KDD vs Data Mining, Normalization in Data Mining and Link Mining: Answers with Worked Examples

Four questions keep coming up for the Data Warehousing and Data Mining papers: how KDD differs from data mining, what normalization in data mining means (with the three formulas), what link mining is, and whether NoSQL is unstructured data. They sit in TU BSc CSIT Data Warehousing and Data Mining (CSC420) unit 2 "Introduction to Data Mining", unit 3 "Data Preprocessing" and unit 8 "Graph Mining and Social Network Analysis", and in BIT Data Warehousing and Data Mining (BIT454) units 2, 3 and 8. The NoSQL part comes from Database Management (CSC265) and BIT Database Management (BIT202) unit 10 "NoSQL and Unstructured Data".

This guide answers all of them in exam form: KDD vs data mining with the process steps, min-max, z-score and decimal scaling normalization with one worked example each, the tasks of link mining, and a short NoSQL answer. Notes are on the CSC420 Introduction to Data Mining page and the CSC420 Data Preprocessing page.

KDD vs data mining

Knowledge discovery in databases (KDD) is the whole process of turning raw data into useful, understandable knowledge. Data mining is one step inside it: the step where algorithms are applied to extract patterns. People often use the two names interchangeably, but in the exam the distinction is expected.

The KDD process has these stages, in order: data cleaning (remove noise and inconsistencies), data integration (combine multiple sources), data selection (retrieve the data relevant to the task), data transformation (consolidate into forms suitable for mining, for example by normalization or aggregation), data mining (apply methods to find patterns), pattern evaluation (identify the truly interesting patterns using interestingness measures), and knowledge presentation (show the result with visualization and reports).

Normalization in data mining

Normalization is a data transformation step that scales attribute values into a small common range so that an attribute with large values (such as income) does not dominate one with small values (such as age) in distance-based methods like k-means and nearest neighbour. Take the data 200, 300, 400, 600, 1000 for all three methods.

Min-max normalization

Min-max normalization maps a value $v$ of attribute A linearly into a new range $[new_min, new_max]$:

$$v' = \frac{v - min_A}{max_A - min_A},(new_max - new_min) + new_min$$

With $min_A = 200$, $max_A = 1000$ and new range $[0, 1]$, the value $v = 400$ gives

$$v' = \frac{400 - 200}{1000 - 200}(1 - 0) + 0 = 0.25$$

and the whole set becomes 0, 0.125, 0.25, 0.5, 1. Its drawback is that a new value outside the original min and max causes an out-of-bounds error, and one outlier compresses the rest.

Z-score normalization

Z-score (zero-mean) normalization uses the mean $\bar{A}$ and standard deviation $\sigma_A$:

$$v' = \frac{v - \bar{A}}{\sigma_A}$$

The mean of the data is $(200 + 300 + 400 + 600 + 1000)/5 = 500$. The squared deviations are 90000, 40000, 10000, 10000 and 250000, which sum to 400000, so the population variance is 80000 and

$$\sigma_A = \sqrt{80000} \approx 282.84$$

For $v = 400$:

$$v' = \frac{400 - 500}{282.84} \approx -0.354$$

and the whole set becomes about -1.061, -0.707, -0.354, 0.354, 1.768. Z-score is useful when the actual minimum and maximum are unknown or when outliers dominate the min-max range.

Decimal scaling normalization

Decimal scaling moves the decimal point of the values by dividing by a power of ten:

$$v' = \frac{v}{10^{j}}$$

where $j$ is the smallest integer such that $\max(|v'|) < 1$. The largest absolute value is 1000, and $1000 / 10^3 = 1$ is not below 1, so $j = 4$. For $v = 400$:

$$v' = \frac{400}{10^4} = 0.04$$

and the whole set becomes 0.02, 0.03, 0.04, 0.06, 0.1.

Link mining

Link mining applies data mining to linked data, such as web pages joined by hyperlinks, people joined in a social network or papers joined by citations. Unlike ordinary mining, where records are independent, here the links between objects carry information. It belongs to the Graph Mining and Social Network Analysis unit, and the same ideas are in BIT454 Social Network Analysis. The main tasks are:

• Link prediction: predicting whether a link will exist between two objects, for example suggesting new friends or likely future collaborations.
• Link-based classification: predicting the category of an object using its own attributes plus the categories and attributes of the objects it links to, for example classifying a web page from the pages that link to it.
• Link-based ranking: ordering objects by importance using the link structure, as in PageRank and HITS, where a page linked from many important pages ranks higher.

Other tasks listed in the syllabus notes include object clustering based on links and group detection (community detection), which finds clusters of densely connected objects.

Is NoSQL unstructured data?

NoSQL is not itself data; it is a family of database systems. A NoSQL database stores semi-structured and unstructured data (JSON documents, key-value pairs, graphs, text) without requiring a fixed table schema, which is why the two ideas are linked. But NoSQL stores can also hold structured data, so the answer is: NoSQL databases are designed to handle unstructured and semi-structured data, but NoSQL and unstructured data are not the same thing.

The characteristics of NoSQL databases are:

• Schema-less (flexible schema), so fields can differ between records.
• Horizontal scaling, adding more servers instead of a bigger server.
• BASE properties (Basically Available, Soft state, Eventually consistent) in place of strict ACID, trading consistency for availability.
• High performance on large volumes of data and distributed storage with replication.

NoSQL types are key-value stores (Redis), document stores (MongoDB), column-family stores (Cassandra) and graph databases (Neo4j). The unit notes are on the BIT202 NoSQL and Unstructured Data page.

How to write it in the exam

For "Differentiate between KDD and data mining", draw the pipeline, mark data mining as one box in it, and give the table. For a normalization numerical, write the formula, list the values you need (min and max, or mean and standard deviation, or $j$), substitute for each value, and state the final normalized list. For link mining, name the three tasks with one example each. For the NoSQL question, give the definition, the characteristics and the four types.

The full subject pages are CSC420 Data Warehousing and Data Mining and BIT454 Data Warehousing and Data Mining.