Big Data Machine Learning

EL
PT
N
Dr. Rajiv Misra
Dept. of Computer Science & Engg.
Indian Institute of Technology Patna
[email protected]

Big Data Computing Vu Pham Big Data Machine Learning

Preface
Content of this Lecture:

In this lecture, we will provide an overview of machine learning

techniques to explore, analyze, and leverage data.

EL
We will also discuss tools and algorithms that you can use to

PT
create machine learning models that learn from data, and to scale
those models up to big data problems.
N

Big Data Computing Vu Pham Big Data Machine Learning

 Machine Learning Overview

EL
PT
N

Big Data Computing VuPham

Vu Pham Big Data Machine Learning
What is Machine Learning?
Learning from data:

Machine learning is the field of study that focuses on computer

systems that can learn from data. That is the system's often

EL
called models can learn to perform a specific task by analyzing
lots of examples for a particular problem. For example, a
machine learning model can learn to recognize an image of a

PT
cat by being shown lots and lots of images of cats.
N

Big Data Computing Vu Pham Big Data Machine Learning

What is Machine Learning ?
No explicit programming: This notion of learning from data
means that a machine learning model can learn a specific task
without being explicitly programmed. In other words, the

EL
machine learning model is not given the step by step
instructions on how to recognize the image of a cat.

PT
Instead, the model learns what features are important in
determining whether it picture contains a cat from the data
N
that has analyzed. Because the model learns to perform this
task from data it's good to know that the amount and quality
of data available for building the model are important factors in
how well the model learns the task.

Big Data Computing Vu Pham Big Data Machine Learning

What is Machine Learning ?
Discovering hidden patterns: Because machine learning
models can learn from data that can be used to discover
hidden patterns and trends in the data.

EL
Data-driven decisions: These patterns and trends lead to
valuable insights into the data. Thus the use of machine

PT
learning allows for data driven decisions to be made for a
particular problem.
So to summarize, the field of machine learning focuses on the
N
study and construction of computer systems that can learn
from data without being explicitly programmed. Machine
learning algorithms and techniques are used to build models,
to discover hidden patterns and trends in the data allowing for
data-driven decisions to be made.
Big Data Computing Vu Pham Big Data Machine Learning
Machine Learning (ML) is an Interdisciplinary Field
In applying machine learning to a
problem, domain knowledge is
essential to the success of end results.
By domain knowledge we mean an
understanding of the application or

EL
business domain.

Knowledge about the application, the

PT
data related to the application, and
how the outcomes will be used are
crucial to driving the process of
N
building the machine learning model.
So domain knowledge is also an
integral part of a machine learning
solution.

Big Data Computing Vu Pham Big Data Machine Learning

Example Application of Machine Learning
Credit card fraud detection: One application of machine learning
that you likely used this past weekend, or even just today, is credit
card fraud detection. Every time you use your credit card, the current
purchase is analyzed against your history of credit card transactions

EL
to determine if the current purchase is a legitimate transaction or a
potentially fraudulent one. If the purchase is very different from your
past purchases, such as for a big ticket item in a category that you

PT
had never shown an interest in or when the point of sales location is
from another country, then it will be flagged as a suspicious activity.
N

Big Data Computing Vu Pham Big Data Machine Learning

Example Application of Machine Learning
Handwritten digit recognition: When you deposit a hand-
written check into an ATM, a machine learning process is used
to read the numbers written on the check to determine the
amount of the deposit. Handwritten digits are trickier to

EL
decipher than typed digits due to the many variations in
people's handwriting.

PT
N

Big Data Computing Vu Pham Big Data Machine Learning

Example Application of Machine Learning
Recommendations on websites: After you buy an item on
a website you will often get a list of related items. Often
this will be displayed as customers who bought this item
also bought these items, or you may also like.

EL
These related items have been associated with the item
you purchased by a machine learning model, and are now
PT
being shown to you since you may also be interested in
them.
N

Big Data Computing Vu Pham Big Data Machine Learning

More Applications of Machine Learning
Targeted ads on mobile apps

Sentiment analysis

EL
Climate monitoring
PT
Crime pattern detection
N
Drug effectiveness analysis

Big Data Computing Vu Pham Big Data Machine Learning

What’s in a Name of Machine Learning?
Machine learning: Machine learning has its roots since statistics, artificial
intelligence, and computer science among other fields. Machine learning
encompasses the algorithms and techniques used to learn from data.
Data mining: The term data mining became popular around the time that the
use databases became common place. So data mining was used to refer to

EL
activities related to finding patterns in databases and data warehouses. There
are some practical data management aspects to data mining related to
accessing data from databases. But the process of finding patterns in data is
similar, and can use the same algorithms and techniques as machine learning.

PT
Predictive analytics refers to analyzing data in order to predict future outcomes.
This term is usually used in the business context to describe activities such as
sales forecasting or predicting the purchasing behavior of a customer.
N
Data science is a new term that is used to describe processing and analyzing
data to extract meaning. Again machine learning techniques can also be used
here. Because the term data science became popular at the same time that big
data began appearing, data science usually refers to extracting meaning from
big data and so includes approaches for collecting, storing and managing big
data.
Big Data Computing Vu Pham Big Data Machine Learning
Machine Learning Models
Learn from data

Discover patterns and trends

EL
Allow for data-driven decisions
PT
Used in many different applications
N

Big Data Computing Vu Pham Big Data Machine Learning

 Categories Of Machine
Learning Techniques

EL
PT
N

Big Data Computing VuPham

Vu Pham Big Data Machine Learning
Categories Of Machine Learning Techniques

Classification

Regression

EL
Cluster Analysis
PT
N
Association Analysis

Big Data Computing Vu Pham Big Data Machine Learning

Classification
Goal: Predict category
In classification, the goal is to predict the category of the input
data. An example of this is predicting the weather as being
sunny, rainy, windy, or cloudy.

EL
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

Classification
Some more examples of classification are classifying a
tumor from a medical image as being benign or
malignant.

EL
Predicting whether it will rain the next day.

PT
Determining if a loan application is high-risk, medium-risk
or low-risk.
N
Identifying the sentiment of a tweet or review as being
positive, negative, or neutral.

