课程详情

In this Machine Learning Interview MCQ Course, students will gain a robust understanding of fundamental and advanced ML concepts, specifically tailored for interview success. Topics covered are:-I. Fundamentals and Core Concepts (Difficulty: Easy to Medium)These topics are essential building blocks and are frequently asked to gauge your basic understanding.1. Introduction to Machine Learning (20 MCQs)What is Machine Learning? Definition, goals, and applications.Types of Machine Learning:Supervised Learning: Definition, examples (classification, regression), labeled data.Unsupervised Learning: Definition, examples (clustering, dimensionality reduction), unlabeled data.Semi-supervised Learning: Brief overview.Reinforcement Learning: Brief overview (focus on the concept, not algorithms).AI vs. ML vs. Deep Learning: Understanding the hierarchy and distinctions.Bias-Variance Trade-off:Bias: Definition (simplifying assumptions), impact on model (underfitting).Variance: Definition (sensitivity to training data), impact on model (overfitting).Trade-off: Why they cannot be simultaneously minimized, strategies to balance.Irreducible Error: Understanding the inherent noise.Overfitting and Underfitting:Causes: Too complex model, insufficient data (overfitting); too simple model, poor features (underfitting).Detection: Learning curves (training vs. validation error).Remedies: Regularization, cross-validation, more data, simpler/complex model, feature engineering.2. Data Preprocessing and Feature Engineering (50 MCQs)Importance of Data Quality: "Garbage in, garbage out."Handling Missing Values:Deletion (Row-wise, Column-wise): When to use, drawbacks.Imputation (Mean, Median, Mode): Pros and cons, when to use.Advanced imputation techniques (K-NN imputation, MICE - conceptual understanding).Handling Outliers:Detection: Z-score, IQR, Box plots.Treatment: Capping/Winsorization, transformation, removal, using robust models (e.g., tree-based).Feature Scaling:Normalization (Min-Max Scaling): Formula, use cases.Standardization (Z-score normalization): Formula, use cases, particularly for algorithms sensitive to scale (e.g., K-NN, SVM, Logistic Regression, Gradient Descent).Robust Scaling: Handling outliers.Handling Categorical Variables:One-Hot Encoding: When to use, dummy variable trap.Label Encoding: When to use, ordinal vs. nominal.Ordinal Encoding: Specific use cases for ordinal data.Binary Encoding, Feature Hashing, Target/Mean Encoding: Conceptual understanding, use cases for high cardinality features.Feature Engineering Techniques:Creating new features from existing ones (e.g., polynomial features, interaction terms, date/time features).Domain knowledge importance.Binning/Discretization: Converting continuous to categorical.Log Transformation: Handling skewed data.Dimensionality Reduction:Curse of Dimensionality: Definition, impact on algorithms.Feature Selection vs. Feature Extraction: Key differences.Feature Selection Methods:Filter Methods: Correlation (Pearson, Spearman), Chi-squared, Information Gain.Wrapper Methods: Recursive Feature Elimination (RFE), Forward/Backward Selection (conceptual understanding).Embedded Methods: Feature importance from tree-based models (e.g., Random Forest, Gradient Boosting).Feature Extraction Methods:Principal Component Analysis (PCA):Goal: Find orthogonal components that capture maximum variance.Eigenvectors and Eigenvalues: Role in PCA.Scree Plot: Determining number of components.Assumptions and limitations.When to use.3. Probability and Statistics for ML (40 MCQs)Descriptive Statistics:Measures of Central Tendency: Mean, Median, Mode.Measures of Dispersion: Variance, Standard Deviation, Range, IQR.Skewness and Kurtosis.Inferential Statistics:Probability Distributions: Bernoulli, Binomial, Poisson, Uniform, Normal (Gaussian) distribution.Central Limit Theorem: Importance and implications.Bayes' Theorem: Formula, intuition, application in Naive Bayes.Hypothesis Testing: Null and alternative hypotheses, p-value, significance level, Type I and Type II errors.A/B Testing: Design and interpretation.Correlation vs. Causation: Understanding the difference.Sampling Techniques: Random sampling, stratified sampling.II. Supervised Learning Algorithms (Difficulty: Medium to Hard)This section covers the most common supervised learning algorithms. For each, you should understand the underlying principles, assumptions, strengths, weaknesses, and hyperparameter tuning.1. Regression Algorithms (70 MCQs)Linear Regression:Assumptions: Linearity, independence of errors, homoscedasticity, normality of residuals, no multicollinearity.Cost Function: Mean Squared Error (MSE), Residual Sum of Squares (RSS).Optimization: Ordinary Least Squares (OLS), Gradient Descent (Batch, Stochastic, Mini-batch - conceptual