Pachinko allocation topics over time model
Model Families AIMech:PachinkoAllocationTopicsOverTimeModel DRAFT
Aliases: PAMTOT
A topic-model family combining pachinko allocation paths with topic-specific continuous timestamp densities at multiple path levels, so that words and observed times jointly inform topics and their co-occurrence structure.
Identity Notes & Concept Scope
Scoped to the 2006 Li, Wang and McCallum workshop model combining PAM and Topics over Time, particularly the demonstrated four-level PAMTOT construction. Both supertopics and subtopics supply time factors, whereas subtopics supply corpus-shared word distributions. This differs from plain PAM, single-topic-level TOT, and dynamic topic models with evolving word probabilities. PAMTOT is the source acronym. The construction models observed timestamp/content associations, not a continuous-time Markov process or causal event graph. 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
#multilevel-time-factors
In four-level PAMTOT each document draws root and supertopic routing vectors, while subtopic word vectors are shared across the corpus. Each token chooses a supertopic/subtopic path, emits a word and generates separate timestamps from Beta densities at both topic levels. Fitting repeats the observed document timestamp in all these time slots. With fixed parameters, the collapsed path weight is the three PAM leave-one-out count factors multiplied by supertopic and subtopic time densities; a common root time factor cancels from assignment odds.
Mechanism
#repeated-time-law
The source acknowledges that forward generation yields distinct times even though fitting repeats one document time. Restricting a product of timestamp densities to a shared time is not automatically a normalized single-time model. Its integral can depend on the chosen path, and can even diverge for otherwise valid individual Beta shapes. A finite two-path example gives assignment probabilities (1/4,3/4) for repeated time factors versus (1/7,6/7) after separately normalizing each path as a single-time density. These are different model interpretations. A coherent general marginal sums complete paths while retaining the product of all their time factors; the printed marginal display on p.2 has inconsistent indexing.
Mechanism
#moment-fitted-parameters
The experimental fitting updates Dirichlet and Beta parameters by empirical moments during Gibbs-style assignment updates. This differs from fixed-target posterior sampling with all parameters integrated. The printed Beta exponent order is (1-t)^(psi1-1)*t^(psi2-1), so a chosen estimator must use a consistent shape convention. This workshop paper does not print its moment formulas; the notation conflict in the separate TOT paper is not asserted as a demonstrated implementation defect here.
⚠️ Curated Limitations & Negative Results
- Fixed topic-word vectors do not represent changing word meanings. Topic correlations and manually labeled temporal plots do not establish causality, future extrapolation or automatic topic birth and death.
- Repeated timestamps and density products require explicit interpretation. Moment fitting also needs policies for empty groups, zero or invalid variance, valid positive Beta shapes, time normalization and endpoints. The general document-local and experimental corpus-shared emission formulations require different corpus marginalization.
- The experiment uses 4454 selected Rexa documents from 1991-2005, near-uniform year subsampling, 50 supertopics and 100 subtopics. The timestamp-prediction comparison lacks an explicit held-out split and complete scoring derivation in this short paper; it is not certified forecasting evidence.
- Selected KL-matched topic examples and reported timestamp metrics are source results with bounded scope. Finite curator calculations distinguish probability laws, not historical software behavior. No author code or benchmark was executed.
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. |
AIO:TimeSeriesAnalysis
Time Series Analysis
|
related_match | The model analyzes observed topical trends and co-occurrences over time, but is not identical to the broad time-series task or a guaranteed forecasting procedure. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
method-source
|
Wei Li, Xuerui Wang, Andrew McCallum
|
primary paper
full_text
|
sha256:8695fa919444…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-12T00:36:41Z
Updated: 2026-10-12T00:36:41Z