Big Data Computing Vu Pham Big Data Machine Learning

Regression
When your model has to predict a numeric value instead of a
category, then the task becomes a regression problem.
An example of regression is to predict the price of a stock. The stock
price is a numeric value, not a category. So this is a regression task

EL
instead of a classification task.
If you were to predict whether the stock price will rise or fall, then
that would be a classification problem. But if you're predicting the

PT
actual price of the stock, then that is a regression problem.
That is the main difference between classification and regression. In
classification, you're predicting a category and in regression, you're
N
predicting a numeric value.

Big Data Computing Vu Pham Big Data Machine Learning

Regression Examples
Estimating the demand of a product based on time or
season of the year.

EL
Predicting a score on a test.

PT
Determining the likelihood of how effective a drug will be
for a particular patient.
N
Predicting the amount of rain for a region.

Big Data Computing Vu Pham Big Data Machine Learning

Cluster Analysis
In cluster analysis, the goal is to organize similar items in your data set
into groups. A very common application of cluster analysis is referred to
as customer segmentation. This means that you're separating your
customer base into different groups or segments based on customer
types.

EL
For example it would be very beneficial to segment your customers into
seniors, adults and teenagers. These groups have different likes and

PT
dislikes and have different purchasing behaviors. By segmenting your
customers to different groups you can more effectively provide
marketing adds targeted for each groups particular interests. Note that
N
cluster analysis is also referred to as clustering.

Big Data Computing Vu Pham Big Data Machine Learning

Cluster Analysis Examples
Some other examples of cluster analysis are:

Identifying areas of similar topography, such as desert

region, grassy areas, mountains etc.

EL
Categorizing different types of tissues from medical images.
PT
Determining different groups of weather patterns, such as
snowy, dry, monsoon and
N
Discovering hot spots for different types of crime from
police reports.

Big Data Computing Vu Pham Big Data Machine Learning

Association Analysis
The goal in association analysis is to come up with a set of rules
to capture associations between items or events. The rules are used to
determine when items or events occur together.
A common application of association analysis is known as
market basket analysis. Which is used to understand customer

EL
purchasing behavior.
For example, association analysis can reveal that banking customers

PT
who have CDs, or Certificates of Deposits, also tend to be interested in
other investment vehicles such as money market accounts.
This information can be used for cross selling. If you advertise money
N
market accounts to your customers with CDs they are likely to open
such an account.

Big Data Computing Vu Pham Big Data Machine Learning

Association Analysis Examples
Some other applications of association analysis are
recommending similar items based on the purchasing behavior
or browsing histories of customers.

EL
Finding items that are often purchased together, such as
garden hose and potting soil, and offer sales on these related

PT
items at the same time to drive sales of both items.

Identifying web pages that are often accessed together so that

N
you can more efficiently offer up these related web pages at
the same time.

Big Data Computing Vu Pham Big Data Machine Learning

Classification of Machine Learning Techniques

So the different categories of machine learning

techniques are classification, regression, cluster
analysis, and association analysis.

EL
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

Supervised vs. Unsupervised
In supervised approaches the target, which is what the model
is predicting, is provided. This is referred to as having labeled
data because the target is labeled for every sample that you
have in your data set.

EL
Referring back to our example of predicting a weather category

PT
of sunny, windy, rainy or cloudy, every sample in the data set is
labeled as being one of these four categories. So the data is
labeled and predicting the weather categories is a supervised
N
task. In general, classification and regression are supervised
approaches.

Big Data Computing Vu Pham Big Data Machine Learning

Supervised vs. Unsupervised
In unsupervised approaches on the other hand, the target that
the model is predicting is unknown or unavailable. This means
that you have unlabeled data.

EL
Remember the cluster analysis example of segmenting
customers into different groups. The samples in your data are
not labeled with the correct group. Instead, the segmentation
PT
is performed using a clustering technique to group items based
on characteristics that they have in common.
N
Thus, the data is unlabeled and the task of grouping customers
into different segments is an unsupervised one. In general,
cluster analysis and association analysis are unsupervised
approaches.
Big Data Computing Vu Pham Big Data Machine Learning
Classification of Machine Learning Techniques

EL
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

Machine Learning Process

EL
PT
N

Big Data Computing VuPham

Vu Pham Big Data Machine Learning
Machine Learning Process
This diagram illustrates the steps in the machine
learning process.
Acquire

It should be kept in mind that all of these steps need to

EL
Prepare be carried out with a clear purpose in mind. That is, the
problem or opportunity that is being addressed must be
Analyze defined with clearly stated goals and objectives.

Report
PT
For example, the purpose of a project may be to study
N
customer purchasing behavior to come up with a more
effective marketing strategy in order to increase sales
Act revenue. The purpose behind the project will drive the
machine learning process.

Big Data Computing Vu Pham Big Data Machine Learning

Step 1: Acquire Data
The first step in the machine learning process is to get all
available data related to the problem at hand. Here, we need
to identify all data sources, collect the data, and finally
integrate data from these multiple sources.

EL
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

Step 2: Prepare Data
This step is further divided into two parts, explore data and pre-
process data.

The first part of data preparation involves preliminary exploration of

EL
the data to understand the nature of the data that we have to work
with. Things we want to understand about the data are its
characteristics, format, and quality. A good understanding of the data

PT
leads to a more informed analysis and a more successful outcome.
Once we know more about the data through exploratory analysis,
the next part is pre-processing of the data for analysis. This includes
N
cleaning data, selecting the variables to use, and transforming data
to make the data more suitable for analysis in the next step.

Big Data Computing Vu Pham Big Data Machine Learning

Step 3: Analyze Data
The prepared data then would be passed on to the analysis step. This
step involves selecting the analytical techniques to use, building a
model using the data, and assessing the results.

EL
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

Step 4: Communicate Results
This includes evaluating the results with respect to the goal set for
the project. Presenting the results in a easy to understand way
and communicating the results to others.

EL
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

Step 5: Apply the Results
The last step is to apply the results. This brings us back to the
purpose of the project. How can the insights from our analysis
be used to provide effective marketing to increase sales
revenue?

EL
Determining actions from insights gained from analysis is the
main focus of the act step.
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

Iterative Process
Note that the machine learning process is a very iterative
one. Findings from one step may require a previous step to be
repeated with new information.

