics toolkit, the old method spotted zero holes; the new one found them.
Why does this matter in practice? If a school wants to know which study groups really form, or a company wants to see which teams truly collaborate, the old tool can quietly invent connections that never happened. That can skew conclusions about how people learn and work together. The distortion is especially sneaky because it disappears when analysts use the fixed cutoffs that most practitioners rely on day to day.
The takeaway: pick the tool that matches your data. For groups actually observed together, the Dowker method is the right choice — and the authors have posted their code free online so others can use it. Their conclusion is blunt: the standard approach fills a real gap in today's software, and for co-occurrence data, it should probably be the default.
- The standard way of analyzing who appears together assumes three people who pair up must also be a trio — which is often false.
- In real data from four online forums, 46% to 91% of those assumed groups never actually happened together.
- The old method also hid 522 to 1,970 genuine connection patterns, which could mislead researchers studying teamwork.
Why It Matters
Could lead to better teamwork insights — and fewer fake 'collaboration' claims in schools and workplaces.