MATH7005 Chap.1 Logistic Models and Sigmoid Scores
Logistic Models and Sigmoid Scores
Three layers of one prediction
Logistic classification begins with a linear score, maps that score through the sigmoid and only then applies an operational threshold. The chapter keeps these layers separate so coefficients are not mistaken for probability changes and thresholds are not mistaken for learned parameters.
It also uses boundary and extreme cases to test signs, interpretation and numerical stability.
What to verify
A complete explanation prints the feature values, parameters, score, sigmoid result and decision rule.
Direction checks catch most errors before arithmetic does: positive coefficients raise the score, negative coefficients lower it, and a score of zero always maps to one half.
Interpretation that survives a question
A coefficient belongs to the log-odds scale. Translate its sign first, then use an odds ratio when a magnitude is needed.
The probability effect depends on the starting record because sigmoid is steep near zero and flat at extreme scores.
Comparing coefficients also requires comparable feature units, so document scaling before describing one variable as more influential than another.
Thresholds belong to the product
The fitted model supplies a probability-like output, while the product team selects the cutoff that triggers an action. Examine false positives, false negatives and operational capacity over several candidate thresholds.
A threshold that looks accurate in aggregate may overload staff or miss the rare cases the product was built to protect. Report the selected rule beside every classification metric.
Implementation evidence
Test a negative score, zero and a positive score against a trusted calculation. Add an extreme input to expose overflow, then confirm that the stable implementation returns a finite result.
Preserve preprocessing parameters learned from training data and apply them unchanged at validation and demonstration time. These checks connect the formula to the behaviour of the working product.
What this chapter covers
- 01
Linear score and intercept
- 02
Sigmoid range and monotonicity
- 03
Probability and odds
- 04
Decision thresholds
- 05
Coefficient interpretation
- 06
Stable extreme-score computation
Worked example · free
Score a record and apply the decision rule
- 2Calculate the linear score.
- 3Apply the sigmoid.
- 2Use the stated threshold.
Key terms
- Linear Score
- The unrestricted weighted sum of features and an intercept before any nonlinear mapping.
- Sigmoid Function
- A monotone map from a real score to a value between zero and one.
- Decision Threshold
- The operational cutoff that converts a model output into a class or action.
- Log Odds
- The logarithm of probability divided by one minus probability, which is linear in logistic regression.
- Intercept
- The parameter that shifts every score by the same amount before features contribute.
Logistic Models and Sigmoid Scores FAQ
Why is logistic regression a classifier?
Its linear score is passed through a sigmoid to produce a positive-class probability, and a separate threshold converts that probability into a class. The word regression refers to the mathematical form, not the final task.
Does a coefficient equal a probability change?
No. A coefficient is a change in log odds for one feature unit with other features fixed. The probability effect depends on the starting score and becomes smaller near the saturated ends of sigmoid.
Why can a sigmoid overflow in code?
For a large negative score, a naive formula evaluates an enormous positive exponential. A stable library routine or a sign-aware branch preserves the same mathematics without asking the computer to represent that huge intermediate value.
How should I choose a decision threshold?
Compare candidate cutoffs on held-out records using the operational costs of false positives and false negatives. Report the resulting workload and important misses, then keep the fitted model fixed while the policy rule is evaluated.
What should I inspect when one group receives different predictions?
Check data coverage, feature coding, score distributions and error rates before attributing the difference to one coefficient. Then test whether preprocessing, model specification or the common threshold creates the disparity, using the same validation period for every group.
What belongs in a reproducible prediction trace?
Preserve the raw feature values, fitted preprocessing parameters, coefficient version, linear score, sigmoid result, selected threshold and final action. Include one boundary case and one extreme score so the trace tests both decision logic and numerical stability.
Can a threshold repair poorly calibrated probabilities?
No. Moving the cutoff changes which actions are taken but does not make predicted probabilities match observed frequencies. Check calibration separately on held-out data, then choose an operational threshold using the costs and capacity of the intended decision.
Assessment move
Begin with a four-column ledger for features, coefficients, products and the final score. Predict whether the result should lie above or below one half before evaluating sigmoid. Repeat with the score set exactly to zero, then with an extreme positive and negative value to test the implementation. Next, keep the model output fixed and apply several thresholds so the distinction between estimation and action becomes concrete.
For project preparation, save one representative record, one boundary record and one numerical-stability record as automated tests. Explain each coefficient on the log-odds scale and convert only one carefully chosen coefficient to an odds ratio. Finish by tracing the same record through preprocessing, score, probability and product action without looking at the printed solution.
Working through Logistic Models and Sigmoid Scores in MATH7005? Sia is AskSia’s AI Computer Science tutor — ask any MATH7005 Logistic Models and Sigmoid Scores question and get a clear, step-by-step explanation grounded in how MATH7005 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.
