Is choosing a team always a combination?

Only if the selected members have no distinct assigned roles in the counting question.

Why divide by k factorial?

The ordered count includes every selected set once for each of its k-factorial orderings.

What if repeats are allowed?

That is a different problem. For an unordered selection of k items from n types with unlimited repetition, the usual count is n plus k minus one choose k, under those specific assumptions.

AI Glossary / Statistics and probability / Beginner

Combinations

Definition

Combinations are selections in which order does not matter. In the usual elementary setting, a combination chooses k distinct objects from n available objects without replacement. The number of possible selections is written n choose k and equals n!/[k!(n - k)!]. Combinations help count possible samples, teams, feature subsets and other unordered selections relevant to statistics and AI.

What it means

Selecting A, B and C gives the same combination as selecting C, A and B. If positions or roles matter, those arrangements are different permutations instead. The distinction depends on the problem, not on the names of the objects. Choosing three reviewers for an equal-role panel is a combination; appointing a chair, deputy and secretary from the same people is an ordered assignment. Repeated selections require a different counting rule from the ordinary distinct-object case.

How it works

First identify the available objects, the required selection size, whether order matters and whether repeats are allowed. For k distinct objects chosen from n distinct objects, there are n(n - 1)...(n - k + 1) ordered selections. Each unordered selection appears k! times in that ordered count, so divide by k! to obtain the binomial coefficient. The boundary cases n choose zero and n choose n both equal one. Constraints such as mandatory inclusions or prohibited pairings require additional counting logic; applying the unconstrained formula alone would count selections that are not allowed.

Three practical examples

Evaluation sampling: count possible sets of benchmark items selected without replacement when their presentation order is irrelevant to the selection.
Feature exploration: count subsets containing exactly a specified number of candidate model features.
Review planning: count possible equal-role panels drawn from a pool of qualified reviewers.

Think of it this way

Think of choosing toppings for a pizza. A set containing mushrooms, olives and peppers is unchanged if you name the toppings in a different order. If you instead assign those ingredients to three labeled compartments, the positions matter and the counting problem changes.

Worked example

Suppose a fictional evaluation team selects three distinct tasks from eight, with no order attached to the selection. There are 8 choose 3 = 8 x 7 x 6/(3 x 2 x 1) = 56 possible sets. Counting ordered selections would give 336, six times as many, because every three-task set has six orderings. Now suppose three of the eight tasks are safety-related and the selection must contain at least one of those three. Count all 56 sets and subtract sets drawn entirely from the other five tasks: 5 choose 3 = 10. The allowed count is therefore 46. If all 56 original sets were equally likely under simple random sampling, the probability of including at least one safety task would be 46/56 = 23/28, approximately 82.14%.

Illustrative case study

Fictional case: Linden Evaluation prepares small review batches from a task bank. An analyst notices that a planning spreadsheet counts the same task set repeatedly whenever its order changes. The team separates unordered batch selection from the later presentation order, which may itself need randomization to control order effects. It also checks constraints ensuring coverage of important task categories. The corrected count helps describe the design space without claiming that all batches provide equal information. This invented case illustrates counting and design, not a measured benchmark result.

Common misconceptions and mistakes

A combination count does not automatically define a probability distribution. The ratio of favorable to total combinations is valid only when those combinations are equally likely under the specified sampling procedure. Sampling with replacement changes the possibilities, and duplicate labels can create confusion about whether objects are actually distinct. An unordered feature subset may still lead to many models through different preprocessing choices and hyperparameters. Conversely, enumerating all subsets can be computationally impractical even when the binomial coefficient is easy to calculate.

Business use

Businesses can use combinations to estimate the scale of sampling, review and model-selection choices. The calculation can reveal when exhaustive evaluation is unrealistic and when constrained designs or sampling are needed. For AI projects, distinguish the number of possible feature sets from the cost of training and validating each one. If many combinations are searched and only the best result is reported, selection can inflate apparent performance. Counting the search space is a useful planning step, but credible evaluation still requires a design that addresses overfitting and multiple comparisons.

Check your understanding

Why does the worked example subtract 5 choose 3 rather than 3 choose 1?

Answer: It counts the complement: all selections with no safety-related task. Those sets must come entirely from the five non-safety tasks. Choosing one safety task alone would not complete the three-task selection and could double-count sets with several safety tasks.

Related terms

Permutations
Combinatorics
Hypergeometric distribution
Permutation
binomial coefficient
factorial
sampling without replacement
simple random sample
complement rule
subset selection
search space

Learn next

compare combinations and permutations using the same objects, then practice counting complements and selections with restrictions. Derive the hypergeometric probability model from equally likely unordered samples. For AI evaluation, separate selection, presentation order and model-search choices so each source of variation is represented accurately.

Probability model

Sources

University of Vermont, Glossary, combinations and combinatorics: https://www.uvm.edu/~statdhtx/fundamentals9/Glossary/Glossary.html Python documentation, math.comb: https://docs.python.org/3/library/math.html#math.comb Python documentation, itertools.combinations: https://docs.python.org/3/library/itertools.html#itertools.combinations

Previous: Column space Next: Combinatorial data analysis