This document provides an introduction to machine learning, including definitions of key concepts like supervised vs. unsupervised learning. It discusses traditional machine learning assumptions like having fully labeled small datasets and stationary data. However, it notes real-world data often has huge amounts of unlabeled or streaming data. It emphasizes tackling these challenges with approaches like semi-supervised learning and stream mining. Overall, the document aims to give an overview of machine learning tasks and algorithms, highlighting the need to select the right methods for problems while working with domain experts on complex real-world data.

1. Machine Learning for Big Data
Prof. Dr. Eirini Ntoutsi
FG Intelligent Systems
Faculty of Electrical Engineering and Computer Science
Leibniz University Hannover & L3S Research Center
Introduction to Machine Learning

2. Overview
 A Machine Learning primer
 Machine Learning in the real world
Prof. Dr. Eirini Ntoutsi - Introduction to Machine Learning

3. A Machine Learning primer
Prof. Dr. Eirini Ntoutsi - Introduction to Machine Learning

4. What is Machine Learning?
 ML “gives computers the ability to learn without being explicitly
programmed” (Arthur Samuel, 1959)
 We don’t codify the solution. We don’t even know it!
 Data is the key & the learning algorithm
Prof. Dr. Eirini Ntoutsi - Introduction to Machine Learning
Algorithms
Models
Models
Automatic decision
making
Data
How can we build computer programs that
automatically improve with experience?

5. How do machines learn?
 A machine is said to learn from experience E with respect to some class of
tasks T and performance measure P, if its performance at tasks in T, as
measured by P, improves with experience E.
Tom Mitchell, Machine Learning book
 Example
 Task T: Recognize good and bad products in a production system (e.g., a
drilling machine)
 Experience E: instances of good and bad products
 Performance measure P: % of correctly identified products
Prof. Dr. Eirini Ntoutsi - Introduction to Machine Learning

6. (Machine) Learning from experience/feedback
 Experience comes in terms of data (the so called, instances or examples)
from the specific problem/ application
 In our example, instances correspond to certain characteristics of the
product, e.g.,
 Shape descriptors
 weight
 Roughness of the surface
 …
 Except for the instance description, we might also have feedback on those
instances from some “teacher”/”expert“
 E.g., whether the produced product is good or bad
Prof. Dr. Eirini Ntoutsi - Introduction to Machine Learning

7. (Machine) Learning from experience/feedback
 Based on the feedback, we can distinguish between:
 Direct-feedback instances
 the correct response /label is provided for each instance by the “teacher”
 e.g., good or bad product
 No-feedback instances
 no evaluation/label of the instance is provided, since there is no “teacher“
 e.g., no information on whether a product is good or bad, just the description of the
product/instance
 Indirect-feedback instances
 less feedback is given, since not the proper action, but only an evaluation of the
chosen action is given by the teacher
Prof. Dr. Eirini Ntoutsi - Introduction to Machine Learning
Supervised learning
Reinforcement learning
Unsupervised learning

8. Unsupervised learning
 Unsupervised learning/ Descriptive:
 Only a description of the instances is available
 No feedback/labels are available
 The goal is to discover groups of similar instances
 Typical examples: clustering, association rules, outlier detection
Prof. Dr. Eirini Ntoutsi - Introduction to Machine Learning
Height[cm]
Width[cm]
Cluster 1Cluster 2
instance width height
1 2,6 4,5
2 3,7 7,3
3 4,1 6,5
4 8,5 8,1
5 9,5 5,5
… … …
nails paper clips

9. Unsupervised learning: Clustering
 A huge variety of clustering algorithms
Prof. Dr. Eirini Ntoutsi - Introduction to Machine Learning
1
2
3
4
5
6
1
2
3 4
5
Partitioning methods
(k-Means)
Grid-based methods
(CLIQUE)
Model-based methods
(DBSCAN)
Hierarchical methods Constraint-based methods
Model-based methods
(EM)

10. Supervised learning
 Supervised learning/ Predictive:
 A description of the instances and their class labels is available
(training set)
 The goal is to learn a mapping from the instances to the class labels,
i.e., given a future unseen instance to predict its class label
 Typical examples: classification, regression, outlier detection
Prof. Dr. Eirini Ntoutsi - Introduction to Machine Learning
Screw
Nails
Paper clips
New object
Height[cm]
Width[cm]
New object
instance width height class
1 2,6 4,5 A
2 3,7 7,3 A
3 4,1 6,5 A
4 8,5 8,1 B
5 9,5 5,5 B
… … … …

