Data Science Cheat Sheet
Models
Support Vector Machine (SVM): A supervised model that finds the optimal hyperplane for class separation, widely used in high-dimensional tasks like text classification (e.g., spam detection).
k-Nearest Neighbors (kNN): A non-parametric method that classifies based on nearest neighbors, often applied in recommendation systems and image recognition.
Decision Tree: A model that splits data into branches based on feature values, useful for interpretable applications like customer segmentation and medical diagnosis.
Linear Regression: A statistical technique that predicts a continuous outcome based on linear relationships, commonly used in financial forecasting and trend analysis.
Logistic Regression: A classification model estimating the probability of a binary outcome, widely used in credit scoring and binary medical diagnostics.
Naive Bayes: A probabilistic classifier assuming feature independence, effective in text classification tasks like spam filtering due to its speed and simplicity.
Confusion Matrix and F1 Score
| Predicted Positive | Predicted Negative | |
|---|---|---|
| Actual Positive | True Positive (TP) | False Negative (FN) |
| Actual Negative | False Positive (FP) | True Negative (TN) |
F1 Score = 2 * (Precision * Recall) / (Precision + Recall)
- 2 * (Positive Predictive Value * True Positive Rate) / (Positive Predictive Value + True Positive Rate)
- 2 * (TP) / (TP + FP + FN)
Key Evaluation Metrics
True Positive Rate (TPR), Sensitivity, Recall
- TPR = Sensitivity = Recall = TP / (TP + FN)
- Application: Measures the model's ability to correctly identify positive cases, useful in medical diagnostics to ensure true positives are detected.
Precision (Positive Predictive Value)
- Precision = TP / (TP + FP)
- Application: Indicates the proportion of positive predictions that are correct, valuable in applications like spam filtering to minimize false alarms.
Specificity (True Negative Rate, TNR)
- Specificity = TNR = TN / (TN + FP)
- Application: Assesses the model's accuracy in identifying negative cases, crucial in fraud detection to avoid unnecessary scrutiny of legitimate transactions.
- FPR = FP / (FP + TN)
- Application: Reflects the rate of false alarms for negative cases, significant in security systems where false positives can lead to excessive interventions.
Negative Predictive Value (NPV)
- NPV = TN / (TN + FN)
- Application: Shows the likelihood that a negative prediction is accurate, important in screening tests to reassure negative cases reliably.
- Accuracy = (TP + TN) / (TP + TN + FP + FN)
- Application: Provides an overall measure of model correctness, often used as a baseline metric but less informative for imbalanced datasets.
Curves & Chart
- X-axis: Percent of data (typically population percentile or cumulative population)
- Y-axis: Lift (ratio of model's performance vs. baseline)
- Application: Helps in evaluating the effectiveness of a model in prioritizing high-response cases, often used in marketing to identify segments likely to respond to promotions.
- X-axis: Percent of data (typically cumulative population)
- Y-axis: Cumulative gain (proportion of positives captured)
- Application: Illustrates the cumulative capture of positive responses at different cutoffs, useful in customer targeting to assess the efficiency of resource allocation.
- X-axis: Percent of data (cumulative population)
- Y-axis: Cumulative response (actual positives captured as cumulative total)
- Application: Evaluates model performance by showing how many true positives are captured as more of the population is included, applicable in direct marketing to optimize campaign reach.
- X-axis: False Positive Rate (FPR)
- Y-axis: True Positive Rate (TPR or Sensitivity)
- Application: Used to evaluate the trade-off between true positive and false positive rates at various thresholds, crucial in medical testing to balance sensitivity and specificity.
- X-axis: Recall (True Positive Rate)
- Y-axis: Precision (Positive Predictive Value)
- Application: Focuses on the balance between recall and precision, especially useful in cases of class imbalance, like fraud detection or medical diagnosis, where positive class accuracy is vital.
