Description
1 Regression (50 points)
The Data
For this portion of the assignment you will be using the Iris Flower dataset, available here.
The dataset consists of 50 samples from each of three species of Iris (Iris setosa, Iris virginica
and Iris versicolor ). Four features were measured from each sample: the length and the width
of the sepals and petals, in centimeters.
The Idea: Using Linear Regression on the Iris Dataset
Often, we may want to predict one feature based upon other features. Your objective here will
be to generate a linear model of one of the features (a continuous variable) in the Iris dataset
using one or more of the remaining features and/or class values. In our case, we’re interested
in finding the best linear model among those that can be generated from this dataset.
What to Do
First, load the Iris dataset. This can be done using the following snippet of code:
import pandas a s pd
fileURL = ’ h t tp : / / a r c hi v e . i c s . u ci . edu/ml/\
machine−l e a r ni n g −d a t a b a s e s / i r i s / i r i s . data ’
i r i s = pd . r e a d c s v ( fileURL , names=[ ’ Sep al Len g th ’ , ’ Sepal Width ’ , \
’ P e t al L e n g t h ’ , ’ Petal Width ’ ,
’ S p e ci e s ’ ] , he ade r=None )
i r i s = i r i s . dropna ( )
Next, you can visualize the correlation between different features using the following snippet of
code, which executes the provided pairs function:
p a i r s ( i r i s )
Some pairs of features tend to be more correlated than others. Try to uncover related features by
using linear regression to model the relationship between pairs of features. In other words, use
one feature as a target (dependent) variable and another feature as a predictor (independent)
variable. To generate a linear regression model, you may use the
linear_model.LinearRegression() function available via the scikit-learn library. To
run the model on the Iris data, first divide the dataset into training and testing sets, then fit
the model on the training set and predict with the fitted model on the testing set. scikit-learn
provides several functions for dividing datasets in this manner, including
cross_validation.KFold and cross_validation.train_test_split.
Several statistics can be generated from a linear model. Given a fitted linear model, the following
code outputs the model coefficients (the parameter values for the fitted model), the residual
sum of squares (the model error), and the explained variance (the degree to which the model
explains the variation present in the data):
# The c o e f f i c i e n t s
p r i n t ” C o e f f i c i e n t s : \n” , r e g r . c o e f
# The mean s q u a r e e r r o r
p r i n t ( ” R e si d u al sum o f s q u a r e s : %.2 f ” %
np . mean ( ( r e g r . p r e d i c t ( i r i s X t e s t ) − i r i s y t e s t ) ∗∗ 2 ) )
# Expl ained v a ri a n c e s c o r e ( 1 i s p e r f e c t p r e d i c t i o n )
p r i n t ( ” V a ri ance s c o r e : %.2 f ” % r e g r . s c o r e ( i r i s X t e s t , i r i s y t e s t ) )
2
You can use these scores to measure the efficacy of a particular linear model.
What to Provide
Your output should contain the following:
• A scatterplot matrix of scatterplots, with one scatterplot for each pairwise combination
of features.
• A plot of the linear regression models generated on each pairwise combination of features,
with corresponding model statistics.
• A plot of the best overall linear regression model you were able to generate using any
combination of features (including the use of multiple features used in combination to
predict a single feature), with corresponding model statistics.
Given this output, respond to the following questions:
1. Based upon the linear models you generated, which pair of features appear to be most
predictive for one another? Note that you can answer this question based upon the output
provided for the linear models.
2. Suppose you tried to generate a classification model on this dataset, but only after removing the feature that you were best able to predict based upon other features. How would
removing this feature affect the classification performance?
3
2 Classification (50 points)
The Data
For this portion of the assignment you will be using a set of images of the works of various
artists. To acquire the large number of images, we used a script that searches for each artist’s
painting on Google Images and downloads it. The provided dataset consists of images obtained
in this manner for several well-known artists. The image are of varying size and quality.
The provided dataset is available here.
You are welcome to use the included function (go_google_image) to search and retrieve
paintings by your preferred artist(s), which may then be used for this assignment.
The Idea: Identifying Artists by their Paintings
Your objective here will be to perform classification on the dataset to discern how reliably
the paintings can be attributed to their respective artists. Several data preprocessing steps,
including dimensionality reduction and clustering, may be applied to generate features more
amenable to this task. Specifically, we would like to perform classification on the images so that
they are correctly attributed to this respective artists. We can assess classification performance
by dividing the dataset into a training set and a testing set, and measuring the error of a
classifier fitted on the training set and evaluated on the testing set.
Intuitively, since we typically view an image as a collection of pixels, we might consider using
the set of pixels as a feature. In other words, each color value for each pixel would be a feature,
with the collection of pixels comprising many features that collectively describe the pixel values
of the image. We call the collection of features that describe a particular aspect of the image a
“feature descriptor.” While this feature descriptor describes each pixel of the image, the specific
locations of particular color values and the size of the image directly affect the derived feature
values, thus making it difficult to use this set of features to directly compare images.
We can supplement or replace the pixel-based features with a histogram of color values. Each
color has 256 possible values, resulting in a histogram of 768 (256*3) color values distributed
over the entire range of pixels. Each value of the histogram is the number of pixels in the image
with the corresponding color value. Here, we would consider each histogram value as a feature
of an image. The histogram provides a representation of the color distribution of an image,
ignoring the specific location of color values within the image. Each value of the histogram then
corresponds to one feature.
We can also use more complicated features, such as histograms of Oriented Gradients (HOGs).
