narrativecluster 0.1.7

Claim-level narrative clustering with HNSW + Leiden + neighbor-vote assignment

narrativecluster

Claim-level narrative clustering for large social-media corpora.

Stack: HNSW (FAISS) → Leiden community detection → neighbour-vote cluster assignment

---

Why this approach?

Clustering millions of social-media posts is an unsupervised problem with no clean ground truth, no fixed number of narratives, and a distribution that shifts every week as new stories emerge. Most standard algorithms break down in one of a few predictable ways.

K-means and its variants require you to specify k up front. For a corpus spanning thousands of distinct claims across a multi-year election cycle, the right k is unknowable in advance — and the algorithm will happily partition noise into fake clusters to fill whatever k you gave it.

Gaussian Mixture Models and DP-Means are theoretically appealing (they can infer k from the data) but brittle at scale. DP-Means is sensitive to its bandwidth parameter in high-dimensional embedding space, where "distance" is dominated by the curse of dimensionality rather than semantic content, and it tends to either over-fragment coherent narratives or lump disparate ones together depending on how that single parameter lands. GMMs have similar issues and become computationally expensive as k grows into the thousands.

DBSCAN / HDBSCAN handle arbitrary cluster shapes well on low-dimensional data but struggle in the intrinsic geometry of dense embedding spaces — nearly every point becomes a "core point" or noise depending on a single eps that is nearly impossible to calibrate in cosine space. They also have no natural mechanism for assigning new documents to existing clusters without re-running the full algorithm.

What narrativecluster does instead

The approach decouples structure discovery (Leiden on a cosine-thresholded KNN graph) from point assignment (neighbour-vote), which buys three things simultaneously: no fixed k, graceful handling of noise, and fast incremental updates.

Structure discovery via graph community detection. Rather than treating each post as a point in metric space, we build a sparse k-NN similarity graph where an edge only exists if two posts exceed a cosine similarity threshold (tau_graph). Leiden community detection then finds densely-connected subgraphs — natural narrative clusters — without any assumption about their number, shape, or size. Leiden is resolution-parameterised, which is semantically interpretable: higher resolution → finer-grained narratives, lower → coarser topical buckets. Critically, isolated posts and borderline cases simply have no edges and fall into small singleton clusters rather than being force-assigned to the nearest centroid.

Neighbour-vote assignment for new data. Once the graph structure is learned, assigning a new post is cheap and explainable: query the HNSW index for its nearest neighbours, filter to neighbours above tau_edge, and take a plurality vote over their cluster memberships. The vote fraction and margin give a per-post confidence score, so you can dial your own precision/recall tradeoff after fitting — without touching the model. This is O(log N) per new post rather than O(N²).

A concrete example. Suppose you have five posts:

A: "Ballots were found in a dumpster in Maricopa County — proof of fraud"
B: "Election officials in Arizona caught destroying ballots"
C: "The mRNA vaccine is causing myocarditis in young athletes"
D: "Pfizer trial data shows spike protein accumulates in organs"
E: "My dog had a great walk this morning :)"

• K-means with k=2 groups {A,B,C,D} vs {E} — conflating two distinct narratives because it must fill both centroids.
• DP-Means may correctly split {A,B} from {C,D}, but will almost certainly pull post E into whichever centroid is geometrically closest rather than leaving it unclaimed.
• narrativecluster: A↔B have high cosine similarity and form one Leiden community; C↔D form another; E has no neighbours above tau_graph and lands in a singleton — it is simply not assigned when you call predict(). That is the correct answer for noise.

Speed. HNSW graph construction is O(N log N) and approximate-neighbour search at prediction time is effectively O(log N). On a 12M-post corpus, fit() takes ~20 minutes on a single CPU; predict() on a new 100K-post batch takes under 2 minutes. DP-Means or a full GMM on the same corpus would be intractable.

---

Install

pip install narrativecluster
# GPU variant (requires faiss-gpu):
pip install "narrativecluster[gpu]"

From source:

git clone https://github.com/patrikgerard/narrativecluster
cd narrativecluster
pip install -e ".[dev]"

---

Quick start

from narrativecluster import NarrativeClusterer, ClusterConfig

cfg = ClusterConfig(
    embed_col="embedding",   # column holding np.ndarray / list embeddings
    text_col="text",
    site_col="platform",
)

model = NarrativeClusterer(config=cfg, checkpoint_dir="./checkpoints")

# Full fit — idempotent: loads from checkpoint if it exists
model.fit(df_train)

# Incremental update — adds new rows, re-runs Leiden, keeps cluster space stable
model.partial_fit(df_new_batch)

# Neighbour-vote assignment for any DataFrame
df_pred = model.predict(df_query)
# df_pred now has: nv_pred_cluster, nv_accepted, nv_vote_frac, …

# Per-site acceptance stats
global_rate, site_tbl = model.site_stats(df_pred)
print(site_tbl)

# Save / load
model.save("./my_model")
model2 = NarrativeClusterer.load("./my_model", config=cfg)

---

CLI

# Full fit + predict
narrativecluster --raw raw_posts.parquet --out ./checkpoints

# Incremental update
narrativecluster --dedup ./checkpoints/df_unique.parquet \
                 --partial_fit new_posts.parquet \
                 --out ./checkpoints

# Force re-predict
narrativecluster --dedup ./checkpoints/df_unique.parquet \
                 --out ./checkpoints \
                 --force_repredict

---

partial_fit semantics

partial_fit is designed for streaming / incremental ingestion:

1. New rows are deduplicated against themselves and all previously seen texts.
2. They are mean-centred using the original µ (the embedding space is not re-centred — this keeps the geometry stable across batches).
3. New vectors are added to the live HNSW index.
4. KNN arrays are extended; Leiden is re-run on the combined graph.

Note: Cluster IDs may shift after each partial_fit because Leiden is non-deterministic and the graph changes. If you need stable IDs across incremental updates, pin random_seed and store a cluster-label mapping separately.

---

Configuration

All hyper-parameters live in ClusterConfig:

Parameter	Default	Description
embed_col	"embedding_aligned"	DataFrame column with embeddings
hnsw_m	32	HNSW M parameter
ef_construction	200	HNSW efConstruction
ef_search	128	HNSW efSearch
k_graph	64	KNN neighbours for Leiden graph
tau_graph	0.68	Cosine-sim threshold for graph edges
do_mutual	False	Keep only mutual edges
cap_m	20	Max out-degree per node (None = uncapped)
leiden_resolution	1.0	Leiden resolution parameter
kq	50	Neighbours queried at prediction time
tau_edge	0.68	Min sim for a vote to count
min_cluster_size	4	Clusters smaller than this are ineligible
vote_margin	2	Min vote margin (top − second) to accept

---

Output columns (predict)

Column	Type	Description
nv_pred_cluster	int32	Predicted Leiden cluster id (-1 = rejected)
nv_accepted	bool	Whether the vote passed quality thresholds
nv_votes_top	int16	Votes for winning cluster
nv_votes_tot	int16	Total eligible neighbour votes
nv_vote_frac	float32	Top votes / total votes
nv_vote_margin	float32	top_votes − second_votes
nv_best_sim	float32	Max cosine similarity among winning-cluster neighbours

---

License

MIT