First of all, it’s easy to say “don’t be nervous”, but honestly, being nervous is a good sign. It means you care about the work you’re doing.
Research can be frustrating precisely because you don’t always get the clean or “expected” results you hoped for. And that’s completely okay. Negative results are still results. What matters is that you do your due diligence to make sure your findings aren’t caused by a bug, a flawed experimental setup, or some preprocessing mistake. Once you’ve ruled those out, the results, whether they confirm your hypothesis or not, are something you can legitimately defend.
Now, about your actual situation:
Is it bad if different classifiers give the same metrics?
Not necessarily. This can happen, and it doesn’t automatically mean something is wrong. In particular:
- If the dataset is very simple, linearly separable, or highly imbalanced, you might see several classifiers converging to the same predictions.
- If your hierarchical setup constrains the labels strongly, the base classifiers may have limited room to differ.
- Ontology-derived features can be very sparse or highly correlated, which often causes multiple models to behave similarly.
- It’s also possible that the signal in the data is weak, and only a certain baseline level of performance is achievable.
Your observation that the Dummy Classifier performs noticeably worse is actually reassuring: it means the rest of your models are learning something real.
What this means for your thesis
If your hypothesis was “these classifiers should behave differently”, but your results say otherwise, that’s not a failure. That’s an interesting finding. It gives you room to explore questions such as:
- Why might different classifiers converge on the same decision boundary?
- Is there some inherent property of ontology-based features that reduces model diversity?
- Does the hierarchical/Bayesian structure constrain the predictions so much that model choice matters less?
- Does this suggest something about the nature of the dataset or the task itself?
This kind of analysis can actually strengthen your thesis. Showing that something didn’t work as expected, and explaining why, demonstrates deeper understanding than just getting the “textbook” results.
Bottom line
You don’t need to be afraid of presenting results that don’t match your initial expectations. Your job is to defend your findings, whatever direction they point in. As long as you have:
- validated your pipeline,
- justified your interpretation,
- and connected your results back to the literature,
you’re in a solid position.
And worst case? You’ve discovered something unexpected, that’s literally what research is.
