Changelog

All notable changes to LaNet-vi will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Unreleased

Added

• --kconn, --kconn-type wide|strict and --kconn-file FILE (kconn, kconn_type in YAML; Network.compute_kconnectivity() / Network.kconnectivity in Python): the k-connectivity of the shells, the C++ -kconn (#47). Port of computeKConnectivityWide / computeKConnectivityStrict of graph_kcores_components.cpp (Beiró, Alvarez-Hamelin & Busch, NJP 2008): a walk from the top shell down that seeds a k-connected set with the first cluster of diameter ≤ 2 (wide: or minimum edge cut ≥ shell; strict: and minimum degree ≥ shell) and grows it with the contracted-diameter and frontier / φ conditions, each shell's clusters examined once in the order of the component tree. Verified against a literal transcription of the C++ loops on 2 400 random graphs. Nodes that are not k-connected are painted black on white / white on black and drawn as squares in the bw / bwi schemes, as the C++ did; the C++ "border color" (a same-radius circle under the node in its SVG, invisible) is not drawn. --kconn-file writes the C++ log/kconn.log (node shell_index k_connectivity, 0 = not k-connected). Refused for k-denses, d-cores, weighted, directed and multigraph inputs, as the C++ did. Not reproduced: the crash on a shell without clusters and the shell == 77 debugging escape of the strict seed.

Documentation

• Issue #25 is closed as won't-fix for the Gomory-Hu connectivity report and the POV-Ray scenes; k-connectivity (-kconn) moves to #47. docs/cpp-migration.md and the parity checklist below say so.

5.2.0 - 2026-09-20

Every item of the September 2026 C++ comparison is now closed except the features listed under #25 (k-connectivity, Gomory-Hu connectivity, POV-Ray): the community flags, --from-layer, --window and the C++ -names forms work, the dead options are gone, the d-core table of the 4.0.0 tool is available, the tests cover 95 % of the package and the documentation site is live.

Added

• --dcore-table FILE and compute_dcore_table (#21): the (k, l)-core table of Giatsidis et al. for a directed graph — for every out-degree threshold l, the largest k with each node in the (k, l)-core — written as the C++ 4.0.0 dcores_list.txt (node k l lines). That driver computed only this table for a directed graph and drew nothing; it also kept a node in the in-degree peeling after its out-degree had dropped below l, reporting a larger k for such nodes, while this implementation follows the definition (checked against a brute-force table in the tests). The picture and the (k_in, k_out) pairs are unchanged.
• Community detection is wired into the picture (#23): --detect-communities (and the community YAML section) detects the communities of the drawn graph with Louvain or greedy modularity (--community-algorithm, --community-resolution, seeded by --seed), colors the nodes and the gradient edges by community (--color-by-community, which hides the index color legend; a colors file takes precedence) and draws a translucent convex hull under each community (--draw-community-boundaries); draw_circles, boundary_alpha and colormap are YAML-only. The CLI prints the number of communities and the modularity. In the API: Network.detect_communities(), Network.communities, Network.colors_by_community, lanet_vi.community.detect_communities(graph, config, seed), a resolution for detect_communities_greedy_modularity, and render_network(communities=, community_config=). community.colormap must name a matplotlib colormap (a short validation error otherwise). The community flags are no longer marked inert.
• --from-layer K (from_layer in YAML) works: the C++ -fromlayer (#23). The subgraph induced by the nodes of index ≥ K is decomposed again (weighted k-cores reuse the strength intervals of the whole graph) and the layout, the picture, the legends and --cores-file describe that subgraph. compute_kcores takes a p_function to reuse. A layer above the maximum index is a usage error.
• --node-labels without --names labels every node with its number, the C++ -names with no file (-names 0) (#23). With a names file only the named nodes are labeled, as before.
• --window HSTART HEND VSTART VEND (window in YAML), the C++ -window: render only that fraction of the frame, measured from the top-left corner, at the full pixel size (#25). PDF and SVG output (by the --output extension) already worked and is now documented.

Removed

• The dead renderer setting (LaNetConfig.renderer, the Renderer enum and the top-level renderer: key of lanet-vi config output): matplotlib is the only renderer and nothing read it (#23). Older YAML files that still have the key load unchanged (unknown keys are ignored).
• The inert spiral layout and spatial index: visualization/spiral_layout.py and visualization/spatial_index.py (nothing called them; the C++ espiral.cpp was never in a release), the --use-spiral-layout, --spiral-K, --spiral-beta and --spiral-separation flags, and the layout.use_spiral_layout, spiral_k, spiral_beta, spiral_separation and use_spatial_hashing settings. YAML files that still contain those keys keep loading (unknown keys are ignored); the flags are now an error. The placement is selected by --coord-distribution alone.

