tasks/ml/manifests/ML17.yaml

# ============================================================================
# T17 — Comparing topic models (LDA, NMF, LSA) on 20newsgroups subsets
# ----------------------------------------------------------------------------
# Unlike paper_replication's P01-P07, the "synthesis" here frames a research
# QUESTION rather than a known paper's method. The model must design the
# experiment (conditions, metrics, datasets) — we only commit to what a
# competent study of this topic would include and what the rubric expects.
# ============================================================================

id: ML17
title: "Topic-model comparison on small 20newsgroups subsets: LDA vs NMF vs LSA"
arxiv_id: null
venue: "ARC-Bench 2026"
paper_asset: null

# The "synthesis" plays the role of the upstream briefing: research question,
# background, why the question matters, what "a reasonable experiment" looks
# like. It deliberately does NOT pre-specify a single method to reproduce.
synthesis: |
  Topic modeling methods are often compared using either intrinsic quality
  measures (such as topic coherence) or extrinsic utility (such as how well
  document representations align with known labels), but the two views can
  disagree. Latent Dirichlet Allocation (LDA), Non-negative Matrix
  Factorization (NMF), and Latent Semantic Analysis (LSA/SVD) each induce
  different assumptions over term-document structure and may therefore rank
  differently depending on metric choice.

  On CPU-constrained benchmarks, a practical study should use a compact text
  corpus and a controlled preprocessing pipeline so that runtime stays short
  while still yielding meaningful distinctions. A small subset of
  20newsgroups is ideal because it has accessible labels for external
  clustering evaluation and enough lexical diversity for coherence analysis.
  Using multiple subset difficulties (well-separated vs more confusable
  categories) helps test robustness of conclusions.

  A credible experiment compares LDA, NMF, and LSA under matched topic counts
  and vectorization settings, then reports both c_v-like coherence and
  adjusted rand index (ARI) from document-cluster assignments against true
  newsgroup labels. Because exact c_v implementations are not native in
  sklearn, an explicit approximation based on sliding-window co-occurrence
  and normalized PMI should be documented and applied consistently across
  methods. Repeated runs for stochastic models are needed to avoid
  over-interpreting single-seed noise.

  The core question is whether better intrinsic coherence implies better
  label alignment on small real-world corpora, and which model offers the
  best trade-off under strict CPU budgets.

  *Do LDA, NMF, and LSA produce different rankings on coherence versus ARI on small 20newsgroups subsets, and does NMF provide the strongest coherence-structure trade-off under CPU limits?*

hypotheses:
  - id: H1
    statement: "NMF achieves higher mean coherence_cv than LSA on at least 2 of 3 evaluated 20newsgroups subsets when using the same number of topics and preprocessing pipeline."
    measurable: true
  - id: H2
    statement: "Across the three methods (LDA, NMF, LSA), the method with the highest coherence_cv is also the highest-ARI method on no more than 1 of the 3 subsets, indicating weak metric agreement."
    measurable: true
  - id: H3
    statement: "LDA achieves mean ARI at least 0.03 higher than LSA on at least 2 of 3 subsets when document clusters are obtained by argmax topic assignment."
    measurable: true

experiment_design:
  research_question: "How do LDA, NMF, and LSA compare on coherence_cv versus document-cluster ARI over small 20newsgroups subsets, and do intrinsic/extrinsic rankings diverge?"
  conditions:
    - name: "lda_k4"
      description: "LatentDirichletAllocation with n_components=4, max_iter=20, learning_method='batch', CountVectorizer features, repeated over 3 seeds."
    - name: "nmf_k4"
      description: "NMF with n_components=4, init='nndsvda', solver='cd', max_iter=300 on TF-IDF features, repeated over 3 seeds."
    - name: "lsa_k4"
      description: "TruncatedSVD (LSA) with n_components=4 on TF-IDF features; document-topic embeddings clustered via argmax after non-negative shift or via KMeans(k=4) with fixed seed."
    - name: "nmf_k6_sensitivity"
      description: "NMF sensitivity condition with n_components=6 to test topic-count robustness for coherence and ARI trends."
  baselines:
    - "lsa_k4 as a linear-algebra baseline without probabilistic topic assumptions"
    - "lda_k4 as a probabilistic-topic baseline"
  metrics:
    - name: "coherence_cv"
      direction: "maximize"
      description: "Approximate c_v coherence computed from top words per topic using sliding-window co-occurrence and NPMI aggregation, averaged across topics and seeds."
    - name: "ari"
      direction: "maximize"
      description: "Adjusted Rand Index between predicted document clusters (topic argmax or equivalent) and true newsgroup labels."
    - name: "runtime_sec"
      direction: "minimize"
      description: "Wall-clock training + inference time per condition/dataset cell on single CPU core."
  datasets:
    - name: "20ng_easy4"
      source: "sklearn.datasets.fetch_20newsgroups with categories=['sci.space','rec.autos','comp.graphics','talk.politics.misc']"
    - name: "20ng_related4"
      source: "sklearn.datasets.fetch_20newsgroups with categories=['comp.graphics','comp.os.ms-windows.misc','comp.sys.ibm.pc.hardware','comp.sys.mac.hardware']"
    - name: "20ng_science4"
      source: "sklearn.datasets.fetch_20newsgroups with categories=['sci.space','sci.med','sci.electronics','sci.crypt']"
  compute_requirements:
    gpu_required: false
    estimated_wall_clock_sec: 720

rubric_path: "experiments/arc_bench/config/ml/rubrics/ML17.json"