Diagnosing Model Failures with Error Analysis

1. Beyond a Single Metric: The "Why" of Error Analysis

Imagine you've built a model with 80% accuracy. While this number gives you a general sense of performance, it hides a critical story. Is the 20% error rate distributed evenly, or is your model failing spectacularly on a specific, important subgroup of your data? Systematic error analysis is the process of answering this question.

By breaking down errors, you can:

• Prioritize your work: Focus your time and resources on fixing problems that will actually improve performance, rather than chasing minor issues.
• Build trust and fairness: Ensure your model performs reliably for all user groups, not just the majority.
• Generate ideas for model improvement: The nature of the errors often points directly to solutions, such as feature engineering, data collection, or addressing data quality issues.

The core idea, as articulated by Andrew Ng, is to turn the "black magic" of model improvement into a systematic engineering discipline.

2. A Practical Workflow for Manual Error Analysis

One of the most powerful and insightful ways to begin error analysis is by manually inspecting your model's mistakes. This hands-on approach builds deep intuition about your model's behavior.

Andrew Ng provides a clear, actionable workflow for this process.

Let's formalize the workflow from the video:

Step 1: Get a Sample of Misclassified Examples
From your development or test set, collect a manageable number of examples that your model got wrong (e.g., 100).

Step 2: Manually Review and Categorize Errors
Create a spreadsheet or table to track your findings. For each misclassified example, look at the input data, the true label, and the predicted label. Then, create columns for potential error categories.

For example, if you were building a credit default prediction model, your table might look like this:

Notice that you can—and should—create new categories as you discover patterns.

Step 3: Quantify and Prioritize
After reviewing your sample, calculate the percentage of errors that fall into each category.

• Self-Employed: 45% of errors
• Short Credit History: 30% of errors
• Low Income: 15% of errors
• Incorrectly entered data: 5% of errors

This analysis provides a clear roadmap. Spending a month developing features to better handle self-employed applicants (45% of errors) has a much higher potential impact than cleaning up data entry issues (5% of errors). You have established a ceiling on how much you can improve the model by tackling each problem.

3. Automating Error Identification

Manual review is fantastic for deep insights, but it can be slow. Tools can help accelerate the identification of problematic data cohorts by automatically slicing the data across different features and calculating error rates for each subgroup.

The output of such a tool often looks like a decision tree, where each leaf node represents a data subset with a specific error rate.

The following reading from Microsoft Research demonstrates this process using their Responsible AI toolkit on the well-known UCI Adult Income dataset.

This automated approach helps you quickly find the "hotspots" in your data where the model is struggling, which you can then investigate more deeply.

4. From Identification to Diagnosis

Once you've identified a failing cohort (e.g., "married individuals with education > 11 years"), the next step is to diagnose the underlying cause. Why is the model failing for this specific group?

The diagnosis could be anything from data quality issues to inherent complexities the model hasn't captured. Let's continue with the Microsoft Research article to see a great example of diagnosis.

This diagnostic step is where you connect the "what" (the failing cohort) with the "why" (the potential root cause). Other common root causes to investigate include:

• Poor Data Quality: Are there more missing values or incorrect labels for this subgroup? (See LINK, "Data" section)
• Feature Representation: Are your features failing to capture important information for this cohort?
• Data Leakage: Did you inadvertently leak information into your training set that doesn't exist for this cohort? (See LINK, "Features" section)
• Concept Drift: Has the underlying data distribution for this cohort changed over time?

5. Error Analysis in Complex Pipelines

So far, we've treated the model as a single block. However, many real-world ML systems are multi-component pipelines. For example, a face recognition system might have separate components for:

1. Background removal
2. Face detection
3. Eye segmentation
4. The final classification model

If the overall system performs poorly, how do you know which component is the bottleneck? You can apply a powerful form of error analysis by simulating a "perfect" version of each component.

This technique is incredibly valuable as it allows you to precisely attribute error and allocate engineering resources in complex systems—a skill highly relevant to building end-to-end projects.

Conclusion

In this lesson, you have learned how to go beyond a single performance metric and systematically investigate your model's failures. This is a cornerstone skill of a professional data scientist.

Key Takeaways:

• Systematic Error Analysis is a structured process to find and understand data subsets where your model performs poorly.
• A practical workflow is to Identify failing cohorts (manually or with tools) and then Diagnose the root causes by forming hypotheses about the data or model.
• Quantifying the "ceiling" for improvement—the maximum gain from fixing a category of errors—is essential for prioritizing your work.
• For multi-component pipelines, you can attribute error by feeding "perfect" ground-truth data into each component sequentially and observing the change in overall accuracy.

You now have the tools to identify where your model is going wrong. In our next lesson, we will focus on the "mitigation" step: using these error analysis findings to guide concrete feature engineering and model improvements.