Monitoring ML Model Performance with Precision and Recall Sampling

PMLE By islever Editorial Last updated

Monitoring, testing, and troubleshooting AI solutions.

Short answer

C Use Precision and recall estimates based on a random sample of raw messages to capture both true positives and false negatives, ensuring a complete view of model performance.

Question

You work for a large social network service provider whose users post articles and discuss news. Millions of comments are posted online each day, and more than 200 human moderators constantly review comments and flag those that are inappropriate. Your team is building an ML model to help human moderators check content on the platform. The model scores each comment and flags suspicious comments to be reviewed by a human. Which metric(s) should you use to monitor the model’s performance?
  1. Number of messages flagged by the model per minute
  2. Number of messages flagged by the model per minute confirmed as being inappropriate by humans.
  3. Precision and recall estimates based on a random sample of 0.1% of raw messages each minute sent to a human for review Correct Answer
  4. Precision and recall estimates based on a sample of messages flagged by the model as potentially inappropriate each minute

Key Takeaways

Sampling only flagged messages (Option D) prevents the calculation of Recall because it ignores False Negatives; a random sample of raw data is required to assess both Precision and Recall accurately.

To monitor an ML model's effectiveness in flagging inappropriate content, use precision and recall estimates derived from a random sample of raw messages. This approach captures both true positives and false negatives, providing a comprehensive performance view.

Common mistake: Candidates often choose D because it focuses on the flagged items, which seems relevant. However, this ignores False Negatives, making it impossible to calculate Recall. Choosing B is also common but only tracks True Positives, lacking context for error rates.

Expert Analysis

The question asks for the appropriate metric to monitor the model’s performance in a real-time setting where humans review flagged comments. The core challenge is estimating precision and recall without labeling the entire dataset, which is impossible.

Option A (Number of messages flagged) measures throughput or volume, not accuracy. It tells you how busy the system is, but not if the flags are correct. Option B (Number of flagged messages confirmed as inappropriate) only counts True Positives. Without knowing False Positives (flagged but safe) or False Negatives (missed bad content), you cannot calculate precision or recall. This is an incomplete picture.

This leaves Options C and D, both of which propose calculating precision and recall from samples. The critical distinction lies in the sampling method:

  • Option D suggests sampling only messages flagged by the model. While this allows you to estimate Precision (TP / (TP + FP)) among flagged items, it completely misses False Negatives. You cannot determine Recall (TP / (TP + FN)) because you have no visibility into the unflagged messages that might be inappropriate. Therefore, Option D fails to provide a complete performance assessment.
  • Option C proposes taking a random sample of raw messages sent to human review. By randomly sampling from the total stream (or a representative subset), you can identify both True Positives (flagged and bad) and False Negatives (unflagged but bad). This allows for the calculation of both Precision and Recall. Although the option mentions sending them "to a human for review," in the context of monitoring, this implies using a ground-truth label set derived from random audits or manual labeling of the sample to compute metrics. The comment section highlights that a random sample avoids bias towards only positive predictions, enabling the detection of missed threats (False Negatives).

Therefore, C is the correct answer because it is the only option that provides a methodology to estimate both key metrics (Precision and Recall) by including potential False Negatives in the evaluation set.

Exam Strategy

When evaluating classification models, always ask: 'Can I calculate both Precision and Recall?' If a sampling strategy excludes unflagged items, it inherently excludes False Negatives, making Recall calculation impossible. Look for options that allow access to the full population distribution (even via sampling) to ensure unbiased metric estimation.

Frequently Asked Questions

Why can't we just count True Positives (Option B)?

Counting only True Positives gives you no information about False Positives or False Negatives. Without these, you cannot derive Precision or Recall, which are essential for understanding model accuracy and completeness.

Is a 0.1% random sample sufficient?

Yes, for large-scale systems like social networks with millions of daily posts, a small random percentage yields a statistically significant number of samples. This allows for reliable estimation of rare events (like inappropriate content) without labeling the entire dataset.

More PMLE FAQ →

Community Votes

C
38%
D
38%
B
24%

38% of anonymous learners picked answer C. Votes are pick records left by other test-takers — they are not the verified answer.

Discussion 44 contributions

Compiled from public community threads and reader submissions, lightly edited for clarity. Handles and vote tallies are shown as recorded, and comments reflect the views of individual contributors.

