MLS-C01 Question 83
Single answerA data science team is building a supervised machine learning model to classify customer reviews as 'Positive', 'Negative', or 'Neutral'. They have a dataset of 10,000 reviews, 7,000 of which are labeled while the remaining 3,000 are unlabeled. The team wants to ensure there is sufficient labeled data before training their model. Which approach would provide the best assessment of whether the labeled data is sufficient?
- A
Analyze the class distribution of the labeled dataset and check for class imbalance.
- B
Use a rule-of-thumb heuristic, such as having at least 1,000 labeled examples per class.
- C
Train the model with the current labeled data and evaluate its performance using a validation set.
- D
Use unsupervised learning to label the remaining 3,000 unlabeled examples and include them in the training set.
Show answer and explanation
Correct answer: C
Explanation
The sufficiency of labeled data should be evaluated by training the model and assessing its performance on a validation set. This approach provides concrete evidence of whether the labeled data supports the model in achieving acceptable accuracy and generalization. Other methods, such as heuristics or analyzing class distribution, can provide additional insights but do not directly address sufficiency in the context of model performance.
- A. Incorrect.
Analyzing class distribution and checking for class imbalance is important but does not directly determine whether the amount of labeled data is sufficient for training the model.
- B. Incorrect.
Using a heuristic like having 1,000 examples per class is too generic and does not consider the specific characteristics of the dataset or the problem being solved.
- C. Correct.
Training the model with the current labeled data and evaluating its performance on a validation set is the most effective way to determine if the labeled data is sufficient. Poor performance may indicate insufficient labeled data.
- D. Incorrect.
Using unsupervised learning to label the remaining data might introduce noise and bias, as these labels would not be manually verified. This does not directly evaluate whether the current labeled data is sufficient.