Data Visualization
Plotting Techniques
iHub-Data-FMML 2023
Data Visualization
Team Round 1 Round 2 Round 3
A 38 15 18
B 58 57 28
C 14 79 49
D 95 43 28
E 68 92 83
iHub-Data-FMML 2023
Data Visualization
X2
Noise
iHub-Data-FMML 2023
Data Visualization
Data visualization deals with a visual representation of data and is part of data analysis.
It is the process of translating data into a chart, graph, or other visual components.
Data visualization can be used for:
Making data engaging and easily digestible
Identifying trends and outliers within a set of data
Highlighting the important parts of a set of data
iHub-Data-FMML 2023
Variables
Variables refer to characteristics, properties, or attributes that can be measured,
observed, or recorded for a particular entity or unit within a dataset
Types of Variables: Dependent Variables, Independent Variables
Based on the nature:
Qualitative (Categorical): It describes the quality of something or someone. It is descriptive
information. For e.g., skin color, eye color gives us qualitative information about a person.
Quantitative (Numerical): It provides numerical information, like how much, how many, or
how often. Can be continuous or discrete. For e.g., the height and weight of a person.
iHub-Data-FMML 2023
Univariate Analysis
Univariate Analysis is used in statistics to describe a data type that contains only one
attribute or characteristic
Data = [0.49,-1.13,2.64,0.15,-1.23,-0.24,1.58,0.77,-0.47, 0.54, -2.46, -0.44, 0.22, 0.76,1.24,
-0.54,-2.11,0.12,0.02,-1.15,....., 0.32,-0.12,0.19]
Includes checking the central tendency (mean, median and mode), the range, the
maximum and minimum values and the standard deviation of a variable
iHub-Data-FMML 2023
Univariate Analysis
Data = [0.49,-1.13,2.64,0.15,-1.23,-0.24,1.58,0.77,-0.47, 0.54, -2.46, -0.44, 0.22, 0.76,1.24,
-0.54,-2.11,0.12,0.02,-1.15,....., 0.32,-0.12,0.19]
Box Plot: Compare the spread of the
Histogram: Frequency distribution
variables and get an insight into outlier
graph
iHub-Data-FMML 2023
Bivariate Analysis
The bivariate analysis is mainly used to compare two
sets of data to find a relationship between the two
variables.
Remember, that if one variable influences the
change in another variable, then you have an
independent and dependent variable.
Ex:- Scatter Plot, Heatmap, Contour Plot, Bivariate
Line Chart, Pair Plot, etc. Scatter Plot: Captures the correlation
between the two
iHub-Data-FMML 2023
Multivariate Analysis
Multivariate analysis is used to reveal the
relationship among several variables simultaneously.
Assists in making informed decisions by considering
multiple variables and their interactions
simultaneously.
Ex: - Grouped Box Plot, Multivariate Scatter Plot, and
3D scatter plot.
iHub-Data-FMML 2023
Visualization Techniques
Distribution of data points: Box plot, Histogram
Comparison of data points: Multi-line chart, Bar plot, Line chart [to show trends in data]
Relationship/Correlation of data points: Scatter plot
Composition of data points: Pie chart, Stacked Area chart, Stacked Bar chart
iHub-Data-FMML 2023
Violin Plot
Data = [0.49,-1.13,2.64,0.15,-1.23,-0.24,1.58,0.77,-0.47, 0.54, -2.46, -0.44, 0.22, 0.76,1.24, -
0.54,-2.11,0.12,0.02,-1.15,....., 0.32,-0.12,0.19]
The box plot is convenient for comparing summary
statistics (such as range and quartiles), but it doesn't
let you see variations in the data.
Are most of the values clustered around the
median, or around the minimum/maximum?
The histogram and kernel density estimation
helps you in seeing the variations in the data,
but you miss the outliers
iHub-Data-FMML 2023
Can we combine both?
Violin Plot Higher Probability
Median
It is a combination of box plot and a 1.5 * IQR
kernel density plot, which shows
peaks in the data
It depicts the summary statistics and IQR
the density of each variable Lower Probability
iHub-Data-FMML 2023
Violin Plot
Vertical violins, grouped by two variables Split violins to take up less space
Plot Courtesy: Seaborn Documentation [[Link]]
iHub-Data-FMML 2023
Pair Plot
Pair plot visualizes given data to find the
relationship between them and plots
pairwise relationships in a data-set
It is used for exploring the relationship
between multiple variables at once
Plots in a matrix format
o Diagonal subplots are the univariate histograms
for each attribute
o Off-diagonal entries are the scatter plots
Plot Courtesy: Seaborn Documentation [[Link]]
iHub-Data-FMML 2023
Joint Plot
Joint plot combines univariate and bivariate
plots to visualize relationship between two
variables
It consists of a scatter plot for the bivariate
relationship, with additional marginal plots
for each variable
Helps in understanding the correlations and
distributions of two variables simultaneously
Plot Courtesy: Seaborn Documentation [[Link]]
iHub-Data-FMML 2023
Heatmap
A heatmap is a color-coded representation
of a 2-dimensional data, representing the
magnitude of individual values within a
dataset
Colours are used to represent the
magnitude, intensity, with the colour
gradient scheme ranging from a lighter
colour (low values) to a darker colour (high
values)
Displays the correlations or relationships in Plot Courtesy: Seaborn Documentation [[Link]]
a correlation matrix
iHub-Data-FMML 2023
Parallel Co-ordinates
Parallel Co-ordinates allows forthe comparison
of multiple data records, by using parallel lines to
connect points based on multiple numerical
variables
o Each vertical line is a dimension
o A data item is connected by line segments
o Large number of samples clutters the
visualization
iHub-Data-FMML 2023
Dimensionality Reduction
iHub-Data-FMML 2023
Dimensionality
The number of input variables or features for a dataset is referred to
as its dimensionality.