EL
For example, during the prepare step, we may find some data
quality issues that may require us to go back to the acquire

PT
step to address some issues with data collection or to get
additional data that we didn't include in the first go around.
N

Big Data Computing Vu Pham Big Data Machine Learning

Goals and Activities in the
Machine Learning Process

EL
PT
N

Big Data Computing VuPham

Vu Pham Big Data Machine Learning
Goals and Activities in the Machine Learning Process

Here we will describe the goals of each step

and the key activities performed in each step.
Acquire

EL
Prepare

PT Analyze
N
Report

Act

Big Data Computing Vu Pham Big Data Machine Learning

Acquire Data
The first step in the data science process is to acquire the data. The
goal of the step is to identify and obtain all data related to the
problem at hand. First, we need to identify all related data and the
sources. Keep in mind, that data can come from different sources

EL
such as files, databases, the internet, mobile devices. So remember
to include all data related to the problem you are addressing.

PT
After you've identified your data and data sources, the next step is to
collect the data and integrate data from the different sources. This
may require conversion, as data can come in different formats. And it
N
may also require aligning the data, as data from different sources
may have different time or spatial resolutions. Once you’ve collected
and integrated your data, you now have a coherent data set for your
analysis.

Big Data Computing Vu Pham Big Data Machine Learning

Prepare Data
The next step after acquiring data is to prepare it to make
it suitable for analysis.

There are two parts to this step, explore data and

EL
preprocess data.

PT
N

Big Data Computing Vu Pham Big Data Machine Learning

Step-2-A: Explore
In data exploration, you want to do some preliminary
investigation in order to gain a better understanding of the
specific characteristics of your data. This in turn will guide the rest
of the process. With data exploration, you'll want to look for
things like correlations, general trends, outliers, etc.

EL
Correlations provide information about the relationship between

PT
variables in your data. Trends in your data will reveal if the
variable is moving in a certain direction, such as transaction
volume increasing throughout the year.
N
Outliers indicate potential problems with the data, or may
indicate an interesting data point that needs further examination.
Without this data exploration activity, you will not be able to use
your data effectively.
Big Data Computing Vu Pham Big Data Machine Learning
Describe your Data
One way to explore your data is to calculate summary statistics to
numerically describe the data.

Summary statistics are quantities that capture various characteristics

of a set of values with a single number, or a small set of numbers.

EL
Some basic summary statistics that you should compute for your
data set are mean, median, mode, range and standard deviation.
Mean and median are measures of the location of a set of values.

PT
Mode is the value that occurs most frequently in your data set, and
range and standard deviation are measures of spread in your data.
Looking at these measures will give you an idea of the nature of your
N
data. They can tell you if there's something wrong with your data.

For example, if the range of the values for age in your data includes
negative numbers, or a number much greater than a hundred,
there's something suspicious in the data that needs to be examined
Big Data Computing Vu Pham Big Data Machine Learning
Visualize Your Data
Visualization techniques also provide
quick and effective ways to explore your
data. Some examples are, a histogram,
such as the plot shown here, shows the
distribution of the data and can show
skewness or unusual dispersion in

EL
outliers.
A line plot, like the one in the lower left,
can be used to look at trends in the data,

PT
such as, the change in the price of a
stock. A heat map can give you an idea of
where the hot spots are.
N
A scatter plot effectively shows
correlation between two variables.
Overall, there are many types of plots to
visualize data. They are very useful in
helping you understand the data you
have.
Big Data Computing Vu Pham Big Data Machine Learning
Step-2-B: Pre-Process
The second part of the prepare step is preprocess. So, after
we've explored the data, we need to preprocess the data to
prepare it for analysis.

EL
The goal here is to create the data that will be used for
analysis.

PT
The main activities on this part are to clean the data, select the
appropriate variables to use and transform the data as needed.
N

Big Data Computing Vu Pham Big Data Machine Learning

Data Cleaning
A very important part of data preparation is to clean the data
to address quality issues. Real world data is nothing. There are
many examples of quality issues with data from real
applications including missing values, such as income in a

EL
survey, duplicate data, such as two different records for the
same customer with different addresses.

PT
Inconsistent or invalid data, such as a six digit zip code. Noise in
the collection of data that distorts the true values. Outliers,
N
such as a number much larger than 100 for someone's age. It is
essential to detect and address these issues that can negatively
affect the quality of the data.

Big Data Computing Vu Pham Big Data Machine Learning

Feature Selection
Feature selection refers to choosing the set of features to use that is
appropriate for the application. Feature selection can involve removing
redundant or irrelevant features, combining features, or creating new
features.
During the exploring data step, you may have discovered that two

EL
features are very correlated. In that case, one of these features can be
removed without negatively effecting the analysis results.

PT
For example, the purchase price of a product and the amount of sales
tax pain are very likely to be correlated. Eliminating the sales tax amount
then will be beneficial. Removing redundant or irrelevant features will
N
make the subsequent analysis simpler.
You may also want to combine features or create new ones. For
example, adding the applicants education level as a feature to a loan
approval application would make sense. There are also algorithms to
automatically determine the most relevant features based on various
mathematical properties.
Big Data Computing Vu Pham Big Data Machine Learning
Feature Transformation
Feature transformation maps the data from one format to
another. Various transformation operations exist. For example,
scaling maps the data values to a specified range to prevent
any one feature from dominating the analysis results.

EL
Filtering or aggregation can be used to reduce noise and

PT
variability in the data.

Dimensionality reduction maps the data to a smaller subset of

N
dimensions to simplify the subsequent analysis. We will discuss
techniques to prepare data in more detail later in this course.

Big Data Computing Vu Pham Big Data Machine Learning

Step-3: Analyze
After preparing the data to address data quality issues and
preprocess it to get it in the appropriate format, the next step in the
machine learning process is to analyze the data. The goals of the staff
are to build a machine learning model, to analyze the data and to

EL
evaluate the results that you get from the model.

The analyze steps starts with this determining the type of problem

PT
you have. You begin by selecting appropriate machine learning
techniques to analyze the data.
N
Then you construct the model using the data that you've prepared.
Once the model is built, you will want to apply it to new data
samples to evaluate how well the model performs. Thus data analysis
involves selecting the appropriate technique for your problem,
building the model, then evaluating the results.
Big Data Computing Vu Pham Big Data Machine Learning
Step-4: Report
The next step in the machine learning process is reporting results from your
analysis. In reporting your results, it is important to communicate your
insights and make a case for what actions should follow.
In reporting your results, you will want to think about what to present, as
well as how to present. In deciding what to present, you should consider

EL
what the main results are, what insights were gained from your analysis,
and what added value do these insights bring to the application.
Keep in mind that even negative results are valuable lessons learned, and

PT
suggest further avenues for additional analysis. Remember that all findings
must be presented so that informs decisions can be made for next steps.
In deciding how to present, remember that visualization is an important
N
tool in presenting your results.
Plots and summary statistics discussing the explore step can be used
effectively here as well. You should also have tables with details from your
analysis as backup, if someone wants to take a deeper dive into the results.
In summary, you want to report your findings by presenting your results and
the value added with graphs using visualization tools.
Big Data Computing Vu Pham Big Data Machine Learning
Step-5: Act
The final step in the machine loading process is to determine what action
should be taken based on the insights gained.

