The Evaluation stage assesses how well the trained AI model performs on new/test data. Its purpose is to measure the model's accuracy and efficiency, identify errors or weaknesses, and improve the model before deployment. It ensures the AI solution produces reliable and acceptable results in real-world conditions.

Source: AI Project Cycle, Chapter 1, Section 1.1

---

• The question is 2 marks, so ~40–50 words is ideal.
• Key points examiners look for: (1) testing on new data, (2) measuring accuracy/performance, (3) improving the model.
• Avoid writing about other stages; stay focused on Evaluation only.
• The word "briefly" confirms a concise answer is expected — no need for bullet elaboration.

(D) A person without the disease tests negative for the disease.

A True Negative (TN) means the test correctly identifies a healthy person as healthy — no disease present, result is negative. This is the "correct rejection." Contrast with False Positive (healthy person tests positive) and False Negative (sick person tests negative).

A confusion matrix is a table used to evaluate the performance of a classification model. It presents actual values on the y-axis and predicted values on the x-axis. Each cell contains the number of predictions falling into that category — True Positive (TP), True Negative (TN), False Positive (FP), and False Negative (FN).

Source: Chapter 3, Section 3.4 – Evaluation Metrics for Classification

---

• Examiners expect the definition to mention: (1) it is a table/matrix, (2) actual vs. predicted values, and (3) the four cells (TP, TN, FP, FN).
• Avoid writing long descriptions of each cell — just name them for a 2-mark answer.
• The phrase "handy presentation of accuracy of a model with two or more classes" from the textbook is a good addition if space allows.

(c) Recall — Recall is defined as the fraction of positive cases that are correctly identified by the model.

Recall (also called sensitivity) = TP / (TP + FN). It measures how many actual positives were correctly caught. Precision focuses on correctness of positive predictions; Accuracy covers overall correct predictions; F1 balances precision and recall.

(b) When a spam email is mistakenly identified as legitimate.

False Negative (FN) means the model wrongly predicts the positive class as negative — i.e., a spam email (positive) is incorrectly classified as not spam (legitimate).

• False Negative = actual positive, predicted negative. In spam detection, "spam" is the positive class. So FN occurs when a spam email slips through and is labelled as legitimate.
• Option (a) describes True Negative; option (c) describes True Positive; option (d) is not a standard confusion matrix term.
• Remember the pattern: False = wrong prediction; Negative = predicted as the negative class (not spam).

(a) Both Statement 1 and Statement 2 are correct.

A confusion matrix is an evaluation metric (Statement 1) and it is also a handy presentation/record that helps evaluate model accuracy (Statement 2).

Both statements are correct because the confusion matrix serves dual purposes: it is listed as a popular classification metric, and it is described as "a handy presentation of the accuracy of a model" — i.e., a record that aids evaluation. Neither statement contradicts the source material.

(b) Both statement 1 and statement 2 are incorrect.

Statement 1 is incorrect because both precision and recall are needed (combined as F1 score) to fully evaluate a model. Statement 2 is incorrect because when Precision = 1 and Recall = 1, F1 = 2(1×1)/(1+1) = 1, not 0.

• F1 = 2 × (Precision × Recall) / (Precision + Recall). When both = 1, F1 = 1 (perfect score).
• Using only precision or only recall gives an incomplete picture; both are needed together for proper evaluation.
• Examiners expect you to know the F1 formula and apply it to verify such statements quickly.

(C) False Positive (FP)

When the model predicts "Fire Present" (positive) but there is actually no fire (negative), it is a False Positive — wrongly predicting the negative class as positive.

FP = model says "yes" but actual answer is "no." Remember: the first word (False) tells correctness, the second word (Positive) tells what the model predicted. Examiners expect the correct option plus a one-line justification.

(B) The ratio of correctly predicted positive observations to total predicted positive observations.

$$\text{Precision} = \frac{TP}{TP + FP}$$

Precision measures how many of the model's positive predictions were actually correct. Examiners expect the formula alongside the definition. Do not confuse with Recall (TP/TP+FN) or Accuracy (correct predictions/total predictions).

