← banlys.comLattice

Banlys · Research

Paper & preprint

RecSys 2026 submission in progress. Methods, ablations, and reproducibility are in the PDF and supplementary material — not summarized in full here.

RecSys 2026 — under reviewPatent pendingarXiv preprint ↗

What the paper covers

Lattice is a confidence-gated composition layer for sequential prediction: an auxiliary behavioral structure is blended in only when support is sufficient; otherwise the backbone model runs unchanged. The contribution is when to compose, not a new clustering algorithm or standalone recommender.

Headline results (see PDF for detail)

  • Primary comparison: LSTM vs LSTM+Lattice on MovieLens (+31.7% gated HR@10, +58.7% ungated, 30 paired seeds)
  • Trade-off: always-on fusion higher on pooled HR (+58.7% vs +31.7% gated); gating trades peak benchmark for auditable ON/OFF control
  • Switch semantics: gate OFF rows match backbone; gate ON rows match always-on fusion on supported rows
  • Backbone-native: transformer +13.3% and SASRec +17.0% gated on MovieLens (same evaluation design)
  • SASRec/BERT4Rec reported as contextual baselines only — not the primary estimand

Controlled shift tests (appendix) illustrate refusal when embeddings leave training support; recommendation benchmarks carry the headline. This site does not reproduce implementation details.

Read on arXivDownload paper PDFDownload supplementary PDFRequest research license

Cite

@misc{bannis2026lattice,
  title={When to Use What You've Learned: Confidence-Gated Archetypes for Sequential Recommendation},
  author={Anonymous Author(s)},
  year={2026},
  note={RecSys 2026 submission; see arXiv for public preprint}
}

Evaluation protocol summary available to research licensees on request, subject to license terms.