What action should be taken based on the results of your analysis? Should

EL
you market certain products to a specific customer segment to increase
sales? What inefficiency is can be removed from your process? What
incentives would be effective in attracting new customers?

PT
Once a specific action has been determined, the next step is to implement
the action. Things to consider here include, how can the action be added to
N
your application? How will end users be affected?

Assessing the impact of the implemented action is then necessary to

evaluate the benefits gained. The results of this assessment determine next
steps, which could suggest additional analysis or further opportunities,
which would begin another cycle of the machine learning process.
Big Data Computing Vu Pham Big Data Machine Learning
Generalization and Overfitting

EL
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

Errors in Classification
Recall that a machine learning model maps the input it receives
to an output. For a classification model, the model's output is
the predicted class label for the input variables and the true
class label is the target.
Then if the classifier predicts the correct classes label for a

EL
sample, that is a success. If the predicted class label is different
from the true class label, then that is an error.

PT
The error rate, then, is the percentage of errors made over the
entire data set. That is, it is the number of errors divided by the
total number of samples in a data set.
N
Error rate is also known as misclassification rate, or simply
error.

Big Data Computing Vu Pham Big Data Machine Learning

Training vs. Testing Phase

EL
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

Errors in Classification
The model is built using training data and evaluated on test
data. The training and test data are two different data sets. The
goal in building a machine learning model is to have the model
perform well on training, as well as test data.

EL
Error rate, or simply error, on the training data is referred to as

PT
training error, and the error on test data is referred to as test
error. The error on the test data is an indication of how well the
classifier will perform on new data.
N

Big Data Computing Vu Pham Big Data Machine Learning

Generalization
This is known as generalization. Generalization refers to how
well your model performs on new data, that is data not used to
train the model.

EL
You want your model to generalize well to new data. If your
model generalizes well, then it will perform well on data sets

PT
that are similar in structure to the training data, but doesn't
contain exactly the same samples as in the training set.
N
Since the test error indicates how well your model generalizes
to new data, note that the test error is also called
generalization error.

Big Data Computing Vu Pham Big Data Machine Learning

Overfitting
A related concept to generalization is overfitting. If your model
has very low training error but high generalization error, then it
is overfitting.

EL
This means that the model has learned to model the noise in
the training data, instead of learning the underlying structure
of the data.
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

Overfitting
These plots illustrate what happens when a model overfits. Training
samples are shown as points, and the input to output mapping that
the model has learned is indicated as a curve. The plot on the left
shows that the model has learned the underlying structure of the
data, as the curve follows the trend of the sample point as well. The

EL
plot on the right, however, shows that the model has learned to
model the noise in a data set.

PT
The model tries to capture every sample point, instead of the general
trend of the samples together. The training error and the
generalization error are plotted together, during model training.
N

Big Data Computing Vu Pham Big Data Machine Learning

Connection between overfitting and generalization

A model that overfits will not generalize well to new data.

So the model will do well on just the data it was trained on, but
given a new data set, it will perform poorly.

EL
A classifier that performs well on just the training data set will
not be very useful. So it is essential that the goal of good

PT
generalization performance is kept in mind when building a
model.
N

Big Data Computing Vu Pham Big Data Machine Learning

Overfitting and Underfitting
Overfitting occurs when the model is fitting to the noise in the
training data. This results in low training error and high test error.
Underfitting on the other hand, occurs when the model has not
learned the structure of the data. This results in high training error
and high test error.

EL
Both are undesirable, since both mean that the model will not
generalize well to new data. Overfitting generally occurs when a

PT
model is too complex, that is, it has too many parameters relative to
the number of training samples. So to avoid overfitting, the model
needs to be kept as simple as possible, and yet still solve the
N
input/output mapping for the given data set.

Big Data Computing Vu Pham Big Data Machine Learning

What Causes Overfitting
In summary, overfitting is when your model has
learned the noise in the training data instead of the
underlying structure of the data. You want to avoid

EL
overfitting so that your model will generalize well to
new data.

PT
N

Big Data Computing Vu Pham Big Data Machine Learning

 Overfitting in Decision Trees

EL
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

Decision Tree Induction
A decision tree, also referred to as tree induction, the
tree repeatedly splits the data in a node in order to
get successively paired subsets of data.

EL
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

Overfitting in Decision Trees
Note that a decision tree classifier can potentially expand its
nodes until it can perfectly classify samples in the training
data.
But if the tree grows nodes to fit the noise in the training
data, then it will not classify a new sample well.

EL
This is because the tree has partitioned the input space
according to the noise in the data instead of to the true

PT
structure of a data. In other words, it has overfit.
N

Big Data Computing Vu Pham Big Data Machine Learning

Avoiding Overfitting in Decision Trees

EL
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

Pre-pruning
With pre-pruning, the idea is to stop tree induction before a
fully grown tree is built that perfectly fits the training data.

To do this, restrictive stopping conditions for growing nodes

EL
must be used. For example, a nose stops expanding if the
number of samples in the node is less than some minimum
threshold.
PT
Another example is to stop expanding a note if the
N
improvement in the impurity measure falls below a certain
threshold.

Big Data Computing Vu Pham Big Data Machine Learning

Post-pruning
In post-pruning, the tree is grown to its maximum
size, then the tree is pruned by removing nodes using a
bottom up approach.

EL
That is, the tree is trimmed starting with the leaf
nodes. The pruning is done by replacing a subtree with a

PT
leaf node if this improves the generalization error, or if
there is no change to the generalization error with this
replacement.
N

Big Data Computing Vu Pham Big Data Machine Learning

Overfitting in Decision Trees
In other words, if removing a subtree does not have a negative
effect on the generalization error, then the nodes in that
subtree only add to the complexity of the tree, and not to its
overall performance.

