gWM 3.1.3

Graph-indexing wavelet matrix for similarity search over large graph databases

gwm

gwm finds every graph in a database whose similarity to a query graph is at least a given threshold — it answers threshold queries, not top-k or approximate nearest-neighbor search.

It encodes each graph as a binary set of Weisfeiler-Lehman (WL) subtree features and indexes that feature space with a wavelet matrix. Queries are answered by traversing the wavelet matrix rather than explicitly scoring every database graph. It has no Python runtime dependencies.

This is the reference implementation of:

Yasuo Tabei and Koji Tsuda. Kernel-based Similarity Search in Massive Graph Databases with Wavelet Trees. Proceedings of the 2011 SIAM International Conference on Data Mining (SDM), pp. 154–163. https://doi.org/10.1137/1.9781611972818.14

@inproceedings{tabei2011kernel,
  author    = {Yasuo Tabei and Koji Tsuda},
  title     = {Kernel-based Similarity Search in Massive Graph Databases with Wavelet Trees},
  booktitle = {Proceedings of the Eleventh SIAM International Conference on Data Mining (SDM)},
  pages     = {154--163},
  publisher = {SIAM},
  year      = {2011},
  doi       = {10.1137/1.9781611972818.14},
}

The package was previously distributed as gWT; gwm is the same algorithm and input format under a new name (the on-disk index format is an internal implementation detail and isn't guaranteed to match gWT's).

Installation

pip install gwm

Prebuilt wheels are provided for Linux (x86_64, aarch64) and macOS (x86_64, arm64) on CPython 3.9–3.14. There are no Windows wheels; other platforms may be able to build from the source distribution, which needs only a C++17 compiler (see Building from source).

Quickstart

Build an index from a database file and search it with a query file, both in gSpan format:

import gwm

index = gwm.Index.build("db.gsp", iteration=2)
index.save("index.bin")

index = gwm.Index.load("index.bin")
for hits in index.search_file("query.gsp", threshold=0.8):
    print(hits)  # [(graph_id, similarity), ...], one list per query graph

graph_id is always the graph's 0-based position in the database (file order, or list order for in-memory graphs) — not the id written on a gSpan file's t # line, which gwm ignores (see Graph file format).

Graphs can also be built and searched entirely in memory. Each graph is a (labels, edges) pair: labels is a 0-based list of integer vertex labels, edges is a list of (u, v, edge_label) triples (also 0-based, undirected):

import gwm

graphs = [
    ([6, 6, 6, 6], [(0, 1, 1), (1, 2, 1), (2, 3, 1), (3, 0, 1)]),
    ([6, 6, 7],    [(0, 1, 1), (1, 2, 2)]),
]

index = gwm.Index.build_from_graphs(graphs, iteration=2)
hits = index.search_graphs(graphs, threshold=0.8)

gwm.from_networkx() converts a networkx.Graph (read from label node/edge attributes by default) into that same (labels, edges) form:

import networkx as nx
import gwm

g = nx.Graph()
g.add_node(0, label=6)
g.add_node(1, label=6)
g.add_edge(0, 1, label=1)

index = gwm.Index.build_from_graphs([gwm.from_networkx(g)], iteration=2)

networkx itself is not a dependency — from_networkx() only relies on the standard .nodes(data=True) / .edges(data=True) protocol.

Index reference

• Index.build(path, iteration=2), Index.build_from_graphs(graphs, iteration=2) — construct.
• Index.load(path) / index.save(path) — the two sides of the on-disk index format.
• index.search_file(path, threshold), index.search_graphs(graphs, threshold) — encode queries with the index's own label dictionary and search in one call.
• index.encode_query_file(path), index.encode_query_graphs(graphs) — encode only, returning each query graph's WL code set (list[list[int]]) for reuse with search_encoded.
• index.search_encoded(labels, threshold) — search a single already-encoded query, where labels is one WL code set returned by encode_query_file/encode_query_graphs.
• index.num_graphs, index.iterations, index.matrix_depth, index.index_length, index.memory_bytes — index statistics.

threshold must be in (0, 1]; it is the cosine-similarity cutoff, not a count.

