Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
Mark Hall
Pentaho Corporation Suite 340, 5950 Hazeltine National Dr. Orlando, FL 32822, USA
Data Mining WEKA - what is it? WEKA UIs Integration with Pentaho Projects based on WEKA
Data Mining
A definition: Extraction of implicit, previously unknown, and potentially useful information from data Goal (business oriented): improve marketing, sales, and customer support operations
Who is likely to remain a loyal customer/jump ship? What products should be marketed to which prospects? What determines whether a person will respond to a certain offer? How can I detect potential fraud?
Central idea: historical data contains information that will be useful in the future
Historical patterns provide useful insight and generalize to future situations
Data Mining: algorithms that automatically detect patterns and regularities in data
Data Mining
Strong patterns can be used to make predictions
Problem 1: a lot of patterns are not interesting Problem 2: patterns may be inexact (or even completely spurious) if data is garbled or missing
Techniques borrowed from statistics, computer science, machine learning research Compared to traditional statistics
Statistics is manual, user driven, top-down - formulate a hypothesis, convert hypothesis into database query, test significance of results Data mining automates the data interrogation Data-driven, self-organizing, bottom-up Automatic examination of a large number of hypothesis
Compared to OLAP
OLAP: data summarization - aggregation via addition # widgets sold in all ZIP codes in the country Data Mining: ratios, patterns and influences Factors influencing the sales of the widgets in those ZIP codes DM can enhance OLAP - suggest dimensions for cube, discretization etc.
Selected data
Preprocesse d data
Transforme d data
Extracted informatio n
Assimilate d knowledge
Transform
Mine
Analyze
What is WEKA?
Hamilton
Develop an open-source state-of-the art workbench of data mining tools Explore fielded applications Develop new fundamental methods
Core Functionality
Support for the whole process of experimental data mining
Preparation of input data Statistical evaluation of learning schemes Visualization of input data and the result of learning
User Interfaces
Explorer - data exploration/visualization, model construction and export, preliminary evaluation Experimenter - large-scale algorithm comparison with statistical tests for significant differences in performance KnowledgeFlow - process model view of data mining, export of DM process
Architecture
Modular, object-oriented architecture
Packages for different types of algorithms: filters, classifiers, clusterers, associations, attribute selection etc. Sub-packages group components by functionality or purpose E.g. classifiers.bayes, filters.unsupervised.attribute
Public interface prescribed by abstract base class or interface for all types of algorithms
Algorithms are Java Beans GUIs use introspection/reflection to dynamically generate editor dialogs at runtime
Explorer
Explorer
Preprocess panel
Load data from various sources (file, SQL database, URL etc.) Apply pre-processing filters to the data Summary statistics & histograms
Classify panel
Apply classification and regression algorithms Evaluate resulting models Numerically via statistical estimation Graphically through visualization (data and model)
Cluster panel
Apply clustering algorithms to the data Visualize the outcome Clusters that represent density estimates can be evaluated based on the statistical likelihood of the data
Associate panel
Learn association rules for market-basket type analysis
Explorer
Select attributes panel
Mix and match algorithms for evaluating the utility of attributes and sets of attributes with different search methods
Visualize panel
Color-coded scatter plot matrix of the data Select, enlarge, zoom in etc.
Knowledge Flow
Define a data mining process
Like the Explorer, all of WEKAs algorithms are available
Data flows through the process from node to node Accommodates both batch-based processing and data streams
Command line interface to WEKA can also train incremental classifiers on data streams
Fully multi-threaded
Accommodates multiple independent flows on the same layout
Knowledge Flow
Experimenter
Automate the process of determining the best method to use
Is an interactive process in the Explorer or Knowledge Flow
Raw output saved to files or databases Analysis results can be export as text, CSV, Gnuplot, LaTeX or HTML Advanced users can distribute the processing load across multiple machines
Experimenter
Extensibility
Plugin mechanisms allow WEKA to be extended without modifying the classes in the WEKA distribution New tabs can be added to the Explorer New visualizations can be added to the pop-up menu in the Explorers Classify panel
Classifier errors, predictions, trees and graphs
Text mining
Judge & IR Utilities - two systems that perform document categorization and clustering
Impact
Has been downloaded more than 1.5 million times since placed on SourceForge in April 2000 Current download rate of more than 20,000 per month Large user-base and active community
2750 people subscribed to the mailing list