Data, Methods,

and Scenarios

Mining Massive Datasets

Prof. Carlos "ChaTo" Castillo (they/them)
https://github.com/chatox/data-mining-course/
 Main Sources
● Data Mining, The Textbook (2015) by Charu Aggarwal
(Chapter 1) + slides by Lijun Zhang
● Mining of Massive Datasets, 2nd edition (2014) by
Leskovec et al. (Chapter 1)
● Data Mining Concepts and Techniques, 3rd edition (2011)
by Han et al. (Chapters 1-2)
 Contents
●Types of data
●Types of problem
●Example scenarios
●Major challenges
Data types
 Nondependency / Dependency
●Nondependency oriented data can be
structured so items are separate
− Relational data, text data
●Dependency oriented data includes
relationships between items
− Graphs, time series
 Mixed attribute data
●
Most attributes we will deal with are numerical, they quantify
something
●
Sometimes attributes are categorical
− Example: elephant, tiger, moose, ...
− Binary (two categories)
●
Example: present, absent
− Ordinal (two or more categories that can be naturally sorted)
●
Example: low, medium, high
●
Real-world datasets include a mixture of types
Binary attributes, sets, dummy vars.
●
Every binary attribute can be used as a marker of
belonging to a set and viceversa
●
One-hot encoding: every categorical attribute
takingName
one of kZip
values
code can be encoded
Parent Capacity as k
“dummy”
Moogbinary attributes
08001 NULL Small
Macarena 08002 NULL Small
Input 08038 NULL Medium
Loft 08018 Razzmatazz Large
Nitsa 08004 Apolo Large
 Question
Suppose you encode capacity using one-hot encoding.
How many columns will your new dataset have?

Name Zip code ParentCo Capacity

Moog 08001 NULL Small
Macarena 08002 NULL Small
Input 08038 NULL Medium
Loft 08018 Razzmatazz Large
Nitsa 08004 Apolo Large
 Textual data
● Text can be represented as:
− As a string
− “Bag of words”: a set of binary
variables, one for each word in
the dictionary, with value True iff
the word belongs to the text
− “Vector space”: a set of
numerical variables indicating
number of occurrences (often
normalized by collection
frequency)
http://uc-r.github.io/creating-text-features
 Time series data
●Contextual attributes
− Timestamps, sequence number, …
●Behavioral attributes
− Readings of a sensor, value of the variable, …

Multivariate time series data has multiple

behavioral attributes
 Spatial data
Girona elevation map
●Two (lat/long) or three
(lat/long/elevation) spatial
attributes
●This is remote sensing data,
including satellite and aerial
photos
 Spatiotemporal data
Two main representations:
●Spatial and temporal attributes are contextual
− Example: sea surface temperature
●Temporal attribute is contextual, spatial attribute
is behavioral
− Example: trajectories
Example: trajectory data aggregation

Bonchi, F., Castillo, C., Donato, D., & Gionis, A. (2009). Taxonomy-driven lumping for sequence mining.
Data Mining and Knowledge Discovery, 19(2), 227-244.
Example: Functional regions in cities

H. Assem, B. Caglayan, T.S. Buda, D. O’Sullivan. ST-DenNetFus: Deep Spatio-Temporal Dense Networks for Network Demand Prediction.
The European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases, 2018
Problem types
 Data mining methods try to find
relationships
●Between columns
− Find associations, correlations, …
− If there is one key column: classification, prediction, ...
●Between rows
− Find clusters
− Detect outliers
 Example:
Association pattern mining
●
Sparse binary databases
representing, e.g., items a person is
interested in

