This is the second exploratory research arc (after
Amortized Bayesian ER), documented for
honesty. Outcome: a negative result. The topology/landscape framing is
novel (a six-angle prior-art scan confirmed it is unoccupied for ER), but under
a fair, calibrated comparison the mechanism is cosmetic — it produces the
same partition as a plain discrete split/merge loop. Source + results in
scripts/research/landscape_er.py and RESULTS-landscape-er.md.The idea
Reframe ER as sculpting a potential landscape: a profiler sketches a coarse terrain of basins (“buckets”); records are “marbles” that roll into basins; after each pass, residuals reshape the terrain — stranded marbles carve new basins (recall), impure basins get a ridge raised to split them (precision) — iterating until every marble settles in a clean basin that represents an entity. The appeal was the symmetric two-move loop and a “global re-flow” after each edit.White-space scan: genuinely open
A six-angle prior-art scan (iterative watershed/Morse, mode-seeking add/split, ER as energy landscape, profiler-driven iterative ER, MDL split-merge, self-growing topology) found the synthesis unoccupied: ER has never been framed as records settling into attractor basins of a sculpted landscape (no Hopfield/energy/attractor ER exists), and the add-for-recall / split-for-precision polarity is unattested in density-landscape clustering. The one crowded angle was the iterative-refinement loop itself (Iterative Blocking, Gruenheid incremental RL, the Sayari pipeline), so the only defensible novelty was the mechanism, not the loop.Kill-criterion: does the mechanism beat a discrete loop?
A fair fight: same affinity graph, same correlation-clustering objective, same Fiedler 2-cut proposals, same init, symmetric move sets. Only the mechanism differed (label-propagation routing + ridge-raising + global re-flow vs. direct relabel). A near-miss worth recording. A first run showed the landscape beating the discrete loop by +0.10 F1 on DBLP-ACM across seeds — but this was a θ-calibration artifact: the threshold θ was set to the median affinity, far below the signal band, so the objective’s optimum was gross over-merge and the landscape merely chased a broken objective less aggressively. Fixing the objective (the canonical correlation-clustering cost with a calibrated unsupervised θ = mean + 2·std of nonzero affinities) erased the win:
In the conflict regime a plain threshold (CC@θ) often beat both refinement loops.
Why it failed
Proximate: given the same objective and the same split proposals, a greedy hill-climb reaches the same local optimum whether a move is expressed as “relabel” or “raise a ridge and re-flow.” The landscape is a different route to the same destination. By sharing the objective and proposals to make the fight fair, the test could only ever measure the optimizer’s route, not the representation’s power. Structural (shared with the first arc):- The benchmarks are saturated — a calibrated threshold already hits ≈1.0 on the subsamples, leaving no headroom for a new mechanism to win.
- It attacked the clustering mechanism, the layer that matters least. On clean data the affinity is good enough that almost any sane clustering recovers the truth; the real bottleneck is representation/affinity quality and operational recall, not the partitioner.
- The topology was old wine: basins = connected components (H₀), ridges = graph cuts, rolling = label propagation. The metaphor brought vocabulary, not new machinery.