From public community threads 30 of 44

  1. hiromi 👍 13 Answered D
    D - https://cloud.google.com/natural-language/automl/docs/beginners-guide - https://cloud.google.com/vertex-ai/docs/text-data/classification/evaluate-model
  2. andresvelasco 👍 10 Answered C
    A. Number of messages flagged by the model per minute => NO, no measure of model performance B. Number of messages flagged by the model per minute confirmed as being inappropriate by humans.=> DONT THINK SO, because we need the total number of messages (flagged?) C. Precision and recall estimates based on a random sample of 0.1% of raw messages each minute sent to a human for review. => I think YES, because as I understand it that would be based on a sample of ALL messages not just the ones that have been flagged. D. Precision and recall estimates based on a sample of messages flagged by the model as potentially inappropriate each minute => I think NO, because the sample includes only flagged messages, meaning positives, so you cannot really measure recall.
  3. ludovikush 👍 5 Answered D
    Precision and recall are critical metrics for evaluating the performance of classification models, especially in contexts where both the accuracy of positive predictions (precision) and the ability to identify all positive instances (recall) are important. In this case: Precision (the proportion of messages flagged by the model as inappropriate that were actually inappropriate) helps ensure that the model minimizes the burden on human moderators by not flagging too many false positives, which could overwhelm them. Recall (the proportion of actual inappropriate messages that were correctly flagged by the model) ensures that the model is effective at catching as many inappropriate messages as possible, reducing the risk of harmful content being missed.
  4. tavva_prudhvi 👍 5
    The main issue with option C is that it uses a random sample of only 0.1% of raw messages. This random sample might not contain enough examples of inappropriate content to accurately assess the model's performance. Since the majority of messages on the platform are likely appropriate, the random sample may not capture enough inappropriate content for a robust evaluation.
  5. guilhermebutzke 👍 4 Answered B
    I think B is a complete answer because the answer says "to be inappropriate for humans". So, having this background information, we have a human-checked sample of messages, and we can use that for the following: - message flagged and confirmed as inappropriate as true positive, - message flagged and confirmed as not inappropriate as false negative, - message not flagged and confirmed as inappropriate as false negative, And indirectly, we can use that to calculate metrics like precision and recall. So I think if we have the information verified by humans, the number of flagged messages has enough information to say if the model is good or bad. Answer D is vaguer to me because it doesn't bring up the use of "confirmed to be inappropriate for humans"
Show 25 more comments
  1. josiejojo 👍 4
    But how can you calculate recall with just flagged samples? How could you get a view of false negatives? This is surely key to a problem like this where we don't want to let inappropriate posts go unflagged.
  2. pmle_nintendo 👍 3 Answered D
    Let's consider below hypothetical scenario: Total number of comments per minute: 10,000 Comments actually inappropriate: 500 If we use a random sample of only 0.1% of raw messages (10 comments) for evaluation, there's a high chance that this small sample may not include any or only a few inappropriate comments. As a result, the precision and recall estimates based on this sample may be skewed, leading to unreliable assessments of the model's performance. Thus, C is ruled out.
  3. pshemol 👍 3 Answered D
    First thought B, but number of messages confirmed isn't any metric. (it's not even a precision alone) Its classification problem so precision and recall metrics are the best to decide on performance.
  4. Sas02 👍 3
    Why not C as we need to look at both precision & recall and both B,D miss that and capture only True/ False +ves? It would be very helpful if someone can explain.
  5. Abhijat 👍 3
    Answer is D
  6. Krish6488 👍 2 Answered D
    A --> Conveys model'a activity levels but nit accuracy B --> Accuracy to some extend but wont give full picture as it does not account False negatives C --> Using a random sample of the raw messages allows you to estimate precision and recall for the overall activity, not just the flagged content. D --> Specifically measures on the subset of data that it flagged Both C & D work well in this case, but the specificity is higher in option D and hence will go with D
  7. frangm23 👍 2
    I see your point, but I think that it is implied that if they are calculating precision and recall, they must know which one are actually inappropriate or not. Plus, I don't have right now the links but it is always recommended to use metrics that are not absolute as they are not that meaningful as a percentage can be. That's why I would answer D.
  8. Voyager2 👍 2 Answered D
    D. Precision and recall estimates based on a sample of messages flagged by the model as potentially inappropriate each minute You will need precision and recall to identify fals positives and false negatives. A very small random sample doesn't help specially becasue probably you will have skewed data. So D.
  9. phani49 👍 2 Answered C
    C is correct: A random sample of raw messages provides an unbiased evaluation of the model's performance across all types of content Option D is problematic because: Creates a biased sample by only reviewing flagged messages Cannot detect false negatives (missed inappropriate content)
