Mining Massive Datasets: Introduction

• The course is an introduction to Mining of Massive Datasets

• Dr. Mehmet Aktaş is the instructor

• Jure Leskovec, Anand Rajaraman, and Jeff Ullman from Stanford University are teaching Mining of Massive Datasets

• Data contains value and knowledge

Extracting Knowledge from Data

• To extract knowledge, the data should be stored, managed, and analyzed
• Data Mining is similar to Big Data, Predictive Analytics, and Data Science

Data Storage

• When a file is stored in a data center, it is stored in multiple servers for replication
• The distributed data management system locates the various copies of that file
• Large files are divided into smaller pieces and stored on multiple servers
• Upon request, these pieces are extracted from storage, combined, and provided to the user as a single file

Data Mining Defined

• Data mining involves discovering patterns and models within large datasets
• Patterns must be valid: hold on new data with some certainty
• Patterns must be useful: actions can be taken based on them
• Patterns must be unexpected: non-obvious to the system
• Patterns must be understandable: interpretable by humans, leading to explainable AI

Data Mining Methods

• Descriptive methods find human-interpretable patterns to describe the data, for example, clustering
• Predictive methods: Use variables to predict unknown/future values of other variables, for example, recommender systems

Meaningful Analysis of Data

• A risk with data mining is that analysts can find meaningless patterns
• Statisticians refer to this risk as Bonferroni’s principle
• If you search too many places for interesting patterns without enough supporting data, you are likely to find irrelevant information

Example of Meaningless Analytic Answers

• Example: Finding unrelated people who stayed at the same hotel on the same day twice
• Assuming a scenario of 1 billion people being tracked for 1,000 days
• Each person stays in a hotel 1% of the time (1 out of 100 days), and hotels hold 100 people (100,000 hotels in total)
• If everyone behaves randomly, one can still detect suspicious activity
• The expected number of "suspicious" pairs of people is 250,000, which is too many combinations to check
• This requires more evidence to find "suspicious" pairs of people more efficiently

Challenges of working with Data

• Usage, Quality, Context, Streaming and Scalability
• Involves ontologies, structured data, networks, text, multimedia, and signals

Overlapping Disciplines

• Data mining overlaps with databases, machine learning, and CS theory
• Databases: large-scale data and simple queries
• Machine learning: small data and complex models
• CS Theory: randomized algorithms
• For a database person, data mining is an extreme form of analytic processing, which queries examine large amounts of data, and the result is the query answer
• For a machine learning person, data mining is the inference of models, and the result are parameters of the model

Course Focus

• The course will focus on both the database and machine learning aspects
• This course overlaps with machine learning, statistics, artificial intelligence, and databases
• There is more emphasis on scalability (big data), algorithms, computing architectures, and automation for handling large data

Course Objectives

• Ways to mine different types of data, including high dimensional data, graph data, never-ending data and labeled data
• Different models of computation, including MapReduce, streams and online algorithms, and single machine in-memory

Solving Real-World Problems

• Solving real-world problems with Recommender systems
• Applying Market Basket Analysis
• Spam and Duplicate document detection