Apriori Frequent Itemsets

This unlocks the ability to find hidden patterns in transaction-like data — which products are often bought together, which symptoms co-occur, which settings cluster. From this you go on to association rule mining, market basket analysis, and the broader family of unsupervised learning techniques like clustering (k-means, DBSCAN) and dimensionality reduction (PCA). In modern AI, the same co-occurrence logic underlies how recommendation systems suggest what to buy next, and how embeddings map similar things close together.

Imagine you run a small grocery store. You have a list of every receipt — what each customer bought. You want to know which items tend to appear together. For example, do people who buy bread also buy butter? Do they also buy jam? You could go through all receipts by hand, but with thousands of receipts that's impossible. Apriori automates this: it scans all transactions, counts how often each item appears, then looks for pairs, triples, and larger groups that appear together more often than a threshold you set. The key trick is that if a single item is rare, any group containing it is also rare — so you can skip whole branches of possibilities. This prunes the search dramatically, making the problem feasible even with millions of transactions.

Think of a party where you want to know which groups of friends tend to arrive together. You have a list of everyone who attended on each night. You set a rule: a group is 'frequent' if they show up together on at least 30% of nights. First, you count how often each person shows up — if someone only appears on 10% of nights, they can't be part of any frequent group. So you ignore them. Then, with the people who pass the 30% threshold, you form all possible pairs. If a pair (say Alice and Bob) shows up together on 35% of nights, they are frequent. Now consider a triple: Alice, Bob, and Charlie. If Alice and Bob are not a frequent pair, then any triple containing both is also not frequent — because if the pair doesn't show up enough, the triple can't show up more often. So you can skip checking that triple entirely. This is the apriori property: if a set is infrequent, all its supersets are infrequent. The party analogy works well for the pruning step, but it breaks down when we talk about 'support' — in a real transaction database, each row represents one basket, and items appear only once per basket. Also, the analogy doesn't capture that we might want rules like 'if Alice comes, Bob usually comes too' — that requires confidence and lift, which the analogy doesn't show.

Apriori is an algorithm that finds all item combinations that appear together in a dataset more often than a user-set support threshold, and then derives association rules from those frequent itemsets.

This is your first taste of unsupervised learning — finding structure without a target variable. It builds on simple counting and arithmetic. The ideas of 'support', 'confidence', and 'lift' will reappear in later topics like recommendation systems and association rule mining. It also connects to clustering (k-means, DBSCAN) in that both find patterns in data, but here the pattern is explicit co-occurrence, not group membership.