understanding).Coefficients Interpretation: What they represent.Evaluation Metrics: MSE, RMSE, MAE, R-squared, Adjusted R-squared.Polynomial Regression: When to use, overfitting risk.Regularization Techniques (for Linear Models):L1 Regularization (Lasso):Penalty term: Sum of absolute values of coefficients.Effect: Feature selection (sparse models), coefficients can become zero.Geometric interpretation.L2 Regularization (Ridge):Penalty term: Sum of squared values of coefficients.Effect: Shrinks coefficients towards zero, prevents overfitting.Geometric interpretation.Elastic Net: Combination of L1 and L2.Why regularization? Combat overfitting, reduce model complexity.2. Classification Algorithms (100 MCQs)Logistic Regression:Not a Regression Algorithm: Why it's a classification algorithm.Sigmoid Function: Role in transforming output to probabilities.Odds and Log-Odds.Cost Function: Cross-Entropy / Log Loss.Decision Boundary.Assumptions: Independence of errors, absence of multicollinearity (less strict on linearity).Multi-class Classification: One-vs-Rest (OvR), One-vs-One (OvO).K-Nearest Neighbors (K-NN):Non-parametric, Lazy Learner.How it works: Distance metrics (Euclidean, Manhattan), K-value selection.Strengths: Simple, no training phase.Weaknesses: Computationally expensive at prediction time, sensitive to scale and outliers, Curse of Dimensionality.Support Vector Machines (SVM):Goal: Find optimal hyperplane to maximize margin.Support Vectors: Definition and role.Hard vs. Soft Margin: Handling noisy data and non-separable classes.Kernel Trick:Why it's needed (non-linear separability).Common Kernels: Linear, Polynomial, Radial Basis Function (RBF/Gaussian).Role of gamma and C parameters.Strengths: Effective in high-dimensional spaces, robust.Weaknesses: Can be slow on large datasets, black-box interpretation (especially with complex kernels).Decision Trees:How they work: Recursive partitioning, splitting criteria (Gini Impurity, Information Gain/Entropy).Tree construction: Root node, internal nodes, leaf nodes.Stopping Criteria: Max depth, min samples split, min samples leaf.Strengths: Interpretable, handles mixed data types, no feature scaling needed.Weaknesses: Prone to overfitting, sensitive to small data changes (high variance).Pruning: Pre-pruning and Post-pruning to avoid overfitting.Naive Bayes:Bayes' Theorem Application: Conditional probability, prior, likelihood, posterior."Naive" Assumption: Conditional independence of features.Types: Gaussian, Multinomial, Bernoulli.Strengths: Simple, fast, works well with high-dimensional data (text classification).Weaknesses: Strong independence assumption rarely holds in reality.Laplace Smoothing: Handling zero probabilities.III. Ensemble Methods (Difficulty: Medium to Hard)Ensemble methods are crucial for improving model performance and robustness.1. General Concepts (30 MCQs)What is Ensemble Learning? Combining multiple models.Advantages: Improved accuracy, reduced overfitting, increased robustness.Types of Ensemble Methods: Bagging, Boosting, Stacking (conceptual differentiation).Weak Learners: Definition, role in boosting.2. Bagging (Bootstrap Aggregating) (40 MCQs)Bootstrap Sampling: Random sampling with replacement.Parallel Training: Independent models.Aggregation: Averaging (regression), Voting (classification).Reduction in Variance: How bagging helps.Random Forest:Building multiple Decision Trees: Each trained on a bootstrap sample.Feature Randomness: Random subset of features considered at each split.Out-of-Bag (OOB) Error: How it works, utility for validation.Feature Importance: How Random Forest calculates it.Strengths: Robust to overfitting, handles high dimensionality, good performance.Weaknesses: Less interpretable than single decision trees, can be computationally expensive.3. Boosting (Sequential Training) (40 MCQs)Sequential Training: Models built iteratively, correcting errors of previous models.Reduction in Bias: How boosting helps.AdaBoost (Adaptive Boosting):Weighting misclassified samples.Combining weak learners (e.g., decision stumps).Gradient Boosting (GBM):Training weak learners (typically decision trees) on the residuals (errors) of previous predictions.Concept of pseudo-residuals.Learning Rate parameter.XGBoost, LightGBM, CatBoost:High-level understanding of their advantages over traditional GBM (e.g., speed, handling categorical features, regularization).Knowing the key differences (e.g., tree growth, handling categorical features).IV. Unsupervised Learning Algorithms (Difficulty: Medium)Focus on the goals, mechanisms, and evaluation of these algorithms.1. Clustering (50 MCQs)What is Clustering? Grouping similar data points without labels.Use Cases: Customer segmentation, anomaly detection, document analysis.K-Means Clustering:Algorithm