Potted meat is viewed differently by different cultures

I've been working on a multi-label email classification model. It's been a frustrating slog, fraught with challenges, including a lack of training data. Labeling emails is labor-intensive and error-prone. Also, I habitually delete certain classes of email immediately after its usefulness has been reduced. I use a CRM-114-based spam filtering system (actually I use two different isntances of the same mailreaver config, but that's another story), which is differently frustrating, but I delete spam when it's detected or when it's trained. Fortunately, there's no shortage of incoming spam, so I can collect enough, but for other, arguably more important labels, they arrive infrequently. So, those labels need to be excluded, or the small sample sizes wreck the training feedback loop. Currently, I have ten active labels, and even though the point of this is not to be a spam filter, “spam” is one of the labels.

Out of curiosity, I decided to compare the performance of my three different models, and to do so on a neutral corpus (in other words, emails that none of them had ever been trained on). I grabbed the full TREC 2007 corpus and ran inference. The results were unexpected in many ways. For example, the Pearson correlation coefficient between my older CRM-114 model and my newer CRM-114 was only about 0.78.

I was even more surprised by how poorly all three performed. Were they overfit to my email? So, I decided to look at the TREC corpus for the first time, and lo and behold, the first spam-labeled email I checked was something I would definitely train all three models with as non-spam, but ham for CRM-114 and an entirely different label for my experimental model.