EL
So those nodes should be removed. In practice, post-pruning

PT
tends to give better results. This is because pruning decisions
are based on information from the full tree. Pre-pruning, on
the other hand, may stop the tree growing process
N
prematurely. However, post-pruning is more computationally
expensive since the tree has to be expanded to its full size.

Big Data Computing Vu Pham Big Data Machine Learning

 Validation Set

EL
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

Training vs. Testing Phase

EL
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

Avoiding Overfitting
Recall that a model that overfits does not generalize well
to new data.

EL
Recall also that overfitting generally occurs when a model
is too complex.

PT
So to have a model with good generalization
performance, model training has to stop before the
N
model gets too complex.

How do you determine when this should occur?

Big Data Computing Vu Pham Big Data Machine Learning

Validation Set
A validation set can be used to guide the training process to
avoid overfitting and deliver good generalization
performance.
We have discussed having a training set and a separate test

EL
set. The training set is used to build a model and the test set is
used to see how the model performs a new data.

PT
N

Big Data Computing Vu Pham Big Data Machine Learning

Training and Validation Error
Now we want to further divide up the training data into a
training set and a validation set.

The training set is used to train the model as before and the

EL
validation set is used to determine when to stop training the
model to avoid overfitting, in order to get the best
generalization performance.
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

Training and Validation Error
The idea is to look at the errors on both training set and validation set
during model training as shown here. The orange solid line on the plot is the
training error and the green line is the validation error. We see that as
model building progresses along the x-axis, the number of nodes increases.
That is the complexity of the model increases. We can see that as the model

EL
complexity increases, the training error decreases. On the other hand, the
validation error initially decreases but then starts to increase.
When the validation error increases, this indicates that the model is

PT
overfitting, resulting in decreased generalization performance.
N

Big Data Computing Vu Pham Big Data Machine Learning

When to Stop Training ?
This can be used to determine when to stop training. Where validation error
starts to increase is when you get the best generalization performance, so
training should stop there. This method of using a validation set to
determine when to stop training is referred to as model selection since
you're selecting one from many of varying complexities. Note that this was

EL
illustrated for a decision tree classifier, but the same method can be applied
to any type of machine learning model.

PT
N

Big Data Computing Vu Pham Big Data Machine Learning

Ways to create and use the validation set
There are several ways to create and use the validation set to
avoid overfitting. The different methods are:

Holdout method

EL
Random subsampling

PT
K-fold cross-validation, and
N
Leave-one-out cross-validation

Big Data Computing Vu Pham Big Data Machine Learning

Holdout Method
The first way to use a validation set is the holdout method. This
describes the scenario that we have been discussing, where
part of the training data is reserved as a validation set. The
validation set is then the holdout set.

EL
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

Holdout Method
Errors on the training set and the holdout set are calculated at each step during
model training and plotted together as we've seen before. And the lowest error on
the holdout set is when training should stop. This is the just the process that we
have described here before. There's some limitations to the holdout method
however.
First, since some samples are reserved for the holdout validation set, the training set

EL
now has less data than it originally started out with.
Secondly, if the training and holdout sets do not have the same data distributions,
then the results will be misleading. For example, if the training data has many more

PT
samples of one class and the holdout dataset has many more samples of another
class.
N

Big Data Computing Vu Pham Big Data Machine Learning

Repeated Holdout Method
Repeating holdout method several times

Randomly select different hold out set each iteration

EL
Average validation errors over all repetitions.
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

K-Fold Cross-Validation
A way to improve on the repeated holdout method is use cross-
validation. Cross-validation works as follows. Segment the data into k
number of disjoint partitions. During each iteration, one partition is
used as the validation set. Repeat the process k times. Each time
using a different partition for validation. So each partition is used for

EL
validation exactly once. This is illustrated in this figure. In the first
iteration, the first partition, is used for validation. In the second
iteration, the second partition is used for validation and so on.

PT
N

Big Data Computing Vu Pham Big Data Machine Learning

K-Fold Cross-Validation
The overall validation error is calculated by averaging the validation
errors for all k iterations.

The model with the smallest average validation error then is

EL
selected. The process we just described is referred to as k-fold cross-
validation. This is a very commonly used approach to model selection
in practice.

PT
This approach gives you a more structured way to divide available
N
data up between training and validation datasets and provides a way
to overcome the variability in performance that you can get when
using a single partitioning of the data.

Big Data Computing Vu Pham Big Data Machine Learning

Leave-One-Out Cross-Validation
Leave-one-out cross-validation is a special case of k-fold cross-
validation where k equals N, where N is the size of your dataset.

Here, for each iteration the validation set has exactly one sample. So

EL
the model is trained to using N minus one samples and is validated
on the remaining sample.

PT
The rest of the process works the same way as regular k-fold cross-
validation.
N
Note that cross-validation is often abbreviated CV and leave-one-out
cross-validation is in abbreviated L-O-O-C-V and pronounced LOOCV.

Big Data Computing Vu Pham Big Data Machine Learning

Uses of Validation Set
Note that the validation error that comes out of this process can also
be used to estimate generalization performance of the model.

In other words, the error on the validation set provides an estimate

EL
of the error on the test set.

PT
N

Big Data Computing Vu Pham Big Data Machine Learning

Uses of Datasets
With the addition of the validation set, you really need three distinct
datasets when you build a model. Let's review these datasets.

The training dataset is used to train the model, that is to adjust the

EL
parameters of the model to learn the input to output mapping.

PT
The validation dataset is used to determine when training should
stop in order to avoid overfitting.
N
The test data set is used to evaluate the performance of the model
on new data.

Big Data Computing Vu Pham Big Data Machine Learning

Uses of Datasets
Note that the test data set should never, ever be used in any way to
create or tune the model. It should not be used, for example, in a
cross-validation process to determine when to stop training.

EL
The test dataset must always remain independent from model
training and remain untouched until the very end when all training
has been completed. Note that in sampling the original dataset to

PT
create the training, validation, and test sets, all datasets must contain
the same distribution of the target classes.
N
For example, if in the original dataset, 70% of the samples belong to
one class and 30% to the other class, then this same distribution
should approximately be present in each of the training, validation,
and test sets. Otherwise, analysis results will be misleading.

