Frequently Asked Questions About Edureka ML Full Course Roadmap Skill
21 answers covering everything from basics to advanced usage.
// Basics
What is machine learning in simple terms?
Machine learning enables computers to act and make data-driven decisions instead of being explicitly programmed for every task. The machine learns and improves over time as it's exposed to new data. Your job as the programmer is to set up the conditions for learning — choosing the right approach, cleaning the data, and evaluating results — not to hard-code every rule the system follows.
What is the difference between AI, machine learning, and deep learning?
Artificial Intelligence is the broad concept of machines carrying out tasks in a smarter way. Machine Learning is a subset of AI that extracts patterns from data and lets machines adapt. Deep Learning is a subset of ML that uses deep neural networks for cases where standard ML underperforms. Treat them as a strict hierarchy: AI ⊃ ML ⊃ Deep Learning. Conflating them leads to wrong tool selection.
What inputs do I need before starting an ML project?
You need a plain-English problem statement describing what you want to predict, classify, cluster, or detect, plus a description of your available data — labeled or unlabeled, structured or unstructured, its size, and feature count. Optionally, provide domain context (banking, healthcare, retail) and a success metric like classification accuracy or mean squared error. These inputs drive every downstream decision in the roadmap.
What is anomaly detection and when do I use it?
Anomaly detection identifies unusual data points or patterns that don't conform to expected behavior — outliers. Use it for fraud detection, intrusion detection, or flagging abnormal medical scans. In the roadmap, anomaly detection is one of four output-driven algorithm families alongside Classification, Regression, and Clustering. You'd choose it when the goal is spotting rare deviations rather than assigning categories or predicting values.
// How To
How do I decide between classification and regression?
Check your output type. If the output is discrete or categorical — spam or not spam, default or no default — use Classification. If the output is continuous or real-valued — house price, temperature, stock value — use Regression. This is a hard decision gate: applying a classification algorithm to a continuous problem, or regression to a binary problem, produces incorrect results by design.
How do I handle class imbalance in my dataset?
First check class distribution with a groupby on your output column before selecting a scoring metric. If one class vastly outnumbers another, accuracy becomes misleading — a model predicting only the majority class can score high while being useless. Handle imbalance with resampling, class weights, or alternative metrics like precision, recall, and F1, and rely on the confusion matrix rather than raw accuracy.
How do I set up 10-fold cross validation correctly?
Split your dataset into 10 equal parts, train the model on 9 parts and test on the remaining 1, then rotate until all 10 combinations are covered. Average the accuracy scores for a robust estimate. Crucially, use the same fixed random seed across all algorithms so the splits are identical, making mean accuracy directly comparable when choosing between models.
How do I explore my data before modeling?
Use univariate plots — box-and-whisker plots to understand distribution and spread, and histograms to spot Gaussian distributions relevant to algorithm assumptions. Then use multivariate plots — a scatter matrix of all attribute pairs to find correlations and structure. Diagonal grouping in scatter plots suggests high correlation, worth noting for feature engineering. Avoid sharing X and Y axes across subplots unless you're intentionally comparing scales.
// Troubleshooting
My model accuracy is high in training but poor on validation — what's wrong?
This is overfitting — high variance where the model memorizes training data instead of learning general patterns. Combat it with regularization, cross validation to reduce variance, and hyperparameter tuning. Confirm you held out ~20% validation data and never touched it during training. If the gap persists, your model may be too complex, or you may have data leakage from features that encode the target.
Why are my algorithm comparison results inconsistent between runs?
You're likely not fixing a random seed before each model run. Without a consistent seed, your train/test splits differ every run, so accuracy scores aren't directly comparable across algorithms. Set a fixed random seed before every split and every model fit. This makes the comparison reproducible and ensures differences in scores reflect the algorithms, not random split variation.
My model worked in testing but fails in production — how do I fix it?
