Scribd
Upload
9 views73 pages

Machine Learning

Uploaded by

Hanin Fitria
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
9 views73 pages

Machine Learning

Uploaded by

Hanin Fitria
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
You are on page 1/ 73

Introduction to Machine

Learning
Why use Orange

 Visual presentation of data


 Does not require programming but if needed can be done in Phython
 Have many machine learning algorithms
 Can perform classification and clustering
 FREE
Machine Learning of
manufacturing data using
orange
SETFI Data

 Simulated Data for Tool Fault Isolation in Wafer Fabrication


SETFI Data
Data description

 First Column is 4000 instances or samples


 Second to 150th column are the different locations of the tools. Each
sample passes through a sequence of tools in different locations. For
example sample 3699 goes through the sequence (1,2)>(2,3)>(3,9)…
etc ie starting at location1 tool2 to location2 tool3 and so on
 The 151st column to the 300th column is the time stamp at each location.
 The last column is the measure of the performance
Importing data into orange
Checking the data
K means
 Given an initial set of k means m1(1),…,mk(1)
 Assign each observation to the cluster whose mean has the least
squared Euclidean distance
 Calculate the new means to be the centroids of the observations in the
new clusters.
 The algorithm has converged when the assignments no longer change.
Visualisation

 Sieve Diagram
 shows the frequencies in a two-way contingency table in relation to expected frequencies under
independence
 highlights the pattern of association between the row and column variables.
 A unit square is divided into rectangles, one for each cell in the contingency table
 The height of each rectangle in row i is proportional to the marginal frequency in that row ( f i+ )
 the width of each rectangle in column j is proportional to the marginal frequency in that column
( f +j)
 the area of each rectangle is proportional to the expected frequency,
 The observed frequency in each cell is shown by the number of squares drawn in each
rectangle.
 the difference between observed and expected frequency appears as the density of shading,
using color to indicate whether the deviation from independence is positive (blue) or negative
(red).
What do we see

 Performance
 <2709.258257
 2709.258257 – 2783.99236
 2783.99236 – 2857.613382
 >2857.613382
 Using a scatter plot
 Low performance C1
 Low performance C2
 More low performance instances in C2 say below 2400
Low performance C1
Low Performance C2
Finding Informative Projections
 Scoring the plots for C2
 Observe that the location 73 seems to feature in all the top score plots and
that most of the lowest performance seems most at tools 3 and 8
High Performance C2
 Location 15 seems to feature in most top scorers and in particular
tool 4
High performance C1
 Tool 3 at loc 73 seems to produce high performance results also.
So can conclude tool 3 is erratic??
 Tool 7 at location 15
Performance Analysis: Beach Ball
Soccer Asian Championship
 Data Collection using apps on android
StatWatch
StatMine in Play Store
Data
Importing Data
Check in Table
Hierarchical Clustering
 One of the easiest techniques to cluster the data.
 Find nearest neighbor by measuring distance for example using
euclidean
 Eucledian distance calculated as

 The shorter the distance the more similar the two instances are.
 In the beginning, all instances are in their own particular clusters.
 Then we seek for the closest instances of every instance in the plot.
 We pin down the closest instance and make a cluster of the original and the
closest instance.
 Now we repeat the process again. What is the closest instances to our new
cluster –> add it to the cluster –> find the closest instance.
 We repeat this procedure until all the instances are grouped in one single
cluster.
Distances Widget
Hierarchical Clustering
Two Clusters
Comparison
Machine Vision Inspection at
Attach Work Center.
Data Description

 The following data was obtained of images classified into 10 different categories
 6310
 6311
 6314
 6316
 6323
 6325
 6326
 6930
 6983
 6994
 Images taken were inconsistent in terms of orientation, illumination and homogeneity
 Each category only consisted between 4 to 6 images
6314

6316
Methodology

 Image embedding was first applied to all images. This generated 4096
features for each image
 Four images were separated as test samples

 The remaining images were trained on 4 models : neural network,


Logistic regression, Naïve bayes and SVM
 Image Prediction on test samples
Image Embedding

Image features before


embedding

Image features
after embedding
Image Training

 The images were


trained on 4 models
 An AUC close to 1
indicates the best
prediction model
and in this case is
Logistic Regression
 However the
classification
accuracy CA is
relatively low
ROC

 It can be observed that the


ROC of Logistic Regression,
Naïve Bayes and Neural
networks for target class
6314 has very good
discrimination capacity whilst
for target class 6326 has very
poor discrimination capacity
Confusion Matrix

 The confusion matrix


for logistic regression
shows the proportion
of actual prediction
for class 6314 is
100%
 The proportion of
actual for 6326 is the
least at 12.6%. The
most confused are
with 6325 and 6323
Confusion Matrix

 The confusion matrix


for logistic regression
shows the proportion
of actual prediction
for class 6314 is
100%
 The proportion of
actual for 6326 is the
least at 12.6%. The
most confused are
with 6325 and 6323
Confused
images of 6326

 6326
The images
of 6326 are
confused
with some of
the images
of 6325 and
6323

6325

6323
Image prediction

 It can be seen
that Logistic
regression and
Naïve Bayes has
predicted
correctly all
images. From the
confusion matrix
these categories
are with the
highest proportion
of actual images
predicted.

You might also like