Big Data Computing Vu Pham Big Data Machine Learning

Validation Set Summary
We have discuss the need for three different datasets in
building model. A training set to train the model, a
validation set to determine when to stop training, and a
test to evaluate performance on new data.

EL
We learned how a validation set can be used to avoid
PT
overfitting and in the process, provide an estimate of
generalization performance.
N
And we covered different ways to create and use a
validation set such as k-fold cross-validation.

Big Data Computing Vu Pham Big Data Machine Learning

 Metrics to Evaluate Model

EL
Performance

PT
N

Big Data Computing Vu Pham Big Data Machine Learning

Class Labels

EL
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

Types of Classification Errors

EL
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

Accuracy Rate

EL
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

Error Rate

EL
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

Precision and Recall

EL
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

Precision and Recall
Precision is considered a measure of exactness because it
calculates the percentage of samples predicted as
positive, which are actually in a positive class.

EL
Recall is considered a measure of completeness, because
it calculates the percentage of positive samples that the
PT
model correctly identified.
N

Big Data Computing Vu Pham Big Data Machine Learning

F-Measure

EL
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

 Confusion Matrix

EL
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

Classification

EL
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

Types of Classification Errors

EL
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

Confusion Matrix

EL
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

 EL
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

 Machine Learning Tool

EL
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

Spark MLlib
Spark MLlib is a distributed machine-learning
framework on top of Spark Core.

EL
MLlib is Spark's scalable machine learning library
consisting of common learning algorithms and utilities,
including classification, regression, clustering,
PT
collaborative filtering, dimensionality reduction.
N

Big Data Computing Vu Pham Big Data Machine Learning

Spark MLlib Component

EL
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

 Classification in Spark

EL
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

Steps for Classification in Spark
1. Load weather data into DataFrame
2. Drop unused and missing data.

EL
3. Create a categorical variable for low humidity days

PT
4. Aggregate features used to make predictions.
5. Split the data into training and test sets.
N
6. Create and train the decision tree.
7. Save the predictions to a CSV file

Big Data Computing Vu Pham Big Data Machine Learning

1. Load weather data into DataFrame
The first cell contains the classes we need to load to run this exercise. Next we
create SQL context and load the weather data CSV into a data frame.
The second cell also prints all the columns in this data frame.
The third cell defines the columns in the weather data we will use for the
decision tree classifier.

EL
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

 EL
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

2. Drop unused and missing data
Use the column name number.
Drop that from the data frame, df = df.drop Number.
Rows with missing data, df = df.na.drop.
Print the number of rows and columns in our resulting data frame,
df.count(), len(df.columns).

EL
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

3. Create a categorical variable for low humidity days
We will enter binarizer = Binarizer (). The first argument specifies a threshold
value for the variable. We want the categorical variable to be 1, if the humidity is
greater than 25%. So we'll enter a threshold=24.9999. The next argument
specifies the column to use to create the categorical variable. We'll input,
inputCol = relative_humidity_3pm. The final argument specifies the new column

EL
name, outputCol = label. A new data frame is created with this categorical
variable. binarizeredDF = binarizer.transform df.

PT
N

Big Data Computing Vu Pham Big Data Machine Learning

4. Aggregate features used to make predictions.
The relative humidity in the first row is greater than 25% and the label is 1. The relative
humidity in the second, third, and fourth rows are less than 25%, and the label is 0. For
aggregating the features we will use to make predictions into a single col, assembler =
VectorAssember( ). The first argument is a list of the columns to be aggregated.
inputCols=featureColumns, and the second argument is the name of the new column
containing the aggregated features, outputCol=features. We can create the new data frame

EL
by running assembled=assembler.transform binarizedDF.

PT
N

Big Data Computing Vu Pham Big Data Machine Learning

5. Split the data into training and test sets.
Next split the data set into two parts, one for training data and one for test
data. We can do this by entering (training Data,
testData)=assembled.randomSplit([0.8, 0.2], seed=13234). We can see the
size of the two sets by running count, trainingData.count(), testData.count
().

EL
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

6. Create and train the decision tree.
Next, we can create a model by training the decision tree. We can do this by
executing it in a pipeline. Enter pipeline=Pipeline (stages=[dt]). We will create
them all by putting a training data, model = pipeline.fit trainingData. Let's run this
Now, we can make predictions using our test data. Enter predictions =
model.transform(testData).

EL
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

Look at the first 10 rows of the prediction by running,
predictions.select('prediction', label') show(10). Here we can
see in the first ten rows, the prediction matches the label.

EL
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

7. Save the predictions to a CSV file
Save the predictions to a CSV file. We can save it by running
predictions.select('prediction', 'label').write.save(path=‘).
Format='com.databricks.spark.csv.

EL
PT
N

Big Data Computing Vu Pham Big Data Machine Learning

Conclusion
In this lecture, we have discussed the machine learning
techniques.

EL
We have also discussed tools and algorithms that you
can use to create machine learning models that learn

PT
from data, and to scale those models up to big data
problems.
N

Big Data Computing Vu Pham Big Data Machine Learning

 Machine Learning Algorithm
Parallel K-means using Map Reduce

EL
for Big Data Analytics

PT
N
Dr. Rajiv Misra
Dept. of Computer Science & Engg.
Indian Institute of Technology Patna
[email protected]
Big Data Computing Vu Pham Machine Learning Classification Algorithm
Preface
Content of this Lecture:
In this lecture, we will discuss machine learning
classification algorithm k-means using mapreduce for
big data analytics

EL
PT
N

Big Data Computing Vu Pham Machine Learning Classification Algorithm

Cluster Analysis Overview
Goal: Organize similar items into groups.

In cluster analysis, the goal is to organize similar items in given

EL
data set into groups or clusters. By segmenting given data into
clusters, we can analyze each cluster more carefully.

PT
Note that cluster analysis is also referred to as clustering.
N

Big Data Computing Vu Pham Machine Learning Classification Algorithm

Applications: Cluster Analysis
Segment customer base into groups: A very common application of
cluster analysis is to divide your customer base into segments based on their
purchasing histories. For example, you can segment customers into those
who have purchased science fiction books and videos, versus those who tend
to buy nonfiction books, versus those who have bought many children's

EL
books. This way, you can provide more targeted suggestions to each different
group.

