Skip to content

chap01.md

Latest commit

 

History

History
73 lines (41 loc) · 3.43 KB

chap01.md

File metadata and controls

73 lines (41 loc) · 3.43 KB

Chapter.1 Basics of ML

Regression & Classification

for Supervise Learning

  • Regression is way of forecasting for continuous variables. Classification predicts for events that have few distinct outcomes.

  • Linear Regression to forecast continuous variables, Logistic Regression to predict discrete variables. Severalother techniques usable for both.

  • In Regression, need to account for generalization over overfitting. Lower the generalization, higher the error rate on unaccounted data.

  • Cross validation on different datasets for training, testing & validation (60%, 20%, 20% data split) provides a check on accuracy of model when hyper-parameters are changed.

Measures of Accuracy

  • Error is measured on testing dataset, as doing so on training dataset would be ambiguous.
  • Absolute Error & RMSE for Continuous variables. Confusion Matrix for Classification, Discrete event prediction.
  • Discrete events happen as probability of an event happening.

  • Absolute Error is absolute value difference of abs(forcast - actual) values. Abs-value help get real overall notion, avoiding mess of +ve & -ve diff-values.

  • RMSE (root mean square error), where getting overall squared (square making all diff values +ve) and calculating sq-root of mean for overall.

  • Confusion Matrix counts number of instances when model predict outcome of an event & measures against actual outcome as TruePositive, TrueNegative, FalsePositive, FalseNegative (TP, TN, FP, FN).

Performance Metrics of Confusion Matrix are:

  • Accuracy: Correct Classification Ratio; as ((TP+TN) / (TP+TN+FP+FN))
  • Precision: Correct Selection on Total Selection; as (TP / (TP+FP))
  • Recall: Correct Selection on Total Required Ratio; as (TP / (TP+FN))
  • F1: F-measure as equlibrium between Precision & Recall; as (2(TP) / (2TP+FP+FN))
  • Specificity: True Negative rate; as (TN / (FP+TN))

AUC Value & ROC Curve

  • AUC (Area Under the Curve): Score input for probability of Positive Classification. Order these in descending.

  • ROC (Receiver Operation Characterisitc) Curve is cumulative number of points considered from descending list that suit TP captured to max.

Unsupervised Learning

  • 2 major techniques: Clustering (row grouping) and PCA (Principle Components Analysis, column grouping).
  • Clustering helps segmenting, can be powerful pre-processing steps in model building.
  • PCA reduces dimensionality/variables of data, thus can speed up model building process.

Approach to building Models

  • Data Source? Filter/fetch required data.

  • Data Preparation, to manage following

  • Missing values in data. Simplest is using average/median of column. Better to use K-nearest neighbors method.

  • Outliers in data. Result in inefficient skews. Typical solution is capping variables at certain % (say 95%).

  • Variable Transformation. Scaling variable based on gradient descent generally cause faster optimization. Log/squared transformations are used when input variable share a non-linear relation with dependent variables.

  • Feature Interaction: process of creating new variables based on combination of variables; need better domain context.

  • Feature Generation: process to get additional features from dataset; need better behavioral linking of data points.

  • Building Models: Once the data is ready and prepared.

  • Production Ready: Depends on workflow for use of model.

  • Continuous Integration: with A/B testing for comparative analysis.