Fixed

• compute_dcores counted a self-loop toward both the in- and the out-degree, so a node with only a self-loop was a (1, 1) node; self-loops are ignored now, as the edge-list reader, the k-core path and the (k, l)-core table already do.
• get_community_colors ignored its colormap argument and, for 21-40 communities, sampled tab20b at fractions that repeat colors. A qualitative colormap is now used entry by entry (up to 10 communities take the tab10 shades instead of tab20's dark/light pairs), a continuous one is sampled evenly, and beyond the entries of a qualitative map the colors are spread over hsv. Community colors follow the community id rather than its position in the list.
• draw_community_boundaries printed Qhull's full diagnostic for a community whose nodes are collinear or share positions; collinear communities now get a sliver hull (Qhull's joggle) and communities with fewer than three distinct positions are skipped.
• Community detection on a graph without edges (isolated nodes) raised a ZeroDivisionError from NetworkX's modularity; the modularity is now 0.
• write_graph_json(include_node_attrs=False) wrote every node with id: 0 (the id was read after the attributes were cleared); the real ids are kept now.
• write_graph_json always writes the edges under "links": NetworkX 3.6 changes the default key of node_link_data to "edges", which would have silently changed the file format.
• Loggers were named lanet_vi.lanet_vi.<module>; get_logger no longer doubles the package prefix. setup_logging closes the handlers of a previous call instead of leaking the log file handle.
• draw_community_circles no longer triggers a matplotlib warning about color overriding edgecolor.

Changed

• Coverage gate raised from 60 % to 80 % (#11) now that metrics/, generators/, community/, visualization/community_viz.py, io/writers.py and logging_config.py have tests (the suite covers 95 %).

Documentation

• Documentation site at https://conexdat.github.io/LaNet-vi/ (#12): MkDocs with the Material theme, the existing guides, the changelog and an API reference generated from the docstrings by mkdocstrings. mkdocs build --strict runs in CI on every PR; .github/workflows/docs.yml deploys to GitHub Pages on every push to main. The docs extra installs the tooling; uv run mkdocs serve previews the site.
• Content pass on the docs: docs/usage.md is now the reference for every CLI option (default, meaning, C++ equivalent, inert ones marked), the configuration file, the input and output formats and the Python API; new docs/cpp-migration.md translates every LaNet-vi 3.x flag; the README is shorter, drops the claims that did not hold (rendering "millions of nodes", community detection as a CLI feature) and links to the guides; docs/concepts.md cites the actual papers (NIPS 2005, NJP 2008, D-cores, k-truss). Every shell block in the docs is now a command that runs as written (the multi-line ones had \\ continuations that broke when pasted).

5.1.0 - 2026-09-18

The first release after a source-level comparison with the original C++ LaNet-vi 3.x. The 5.0.0 rewrite had replaced most of the algorithms with approximations; 5.1.0 ports them faithfully, so the pictures are the LaNet-vi pictures again: the placement of Alvarez-Hamelin, Dall'Asta, Barrat & Vespignani (NIPS 2005), the triangle peeling of k-dense, the weighted k-core peeling, the color scale of types.cpp with red for the top core, the per-edge sampling seeded by --seed, and the pow/log circle packing of sibling components. It also fixes the crashes and CLI problems found on the way (d-cores through the CLI, tab-separated input, self-loops, --multigraph, YAML round trip, --config precedence, --names), drops Python 3.9, resolves every open security alert and adds CITATION.cff. The details, by category:

Documentation

• American English is now the mandatory spelling for identifiers, docstrings, comments and docs (AGENTS.md, CONTRIBUTING.md, PR template). Existing British spellings in comments, docstrings and docs were corrected; the only renamed identifiers are private (_Placer.place_centre → place_center) and six test names, so no public API changed.
• CITATION.cff (validated with cffconvert): software citation with the package authors, version and release date, plus the NIPS 2005 and New Journal of Physics 2008 papers as references; the release checklist now updates it with each version.
• README now has a Heritage section crediting the original C++ LaNet-vi (Beiró, Alvarez-Hamelin et al., 2005–2016), its SourceForge distribution under the Academic Free License 3.0 and the original homepage; the License section states the relation between the MIT-licensed rewrite and the AFL-licensed original.
• The 5.0.0 entry below overstated feature parity with the C++ version; it now carries a note pointing to the tracking issues (#18–#26).

Security

• Refreshed uv.lock: resolves all 79 open Dependabot alerts (1 critical in jupyter-server, 42 high across Pillow, tornado, urllib3, mistune, jupyterlab and others). Only requests and Pillow (via matplotlib) are runtime dependencies; the rest were pulled in by the dev extra.
• New Security workflow: pip-audit over the locked tree on every lockfile change and weekly, plus zizmor over the workflow files.
• GitHub Actions pinned to commit SHAs, persist-credentials: false on every checkout, and no cache restore in the release workflow.
• SECURITY.md now states supported versions, the private reporting channel, response targets and the coordinated-disclosure process.

