Generalized component analysis
Model Families AIMech:GeneralizedComponentAnalysis DRAFT
Aliases: GCA
An undirected latent-component model family coupling categorical text and heterogeneous document attributes to positive lognormal hidden variables, with variables conditionally independent within either layer given the other and shared attribute-specific interaction weights.
Identity Notes & Concept Scope
Scoped to Wang, Pal and McCallum's KDD2007 GCA harmonium for words, author positions and discrete time bins. Its positive hidden coordinates couple through their logarithms, giving an equivalent Gaussian-hidden representation under a change of variables. This differs from the native Replicated Softmax model with binary hidden variables and its length-scaled bias. It is not ordinary PCA, a simplex topic mixture, or every method subsequently named generalized component analysis. No normalized candidate identity matched all 442 pinned AIO IDs, labels and scoped synonyms or 438 preceding native IDs, names and aliases. Full-checkout searches included ignored and hidden files in both checkouts; eleven matching lines per checkout came from four unique prior discovery-ledger contents with pending leads and bibliography context. All matching excerpts were read. All 205 semantic label/synonym hits and 21 annotated AIO nodes with outgoing edges were byte-verified against prior readings. Twenty-nine contextual YAMLs comprise 28 unchanged prior readings and one fully read final round103 native target verified against its saved review inputs. No candidate record is present within this bounded pinned-snapshot and maintained-corpus search.
Claims & Grounded Evidence
3 Claims
Mechanism
#log-hidden-harmonium
For fixed document length and author-position count, word tokens, author tokens and a discrete timestamp contribute shared matrix columns to a vector S. The positive-hidden joint contains exp(sum_i[-log(t_i)-log(t_i)^2/2+S_i*log(t_i)]). Changing variables to y=log(t), including the Jacobian, gives y conditional on all visible observations distributed as N(S,I), and visible mass proportional to exp(||S||^2/2). The positive-hidden mean is exp(S_i+1/2), not S_i; positive coordinates do not by themselves create a normalized simplex of topics. Missing visible modalities generally produce mixtures after marginalization rather than the same single Gaussian with their columns set to zero.
Mechanism
#contrastive-training-statistics
The source fits approximate joint or conditional likelihood using contrastive divergence and Gaussian weight shrinkage. Conditional training clamps words and time while reconstructing authors. Replicated categorical author positions use occurrence counts as sufficient statistics, including when a reconstruction repeats an author. Equation 8 instead displays a presence indicator and prints A-1 as a word-sum limit; the general statistic requires author counts and the vocabulary dimension. The exact fixed-length likelihood gradient contrasts observed count-times-S statistics with their model expectation, with an additional prior gradient if regularized. Short-chain CD approximates that expectation and is not exact maximum likelihood.
Limitation
#latent-rotation-boundary
A common orthogonal transformation of the hidden rows of all interaction matrices preserves ||S||^2 and therefore every fixed-size visible probability. It also preserves an isotropic Gaussian weight penalty. Such shrinkage consequently does not remove all latent rotation ambiguity or guarantee a uniquely interpretable set or number of components. The source's suggestion that the prior addresses nonidentifiability or enables automatic component selection has this mathematical boundary.
⚠️ Curated Limitations & Negative Results
- Finite categorical sample spaces make the partition function finite for each fixed number of word and author positions and finite weights. The source does not specify a normalized distribution over unbounded document lengths. Independent author positions can repeat authors; this is not automatically a distribution over unique author sets.
- The evaluated timestamps are discrete years or months. Proposed Gaussian or Beta visible attributes require their own compatible joint and normalization analysis. Broad PCA comparisons do not establish equivalence to probabilistic PCA with Gaussian observations.
- The reported author/recipient evaluation retrieves training documents by cosine similarity in latent space and scores the first result sharing an author or recipient. This is not exact author-set conditional-probability ranking. Treatment of missing target-modality features in embeddings is underspecified; no data-leakage allegation or unprovided implementation is inferred.
- Selected NIPS and single-user email data, random splits, twenty components and best reported iteration scores provide bounded empirical evidence. Finite checks verify fixed-size probabilities, a gradient, rotation invariance and missing-modality mixtures, not historical software, convergence or the reported benchmark.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:Model
Model
|
broad_match | This source-defined probabilistic model family is narrower than the pinned Model class. |
AIO:ProbabilisticTopicModel
Probabilistic Topic Model
|
related_match | The family supports probabilistic topic modeling; the pinned target is task-scoped, so related match preserves the model/task distinction. |
AIO:RestrictedBoltzmannMachine
Restricted Boltzmann Machine
|
related_match | GCA has the bipartite conditional-independence structure of an exponential-family harmonium, but its lognormal/Gaussian hidden law and heterogeneous categorical positions are not asserted identical to the underspecified pinned RBM class. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
method-source
|
Xuerui Wang, Chris Pal, Andrew McCallum
|
primary paper
full_text
|
sha256:e0c6abaef231…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-12T00:49:25Z
Updated: 2026-10-12T00:49:25Z
Source File:
data/curated/GeneralizedComponentAnalysis.yaml