Production data drifts from training data over time, degrading performance. This is why post-deployment validation is mandatory — users must validate model performance in production, and you retrain when it degrades. Apply MLOps practices: monitoring, lifecycle management, and CI/CD pipelines using tools like MLflow, Docker, or AWS SageMaker. Fix issues before re-releasing, and set up ongoing monitoring rather than deploying once and walking away.
My code breaks with library errors — what should I check?
Check library version compatibility first. Version mismatches in scikit-learn, NumPy, Pandas, or matplotlib cause silent API changes that break code or shift behavior. Verify the exact versions at the start of every project and pin them in a requirements file. Many confusing errors and unexpected results trace back to a library version that differs from the one the tutorial or example assumed.
// Comparisons
How does this roadmap compare to using AutoML tools?
AutoML automates algorithm selection and tuning, but this roadmap builds the understanding AutoML hides — why an output type maps to an algorithm family, why you split validation data, and how bias-variance tradeoff affects results. Use the roadmap to frame the problem correctly and interpret AutoML output critically. AutoML can search faster, but it won't stop you from feeding it a wrong problem definition.
Should I use logistic regression or linear regression for a yes/no prediction?
Use logistic regression. Linear regression entertains values below 0 and above 1, violating the binary constraint of a yes/no output. Logistic regression uses a sigmoid function to squash output into a 0-to-1 probability, then applies a threshold (default 0.5) to produce a discrete class. Using linear regression for classification produces meaningless out-of-range predictions and is a classic mistake this roadmap warns against.
How is clustering different from classification?
Classification is supervised — it uses labeled data to predict a known discrete category. Clustering is unsupervised — it groups unlabeled data by intrinsic similarity without knowing what the groups mean. A clustering algorithm knows which instances are similar but cannot label the clusters; a human expert must interpret them. Choose clustering when you have no labels and want to discover structure, classification when labels already exist.
// Advanced
What's the difference between explicit and implicit data?
Explicit data is entered directly by users — ratings, comments, reviews. Implicit data is generated passively by behavior — purchase history, search history, card details. Both feed recommendation engines and behavioral models. In the roadmap's data acquisition step, identifying which type you have shapes how you clean it and which patterns you can extract, especially for recommendation and personalization use cases.
How does gradient descent minimize the cost function?
Gradient descent iteratively updates model parameters (B0, B1) by computing partial derivatives of the cost function — Mean Squared Error — and moving in the direction that reduces error, continuing until MSE reaches its minimum. Libraries like scikit-learn handle this automatically, but understanding the mechanism helps you diagnose underfitting and overfitting and interpret why a model converges to particular parameter values.
How much project time should I budget for data cleaning?
Budget 50–80% of total project time for data processing and cleaning. Data cleaning is not a one-time step — it's iterative. After importing data into the ML pipeline, duplicate values, null values, and inconsistencies reappear and must be removed again because they cause wrongful predictions. Use describe() to check count, mean, std, min, max, and percentiles to detect anomalies at each pass.
What is the bias-variance tradeoff and how do I manage it?
It's the balance between underfitting (high bias — model too simple, misses patterns) and overfitting (high variance — model too complex, memorizes training data). Both extremes hurt performance on unseen data. Manage it with cross validation to reduce variance, regularization to combat overfitting, and careful hyperparameter tuning. For linear regression prone to multicollinearity, apply dimensionality reduction. Aim for the sweet spot where the model generalizes.
How do I evaluate an unsupervised model when there are no labels?
You can't use accuracy because there's no ground truth to compare against. Instead, evaluate cluster quality with within-cluster similarity metrics that measure how tightly grouped points are and how separated clusters are from each other. A human expert must interpret and validate the resulting groups, since the algorithm knows which instances are similar but cannot assign meaning to the clusters it forms.
What does MLOps add after I deploy a model?
MLOps covers the best practices for maintaining models in production: monitoring for performance drift, lifecycle management, CI/CD pipelines, and automation using tools like Docker, Kubernetes, MLflow, and AWS SageMaker. It ensures deployed models are validated on live data, retrained when performance degrades, and released through controlled pipelines. Without MLOps, models silently decay as production data drifts from the data they were trained on.