PT
Characterize different weather patterns for a region: Some other
examples of cluster analysis are characterizing different weather patterns for
a region.
N
Group news articles into topics: Grouping the latest news articles into
topics to identify the trending topics of the day.
Discover crime hot spots: Discovering hot spots for different types of
crime from police reports in order to provide sufficient police presence for
problem areas.

Big Data Computing Vu Pham Machine Learning Classification Algorithm

Cluster Analysis
Divides data into clusters: Cluster analysis divides all the samples in a data set
into groups. In this diagram, we see that the red, green, and purple data points
are clustered together. Which group a sample is placed in is based on some
measure of similarity.
Similar items are placed in same cluster: The goal of cluster analysis is to

EL
segment data so that differences between samples in the same cluster are
minimized, as shown by the yellow arrow, and differences between samples of
different clusters are maximized, as shown by the orange arrow. Visually, we

PT
can think of this as getting samples in each cluster to be as close together as
possible, and the samples from different clusters to be as far apart as possible.
N

Big Data Computing Vu Pham Machine Learning Classification Algorithm

Similarity Measures
Cluster analysis requires some sort of metric to measure
similarity between two samples. Some common
similarity measures are Euclidean distance, which is the
distance along a straight line between two points, A and
B, as shown in this plot.
Manhattan distance, which is calculated on a strictly

EL
horizontal and vertical path, as shown in the right plot.
To go from point A to point B, you can only step along
either the x-axis or the y-axis in a two-dimensional case.

PT
So the path to calculate the Manhattan distance consists
of segments along the axes instead of along a diagonal
path, as with Euclidean distance.
N
Cosine similarity measures the cosine of the angle
between points A and B, as shown in the bottom plot.
Since distance measures such as Euclidean distance are
often used to measure similarity between samples in
clustering algorithms, note that it may be necessary to
normalize the input variables so that no one value
dominates the similarity calculation.
Big Data Computing Vu Pham Machine Learning Classification Algorithm
Another natural inner product measure

EL
PT
N

Big Data Computing Vu Pham Machine Learning Classification Algorithm

Another natural inner product measure

EL
PT
N

Big Data Computing Vu Pham Machine Learning Classification Algorithm

Cosine similarity- normalize

EL
PT
N

Big Data Computing Vu Pham Machine Learning Classification Algorithm

Cosine similarity- normalize

EL
PT
N
Not a proper distance metric
Efficient to compute for
sparse vecs
Big Data Computing Vu Pham Machine Learning Classification Algorithm
Normalize

EL
PT
N

Big Data Computing Vu Pham Machine Learning Classification Algorithm

Cosine Normalize

EL
PT
In general, -1 < similarity < 1
N
For positive features (like tf-idf)
0 < similarity < 1

Define distance = 1-similarity

Big Data Computing Vu Pham Machine Learning Classification Algorithm
Cluster Analysis Key Points
Cluster analysis is an unsupervised task. This means that
Unsupervised there is no target label for any sample in the data set.

In general, there is no correct clustering results. The best

set of clusters is highly dependent on how the resulting
There is no clusters will be used.

EL
‘correct’ clustering Clusters don't come with labels. You may end up with
five different clusters at the end of a cluster analysis
Clusters don’t
PT
process, but you don't know what each cluster
come with labels represents. Only by analyzing the samples in each cluster
can you come out with reasonable labels for your
clusters. Given all this, it is important to keep in mind
N
that interpretation and analysis of the clusters are
required to make sense of and make use of the results of
cluster analysis.

Interpretation and analysis required to

make sense of clustering results!
Big Data Computing Vu Pham Machine Learning Classification Algorithm
Uses of Cluster Results
Data segmentation
Analysis of each segment can provide insights: There are several ways
that the results of cluster analysis can be used. The most obvious is data segmentation
and the benefits that come from that. If we segment your customer base into different

EL
types of readers, the resulting insights can be used to provide more effective marketing
to the different customer groups based on their preferences. For example, analyzing
each segment separately can provide valuable insights into each group's likes, dislikes
and purchasing behavior, just like we see science fiction, non-fiction and children's
books preferences here.
PT Science fiction
N
non-fiction

children’s

Big Data Computing Vu Pham Machine Learning Classification Algorithm

Uses of Cluster Results
Categories for classifying new data
New sample assigned to closest cluster: Clusters can also be used to
classify new data samples. When a new sample is received, like the orange sample
here, compute the similarity measure between it and the centers of all clusters, and

EL
assign a new sample to the closest cluster. The label of that cluster, manually
determined through analysis, is then used to classify the new sample. In our book
buyers' preferences example, a new customer can be classified as being either a
science fiction, non-fiction or children's books customer depending on which cluster

PT
the new customer is most similar to.
N
Label of closest
cluster used to
classify new sample

Big Data Computing Vu Pham Machine Learning Classification Algorithm

Uses of Cluster Results
Labeled data for classification
Cluster samples used as labeled data: Once cluster labels have been
determined, samples in each cluster can be used as labeled data for a classification
task. The samples would be the input to the classification model. And the cluster label

EL
would be the target class for each sample. This process can be used to provide much
needed labeled data for classification.

PT Label of closest
cluster used to
N
classify new sample

Labeled samples for

science fiction
customers
Big Data Computing Vu Pham Machine Learning Classification Algorithm
Uses of Cluster Results
Basis for anomaly detection
Cluster outliers are anomalies: Yet another use of cluster results is as a basis
for anomaly detection. If a sample is very far away, or very different from any of the
cluster centers, like the yellow sample here, then that sample is a cluster outlier and

EL
can be flagged as an anomaly. However, these anomalies require further analysis.
Depending on the application, these anomalies can be considered noise, and should be
removed from the data set. An example of this would be a sample with a value of 150
for age.

