balanced reviews
Machine learning, without the theatre
Can you trust
what you read?
An honest detector for computer-generated product reviews—built to show its reasoning, its limits, and the experiment behind every result.
Étude № 01
Sep. 2026
94.88%
Held-out accuracy
across 8,087 unseen reviews
Private by design
Inference happens locally. No account, API, cookie, or uploaded text.
139,295 learned termsThe evidence
Representation beats
optimisation.
Five controlled configurations. One held-out split. Changing the text representation moves accuracy by 41 points; changing the learning algorithm moves it by less than one.
02
unseen reviews
best model
surface → TF-IDF
Held-out accuracy
The comparison
The finding
Features mattered roughly fifty times more than the optimiser.
The from-scratch Adaline and scikit-learn baseline finish within 0.41 percentage points on identical TF-IDF features. That validates the custom gradient descent while keeping the conclusion appropriately modest.
Representation~41 pp
Algorithm<1 pp
Error analysis
Where the best model misses.
TF-IDF + Adaline · 8,087 held-out reviews
Predicted genuine
Predicted fake
Actually genuine
3,900correct
144false positive
Actually fake
270false negative
3,773correct
*
Conservative by accident,
Conservative by accident,
not by design.
The model produces fewer false positives than false negatives—a useful tendency for moderation, where accusing a real customer is the costlier error. A production system should tune that trade-off explicitly.
Under the bonnet
Four lines of maths.
No black box.
Adaline learns a linear decision boundary by minimising mean squared error. The optimisation is written from scratch in NumPy; scikit-learn supplies only pipeline plumbing and the comparison baseline.
03
Represent
Convert reviews into sparse word and phrase weights with TF-IDF.
→
Learn
Update 139,295 weights with transparent batch gradient descent.
→
Decide
Threshold the continuous Adaline activation at 0.500.
Objective
L(w,b) = 1/n Σ (yᵢ − (xᵢᵀw + b))²
A differentiable, convex loss surface with a global optimum—provided the learning rate is stable.
adaline.pyPython
errors = y - self.activation(
self.net_input(X)
)
self.w_ += self.lr * 2.0 * (
X.T @ errors
) / n_samples
self.b_ += self.lr * 2.0 * errors.mean()Sparse by necessity
The TF-IDF matrix would need roughly 35 GB if densified. Every operation preserves SciPy sparse matrices end to end.
Divergence fails loudly
A large learning rate can explode while remaining a finite float. Explicit guards reject the run instead of returning silent garbage.
Limits are part of the result
One dataset, one GPT-2 generator, one split. This is a research demonstration—not a universal truth machine.