Classification Expectation Maximization
Algorithms AIMech:ClassificationExpectationMaximization DRAFT
Aliases: Classification EM, CEM (classification expectation maximization)
A model-based hard-clustering algorithm that computes mixture responsibilities, assigns every observation to a highest-responsibility component, and maximizes the completed-data classification likelihood for that assignment.
Identity Notes & Concept Scope
CEM optimizes a joint parameter-and-partition classification criterion rather than ordinary marginal-mixture likelihood. The historical name refers to clustering unlabeled observations, not necessarily supervised classification. It differs from posterior-label sampling in SEM, cross-entropy optimization, and soft-responsibility EM. k-means is a restricted equal-weight spherical-Gaussian specialization, not an alias for every CEM instance. Bare CEM is not asserted as a unique alias. No normalized identity matches among 442 pinned and 473 maintained records, including the bare acronym probes. Existing supplemented EM and cross-entropy optimization are distinct methods; qualified aliases avoid claiming SEM or CEM as unique names. ICE, MCEM, SAEM, deterministic tempered EM, simulated annealing and k-means provide neighboring but different scopes. Existing stochastic-EM frontier and incidental references are not records of the new identities. The absence assessment is bounded to this snapshot and corpus. Identity-search counts: 227 matched lines in the worktree and 245 in main; all 137 distinct matching-text windows were read, with ignored and hidden files included.
Claims & Grounded Evidence
4 Claims
Mechanism
#e-c-m
Compute t_ik proportional to p_k f_k(x_i) at the current fit. Assign observation i to a maximizing component; the inspected report chooses the smallest component index on ties. Fit parameters by maximizing the hard-assignment completed log likelihood. When mixing weights are estimated, p_k becomes n_k/N and each component fit uses its assigned data. Exact admissible C and M steps make that classification criterion nondecreasing; undefined or singular maximizers need an explicit policy.
Scope
#criterion-and-kmeans
The mixture classification criterion is sum_i log[p_(z_i) f_(z_i)(x_i)], which includes sum_k n_k log p_k. The separate-sampling criterion omits that weight term. With fixed equal weights and a common spherical Gaussian covariance, assignment to the nearest mean and centroid refitting recover the k-means variance objective. Estimating arbitrary mixing weights or using unequal covariances changes the assignment geometry and objective.
Limitation
#marginal-and-geometry
Classification-likelihood ascent is not marginal-mixture likelihood ascent: the saved two-component Bernoulli example increases the former and decreases the latter in one nonempty CEM step. The report’s general claim that CEM classes have disjoint convex hulls also fails for unequal-variance Gaussian components; the saved one-dimensional example assigns outer points to a broad component and inner points to a narrow one. Those limitations do not negate the restricted equal-covariance k-means case.
Scope
#convergence-scope
The finite-partition argument for stabilization needs admissible finite profiled objectives and well-defined maximizer selection; it is not a global-optimum guarantee. Initial partitions and ties can affect the solution. The local-rate proposition invokes stronger curvature and coordinate-optimization assumptions. Because the relaxed objective is affine in the assignment matrix, no generic negative-definite assignment Hessian or universal linear-rate claim is certified here.
⚠️ Curated Limitations & Negative Results
- The inspected primary version is the January 1991 author report, not the inaccessible 1992 journal full text. Eq2.2 prints a malformed marginal likelihood without the component sum; the mixture law in Eq2.1 determines the correct normalization. Eq2.7 doubles the variable-dependent Gaussian log-likelihood terms, leaving the displayed optimizer but not literal likelihood values unchanged.
- Hard reassignment can empty a component; unrestricted component likelihoods can become singular or unbounded. Fixed component count, nonempty partitions and finite M-step solutions are not guaranteed by an argmax assignment alone.
- The source’s local-rate proof and its external grouped-coordinate theorem were not independently repaired. Affine assignment dependence and boundary solutions require care before importing smooth interior curvature claims.
- The numerical experiments evaluate a variance criterion under selected mixture and serum-data settings. A practical SEM-then-CEM hybrid is distinct from pure CEM; no universal superiority, generalization or semantic-cluster claim is curated.
- No author implementation or historical numerical experiment was reproduced. Finite curator calculations check formula boundaries, not empirical performance or a repaired convergence proof.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:UnsupervisedClustering
Unsupervised Clustering
|
related_match | The defining procedure estimates a hard grouping of unlabeled observations alongside fitted component parameters. The pinned target is a clustering task, so the algorithm is related without an is-a assertion. The word classification in the historical name does not identify the explicitly supervised AIO:Classification class. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
cem-source
|
Gilles Celeux, Gérard Govaert
|
author technical report
full_text
|
ID: inria-00075196v1
sha256:30bd7f969d0f…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-12T04:22:35Z
Updated: 2026-10-12T04:22:35Z