An Overlooked Baseline Artifact in Comparing Specialized and Pooled Classifiers

When a system must detect several related conditions over the same inputs, its designer can train a specialized classifier for each condition or one pooled model for all of them. Published comparisons of these designs contradict one another, and certain dataset properties, class imbalance above all, usually take the blame. We re-examine the comparison on code-smell detection, a domain where the contradiction is well documented, testing three ways of building the pooled baseline across five code-smell datasets, four generic benchmarks, four classical classifier families, and a neural network, with matched cross-validation folds and threshold-independent scoring. Much of the disagreement turns out to be manufactured by the comparison itself. The conventional pooled baseline appends a task indicator to the features, which quietly forces linear models, and neural networks conditioned only through the input, to share one set of feature weights across tasks; tree ensembles and softmax classifiers, which carry per-task structure of their own, are unaffected. Fixed-threshold F1 scoring accounts for the rest. Once the pooled model receives per-task weights, the gap closes on every dataset, and controlled positive-rate sweeps show that class imbalance moderates nothing. For tabular task families of this kind, the choice between the designs is an engineering decision, not an accuracy one; we close with a short protocol for making such comparisons fairly.