Baseline Comparisons

A metric needs a reference point. A baseline is a simple, auditable method that answers: what performance is available without the proposed complexity?

Useful baselines depend on the task:

regression       -> predict the training-target mean or median
classification   -> predict the training majority class
tabular features -> linear/logistic model
local structure  -> nearest-neighbor model

Compute every baseline with training information only, then evaluate it on the same held-out examples and metric as the neural model. Using the validation mean or majority label would leak held-out targets.

Two Baselines by Hand

Training regression targets are:

[2, 4, 6]
training mean = 4

On validation targets [3, 7], the mean baseline predicts [4, 4]:

absolute errors = [1, 3]
MAE = (1 + 3) / 2 = 2

For classification, suppose training labels contain 70 cats and 30 dogs. The majority baseline always predicts cat. Its validation accuracy can be high on an imbalanced set while its recall for dogs is zero. That is why the baseline and metric must be interpreted together.

What a Baseline Failure Means

If a neural network does not beat a well-chosen simple baseline, do not assume the network needs more layers. First inspect:

• data and label alignment;
• feature preprocessing and leakage;
• loss/metric agreement;
• optimization and implementation;
• whether the task contains signal beyond the baseline.

Failure to beat a relevant, correctly implemented baseline is evidence against the added complexity—not an instruction to make the network larger. Beating it is still not proof of usefulness: the gain must be large and stable enough to justify added compute, latency, maintenance, and failure modes.