(A) Precision

Precision is defined as the percentage of true positive cases out of all cases where the prediction is positive (predicted true).

Precision = True Positives / (True Positives + False Positives). It measures how many of the predicted positive cases are actually positive. Recall, by contrast, measures true positives out of all actual positives. Students often confuse the two — remember: Precision is about the quality of positive predictions.

True Positive (TP): When the model correctly predicts a positive outcome — i.e., the actual value is positive and the model also predicts it as positive.

This is a 1-mark definition question from the Confusion Matrix topic (Chapter 3). Examiners expect the key idea: correct prediction of a positive class. Mention both conditions — actual is positive AND predicted is positive. Avoid confusing it with False Positive (actual negative, predicted positive).

F1 Score is the harmonic mean of Precision and Recall, used when the dataset is unbalanced and we cannot decide whether FP or FN is more important.

$$\text{F1 Score} = \frac{2 \times \text{Precision} \times \text{Recall}}{\text{Precision} + \text{Recall}}$$

Source: Chapter 3, Section 3.4 – F1 Score

Examiners expect the definition and the formula for full marks. Mention that it combines both Precision and Recall into a single measure — that's the key idea. Avoid writing a long paragraph; one sentence + formula is sufficient for 1 mark.

From the confusion matrix:

• TP = 50, FP = 50, FN = 0, TN = 0

Step 1 – Calculate Precision:

$$\text{Precision} = \frac{TP}{TP + FP} = \frac{50}{50 + 50} = \frac{50}{100} = 0.5$$

Step 2 – Calculate Recall:

$$\text{Recall} = \frac{TP}{TP + FN} = \frac{50}{50 + 0} = \frac{50}{50} = 1$$

Step 3 – Calculate F1 Score:

$$\text{F1 Score} = \frac{2 \times \text{Precision} \times \text{Recall}}{\text{Precision} + \text{Recall}} = \frac{2 \times 0.5 \times 1}{0.5 + 1} = \frac{1}{1.5} = 0.67$$

The F1 Score for this model is 0.67.

Source: AI Chapter – Evaluating AI Models (Confusion Matrix & Performance Metrics)

---

• Examiners expect all three steps to be shown: Precision → Recall → F1.
• Identify TP, FP, FN, TN correctly from the confusion matrix first — that is the foundation.
• Write the formula for each metric and substitute values clearly; step-by-step working earns full marks.
• F1 Score is the harmonic mean of Precision and Recall, used when there is a class imbalance — relevant here since TN = 0.

(b) F1 Score

F1 Score is defined as the harmonic mean of precision and recall, combining both into a single metric. The source explicitly states: "F1-Score provides a way to combine both precision and recall into a single measure that captures both properties." Accuracy measures correct predictions overall; reliability and punctuality are not ML evaluation metrics.

(b) (ii) and (iii)

Recall = TP / (TP + FN). It considers True Positives (ii) and False Negatives (iii), measuring how well the model identifies all actual positives.

---

Recall focuses on minimising missed positives (False Negatives). The formula TP/(TP+FN) uses only True Positives and False Negatives — not False Positives or True Negatives. Students often confuse recall with precision (which uses TP and FP), so remember: Recall = catching all actual positives; Precision = correctness of positive predictions.

When a model is evaluated on the training data, it simply remembers the whole training set and always predicts correctly. This is known as overfitting.

The source passage explicitly states: "our model will simply remember the whole training set, and will therefore always predict the correct label for any point in the training set. This is known as overfitting." Examiners expect the exact term overfitting as the answer.

(C) Statement 1 is correct but Statement 2 is incorrect.

Statement 1 is correct as overfitting means the model memorizes training data. Statement 2 is incorrect because using the same data for training and evaluation leads to overfitting, not accurate results.