HOGs are feature descriptors used in computer vision and image processing for the purpose of
object detection that count the occurrences of gradient orientation in localized portions of an
image. HOG descriptors are based upon the premise that local object appearance and shape
within an image can be described by the distribution of intensity gradients or edge directions.
These properties can derived by dividing the image into small connected regions or “cells,”
and compiling a histogram of gradient directions or edge orientations for the pixels within each
cell. The combination of these histograms then represents the HOG descriptor, which has been
shown to be quite effective for the purposes of classifying images.
While these are several of the popular feature descriptors used for image classification, other
feature descriptors could be incorporated in the final set of feature used to train a classification
model. In addition, each of these feature descriptors could be further preprocessed for further
enhance their descriptive power.
4
What to Do
The IPython Notebook provided with this assignment includes functions to compute the histograms and plot the images within the transformed (2-dimensional) space (load_images and
plot_image_space, respectively). There are also functions to generate and plot the color
palettes associated with each image (cluster_image_colors and plot_color_palette,
respectively); the palettes are generated via (k-means) clustering of the pixel color values, and
may be investigated at your own leisure—they are not needed to complete the assignment.
The images can be loaded and the histograms generated by running the following code snippet
(which imports pandas as pd). Please ensure that the directory provided to the load_images
function is correct. For example, if you have placed all the images in your base IPython
Notebook directory in a folder labeled images, with the paintings images in a subfolder labeled
paintings, then the (relative) path to the painting images would be images/paintings.
The following code snippet loads all images from the ‘Pablo Picasso Paintings’ and ‘Vincent
van Gogh Paintings’ subdirectories (the painters_subdirs variable can be changed to your
painters of interest, or omitted to load all of the subdirectories) and generates feature descriptors
based upon the loaded images:
import pandas a s pd
# Load images and g e n e r a t e i n i t i a l f e a t u r e d e s c r i p t o r s
p a i n t e r s d i r = ’ /path/ t o / p ai n ti n g im a g e s ’ # d i r e c t o r y path
p a i n t e r s s u b d i r s = [ u ’ Pablo Pi c a s s o P ai n ti n g s ’ ,
u ’ Vincen t van Gogh P ai n ti n g s ’ ]
data = f e t c h p a i n t i n g s ( p a i n t e r s d i r , p a i n t e r s s u b d i r s )
g e n p i x e l f d ( data ) # g e n e r a t e pi x el −based f e a t u r e d e s c r i p t o r
g e n h i s t f d ( data ) # g e n e r a t e c o l o r hi s t o g r am f e a t u r e d e s c r i p t o r
g e n h o g f d ( data ) # g e n e r a t e HOG f e a t u r e d e s c r i p t o r
Each of these feature descriptors has a different representational format. However, the classification models available via scikit-learn only accept data into a single-dimensional vector format.
Accordingly, these features need to be flattened into a single long array. The following code
snippet accomplishes these tasks:
# Arrange f e a t u r e d e s c r i p t o r s i n t o a s i n g l e f e a t u r e v e c t o r
images = [ pp [ ’ image ’ ] f o r p i n data . key s ( ) f o r pp i n data [ p ] ]
X = pd . DataFrame ( [ pp [ ’ f e a t u r e s ’ ] \
f o r p i n data . key s ( ) f o r pp i n data [ p ] ] )
y = pd . DataFrame ( [ c l s f o r c l s i n [ pp [ ’ c l a s s ’ ] \
f o r p i n data . key s ( ) f o r pp i n data [ p ] ] ] )
f o r inde x i n pd . i s n u l l (X ) . any ( 1 ) . n onze r o ( ) [ 0 ] : d el images [ inde x ]
y = y . drop ( y . inde x [ pd . i s n u l l (X ) . any ( 1 ) . n onze r o ( ) [ 0 ] ] )
X = X. dropna ( )
Given this feature vector, visualize the data into two-dimensions using a method of your choice.
Subsequently classify the data using 5-fold cross-validation, using one or more classifiers or your
choice.
What to Provide
Your output should contain the following:
• For a pair of artists of your choice, generate a two-dimensional visualization.
• For a pair of artists of your choice, generate at least three classification models that
distinguish between the artists’ paintings.
5
• Generate at least one feature descriptor (in addition to those provided) and include it in
the set of features used for your classifiers. Rerun the classifiers with the newly-included
feature descriptor.
Given this output, respond to the following questions:
1. What is the highest classification accuracy you achieve?
2. Of the classification models you tried, did any tend to perform better or worse than others?
Any thoughts why?
3. What feature descriptor(s) did you generate? Did they improve the classification performance? If so, by how much?
3 Extra Credit Portion (+10 points)
The Idea: Good Classifiers Generalize Well
A good classifier should generalize well to data it has not yet seen. When performing any classification task, the typical objective is to minimize the generalization error that a classification
model will produce on new data.
What to Do
Generate the best classifier you can for distinguishing between two different artists’ paintings,
per the classification problem outlined above. We define the best-performing model here as the
model with the highest accuracy. We will use a new dataset from two unprovided artists to test
each student’s model using 5-fold cross-validation. The student with the best-performing model
(the model with the highest accuracy on this new dataset) will earn 10 points extra credit; the
student with the second-best model will earn 9 points; and so on, with the tenth-best model
earning 1 point. Thus up to ten students will receive some degree of extra credit on this task.
What to Provide
If you’re interested in participating, ensure that your model is properly delineated in and
runnable from your notebook. All content needed to generate your classification model should
be included in your assignment submission.
6
