Ad
More Related Content
Similar to machine learning types methods classification regression decision tree (20)
Recently uploaded (20)
Ad
machine learning types methods classification regression decision tree
- 1. Role of Machine Learning in Telecommunication Dr. Mohamad Abou Taam
- 2. WHAT IS MACHINE LEARNING? Machine learning is a subfield of computer science that studies and develops algorithms that can learn from data without being explicitly programmed Computer Science Artificial Intelligence Machine Learning Deep Learning Machine learning algorithms can detect patterns in data and use them to predict future data
- 3. Machine learning Data Rules / Model Traditional software: applying given rules to data Traditional software Rules Data Answers / Actions Machine learning – how is it different? M a c h i n e l
- 4. Model design, training and testing (model building, feature engineering) Historical Data Machine Learning Model 1 Model application (model scoring) New Data Model Predictions 2
- 5. TRIAD OF ALGORITHMS, DATA AND TRAINING Data Machine learning Algorithms Training "Learning"is the process of estimating an unknown dependency or structure of a system (building a model) from a limited number of observation (data points) and ability to generalize it onto previously unseen data
- 6. Inferential Statistics Descriptive Statistics • Sample should be representative of population • Generalization – extrapolation to entire population • Watch for population drift! Inference THE "CENTRAL DOGMA" OF STATISTICS Machine learning == statistical learning Sampling principle Probability Population Learning on sample Sample
- 7. THREE TYPES OF MACHINE LEARNING Reinforcement Learning The goal is to optimise actions in a way that maximises cumulative reward. no explicitly labeled data is given, but “rewards” and “punishment” signals are provided X – input data /independent variable Unsupervised Learning The goal is to learn patterns and structure in data given only inputs X. (no output Y information given at all) Supervised Learning The goal is to learn mapping from given inputs X to outputs Y, given a labeled set of input-output (X-Y) pairs . X – input data / independent variable Y – response/ dependent variable
- 9. SUPERVISED LEARNING: REGRESSION Response variable Y – real valued Years of Education S e n i o r i t y I n c o m e 0 50 100 200 300 5 10 15 20 25 TV Sales Sales multivariate univariate
- 10. SUPERVISED LEARNING: CLASSIFICATION Response variable Y – categorical binary multiclass
- 11. REGRESSION AND CLASSIFICATION ARE SIMILAR Regression Predict a numeric variable Classification Predict a binary (or categorical) outcome 0 Y 5 10 15 20 25 X 15 5 0 10 0.0 0.2 0.4 0.6 0.8 1.0 -2 -1 0 1 2 X Probability of event Data are 1s and 0s – event either happens or doesn't happen
- 12. MODEL OVERFITTING Regression Too simple Too complex Just right Predictions will have high "bias" – from inadequate assumptions Predictions will have high "variance" – driven by noise in the training data Model complexity is appropriate given the noise
- 13. MODEL OVERFITTING Classification overfit boundary just right two classes
- 14. 14 PREDICTION ACCURACY VS EXPLAINABILITY Model explainability Prediction accuracy White box models • Interpretable by design • Easy to explain • Quick to run • Limited tuning needed Black box models • Lots of work to get insights Better predictive performance • Potential for overfitting • Often lot of tuning required • Linear / logistic regression • Decision trees Model properties Algorithm examples • Random forests • Gradient boosting • Neural networks • Deep learning
- 17. REGRESSION EVALUATION Quality metrics Standard quality metrics Mean absolute error: Mean squared error: Root mean squared error: R-squared:
- 20. CLASSIFICATION EVALUATION Quality metrics Actual Yes (or 1) No (or 0) True positives TP False Positives FP False Negatives FN True negatives TN Predicted Yes (or 1) No (or 0) True positive = Predict event and event happens True negative = Predict event does not happen, nothing happens False positive = Predict event and event does not happen (false alarm) False negative = Fail to predict event that does happen (missed alarm)
- 21. TRAINING AND TESTING Train-test split • 70%-90% of the data • Used to build the model • 10%-30% of the data • Used to check the performance of the model on unseen data Train & Test split • Measure algorithm performance on both train and test sets! • Performance will be worse on the test set • Algorithms hyperparameter tuning can be used to improve test set performance • Avoid overfitting! • Actual performance of the algorithm in production will not be better than on test set!
- 22. TRAINING AND TESTING Cross-validation • Makes best use of the data • Data split in to N "folds" at random • N models built. On each model, N-1 folds are used for training and one is used for testing • Evaluation criteria averaged across folds • Allows use of eg 90% training data / 10% test data splits for 10-fold cross validation • More data for training increases predictive power • Reduces the chance of getting lucky/unlucky just due to the way a single train/test split is done • More time/computer resources consuming average Cross-validation 5-fold cross-validation
- 23. TYPICAL SUPERVISED LEARNING PIPELINE Model training Model application regression model value value and testing
- 24. A SUPERVISED MACHINE LEARNING WORKFLOW Prepare data Model and predict Impact business Define problem and potential solution Get the data Understand the data Clean the data Feature engineering Build and test model Understand the model What does it mean for the business? What are we going to change? Productionise Iterate Ongoing monitoring and improvements