Command line

The pip package installs a gwm command:

# build an index (-iteration = number of WL relabeling rounds, default 2)
gwm build -iteration 2 db.gsp index.bin

# search (-kthreshold = cosine similarity threshold, default 0.8)
gwm search -kthreshold 0.8 index.bin query.gsp

The single-dash -iteration / -kthreshold flags (not --iteration) match the legacy gWT CLI. Building from source (see below) also produces standalone gwm-build / gwm-search binaries with no external library dependencies and the same options; these are not part of the pip package.

Graph file format

Both the CLI and Index.build/search_file read databases and queries in gSpan format — one record per graph, starting with a t # line; blank lines between records are optional (a new t # line is itself enough to end the previous record):

t # <anything ignored by gwm>
v <vertex-id> <vertex-label>
e <from> <to> <edge-label>

The tokens after t # (conventionally a graph id, a class label, and a name) are ignored by the index — as noted above, a graph's graph_id in results is its 0-based position in the file, not this field. Vertex ids are 1-based positive integers and must be declared with v before they appear in an e line; vertex and edge labels must be non-negative integers. Vertex ids should be contiguous starting at 1 — gaps are not rejected, but are silently filled with extra zero-label, edgeless vertices that then contribute their own WL features. Graphs are treated as undirected, and self-loops and repeated edges are read as given (not rejected or merged). gwm.read_gspan(path) parses this format into a list of (labels, edges) graphs used by the in-memory API.

How it works

For each graph, the initial vertex labels are used as round-0 features; each subsequent WL relabeling round then produces one signature string per vertex (its own label, then its sorted (neighbor label, edge label) multiset). Each distinct signature is assigned a dense integer code shared across the whole database. A graph's feature set is the deduplicated set of codes seen across all rounds. Unlike the counted feature vectors some WL subtree-kernel implementations use, this is a binary feature vector — a code is either in a graph's set or it isn't — so a feature set's squared L2 norm is simply its size.

iteration (default 2) is the number of WL relabeling rounds run after the initial vertex labels, so iteration=2 contributes three rounds of features per graph: the initial labels, plus two rounds of relabeling.

At build time, every code's posting list (the graphs containing it) is concatenated into one sequence and stored as a wavelet matrix. At query time, each of the query's codes contributes its interval in that sequence; a DFS over the matrix intersects those intervals level by level, pruning any subtree whose surviving range is already too small to reach the threshold. At a leaf, the surviving range size is the number of codes shared between the query's feature set Q and the graph's feature set G, giving the cosine similarity between them directly: |Q ∩ G| / sqrt(|Q| · |G|), without explicitly scoring every database graph one by one.

Limitations

• Graphs are treated as undirected; edges are stored symmetrically regardless of a gSpan file's from/to order.
• Similarity is computed over binary WL subtree-feature sets (present or absent), not per-code counts.
• gwm supports threshold search, not top-k or approximate nearest-neighbor search.
• The on-disk index format is an internal detail and isn't guaranteed to be stable across gwm versions or compatible with indexes produced by the original gWT package.

When to use

gwm is useful when you have a mostly-static graph database and need to run many threshold-similarity queries against it. For a one-off query against a small database, a plain linear scan may be simpler and just as fast.

Building from source

make                  # from the repo root; produces src/gwm-build and src/gwm-search
make test             # build + output-parity regression

or with CMake:

cmake -S . -B build && cmake --build build

Both require only a C++17 compiler; all non-standard code the build uses is vendored in the repository (see License).

License

gwm's own code is MIT licensed (see LICENSE). It vendors Sebastiano Vigna's sux rank/select library (src/rank9sel.{h,cpp}, src/macros.h, src/popcount.h), which is LGPL-2.1-or-later (see LICENSES/LGPL-2.1-or-later.txt). The distribution as a whole is therefore MIT AND LGPL-2.1-or-later; full source for every component, including the vendored library, ships in the source distribution.