The Knowledge Discovery (KDD) Process

1Data mining = "knowledge mining from data"

In simple terms, data mining is extracting ("mining") knowledge from large amounts of data.

There are two views of how data mining relates to KDD (Knowledge Discovery from Data):

2The 7 steps of the KDD process

Knowledge discovery is an iterative sequence of these steps (Figure 2-1):

3Architecture of a typical data mining system

Figure 2-2 shows the components as layers. From the bottom up:

                ┌──────────────────────────┐
                │      User interface      │
                └────────────┬─────────────┘
                ┌────────────┴─────────────┐   ┌──────────────┐
                │ Pattern evaluation module│◄─►│              │
                └────────────┬─────────────┘   │  Knowledge   │
                ┌────────────┴─────────────┐   │    base      │
                │   Data mining engine     │◄─►│              │
                └────────────┬─────────────┘   └──────────────┘
                ┌────────────┴─────────────┐
                │ Database / data warehouse│
                │          server          │
                └────────────┬─────────────┘
   data cleaning, integration & selection
     Database   Data warehouse   WWW   Other repositories

4What kinds of data can be mined?

In principle data mining applies to any data repository and also to transient data such as data streams.

Relational databases

A DBMS = a collection of interrelated data (the database) plus software to manage and access it. The software handles defining structures, storage, concurrent/shared/distributed access, and consistency and security, even through system crashes or attempts at unauthorised access.

Data warehouses

A data warehouse is a repository of information collected from multiple sources, stored under a unified schema, usually at a single site. It is built through:

Transactional databases

A file where each record is a transaction: a unique trans ID plus the list of items in the transaction (e.g. items bought). Extra tables can hold the date, customer ID, salesperson ID, branch and so on.

This is the data used for market basket analysis (Apriori, Week 9).

Advanced database systems

Details worth remembering

• Spatial data formats
• Raster: n-dimensional bit/pixel maps (e.g. each pixel of a satellite image = rainfall in that area)
• Vector: roads, bridges, lakes as unions of points, lines and polygons
• Text data structure levels
• Unstructured: some web pages
• Semi-structured: e-mail, HTML/XML pages
• Well structured: library catalogues, which can be stored in relational DBs
• Multimedia needs support for large objects (gigabytes) and special storage and search. Audio and video need real-time retrieval at a steady rate (to avoid gaps and buffer overflows), so they are called continuous-media data. Promising approaches are multimedia data cubes, feature extraction and similarity-based pattern matching.
• Web usage mining (Weblog mining) means capturing user access patterns on the web. It helps system design (efficient access between related objects) and marketing (place ads on frequently visited pages, better customer classification).

5Data mining task primitives

A user specifies a data mining task as a data mining query. The query is defined in terms of task primitives, which let the user interactively communicate with the system to direct the mining and look at results from different angles.

There are five primitives (Figure 2-4):

6Quick revision sheet