Summary
With random_state set, the docs state UMAP is reproducible ("prevent any stochastic behavior"). This holds for all-unique inputs, but breaks when the input contains duplicate (tied) rows: two seeded fit_transform calls on the identical array return different coordinates. Setting random_state already forces n_jobs=1 (UMAP warns it does), so this is not thread-count related.
Environment
python 3.13.7
platform macOS-26.5-arm64-arm-64bit
umap-learn 0.5.12
numba 0.65.1
numpy 2.4.6
scikit-learn 1.9.0
sentence-transformers 5.5.1
torch 2.12.0
Reproduction
Public data only (20 Newsgroups + all-MiniLM-L6-v2). One set is all-unique (control); the other is the same size but built from a 1500-doc base with 300 docs repeated 9x (2700 duplicate rows). Embeddings are computed once and reused, so the embedding model is not the variable.
import numpy as np
from sentence_transformers import SentenceTransformer
from sklearn.datasets import fetch_20newsgroups
from umap import UMAP
def fit(x):
return UMAP(n_neighbors=15, n_components=5, min_dist=0.0,
metric="cosine", random_state=42).fit_transform(x)
data = fetch_20newsgroups(subset="all", remove=("headers", "footers", "quotes"))["data"]
docs, seen = [], set()
for d in data:
d = d.strip()
if d and d not in seen:
seen.add(d); docs.append(d)
if len(docs) == 4200:
break
model = SentenceTransformer("all-MiniLM-L6-v2")
control = model.encode(docs) # 4200 unique
base = docs[:1500]
treat = model.encode(base + [d for d in base[:300] for _ in range(9)]) # 1500 unique + 2700 dupes
a, b = fit(control), fit(control)
print("unique: ", np.allclose(a, b), float(np.abs(a - b).max()))
c, e = fit(treat), fit(treat)
print("duplicated:", np.allclose(c, e), float(np.abs(c - e).max()))Expected
Both lines print True 0.0 (identical output for identical seeded input).
Actual
unique: True 0.0
duplicated: False 21.819...
The unique set is bit-identical; the set with duplicates is not (the exact max diff varies per run, but it is consistently nonzero). Deduplicating the rows restores determinism.
Likely cause
The approximate nearest-neighbor build (pynndescent) appears to break distance ties between identical points nondeterministically, and this is not controlled by random_state or n_jobs=1. Worth either fixing the tie handling or documenting that random_state does not guarantee reproducibility in the presence of duplicate rows.
Summary
With
random_stateset, the docs state UMAP is reproducible ("prevent any stochastic behavior"). This holds for all-unique inputs, but breaks when the input contains duplicate (tied) rows: two seededfit_transformcalls on the identical array return different coordinates. Settingrandom_statealready forcesn_jobs=1(UMAP warns it does), so this is not thread-count related.Environment
Reproduction
Public data only (20 Newsgroups + all-MiniLM-L6-v2). One set is all-unique (control); the other is the same size but built from a 1500-doc base with 300 docs repeated 9x (2700 duplicate rows). Embeddings are computed once and reused, so the embedding model is not the variable.
Expected
Both lines print
True 0.0(identical output for identical seeded input).Actual
The unique set is bit-identical; the set with duplicates is not (the exact max diff varies per run, but it is consistently nonzero). Deduplicating the rows restores determinism.
Likely cause
The approximate nearest-neighbor build (pynndescent) appears to break distance ties between identical points nondeterministically, and this is not controlled by
random_stateorn_jobs=1. Worth either fixing the tie handling or documenting thatrandom_statedoes not guarantee reproducibility in the presence of duplicate rows.