PT
N
Anomalies that require
further analysis

Big Data Computing Vu Pham Machine Learning Classification Algorithm

Cluster Analysis Summary

Organize similar items into groups

EL
Analyzing clusters often leads to useful insights about
data

PT
Clusters require analysis and interpretation
N

Big Data Computing Vu Pham Machine Learning Classification Algorithm

k-Means Algorithm
Select k initial centroids (cluster centers)

Repeat

EL
Assign each sample to closest centroid
Calculate mean of cluster to determine new centroid

PT
Until some stopping criterion is reached
N

Big Data Computing Vu Pham Machine Learning Classification Algorithm

k-Means Algorithm: Example

EL
Original samples Initial centroids Assign samples

PT
N
Re-calculate Assign samples Re-calculate
centroids centroids

Big Data Computing Vu Pham Machine Learning Classification Algorithm

Choosing Initial Centroids
Issue:
Final clusters are sensitive to initial centroids

EL
Solution:
Run k-means multiple times with different random

PT
initial centroids, and choose best results
N

Big Data Computing Vu Pham Machine Learning Classification Algorithm

Evaluating Cluster Results

error=distance between sample & centroid

EL
Squared error = error2

Sum of squared errors PT

between all samples & centroid
N
Sum over all clusters ➔ WSSE
(Within-Cluster Sum
of Squared Error)
Big Data Computing Vu Pham Machine Learning Classification Algorithm
Using WSSE

WSSE1 < WSSE2 ➔ WSSE1 is better numerically

EL
Caveats:
Does not mean that cluster set 1 is more ‘correct’ than
cluster set 2 PT
N
Larger values for k will always reduce WSSE

Big Data Computing Vu Pham Machine Learning Classification Algorithm

 Choosing Value for k k=?

Approaches: Choosing the optimal value for k is always a

big question in using k-means. There are several methods
to determine the value for k.
• Visualization: Visualization techniques can be used to determine the

EL
dataset to see if there are natural groupings of the samples. Scatter plots and the
use of dimensionality reduction are useful here, to visualize the data.

Application-Dependent:
•

PT A good value for k is application-dependent.

So domain knowledge of the application can drive the selection for the value of k.
For example, if you want to cluster the types of products customers are
purchasing, a natural choice for k might be the number of product categories that
N
you offer. Or k might be selected to represent the geographical locations of
respondents to a survey. In which case, a good value for k would be the number of
regions your interested in analyzing.

• Data-Driven: There are also data-driven method for determining the value
of k. These methods calculate symmetric for different values of k to determine the
best selections of k. One such method is the elbow method.
Big Data Computing Vu Pham Machine Learning Classification Algorithm
 Elbow Method for Choosing k

“Elbow” suggests
value for k should be 3

EL
PT
The elbow method for determining the value of k is shown on this plot. As we saw in the
previous slide, WSSE, or within-cluster sum of squared error, measures how much data
samples deviate from their respective centroids in a set of clustering results. If we plot WSSE
N
for different values for k, we can see how this error measure changes as a value of k changes
as seen in the plot. The bend in this error curve indicates a drop in gain by adding more
clusters. So this elbow in the curve provides a suggestion for a good value of k.

Note that the elbow can not always be unambiguously determined, especially for complex
data. And in many cases, the error curve will not have a clear suggestion for one value, but
for multiple values. This can be used as a guideline for the range of values to try for k.
Big Data Computing Vu Pham Machine Learning Classification Algorithm
Stopping Criteria
When to stop iterating ?
• No changes to centroids: How do you know when to stop
iterating when using k-means? One obviously stopping criterion is when
there are no changes to the centroids. This means that no samples would

EL
change cluster assignments. And recalculating the centroids will not result
in any changes. So additional iterations will not bring about any more
changes to the cluster results.
•
PT
Number of samples changing clusters is below
threshold: The stopping criterion can be relaxed to the second
N
stopping criterion: which is when the number of sample changing clusters
is below a certain threshold, say 1% for example. At this point, the
clusters are changing by only a few samples, resulting in only minimal
changes to the final cluster results. So the algorithm can be stopped here.

Big Data Computing Vu Pham Machine Learning Classification Algorithm

Interpreting Results
Examine cluster centroids
• How are clusters different ? At the end of k-means we have a set of
clusters, each with a centroid. Each centroid is the mean of the samples assigned
to that cluster. You can think of the centroid as a representative sample for that
cluster. So to interpret the cluster analysis results, we can examine the cluster

EL
centroids. Comparing the values of the variables between the centroids will reveal
how different or alike clusters are and provide insights into what each cluster
represents. For example, if the value for age is different for different customer
clusters, this indicates that the clusters are encoding different customer segments

PT
by age, among other variables.
N

Big Data Computing Vu Pham Machine Learning Classification Algorithm

K-Means Summary

Classic algorithm for cluster analysis

EL
Simple to understand and implement and is efficient

PT
Value of k must be specified
N
Final clusters are sensitive to initial centroids

Big Data Computing Vu Pham Machine Learning Classification Algorithm

Parallel K-means using Map
Reduce on Big Data Cluster

EL
Analysis
PT
N

Big Data Computing Vu Pham Machine Learning Classification Algorithm

MapReducing 1 iteration of k-means
Classify: Assign observations to closest cluster center

Map: For each data point, given ({μj},xi), emit(zi,xi)

EL
Recenter: Revise cluster centers as mean of assigned
observations
PT
N
Reduce: Average over all points in cluster j (zi=k)

Big Data Computing Vu Pham Machine Learning Classification Algorithm

Classification step as Map
Classify: Assign observations to closest cluster center

EL
PT
N

Big Data Computing Vu Pham Machine Learning Classification Algorithm

Recenter step as Reduce
Recenter: Revise cluster centers as mean of
assigned observations

EL
reduce(j, x_in_cluster j : [x1, x3,…, ])
sum = 0
count = 0 PT
for x in x_in_cluster j
N
sum += x
count += 1
emit(j, sum/count)

Big Data Computing Vu Pham Machine Learning Classification Algorithm

Distributed KMeans Iterative Clustering

EL
PT
N

Big Data Computing Vu Pham Machine Learning Classification Algorithm

Distributed KMeans Iterative Clustering

EL
Find Nearest Center
PT Key is Center, Value is Movie
N
Average Ratings

Big Data Computing Vu Pham Machine Learning Classification Algorithm

Summary of Parallel k-means using MapReduce

Map: classification step;

data parallel over data points

EL
PT
Reduce: recompute means;
data parallel over centers
N

Big Data Computing Vu Pham Machine Learning Classification Algorithm

Some practical considerations
k-means needs an iterative version of MapReduce
Not standard formulation

EL
Mapper needs to get data point and all centers
A lot of data!
PT
Better implementation:
N
mapper gets many data points

Big Data Computing Vu Pham Machine Learning Classification Algorithm

Conclusion

In this lecture, we have given an overview of cluster

analysis and also discussed machine learning
classification algorithm k-means using Mapreduce for

EL
big data analytics

PT
N

Big Data Computing Vu Pham Machine Learning Classification Algorithm