The source passage explicitly states: "It's not recommended to use the data we used to build the model to evaluate it… the model will simply remember the whole training set… This is known as overfitting." So Statement 2 is clearly wrong. Statement 1 matches the textbook definition of overfitting directly.

(c) To assess how well the chosen model will work in future.

Model evaluation helps find the best model that represents our data and how well the chosen model will work in the future.

Source: Chapter 3, Introduction & Section 3.1

---

The passage explicitly states: "Model Evaluation… helps to find the best model that represents our data and how well the chosen model will work in the future." This directly matches option (c). For MCQs, always quote or paraphrase the source to show reasoning, but keep it to one line.

(B) To evaluate the model's accuracy.

The testing dataset is used to evaluate the performance of a trained model by comparing its predictions against known expected values.

Source: Chapter 3, Section 3.2 (Train-test split)

The key idea from the passage is that the test dataset's input is fed to the trained model, predictions are made, and those predictions are compared to expected values — this is evaluation. The training dataset is what teaches/trains the model (ruling out option A). Options C and D are unrelated to the purpose of a test split.

Evaluation of an AI Model: Model evaluation is the process of using different evaluation metrics to understand a machine learning model's performance. It acts like a report card, helping identify a model's strengths, weaknesses, and suitability for a task, ensuring it is trustworthy and reliable.

Overfitting: Overfitting occurs when a model performs very well on training data but poorly on new/test data. It means the model has memorised the training data rather than learning general patterns, making it unreliable for real-world use.

Source: Chapter 3 – Model Evaluation, Section 3.1

---

• Evaluation = metrics-based performance check (think: report card analogy — examiners love this).
• Overfitting is a key evaluation concept: the model is "too fitted" to training data. Contrast with training accuracy vs. test accuracy to score full marks.
• Even though overfitting isn't explicitly detailed in the given passages, it is a standard concept tied to train-test split in this chapter — always link it to poor performance on unseen data.

Train-test split is a technique used to evaluate the performance of a machine learning algorithm. It divides the dataset into two subsets:

• Training dataset – used to train/teach the model.
• Testing dataset – input is given to the trained model, predictions are made, and compared with expected values.

This technique is suitable when a sufficiently large dataset is available.

Source: Chapter 3, Section 3.2

Examiners expect two things: (1) a clear definition stating it is an evaluation technique, and (2) the role of each subset (train = learning, test = comparing predictions vs actual). Mentioning that it avoids overfitting is a bonus but not mandatory for 2 marks. Keep it concise.

(A) Both Statement 1 and Statement 2 are correct.

Overfitting means the model memorises training data, so it performs well on training data but poorly on unseen data — making it unsuitable for model evaluation. Both statements correctly describe this problem.

Model Evaluation helps to find the best model that represents our data and how well the chosen model will work in the future.

Source: Chapter 3, Section 3.1 – Importance of Model Evaluation

---

The exact phrase comes directly from the introduction of Chapter 3. Examiners expect the term "Model Evaluation" as the answer. Remember: model evaluation acts like a report card for your AI model, measuring its performance using metrics like accuracy, precision, and recall.

(i) True Negative (TN):
True Negative means the model predicted No earthquake, and in reality there was No earthquake.
From the table: TN = 20

(ii) Calculations:

From the confusion matrix:

• TP = 50, FP = 5, FN = 25, TN = 20
• Total = 100

$$\text{Precision} = \frac{TP}{TP + FP} = \frac{50}{50 + 5} = \frac{50}{55} \approx 0.909$$

$$\text{Recall} = \frac{TP}{TP + FN} = \frac{50}{50 + 25} = \frac{50}{75} \approx 0.667$$

$$\text{F1 Score} = \frac{2 \times \text{Precision} \times \text{Recall}}{\text{Precision} + \text{Recall}} = \frac{2 \times 0.909 \times 0.667}{0.909 + 0.667} \approx \frac{1.212}{1.576} \approx \mathbf{0.769}$$

Source: Confusion Matrix (AI), Class 10 IT/CS

---

• TN is always the "Predicted No & Reality No" cell — easy 1 mark.
• For part (ii), correctly identify all four values (TP, FP, FN, TN) first — examiners award method marks for this.
• Show the formula, substitution, and final value for each metric to secure full marks.
• F1 Score is the harmonic mean of Precision and Recall — remember the formula exactly.