  10. JamesDoe 👍 2
    - But I guess to implement D we have to fetch the "human result / label", so I guess that's the better option since it uses real world performance but better or more standard metrics? But B or D at least, since we want to monitor real-world performance in real time.
  11. pshemol 👍 2
    It's confusing.. " as potentially inappropriate each minute" so it could be TP or FP only, then it's possible to calculate precision but recall can't be calculated from this subset :)
  12. Werner123 👍 2 Answered D
    C does not make sense to me since it is a very small random sample. It is also only messages that have been sent to humans for review meaning that there is bias in that result set.
  13. PST21 👍 2
    Question is to measure model performance so has to be precision & recall , hence D.
  14. pshemol 👍 2
    then C is best now for me
  15. ares81 👍 2 Answered B
    B is the only way to go!
  16. Mickey321 👍 2 Answered D
    In favor of D
  17. M25 👍 2 Answered D
    Went with D
  18. MithunDesai 👍 2 Answered B
    I think B
  19. tavva_prudhvi 👍 1
    The issue with option B is that it only focuses on the number of messages flagged by the model per minute that are confirmed as being inappropriate by humans, which corresponds to true positives. However, it does not account for other important aspects of the model's performance, such as false positives and false negatives, which are crucial for calculating precision and recall. In option B, we don't have information about: Messages flagged by the model but confirmed as not inappropriate by humans (false positives) Messages not flagged by the model but confirmed as inappropriate by humans (false negatives) Without these pieces of information, we cannot calculate precision and recall to fully evaluate the model's performance.
  20. OpenKnowledge 👍 1 Answered B
    About C... too few and random samples -> possibility of not having enough samples (or even missing samples) for True Positive (i.e. correctly flagged), False Positive (i.e. incorrectly flagged) and False Negative (i.e. incorrectly NOT flagged). So, C is out. About D... only flagged samples are used. Although Precision can be calculated from those samples by using True Positive (i.e., correctly Flagged) and False Positive (i.e. Incorrectly flagged), Recall cannot be calculated from those flagged samples because Recall calculation needs False Negative (i.e., Incorrectly NOT flagged). So, D is out. So, B is the answer
  21. hghdh5454 👍 1
    You make a valid point. Answer B is indeed a more complete answer, as it specifically mentions using the number of messages flagged by the model and confirmed as inappropriate by humans to monitor the model's performance. This approach would provide a more accurate evaluation of the model's precision, which is the proportion of flagged messages that are actually inappropriate. Additionally, the recall metric, which is the proportion of inappropriate messages that are flagged by the model, could also be calculated using this approach. Therefore, answer B is the best choice for monitoring the model's performance.
  22. tavva_prudhvi 👍 1
    This option indeed focuses on true positives (messages flagged by the model and confirmed as inappropriate). However, it does not provide explicit information about false positives (messages flagged by the model but not inappropriate) or false negatives (inappropriate messages not flagged by the model). While you can infer the number of true positives from option B, you cannot directly calculate precision and recall without additional information about false positives and false negatives.
  23. Mohamed_Tareq 👍 1 Answered C
    To measure both precision and recall correctly, you need an unbiased sample of the overall data. By sampling from raw messages (not just flagged ones): -You can identify false negatives (bad comments the model missed) -You can compute true recall Why 0.1% can be sufficient When you have millions of comments per day, 0.1% is actually a large absolute number: 1,000,000 comments/day → 1,000 samples/day 10,000,000 comments/day → 10,000 samples/day
  24. amene 👍 1 Answered B
    I went with B. Remember how to calculate Recall: TP/(TP+FN). Since "sample of messaged flagged by the model" are only P cases, you won't have your F cases reviewed by a human, therefore you won't have FN, therefore it's not D. I also believe that 0.1% of raw messages is going to have too little P cases, therefore not C. And then we remain with option B, which is not optimal, but it is the best we can do in this situation.
  25. d6c984b 👍 1 Answered C
    In order to estimate recall, you need to sample from ALL potential messages. The system is helping a team of 200 human reviewer who were previously performing the job manually. Even by sparing 1 human reviewer (hopefully in rotation), it's possible to review 0.5% of the original throughput capacity. Obviously, once the system is deployed and proves to be efficient, the team of 200 reviewers will shrink.

Reader comments & corrections

No reader corrections yet — spotted an error or have a note? Share it below.

Log in to comment, report an error, or add a note about this question.

Optimizing ML Model Recall for Compliance Risk

S-grade

How to Minimize Development Time for Weekly TensorFlow Retraining in Vertex AI?

S-grade

How to Handle Categorical Features in BigQuery ML for Fast Deployment?

S-grade

How to preprocess BigQuery data for TensorFlow?

S-grade

Practice All PMLE Questions

Access 65 questions with complete answers and detailed explanations.

View Full PMLE Practice Test →

← Back to PMLE Study Guide