●
The relative frequency of a pattern is
its support
https://cs.nju.edu.cn/zlj/Course/DM_15.html
 Association pattern mining (cont.)
●Given a binary n × d data matrix D,
− determine all subsets of columns such that all the
values in these columns take on the value True for at
least a fraction min_support of the rows in the matrix.
●The relative frequency of a pattern is referred to
as its support
 Association pattern mining (cont.)
●The confidence of a rule A→B is
− support(A U B) / support(A)
●Example:
− { Chips, Olives } → { Beer }
 Exercise
The confidence of a rule A→B is
support(A U B) / support(A)
Suppose
10 people buy only Chips and Beer
20 people buy only Chips and Olives
30 people buy only Olives and Beer
40 people buy all three: Chips, Olives, and Beer.
What is the confidence of the rule
{ Chips, Olives } → { Beer } ?
 Clustering
●Partition records/rows in a way that
− elements in the same partition are similar
− elements in different partitions are different
●Applications:
− Segmentation, summarization, …
− Sometimes a step in a larger DM algorithm
Image credit: http://www.sthda.com/english/articles/tag/pam-clustering/
 Clustering is not easy
●What does it mean to be similar?
●How many sets?
●Can a record/row belong to more than one set?
●Can a record/row belong to no set at all? ...

Image credit: sthda.com

 Outlier detection
●Given a database, find
records/rows that are different
from the rest of the database
●Applications:
− Intrusion detection, credit card fraud, interesting
sensor events, medical diagnosis, ...

Image credits: https://www.kdnuggets.com/2017/01/3-methods-deal-outliers.html

 Outlier detection
is not easy
●How different should they be?
●How many can be different?
●What does it mean to be different?
●What should we do with outliers?

Image credits: https://www.kdnuggets.com/2017/01/3-methods-deal-outliers.html

One of my
favorite
outliers
August Landmesser

August Landmesser in 1936

 Data classification
●Sometimes data has a feature known as a
class label
●A model can learn from previous data to
associate a record/row to a class label
●One of the most useful tools in your belt!
 Tasks with complex data types
●Frequent temporal patterns
●Time series motifs
●Graph motifs
●Trajectory clusters
●Collective classification
●...
Data types x Prototypical problems

Data Mining, The Textbook (2015) by Charu Aggarwal

Example scenarios
 Example scenario 1
●Organize products to maximize co-
purchases of items frequently bought
together
− Input data: baskets
− Output: similar pairs
− Algorithm: frequent pattern mining
 Example scenario 2
●Recommend media to users (movies,
series, music, books, podcasts, …)
− Input data: viewing history
− Output: recommendations
− Simple algorithm: k nearest neighbors
 Example scenario 3
●Help diagnose if an electrocardiogram is
associated to a health problem
− Input data: time series, possibly multi-
dimensional
− Output: binary label or risk score
− Algorithms: outlier detection or classification
 Example scenario 4
●Help a sysadmin determine if an intruder
is trying or has accessed the network
−Input data: time series of event records
−Output: binary label or risk score
−Algorithms: event detection
 Exercise
Which ones are data mining tasks?
A) Dividing the customers of a company by postal code
B) Finding credit card scammers among customers of a
company
C) Computing the total sales of a company
D) Sorting a student database by student identification
number
E) Predicting the future stock price of a company using past
records
F) Determine when a complex machine needs to be repaired
G) Extracting the frequencies of a sound wave
Major challenges
 Methodological challenges
●Mining high-dimensional data
●Handling uncertainty, noise, incompleteness, ...
●Mining data from a domain in which you do not
have expertise, or worse, in which you believe
you have expertise
− Conclusions are often worthless if you do not talk
with domain experts
 User interaction challenges
●Users should ask questions that matter to them
●Performing interactive mining
●Presenting and visualizing data mining results
 Efficiency and scalability
●Even for polynomial-running-time algorithms, a
process can become
○unreasonably slow or
○require an unreasonable amount of space
●Streaming and/or distributed mining algorithms
can help to some extent
 Diversity of database types
●Real databases are
○high dimensional and involve a
○mixture of various data types
●Sometimes you need to integrate from
○dynamic, or
○distributed data sources
 Data mining can be harmful
●Social impacts of data mining
− Who wins? And more importantly, who loses?
●Privacy-preserving data mining
− Avoid invisible, pervasive, invasive data mining
Summary
 Things to remember
●Types of data
●Types of data mining methods
●Prototypical data mining scenarios
●Typical challenges of data mining
 Exercises for this topic
●Section 1.9 of Data Mining, The Textbook
(2015) by Charu Aggarwal
● Exercises 1.7 of Introduction to Data Mining,
Second Edition (2019) by Tan et al.