Confusion Matrix:

| | Predicted Positive | Predicted Negative |
|---|---|---|
| Actual Positive | TP = 200 | FN = 55 |
| Actual Negative | FP = 35 | TN = 50 |

• Total predictions = 200 + 50 + 35 + 55 = 340

Source: Evaluation metrics for Classification, Chapter 3

---

• The confusion matrix is a 2×2 table with Actual values on rows and Predicted values on columns.
• Place TP and TN on the diagonal (correct predictions); FP and FN off-diagonal (wrong predictions).
• Examiners expect the correct placement: TP (top-left), FN (top-right), FP (bottom-left), TN (bottom-right).
• Mentioning total is optional but shows understanding.

(a) True Positives (TP) = 300
(Cases where the model correctly predicted a review as Positive, and it actually was Positive.)

---

(b)

Given: TP = 300, FP = 60, FN = 40, TN = 100

$$\text{Precision} = \frac{TP}{TP + FP} = \frac{300}{300 + 60} = \frac{300}{360} \approx 0.833$$

$$\text{Recall} = \frac{TP}{TP + FN} = \frac{300}{300 + 40} = \frac{300}{340} \approx 0.882$$

$$\text{F1-Score} = \frac{2 \times \text{Precision} \times \text{Recall}}{\text{Precision} + \text{Recall}} = \frac{2 \times 0.833 \times 0.882}{0.833 + 0.882} = \frac{1.469}{1.715} \approx \mathbf{0.857}$$

---

• Part (a): TP is simply read directly from the confusion matrix cell (Reality Yes, Predicted Yes).
• Part (b): Remember the three formulas — Precision measures how many predicted positives are correct; Recall measures how many actual positives were caught; F1-Score balances both. Always substitute values clearly step-by-step for full marks.

(B) False Negative (failing to detect actual danger)

In safety-critical systems, missing actual danger (False Negative) is most critical to minimize, as undetected hazards can cause accidents or loss of life.

• The source passage highlights that in medical diagnosis, avoiding missed detections (false negatives) is prioritised over overall accuracy — the same principle applies to autonomous vehicle safety.
• False Negative = model predicts "no danger" when danger actually exists → catastrophic consequence.
• False Positive = model detects danger when there isn't any → inconvenient but not life-threatening (vehicle brakes unnecessarily).
• Examiners expect you to link the concept of minimising error to the real-world consequence of the error type, not just define the term.

Source: Chapter 3, Section 3.3 — Accuracy and Error

True.

In machine learning, the error is used to see how accurately the model can predict data it uses to learn new, unseen data.

The passage under the Error section explicitly states: "In Machine Learning, the error is used to see how accurately our model can predict data it uses to learn new, unseen data." So the statement is directly supported by the source. Just quote or paraphrase this line for full marks.

(C) 800

Total correct predictions = TP + TN = 200 + 600 = 800

Correct predictions include True Positives (correctly predicted positive) and True Negatives (correctly predicted negative). FP and FN are incorrect predictions and are not counted. Always use: Correct = TP + TN.

(A) True Negatives (TN):

True Negative = cases where the model predicted No (Real News) and the reality was also No (Real News).

