feat(algo): descriptor-template GDS layer — GDS_LEIDEN, GDS_LOUVAIN, GDS_PPR - #56
Merged
Merged
Conversation
Louvain community detection via the icebug bridge: projected graph -> InMemGraph CSR -> Arrow UInt64 arrays (system pool, see previous commit) -> zero-copy NetworKit::GraphR -> PLM -> community assignments streamed back through the GDS result pipeline. Coexists with the hand-rolled LOUVAIN under the GDS_ prefix, same as GDS_PAGE_RANK. Test asserts community structure (two 4-cliques joined by a bridge edge), not raw IDs, since PLM is parallel and nondeterministic. Co-Authored-By: Claude Fable 5 <[email protected]>
Contributor
Author
|
Rebased over #47's merge (the gated source list and registration both conflicted — resolved as the union). 75/75 locally with node2vec and Louvain together. |
zachwinter
force-pushed
the
feat/gds-louvain-helper
branch
from
August 13, 2026 18:34
c76a58f to
3f4d6cf
Compare
Contributor
|
Much of gds_xxx.cpp seems to be boilerplate. Is it possible to use some templates/reflection/codgen mechanism to automate it? Or we're looking at 100+ PRs. Also |
Random walk with restart from caller-supplied source nodes:
CALL GDS_PPR('G', [rowid, ...]) — teleportation is uniform over the
sources (NetworKit::PageRank::forSources, memory-efficient: no n-sized
personalization vector), so scores measure standing relative to the
anchors rather than globally. Consumes the shared zero-copy CSR bridge.
Optional params dampingFactor/tolerance as GDS_PAGE_RANK. Sources
validated non-empty and non-negative at bind, range-checked against the
projected node count at run time.
Tests: star anchored at a leaf (asymmetric distribution vs the plain
PageRank tie), and a two-clique trust-closure case — anchored in one
clique, mass stays with the anchor's community (0.17-0.29 vs 0.03-0.07
across the bridge) — plus empty-list and out-of-range error cases.
Co-Authored-By: Claude Fable 5 <[email protected]>
Per review: the per-algorithm boilerplate is gone. GDSPerNodeScalarAlgo (gds_algo_template.h) supplies bind, the shared zero-copy CSR bridge, GraphR construction, and vertex-compute streaming for the whole per-node-scalar equivalence class; an algorithm is now a descriptor — name, output column, optional params, extra-positional-arg policy (NoExtraArgs / SourceListArg), and the NetworKit invocation. GDS_PAGE_RANK, GDS_LOUVAIN, and GDS_PPR are rewritten as descriptors (unchanged CALL surfaces, unchanged test expectations). GDS_LEIDEN is the first new algorithm to ride the template: ParallelLeidenView with gamma/iterations optional params, ~100 lines including both param structs. Per-node-vector outputs (GDS_NODE2VEC) stay bespoke until a second member of that class motivates the template. Co-Authored-By: Claude Fable 5 <[email protected]>
zachwinter
changed the title
Contributor
Author
|
Both addressed in d9ece89 (+d90630c). Boilerplate: descriptor templates per algorithm shape — the per-node-scalar template now serves PAGE_RANK, LOUVAIN, LEIDEN, and PPR; a new algorithm in this class is a ~30–100 line descriptor. Leiden: GDS_LEIDEN via ParallelLeidenView with gamma/iterations params, first rider on the template. PPR came along as the extra-args policy's proof (sources = teleportation anchors — the trust primitive from the Discord thread). 78/78. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Reworked per review — both points:
1. Boilerplate → descriptor templates.
gds_algo_template.hintroducesGDSPerNodeScalarAlgo<DESC>: one class template per algorithm shape (per-node-scalar is the largest equivalence class — centrality, community detection). The template owns bind, the shared zero-copy CSR bridge,GraphRconstruction, and result streaming; an algorithm is now a descriptor — name, output column, optional params, an extra-positional-args policy (NoExtraArgs/SourceListArg), and the NetworKit invocation. The 100+ algorithms become ~30–100 line descriptor files, not PRs full of copied plumbing.GDS_PAGE_RANKis refactored onto it in this PR (CALL surface and test expectations unchanged). Per-node-vector outputs (GDS_NODE2VEC) stay bespoke until a second member of that class motivates the template.2. Leiden first.
GDS_LEIDENviaParallelLeidenView(gamma + iterations optional params) — ~100 lines including both param structs, the first new algorithm to ride the template.GDS_LOUVAINstays (a ~55-line descriptor now) since the two are useful to compare, but the Leiden file carries the "prefer this" note.Bonus:
GDS_PPR—CALL GDS_PPR('G', [rowid, ...])viaPageRank::forSources, theSourceListArgpolicy's first user. Its test is worth a look: two 4-cliques joined by a bridge, anchored in one clique — trust mass stays with the anchor's community (0.17–0.29 vs 0.03–0.07 across the bridge). That's the confidence primitive from the Discord trust/viewpoint discussion, as a query.Suite: 78/78 locally (all existing GDS test expectations unchanged by the refactor).
🤖 Generated with Claude Code