Hierarchical pachinko allocation model
Model Families AIMech:HierarchicalPachinkoAllocationModel DRAFT
Aliases: hPAM
A pachinko topic-model family that associates word-emission distributions with internal topic nodes as well as terminal topics, allowing document tokens to use multiple paths and emit from different levels of a topic DAG.
Identity Notes & Concept Scope
Scoped to Mimno, Li and McCallum's ICML2007 hierarchical PAM family and its two explicit variants. Internal vocabulary emissions distinguish it from the demonstrated four-level PAM with word emissions only at subtopics. Token-specific DAG paths and shared subtopics distinguish it from hierarchical LDA's single document path in a tree. The path-level and local-exit variants remain one source-defined family with explicit differences, not asserted interchangeable parameterizations. This paper uses fixed topic counts; nonparametric PAM is a separate construction. No normalized candidate identity matched all 442 pinned AIO IDs, labels and scoped synonyms or 435 preceding native IDs, names and aliases. Full-checkout searches included ignored and hidden files in both checkouts; two matching lines per checkout came from one unique prior discovery-ledger content with a source URL and pending extension cohort. All matching excerpts were read. All 205 semantic label/synonym hits and 18 annotated AIO nodes with outgoing edges were byte-verified unchanged from prior readings. Twenty-six contextual YAMLs comprise 24 unchanged prior readings and two fully read final round104 native targets verified against saved review inputs. No candidate record is present within this bounded pinned-snapshot and maintained-corpus search.
Claims & Grounded Evidence
3 Claims
Mechanism
#path-level-variant
In model 1 each document draws a root-to-supertopic vector and a subtopic vector for each supertopic. Each token chooses a supertopic and subtopic, then a level from a corpus-shared distribution specific to that path. It emits from the shared root, supertopic or subtopic word vector according to the selected level. Even root-emitted tokens have a latent complete path and contribute routing counts. With fixed priors, collapsed assignment weights combine document routing, corpus path-level and actual emitting-node word counts. Word counts must pool all paths that use the same emission vector.
Mechanism
#local-exit-variant
Model 2 adds an exit coordinate to each document-local routing vector. A root exit emits from the root; otherwise a supertopic is selected and either exits to its word distribution or routes to a subtopic. The path stops at the emitting node, so unused lower routing choices are not sampled. There is no model 1 path-specific level vector. The variants have different cross-document sharing: in a one-path example a shared Beta(1,2) root-level probability gives joint root emission probability 1/6 across two documents, whereas independent document-local Beta(1,2) exits give 1/9 despite the same marginal probability.
Mechanism
#collapsed-fitting
The paper describes Gibbs sampling with a factorization that first marginalizes paths to sample an emitting node, then samples a compatible path conditional on that node. This preserves the conditional when those probabilities are computed exactly. Document routing hyperparameters are periodically fitted from sampled counts by a Gibbs-EM procedure using fixed-point Dirichlet estimation. Such adaptation is distinct from a chain targeting one fixed-hyperparameter posterior throughout training.
⚠️ Curated Limitations & Negative Results
- Experiments use fixed three-level topic structures and finite topic counts. Probabilistic routing hierarchies are not automatically ontological subsumption or causal graphs. Private/public children, separate Beta exits and other proposed DAG variations are sketches outside the two evaluated constructions.
- Collapsed counts exclude the current token and respect shared emitting nodes. A literal independent word distribution per path/level triple would change the model. Finite calculations verify both variants' conditionals on synthetic examples, not historical software or chain convergence.
- Selected Medline and newsgroup studies use particular hyperparameters, finite topic grids and simulated-document likelihood estimates with plug-in shared parameters. Topic/journal mutual information is not direct taxonomy recovery or document classification accuracy. Some numeric prose and table summaries disagree; no disputed summary is adopted.
- Reported speed and performance comparisons depend on author implementations and chosen datasets. No author software, benchmark, universal superiority or exact marginal evidence is certified.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:Model
Model
|
broad_match | This source-defined latent-variable model family is narrower than the pinned Model class. |
AIO:ProbabilisticTopicModel
Probabilistic Topic Model
|
related_match | The model supports probabilistic topic modeling; the pinned target is task-scoped, so related match preserves the model/task distinction. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
method-source
|
David Mimno, Wei Li, Andrew McCallum
|
primary paper
full_text
|
sha256:d6a539a22864…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-12T00:36:41Z
Updated: 2026-10-12T00:36:41Z