TN = 420

---

(B) Precision, Recall and F1-Score:

From the matrix:

• TP = 45, FP = 15, FN = 20, TN = 420

$$\text{Precision} = \frac{TP}{TP + FP} = \frac{45}{45 + 15} = \frac{45}{60} = \textbf{0.75}$$

$$\text{Recall} = \frac{TP}{TP + FN} = \frac{45}{45 + 20} = \frac{45}{65} \approx \textbf{0.692}$$

$$\text{F1-Score} = \frac{2 \times \text{Precision} \times \text{Recall}}{\text{Precision} + \text{Recall}} = \frac{2 \times 0.75 \times 0.692}{0.75 + 0.692} = \frac{1.038}{1.442} \approx \textbf{0.72}$$

---

• TN = correctly predicted negatives (Real News predicted as Real News) → 420.
• Precision measures how many predicted "Fake" were actually fake (avoids false alarms).
• Recall measures how many actual "Fake" were correctly caught (avoids misses).
• F1-Score is the harmonic mean of Precision and Recall — useful when classes are imbalanced.
• Always identify TP, FP, FN, TN first before computing metrics; examiners award 1 mark per formula+substitution+result.

(A) Confusion Matrix:

| | Predicted: Attack | Predicted: Not Attack |
|---|---|---|
| Actual: Attack | 1000 (TP) | 50 (FN) |
| Actual: Not Attack | 200 (FP) | 250 (TN) |

(B) True Negative (TN):

True Negative = cases correctly identified as not cyber attacks = 250

(C) Precision:

$$\text{Precision} = \frac{TP}{TP + FP} = \frac{1000}{1000 + 200} = \frac{1000}{1200} \approx 0.833 \text{ (or 83.3\%)}$$

---

• TP (True Positive): correctly predicted attacks → 1000
• TN (True Negative): correctly predicted non-attacks → 250
• FP (False Positive): predicted attack, actually not → 200
• FN (False Negative): predicted not attack, actually was → 50
• Precision measures "of all predicted positives, how many were actually positive" — examiners expect the formula and substitution shown clearly.

(A) False Positive

A False Positive occurs when the model predicts a positive outcome, but the actual/true value is negative (incorrect positive prediction).

Since the diagram is not fully visible, the standard exam approach is: False Positive (FP) is the most commonly tested condition in MCQs of this type, where the model predicts "Yes/Positive" but the actual label is "No/Negative." Examiners expect students to know all four confusion matrix conditions — TP, TN, FP, FN — and identify them from diagrams showing predicted vs. actual labels.

(B) True positive and True negative

In both cases, the model's prediction matches reality — correctly predicting positive as positive, and negative as negative.

"True" in True Positive and True Negative means the prediction matches the actual outcome. "False" (as in False Positive/Negative) means there is a mismatch. Examiners expect you to know these definitions from the confusion matrix topic.

(A) Viral Disease Outbreak

A viral disease outbreak has a high FN cost because missing an infected person (predicting negative when actually positive) can lead to undetected spread of disease, causing serious public health consequences.

• False Negative (FN) = model predicts "No Disease/No Risk" but the actual case IS positive.
• In a viral outbreak, failing to identify an infected person is extremely dangerous (they spread the disease further), so FN cost is very high — Recall matters most here.
• Spam (FP cost is higher), Mining and Image Search are not life-critical, so FN cost is comparatively low.
• Examiners expect you to correctly identify the option and give a one-line justification.

Accuracy is an evaluation metric that measures the total number of predictions a model gets right. The accuracy of the model and its performance are directly proportional — the better the performance, the more accurate the predictions.

Formula:

$$\text{Accuracy} = 1 - \text{Error Rate}$$

where $\text{Error Rate} = \dfrac{\text{Error (Actual - Predicted)}}{\text{Actual Value}}$

Accuracy% = Accuracy × 100%

Source: Chapter 3, Section 3.3 — Accuracy and Error

---

• The examiner expects two things: a definition of accuracy AND the formula — both are needed for full 2 marks.
• The formula is presented in the textbook's activity table, so use the step-by-step form: Error → Error Rate → Accuracy = 1 − Error Rate.
• Do not confuse this regression-style accuracy formula with classification accuracy (correct predictions / total predictions) — stick to what the source passage shows.

False Positive (FP): A legitimate transaction is incorrectly flagged as fraudulent. This causes inconvenience to genuine customers whose valid transactions are blocked unnecessarily.

