[Song ngữ Việt-Anh] Buổi 3 lớp FIT-LAB Spring 2026 về AI, ML, DL, CV, NLP và LLM tại NEU

Key Concepts

• Data Splitting & Cross-Validation: Dividing datasets into training, validation, and test sets, and utilizing K-Fold Cross-Validation for robust model evaluation.
• Evaluation Metrics Beyond Accuracy: Recognizing the limitations of accuracy, especially with imbalanced datasets, and focusing on Precision, Recall, and the F1 Score.
• Precision vs. Recall Trade-off: Understanding the inherent conflict between maximizing the quality of positive predictions (Precision) and maximizing the detection of all positive instances (Recall).
• Contextual Prioritization: Recognizing that the optimal balance between Precision and Recall depends on the specific problem and its associated costs and societal implications.
• Importance of Evaluation Before Training: Advocating for a thorough understanding of evaluation metrics before attempting to build and train machine learning models.

---

Data Preparation & Model Evaluation Fundamentals

The discussion begins with foundational concepts in machine learning, emphasizing the critical importance of proper data splitting. A dataset should be divided into three distinct subsets: a training set (for model learning, likened to studying and practice), a validation set (for monitoring model performance during training, like a final exam), and a test set (for final, unbiased evaluation, like a standardized exam). The test set is crucial for comparing different models. To maximize data utilization, K-Fold Cross-Validation (K=5) is introduced. This technique involves dividing the data into k folds, iteratively training the model on k-1 folds and validating on the remaining fold, ultimately averaging the results for a more robust performance estimate. The final training uses the entire dataset.

The Pitfalls of Accuracy & Imbalanced Datasets

A central argument is that accuracy, while seemingly straightforward, is a misleading metric when dealing with imbalanced datasets – where one class has significantly fewer samples than the other. Examples like cancer detection (where prevalence is approximately 2-3 out of 100 people in Vietnam) and spam email detection illustrate this point. A model achieving 97-98% accuracy in cancer detection might still fail to identify a significant portion of actual cancer cases, rendering it practically useless. This highlights the need to move beyond simple accuracy and consider metrics that are more sensitive to the minority class. Supervised learning, encompassing binary classification, multiclass classification, and multilabel classification, is briefly introduced.

Precision, Recall, and the F1 Score

The discussion then delves into Precision (true positives / (true positives + false positives)) and Recall (true positives / (true positives + false negatives)). These metrics represent conflicting goals: maximizing quality (Precision) versus maximizing quantity (Recall). The F1 Score (2 * (precision * recall) / (precision + recall)) is presented as a harmonic mean of Precision and Recall, offering a compromise, but less useful when a clear prioritization is needed. F2 and F0.5 scores are mentioned as less practical alternatives. The concepts are linked to Type I errors (false positives) and Type II errors (false negatives).

Real-World Applications & Contextual Considerations

Numerous real-world examples are used to illustrate the Precision/Recall trade-off. A restaurant recommendation scenario highlights the choice between a short list of guaranteed good restaurants (high Precision) versus a longer list with a higher chance of including bad options (high Recall). During the COVID-19 pandemic, prioritizing Recall was crucial to identify all positive cases, even at the cost of false positives. In criminal justice, the higher cost of a false positive (wrongfully convicting an innocent person) suggests prioritizing Precision. The segment also addresses potential confusion arising from Vietnamese translations of “regression” and “recurrence,” suggesting “hồi tiếp” for recurrence.

The Legal System as a Machine Learning Analogy

The final segment draws a compelling analogy between the legal system and machine learning. The legal process, with its initial court rulings and subsequent appeals, is presented as a multi-stage error correction mechanism. The discussion centers on whether to prioritize minimizing false positives (Precision – avoiding wrongful convictions) or minimizing false negatives (Recall – catching all guilty individuals). The argument is made that in a stable society, prioritizing Precision is paramount, as exemplified by the case of identical twins where lack of conclusive evidence led to their release to avoid a wrongful conviction. The instructor emphasizes that societal stability influences this prioritization; in chaotic times, prioritizing Recall might be considered.

Conclusion

The overarching takeaway is that effective machine learning evaluation requires a nuanced understanding of evaluation metrics beyond simple accuracy. The choice between prioritizing Precision and Recall is not universal, but rather depends entirely on the specific problem, its associated costs, and the broader societal context. Furthermore, the ability to articulate and defend one’s reasoning, even in the absence of a single “correct” answer, is a crucial skill for any data scientist. The legal system serves as a powerful analogy, demonstrating the real-world consequences of these trade-offs and the importance of a robust process for minimizing errors.