A3.4.3 · Alternative databases and data warehouses
OLAP and data mining for business intelligence
A3.4.2's warehouse can answer "what were last quarter's numbers." It cannot, on its own, answer "why did they change," "what will happen next," or "which of these ten thousand customers are secretly alike." That gap is exactly what OLAP and data mining exist to close.
Explainer
Business intelligence is the umbrella term for everything an organization does to turn scattered data into decisions: the tools that pull it together, the analysis that makes sense of it, and the dashboards and reports that put it in front of the people who actually decide things. OLAP and data mining are the two technologies this statement asks about, and they do different jobs within that umbrella.
OLAP lets a person explore data that is already there, from different angles, quickly. Data mining goes further: it searches for patterns a person would never think to look for, using statistics, machine learning and AI. One is exploration, the other is discovery, and a mature business intelligence setup genuinely needs both.
Explainer
OLAP's role: fast answers from many angles
Online analytical processing (OLAP) is what lets a person turn a warehouse's data over in their hands, viewed by time, by product, by region, or any combination, without waiting for someone else to build that specific view first. Because the totals it works from are already calculated in advance, a query that would otherwise mean scanning years of raw rows comes back almost instantly, fast enough for calculations like a year-over-year comparison, a chart, or a dashboard to feel interactive rather than requested. That speed is also what makes an ad hoc query possible at all: a manager can chase a question the moment it occurs to them, rather than filing a request and waiting for a predefined report.
One OLAP slice: Fernbridge Community Bank's loan volume by branch and quarter
| Branch | Q1 2027 | Q2 2027 | Q3 2027 |
|---|---|---|---|
| Northgate | 1.8m | 2.1m | 2.4m |
| Riverside | 0.9m | 0.85m | 0.8m |
| Harborview | 3.2m | 3.1m | 3.6m |
This is one slice of a larger cube that also has a Loan Type dimension. A manager can pivot the same underlying data to view it by Branch × Quarter (above), or instead by Loan Type × Quarter, without writing a new query each time, that pivoting is what "multidimensional" means in practice.
Why this counts as explaining OLAP's role, not just describing it: the pre-aggregated view exists so that a manager gets Riverside's declining trend in seconds rather than minutes, the ad hoc query capability exists so an analyst can chase a follow-up question (why did Harborview dip in Q2?) the moment it occurs, without submitting a request to a reporting team. Every OLAP feature traces back to making exploration fast enough to keep up with a human's train of thought.
Explainer
Data mining's role: finding what nobody asked to see
Data mining takes a different approach. Rather than answering a question someone already thought to ask, it runs statistical methods, machine learning and AI over a large data set looking for patterns and relationships that were never explicitly requested, connections a person would have had no reason to go looking for. Unlike OLAP, nobody has to know what question to ask first, the technique itself surfaces the pattern. Six techniques cover most of what it does. Click through each one.
The syllabus notes this explicitly: these techniques link to A4 Machine learning. Classification and regression are supervised learning tasks (they learn from labelled examples), clustering is unsupervised learning (it finds structure with no labels at all). Data mining is not a separate field from machine learning here, it is machine learning applied specifically to extracting business insight from large data sets.
Worked example
One suspicious transaction, three techniques working together
Fernbridge's overnight batch job checks every transaction. Techniques rarely work alone, watch how three of them combine on a single case.
Transaction TXN-88231
| Customer | Usual monthly spend | This transaction | Location |
|---|---|---|---|
| Dana Okafor | 180 | 4,200 | Overseas ATM |
Anomaly detection flags it first. 4,200 is nowhere near Dana's usual pattern, and the system does not need to know anything about fraud specifically, it only needs to know this is unusual.
Classification decides what to do about it. Trained on thousands of past transactions labelled "fraud" or "genuine", it weighs the transaction size, the overseas location and the time of day, and outputs a label: Review.
Clustering explains why the classifier was right to be suspicious. Dana's normal spending pattern places her in a cluster of low-travel domestic customers. This transaction's profile looks nothing like anyone in that cluster, and much more like the cluster of customers whose cards were reported stolen last month.
Three techniques, three different questions: is this unusual (anomaly detection), what should happen next (classification), and who does this actually look like (clustering). None of the three alone tells the whole story.
