How to Approach ML Problems Step by Step !!

For THEORY PART, please refer to https://github.com/sandipanpaul21/Machine-Learning-Notes-Daywise

01 Know about the Dataset

• What Dataset is all about? Problem Objective ?
• Number of Rows and Columns. Data Stored is in which type? Is it a Data-frame or a Dictionary?

02 Univariate Analysis

• Uni means one, so Single Variable Analysis
• Mainly deals with Numerical Measures used in a dataset (Total 14 Numerical Measures)
• Measure of Central Tendency : 1. Mean, 2. Median, 3. Mode
• Measure of Data Spread : 4. Quartile, 5. Percentile, 6. Range, 7. IQR, 8. Boxplot, 9. Variance, 10. Standard Deviation
• Variation between Variables : 11. Covariance, 12. Correlation Coefficient (Pearson and Spearman)
• Measure Distribution and Peakness : 13. Skewness and 14. Kurtosis

03 Bivariate Analysis

• Bi means two, so Two Variable Analysis
• There are majorly two types of Data Variable: Continuous & Categorical Variable
• So 3 possible combinations for Bivariate Analysis
  1. Continuous vs Continuous : Correlation Coefficient
  2. Categorical vs Categorical : Chi Square Test
  3. Continuous vs Categorical : T Test (n < 30), Z Test (n > 30) and ANOVA Test

04 Outlier & Missing Values

• Outlier are data points that differs significantly from other observations
• Techniques to Detect Outliers : 1. Box Plot and 2. Z-Score
• Technqiues to Remove Outliers : Capping Based on Upper and Lower Range
• Missing Values in the Dataset cause concern for Machine Learning Model
• Techniques for Imputing Missing Values
  1. Continuous Data : Median Imputation
  2. Categorical Data : Mode Imputation
  3. KNN Imputation (why better than Median and Mode imputation)

05 Feature Engineering

• Tweaking the features, to increase the efficiency of the Model
• 3 Major Steps in Feature Engineering : 1.Transformation, 2.Scaling & 3. Construction
• Feature Transformation
  1. Feature transformation is performed to normalize the data
  2. Methods Used : 1.Log Transformation, 2.Square Root, 3.Cube Root & 4.Box-Cox Transformation
• Feature Scaling
  1. Feature scaling is conducted to standardize the independent features
  2. Method Used : Mix-Max Scaler
• Feature Construction
  1. It is a process of creating features based on the original descriptors
  2. Methods Used : 1.Binning and 2.Encoding

06 Regression Assumption

• With Boston Dataset, all 5 regression assumptions checked (Why, What and How)
  1. Linearity between Target & Features : Plot Predicted & Target
  2. Normality of Error Term : Anderson-Darling Test & Skewness in Error Term
  3. Multicollinearity among Predictors : Correlation & VIF
  4. Autocorrelation among Error Term : Durbin-Watson Test
  5. Homoscedasticity,same variance within error terms : Residual Plot

10 K Means Clustering

• Define number of clusters, take centroids and measure distance
• Euclidean Distance : Measure distance between points
• Number of Clusters defined by Elbow Method
• Elbow Method : WCSS vs Number of Cluster
• Silhouette Score : Goodness of Clustering

11 Hierarchical Clustering

• Group similar objects into groups
• Type of HC
  1. Agglomerative : Bottom Up approach
  2. Divisive : Top Down approach
• Number of Clusters defined by Dendogram
• Dendogram : Joining datapoints based on distance & creating clusters
• Linkage : To calculate distance between two points of two clusters
  1. Single linkage : Minimum Distance between two clusters
  2. Complete linkage : Maximum Distance between two clusters
  3. Average linkage : Average Distance between two clusters

12 DB Scan Clustering

• No need to give pre-define clusters
• Distance metric is Euclidean Distance
• Need to give 2 parameters
  1. eps : Radius of the circle
  2. min_samples : minimum data points to consider it as clusters

13 Gausian Mixture Model

• Probablistic Model
• Uses Expectation-Minimization (EM) steps:
  1. E Step : Probability of datapoint of each cluster
  2. M Step : For each cluster,revise parameter based on proabability