Steps: Initialization of centroids, assignment, update.Distance Metric: Euclidean distance.Choosing K: Elbow method, Silhouette Score.Strengths: Simple, fast, scalable.Weaknesses: Sensitive to initial centroids, assumes spherical clusters, sensitive to outliers, requires pre-defined K.Hierarchical Clustering:Agglomerative (Bottom-up) vs. Divisive (Top-down).Dendrogram: Interpretation.Linkage Methods: Single, Complete, Average.Strengths: No need to pre-define K, visual representation.Weaknesses: Computationally intensive for large datasets.DBSCAN (Density-Based Spatial Clustering of Applications with Noise):Density-based: No assumption on cluster shape.Core points, Border points, Noise points: Definitions.Parameters: eps (radius), min_samples.Strengths: Finds arbitrarily shaped clusters, robust to outliers.Weaknesses: Difficulty with varying densities, parameter sensitivity.Evaluation Metrics for Clustering (Intrinsic vs. Extrinsic):Silhouette Score: Measures how similar an object is to its own cluster compared to other clusters.Davies-Bouldin Index.Adjusted Rand Index (if ground truth available).V. Model Evaluation and Selection (Difficulty: Medium)Crucial for understanding how to assess and compare models.1. General Concepts (30 MCQs)Training, Validation, and Test Sets: Why and how to split data.Cross-Validation:Purpose: Robust estimation of model performance, reducing overfitting.K-Fold Cross-Validation: Steps, advantages.Stratified K-Fold: For imbalanced datasets.Leave-One-Out Cross-Validation (LOOCV): Conceptual understanding.Time Series Cross-Validation: Specific considerations.Hyperparameter Tuning:Difference between parameters and hyperparameters.Grid Search: Brute-force search.Random Search: More efficient, explores wider range.Bayesian Optimization (conceptual).2. Metrics for Classification (80 MCQs)Confusion Matrix:True Positives (TP), True Negatives (TN), False Positives (FP), False Negatives (FN).Accuracy: Formula, limitations (especially for imbalanced datasets).Precision: Formula, intuition (of all predicted positives, how many are actual positives).Recall (Sensitivity, True Positive Rate): Formula, intuition (of all actual positives, how many did we correctly identify).F1-Score: Formula, harmonic mean of precision and recall, balances both.Specificity (True Negative Rate): Formula, intuition.ROC Curve (Receiver Operating Characteristic):Plotting TPR vs. FPR at various thresholds.Interpretation: Trade-off between sensitivity and specificity.AUC (Area Under the ROC Curve):Interpretation: Probability that a randomly chosen positive instance is ranked higher than a randomly chosen negative instance.Comparison of models.Precision-Recall Curve: When to use instead of ROC (imbalanced datasets).Log Loss (Cross-Entropy): For probabilistic models.Handling Imbalanced Datasets:Resampling techniques: Oversampling (SMOTE), Undersampling.Cost-sensitive learning.Changing evaluation metrics.3. Metrics for Regression (20 MCQs)Mean Absolute Error (MAE): Intuition, less sensitive to outliers.Mean Squared Error (MSE): Intuition, penalizes larger errors more.Root Mean Squared Error (RMSE): Same units as target variable, commonly used.R-squared (Coefficient of Determination): Proportion of variance explained, limitations.Adjusted R-squared: Accounts for number of predictors.VI. Model Interpretability and Explainability (Difficulty: Medium)Understanding how to explain model predictions is increasingly important.1. General Concepts (10 MCQs)Why Interpretability? Trust, debugging, fairness.White-box vs. Black-box models.2. Techniques (20 MCQs)Feature Importance: From tree-based models (e.g., Random Forest, Gradient Boosting).Coefficients: For linear and logistic regression.Partial Dependence Plots (PDP): Show the marginal effect of one or two features on the predicted outcome.LIME (Local Interpretable Model-agnostic Explanations): Explaining individual predictions.SHAP (SHapley Additive exPlanations): Game theory approach to explain individual predictions.VII. Practical Considerations and Best Practices (Difficulty: Medium to Hard)These questions assess your ability to apply ML in real-world scenarios.1. Workflow and MLOps Concepts (20 MCQs)CRISP-DM or similar ML lifecycle.Data Collection, Cleaning, Transformation.Model Training, Evaluation, Deployment, Monitoring.Version Control for Data and Models.Reproducibility.2. Common Challenges and Solutions (20 MCQs)Data Leakage: Definition, causes, prevention.Concept Drift: Definition, detection, mitigation.Model Drift.Cold Start Problem: In recommendation systems (conceptual).Scalability: Handling large datasets.3. Ethical Considerations (10 MCQs)Fairness and Bias in ML Models: Sources of bias, mitigation strategies.Transparency.Privacy.And Much More!!!

课程标签

0人关注该课程

主题相关的课程