Check your understanding
Fernbridge notices that customers who open a savings account also apply for a credit card within 30 days, far more often than chance would suggest. Which technique found this?
Common mistake
A3.4.3RecurringClassification and clustering get confused constantly, because both sort data into groups. The distinction candidates keep missing: classification is supervised, it needs pre-labelled examples to learn from and predicts a known label for new data. Clustering is unsupervised, it has no labels at all, it just groups similar records together and leaves a human to interpret what each group represents.
What a weak answer looks like: "Classification and clustering both sort data into categories to find patterns." True of both, and therefore useless for telling a marker which one is actually being described.
What it should say: "Classification is supervised: it is trained on data already labelled fraud or genuine, and predicts that same label for new transactions. Clustering is unsupervised: it groups customers by similarity with no labels at all, and the bank has to interpret afterwards what each resulting group represents." Supervised-with-labels versus unsupervised-with-no-labels is the one distinction that actually separates the two.
Linking question
How could machine learning be applied to databases?
This page's own six techniques are the direct answer. Classification and regression are supervised machine learning tasks, trained on a database's historical, labelled records. Clustering is unsupervised machine learning, finding structure in a database with no labels at all. Applying machine learning to a database is not a separate idea from data mining, on this statement they are the same activity, described from two different angles.
Revision questions
Past-paper style. Write an answer, then reveal the marking points and mark yourself.
[2]Fernbridge wants to forecast next quarter’s deposit volume from years of historical figures. Identify the data mining technique this needs, and describe how it works.
Marking points
评分要点
- Regression (or time-series forecasting).
- It fits a model to the historical deposit figures, learning the relationship between time and volume, then extends that relationship to predict the next quarter's value.
- 回归(或时间序列预测)。
- 它把一个模型拟合到历史存款数据,学习时间与存款量之间的关系,然后把该关系外推以预测下一季度的值。
[5]Fernbridge uses one technique to label a transaction as fraudulent or genuine, and another to find the usual order of events before an account is closed. Compare the two techniques.
Marking points
评分要点
- Labelling a transaction fraudulent or genuine is classification: it assigns each record to one of a fixed set of known categories using a model trained on labelled examples.
- Finding the usual order of events before an account is closed is sequence (or association) mining: it discovers patterns and typical orderings, with no predefined target category.
- Classification is predictive and supervised; sequence mining is descriptive and unsupervised.
- Both work over the same historical data but answer different kinds of question: which class is this, versus what tends to happen together and in what order.
- 把交易标记为欺诈或正常是分类:它用在带标签样例上训练的模型,把每条记录归入一组固定的已知类别之一。
- 找出账户关闭前事件的通常顺序是序列(或关联)挖掘:它发现模式和典型顺序,没有预先定义的目标类别。
- 分类是预测性的、有监督的;序列挖掘是描述性的、无监督的。
- 两者都作用于相同的历史数据,但回答不同类型的问题:这是哪一类,与什么倾向于一起发生、以什么顺序发生。
[6]Explain the role of OLAP and data mining for business intelligence, using an example of your own to illustrate each.
Marking points
评分要点
- OLAP lets analysts interactively slice, aggregate and drill through historical data to answer questions they already have, for example sales by region by quarter.
- Data mining automatically discovers patterns and predictive rules the analyst did not think to ask for.
- OLAP example: a manager drills from yearly to monthly sales to see which month a dip occurred.
- Data mining example: the system finds that customers who buy product A within a week of joining are far more likely to renew.
- Together they support business intelligence: OLAP for guided exploration, data mining for discovery and prediction.
- OLAP 让分析师交互式地对历史数据切片、汇总和钻取,以回答他们已有的问题,例如按地区、按季度的销售额。
- 数据挖掘自动发现分析师没想到要问的模式和预测规则。
- OLAP 例子:经理从年度销售钻取到月度销售,看下滑发生在哪个月。
- 数据挖掘例子:系统发现加入一周内购买产品 A 的顾客续订的可能性要高得多。
- 两者共同支持商业智能:OLAP 用于有引导的探索,数据挖掘用于发现和预测。
Flashcards
Active recall for this statement's key terms.
Keep this statement's practice with you
Free account: unlock every revision question and flashcard for A3.4.3, and every other IB CS 2027 statement.