The difficulties related to training machine learning models due to
high dimensional data is referred to as ‘Curse of Dimensionality’.
When dealing with high dimensional data, it is often useful to reduce the
dimensionality by projecting the data to a lower dimensional subspace
which captures the “essence” of the data. This is called dimensionality
reduction.
— Page 11, Machine Learning: A Probabilistic Perspective, 2012.
iHub-Data-FMML 2023
Dimensionality Reduction
Feature Selection
o Select the most relevant subset of features
o Reducing the number of irrelevant or redundant features
Feature Extraction
o extracting/deriving information from the
original features set to create a new features
𝟏
subspace
𝟏 𝟐
o compress the data with the goal of 𝟐 𝟑
maintaining most of the relevant information 𝟒
iHub-Data-FMML 2023
Feature Selection Techniques
Forward Feature Selection
Wrapper Method Backward Feature Elimination
Supervised Bi-directional Elimination
Techniques
Information Gain
Filter Method Chi-square Test
Feature Selection Fischer’s Score
Techniques
Embedded Method Regularization
Random Forest
Unsupervised Correlation based
Techniques Feature Selection
iHub-Data-FMML 2023
Forward-Feature Selection
Iteratively selects one feature at a time, evaluating the model's performance after adding
each feature and keeping the best subset of features that maximizes or minimizes the
chosen performance metric.
Step 1: Initialization: Initialize an empty set S to store the selected features
Step 2: Loop over Features: For each feature Xi to be added to S
Find Xi that best improves the model's performance metric when added to S
Update the selected Xi to S
Step 3: Termination:
Repeat Step 2 until adding any remaining feature does not improve the model's
performance.
Return S as the selected feature subset
iHub-Data-FMML 2023
Forward Feature Selection
Sr. no Feature_idx Avg_Score
1 (10,) 0.541
2 (5,10) 0.638
3 (5,8,10) 0.682
4 (5,7,8,10) 0.696
5 (4,5,7,8,10) 0.715
6 (3,4,5,7,8,10) 0.721
7 (3,4,5,7,8,9,10) 0.724
8 (1,3,4,5,7,8,9,10) 0.728
9 (0,1,3,4,5,7,8,9,10) 0.729
10 (0,1,2,3,4,5,7,8,9,10) 0.730
11 (0,1,2,3,4,5,6,7,8,9,10) 0.732
iHub-Data-FMML 2023
Backward Feature Elimination
BackwardFeatureElimination(X, y):
S = {all features} # Initialize with all features
Start with all available features, iteratively best_score = EvaluateModel(X, y, S) # Evaluate initial model
remove one feature at a time, and using all features
evaluate the model's performance. while there are remaining features in S:
for feature in S:
If the performance improves, we keep the Remove feature from S
Train a model using the features in S
feature removed; otherwise, we add it Evaluate model performance using a chosen metric
back. If model performance improves compared to
best_score:
Update best_score to the new performance
The final set of features that maximizes or else:
minimizes the chosen performance metric Add feature back to S
is returned as the selected feature subset.
return S # Return the remaining features after elimination
iHub-Data-FMML 2023
Backward Elimination
Sr. no Feature_idx Avg_Score
11 (0,1,2,3,4,5,6,7,8,9,10) 0.732
10 (0,1,2,3,4,5,7,8,9,10) 0.730
9 (0,1,3,4,5,7,8,9,10) 0.729
8 (1,3,4,5,7,8,9,10) 0.728
7 (3,4,5,7,8,9,10) 0.724
6 (3,4,5,7,8,10) 0.721
5 (4,5,7,8,10) 0.715
4 (5,7,8,10) 0.696
3 (5,8,10) 0.682
2 (5,10) 0.638
1 (10,) 0.541
iHub-Data-FMML 2023
Bi-directional Elimination
Combines forward and backward feature selection techniques to iteratively select a subset
of features that optimizes a model performance metric
Step 1: Initialization: Initialize an empty set S to store the selected features and choose
direction
Step 2: Loop over Features: For direction chosen as ‘forward’ or ‘backward’
If ‘forward’: Perform forward selection adding the best feature that
improves the model's performance metric
If ‘backward’: Perform backward elimination, removing the least significant
feature
Step 3: Termination:
Repeat Step 2 until adding/eliminating any remaining feature does not improve the
model's performance.
Return S as the selected feature subset
iHub-Data-FMML 2023
Feature Extraction
Aims to reduce the number of features in a dataset by creating new features from the
existing ones (discarding the original ones)
Feature Extraction Techniques:
o Principal Component Analysis: Linear transformation techniques by finding
orthogonal axes that capture the most variance
Image Source: [Link]
iHub-Data-FMML 2023
Feature Extraction
Aims to reduce the number of features in a dataset by creating new features from the
existing ones (discarding the original ones)
Feature Extraction Techniques:
o Isomap, t-SNE(t-distributed Stochastic Neighbor Embedding): Non-linear
dimensionality reduction technique that emphasizes the local structure of the data
Image Source : [Link]
iHub-Data-FMML 2023