11. Supervised learning: classification
 A huge variety of classification algorithms
Prof. Dr. Eirini Ntoutsi - Introduction to Machine Learning
Decision trees k nearest neighbours Support vector machines
Neural networks Bayesian classifiers Ensembles

12. Supervised learning: classification
 Different methods different partitionings of the feature space
Prof. Dr. Eirini Ntoutsi - Introduction to Machine Learning
http://scikit-learn.org/stable/auto_examples/classification/plot_classifier_comparison.html#sphx-glr-auto-examples-classification-plot-classifier-comparison-py

13. Supervised learning: regression
 Similar to classification, but the feature-result to be learned is continuous rather
than discrete.
 Goal: Predict a value of a given continuous valued variable based on the values of
other variables, assuming a linear or nonlinear model of dependency.
Prof. Dr. Eirini Ntoutsi - Introduction to Machine Learning
Given this data, a friend has a house 750
square feet - how much can they be
expected to get?

14. Reinforcement learning
 The learning machine interacts with its environment via actions
 Minimal feedback is provided regarding how the learning machine is
performing
 Feedback in terms of reward
 The goal of the agent is to learn a policy so as to maximize the expected
rewards
Prof. Dr. Eirini Ntoutsi - Introduction to Machine Learning
Source: https://en.wikipedia.org/wiki/Reinforcement_learning

15. Machine Learning in the Real World
Prof. Dr. Eirini Ntoutsi - Introduction to Machine Learning

16. ML in the real world
 Traditional ML assumptions
 The datasets are small and fit in memory
 Data is of a single type (e.g., numerical or text or images)
 For supervised learning
 The classes are well represented in the population (class balance)
 Labels are available for all instances (fully supervised learning)
 The quality of data (and labels) is good
 Data is stationary (the whole dataset is known in advance and there are no
changes in the data characteristics with time)
 …
Prof. Dr. Eirini Ntoutsi - Introduction to Machine Learning

17. Real world data
 The data manifest all Vs of big data
Prof. Dr. Eirini Ntoutsi - Introduction to Machine Learning
Source: http://blog.eoda.de/wp-content/uploads/2013/10/dv1.jpg

18. Big Data
Tackling the traditional restrictive ML assumptions
 The myth: we have big (labeled) data
 The reality:
 Huge amounts of unlabeled data
 Only a few labeled data
 Goal: Use both labeled and unlabeled
data for training
 Related ML areas
 Semi-supervised learning
Prof. Dr. Eirini Ntoutsi - Introduction to Machine Learning
Unlabeled
Labeled
Iosifidis&Ntoutsi, Large scale sentiment annotation with limited labels, KDD 2017

19. Tackling the traditional restrictive ML assumptions
 The myth: data is stationary
 The reality:
 Data are collected over time and their characteristics
might change  data streams
 Goal: maintain valid models of the population
 Related ML areas:
 stream mining, adaptive ML
Prof. Dr. Eirini Ntoutsi - Introduction to Machine Learning
1
0
1
1
1
0
1
0
0
1
1
[Zhang et al, Journal Neurocomputing 2012]

20. Tackling the traditional restrictive ML assumptions
 Similar example for stream clustering
Prof. Dr. Eirini Ntoutsi - Introduction to Machine Learning
Time T1 Time T2
Time T3
Cluster expands
Cluster shrinks
Cluster is split
Spiliopoulou, Ntoutsi, Theodoridis & Schult, MONIC and Followups on
Modeling and Monitoring Cluster Transitions, ECML PKDD 2013

21. Summary
 Machine learning is an exciting field with a huge variety of learning tasks
and algorithms for each task
 Different methods come with different assumptions, strengths and
limitations.
 The selection of the right method (and correct parameterization) is important
and requires a deep understanding of the methods and of the problem at hand
 A close cooperation with domain experts is required.
 Production systems (industry in general) impose new challenges for ML
due to their data complexity (volume, velocity, veracity, variety, value, …)
 “Factories are AI’s next frontier”, Andrew Ng
 https://www.technologyreview.com/s/609770/andrew-ng-says-factories-are-
ais-next-frontier/
 Landing AI startup to closely work with manufacturers like Foxconn, the
world’s largest contract manufacturer and maker of Apple’s iPhones
Prof. Dr. Eirini Ntoutsi - Introduction to Machine Learning

22. Thank you for your attention!
Questions/Comments?
Prof. Dr. Eirini Ntoutsi - Introduction to Machine Learning
Prof Dr. Eirini Ntoutsi
FG Intelligent Systems
Faculty of Electrical Engineering and Computer Science
Leibniz University Hannover & L3S Research Center
http://www.kbs.uni-hannover.de/~ntoutsi/
ntoutsi@l3s.de