Classification of Data Mining Systems

Types of Data Mining Systems

Data mining extracts useful and meaningful information from large amounts of raw data, identifying patterns, trends, and relationships using specialized software tools.

As the field has grown, so has the variety of data mining systems available. They differ in the data they work with, the knowledge they produce, the methods they use, and the problems they are built for. Classifying them along these lines is what makes it possible to compare systems and choose an appropriate one.

A note on the word "classification". This lesson is about categorizing data mining systems. That is a different thing from classification the technique — the supervised learning method that assigns records to predefined classes, covered in this series' lesson on data mining techniques. The two share a word and nothing else. Keeping them apart avoids a great deal of confusion.

Four Ways to Classify Data Mining Systems

Data mining systems are conventionally classified by:

• The type of database mined
• The type of knowledge mined
• The techniques used
• The application area served

1. Classification Based on the Databases Mined

Systems can be classified by the type of database they work with. Different databases store data in different formats, so the mining approach varies accordingly.

• Relational databases — data stored in tables of rows and columns
• Transactional databases — records of transactions such as sales or purchases
• Object-relational databases — combining relational structure with object-oriented features
• Data warehouse systems — large stores of historical data organized for analysis

This can be extended by the kind of data rather than the storage system: systems specialized for text, multimedia, spatial, time-series, web, or streaming data each require different handling. A system built for relational tables will not cope with a stream of sensor readings arriving continuously, and vice versa.

2. Classification Based on the Type of Knowledge Mined

Systems can also be classified by what kind of knowledge they are designed to discover:

• Characterization — summarizing the general characteristics of a group of data
• Discrimination — comparing the features of one data class against another
• Association and correlation analysis — finding relationships between variables
• Classification — assigning data to predefined categories
• Prediction — forecasting future outcomes from past data
• Outlier analysis — detecting unusual or abnormal records
• Evolution analysis — studying how data changes over time

Most practical systems provide several of these rather than one. A system offering only a single function is usually a specialized tool rather than a general platform.

3. Classification Based on the Techniques Used

Systems can be categorized by the methods they apply:

• Statistical methods — regression, hypothesis testing, probability models
• Machine learning algorithms — decision trees, neural networks, clustering methods
• Pattern recognition — identifying regularities in data, including image and signal data
• Database-oriented techniques — methods that exploit database structure and indexing for efficiency
• Visualization methods — presenting data graphically so patterns become visible to a human analyst
• Information science approaches — retrieval and indexing methods

Two further distinctions are often drawn within this category. Systems differ in their degree of user interaction — some are query-driven, requiring the analyst to direct each step; some are autonomous, running without supervision; some are interactive, allowing the analyst to guide the process as it unfolds.

The choice of technique depends on the type of data and the goal of the analysis, and most capable systems combine several rather than committing to one.

4. Classification Based on Applications

Data mining systems are also classified by the domain they serve, since each field brings its own data types, regulations, and evaluation standards:

• Finance — risk analysis and fraud detection
• Telecommunications — customer usage pattern analysis
• DNA and bioinformatics — genetic sequence and expression analysis
• Stock market — market trend prediction
• Email systems — spam detection and filtering
• Healthcare — diagnosis support and treatment analysis
• Retail — market basket analysis and customer segmentation

A domain-specific system embeds knowledge a general-purpose tool lacks. A bioinformatics system understands sequence data structurally; a general clustering tool sees only numbers. This is why specialized systems often outperform general ones within their domain while being useless outside it.

Integration with Database and Data Warehouse Systems

A separate and practically important question is how tightly the mining system is coupled to the database or warehouse holding the data. Four levels are conventionally distinguished.

1. No coupling

The mining system works entirely independently, using no database or warehouse functionality. Data must be fetched or loaded from files by other means. This is the simplest to build and the least efficient, since it ignores everything the database could do — indexing, query optimization, and the ability to process data where it already sits.

2. Loose coupling

The mining system uses some database facilities: it retrieves data through the database, performs mining separately, and stores results independently. Better than no coupling, but data still moves out of the database to be processed, which limits scalability on large datasets.

3. Semi-tight coupling

Some mining operations are performed directly inside the database system — commonly frequently-needed primitives such as sorting, aggregation, histograms, and precomputed statistics. Pushing these into the database avoids transferring large volumes of data and takes advantage of database optimizations.

4. Tight coupling

The mining system is fully integrated with the database or warehouse, treated as one of its functions. Queries and mining operations are optimized together, data never leaves the system, and performance is best. It is also the hardest to build, which is why it is usually found in commercial database platforms with built-in analytics rather than in standalone tools.

These terms overlap confusingly with the data integration approaches covered in this series' lesson on data integration, where "tight coupling" means physically consolidating data into a warehouse and "loose coupling" means leaving it in place. Same words, different subject: here they describe how a mining system connects to its data store; there they describe how sources are combined.

Why This Classification Matters

Working through these categories is a practical way to specify what a project actually needs before evaluating tools. The questions fall out directly:

• What kind of data is it, and where does it live?
• What kind of knowledge is wanted — categories, predictions, associations, anomalies?
• Which techniques suit that goal and that data?
• Is there a domain-specific system that already understands this field?
• How large is the data, and does that force mining closer to where it is stored?

A system that scores well on every dimension except the one that matters for a given project is the wrong choice, and this framework is what makes that visible before the work starts rather than after.

Related Concepts

The knowledge types listed in section 2 — characterization, association, classification, prediction, outlier analysis — are covered individually in this series' lesson on data mining techniques. Specific systems classified this way, including Orange, SAS, RapidMiner, Rattle, and DataMelt, appear in the data mining tools lesson. The databases described in section 1 are covered in the introduction lesson, and the architecture that connects these components appears in the data mining architecture lesson.