False Negative (FN): A fraudulent transaction is incorrectly classified as legitimate. This is more dangerous as actual fraud goes undetected, causing financial loss.

The examiner expects you to clearly define both terms in context (fraud detection), not in general. Always mention the real-world consequence — for fraud detection, FN is costlier because actual fraud is missed. This shows conceptual understanding, which CBSE rewards. Remember: FP = wrongly accused legitimate; FN = missed actual fraud.

(A)
True Negative (TN) = 410
These are cases where the model predicted No and the reality was also No.

---

(B)

Given: TP = 110, FP = 60, FN = 50, TN = 410

Precision = TP / (TP + FP) = 110 / (110 + 60) = 110 / 170 ≈ 0.647

Recall = TP / (TP + FN) = 110 / (110 + 50) = 110 / 160 ≈ 0.6875

F1 Score = 2 × (Precision × Recall) / (Precision + Recall)
= 2 × (0.647 × 0.6875) / (0.647 + 0.6875)
= 2 × 0.4448 / 1.3345
≈ 0.667

Source: Confusion Matrix & Model Evaluation, AI Chapter

---

• TN is straightforward — read directly from the matrix (Prediction No / Reality No).
• For Precision, use only the "Prediction Yes" row (TP + FP in denominator).
• For Recall, use only the "Reality Yes" column (TP + FN in denominator).
• F1 Score is the harmonic mean of Precision and Recall — always show the formula first, then substitute values. Examiners award marks for correct formula + correct substitution + correct answer.

Two conditions when prediction matches with the reality are True Positive and True Negative.

The confusion matrix has four outcomes: TP, TN, FP, FN. Only TP (predicted positive, actually positive) and TN (predicted negative, actually negative) are cases where prediction matches reality. FP and FN are errors.

(B) Accuracy

Accuracy is defined as the percentage of correct predictions out of all the observations.

The question directly defines Accuracy — the ratio of correct predictions to total predictions, expressed as a percentage. Precision = TP/(TP+FP); Recall = TP/(TP+FN); F1 combines both. Don't confuse them.

When the prediction matches the reality, the condition is termed as True Positive (TP) (or a correct prediction / accurate prediction).

Source: Chapter 3, Section 3.3 – Accuracy and Error; Section 3.4 – Confusion Matrix

---

The passage defines True Positive as "the outcome of the model correctly predicting the positive class," meaning prediction matches reality. More broadly, the chapter links this to accuracy — when predictions match actual values, they are counted as correct (TP or TN). Examiners accept either "True Positive" (if context is confusion matrix) or "accurate/correct prediction" (if context is general accuracy). Write whichever fits the context of your question paper.

(D) When a defective product is mistakenly identified as non-defective.

A false negative means the model fails to detect a positive (defective) case, incorrectly labelling it as negative (non-defective).

• False Negative (FN): The actual condition is positive (defective), but the model predicts negative (non-defective) — the defect is missed.
• Option (A) = True Positive; Option (B) = False Positive; Option (C) = True Negative.
• Remember the key rule: "False Negative = missed detection of a real positive case." This concept is directly linked to Recall/Sensitivity, which penalises false negatives.

(a) Accuracy

Accuracy measures how true (correct) the predictions made by a model are — i.e., the total number of predictions a model gets right.

The passage explicitly states: "Accuracy is an evaluation metric that allows you to measure the total number of predictions a model gets right." Examiners expect you to recall this definition directly. Reliability is not a standard ML metric; Recall measures how many actual positives were found; F1 score balances precision and recall — none of these describe "how true" predictions are.

The correct answer is (B) It is a Model Training Matrix.

A Confusion Matrix is used for model evaluation, not model training. It records comparisons between predictions and actual values, helping us understand and evaluate model results.

The examiner expects you to identify the false statement about a Confusion Matrix. Options A, C, and D are all true — the Confusion Matrix helps understand predictions, evaluates ML models, and compares prediction vs. reality. Option B is false because a Confusion Matrix is an evaluation tool, not a training tool. Remember this distinction clearly.