Added

• The pow and log coordinate distributions of the C++ (--coord-distribution, #18): the network is a unit disc, the children of a component share a disc of radius ((T - S) / T)^(1/8) R capped at 0.96 R (log: sqrt((n - s) / n) R) and sibling components are packed as non-overlapping discs whose areas follow alpha w^beta of their sum(log(1 + d)^(2 / beta)) share, inflated 1 % per round until they no longer fit (distribute_components.cpp, seeded by --seed). Node radii in these modes are 0.007 ratio_constant log(1 + d)^1.5 (computeHostRatio), with the new --ratio-constant (C++ -ratioConstant, default 1). K-dense pictures use the variant of kdenses_component.cpp (fixed 0.92 shrink, same-index neighbors in formula (1) and the angle, epsilon scaled by tau, no u, node radius ratio_constant sqrt(log(1 + d)) with ratio_constant auto-adjusted by the top cores). --alpha and --beta now act; alpha defaults to the C++ 0.3 (was an inert 1.0). Deliberate deviations: inflated discs keep their angle and are pulled towards their container's center (the C++ pulled them towards the picture's origin, and overwrote x before computing y from it, skewing the angle), unplaced neighbors are skipped in the angle. classic stays the default for every decomposition, although the C++ k-dense only ever had this placement (its default flags behaved as pow).

Deprecated

• use_spatial_hashing is accepted and ignored: the circle packing follows the C++ algorithm (#18).
• edge_alpha (--edge-alpha) is an alias of opacity (--opacity, the C++ -opacity, default 0.2); min_edge_width / max_edge_width (--min-edge-width, --max-edge-width) are accepted and ignored (#24). lanet-vi config no longer writes the deprecated aliases (edge_alpha, show_size_legend) into the template, where they were ignored next to the current field.

Removed

• visualization.circular_average and layout.compute_hierarchical_layout (replaced by lanet_layout); layout.distribute_components is now the faithful port used by the pow/log modes (the earlier version lacked the growth loop) and layout.SpatialHashGrid is gone with it.
• Python 3.9 support. 3.9 reached end of life in October 2025 and the patched releases of Pillow, jupyter-server and others require 3.10+. requires-python is now >=3.10; the code base uses X | None unions and zip(..., strict=True).

Fixed

• Rendering follows the C++ LaNet-vi again (#24). Color scale from types.cpp: the rainbow runs magenta → blue → cyan → green → yellow → red (the maximum index was magenta), the black-and-white scale runs white → gray → black (it was inverted) and bwi interlaces that same scale; consecutive shells alternate a luminosity of 0.7 and 1.2 (was 0.7/1.0), with the k-dense rules of graphics_kdenses.cpp (constant 0.9 on a white background; with measure = mcore, the default, --color-scale-max is read in m-core units and the legend labels each k-dense as k - 2 under the title m-core; --measure kdense keeps the k-dense numbers). Nodes absent from a --colors-file are white on black / black on white, and the color legend is hidden in that case.
• Edges: every edge is kept with probability max(edges_percent, min_edges / E) (the C++ per-edge Bernoulli) using the layout seed, so --seed makes pictures reproducible (the stratified sampler used the unseeded random module). Edge colors darken by 0.75 in col pictures and lighten by 1.2 in bw/bwi ones; k-dense edges are one color, their own dense index darkened by 0.5 (the flat gray the 3.0.1 release used for edges between clusters was removed in the 3.0.2 and 4.0.0 drivers and is not reproduced). Edge width is 0.2 host radii of the smaller endpoint degree (the C++ ratioEdge), in layout units with a one-pixel floor; edges are drawn under the nodes in increasing index order. Nodes are opaque (the >1000-node path drew them at alpha 0.9).
• Legends are drawn in layout units where generateNetworkFile put them, so they scale with the picture: the color legend has one circle per index (from 1, or 2 for k-denses), labeled in the index color every max // 15 + 1 indices; the degree legend shows ceil(dmax / 4^i) (down to 2, at most five) with the radius those nodes have in the picture (it used its own size formulas and ignored node_size_scale), or strengths for weighted graphs. Weighted graphs whose strengths never exceed 1 (where the C++ strength law divides by log(max) <= 0) now size nodes by the degree law instead of a constant radius, and the legend follows.
• The PNG is exactly width x height pixels (bbox_inches="tight" cropped it); the frame is scaled uniformly to fit and centered, so any aspect ratio works and the validator that rejected sizes such as 3200x800 is gone.
• Weighted k-cores (#20) now peel, as the C++ findCores weighted branch does: every node starts in the strength interval of its total strength and, when a shell is removed, its neighbors are re-binned using only the strength they still receive from nodes above that shell. The previous code only binned total strengths (a histogram, not a core decomposition; it disagreed with the C++ on 1,570 of the 2,070 random configurations checked, see PR #31). Default granularity is the maximum degree again (the cap at 100 is gone). Indices run 1..granularity in every interval method (0 for isolated nodes); the C++ 3.0.1 ran 2..granularity+1 in equalIntervalSize and equalNodesPerInterval because of a duplicated 0.0 boundary. equalLogIntervalSize starts from the smallest positive strength instead of dividing by zero. The remaining strength of each node is kept incrementally while peeling (the C++ re-summed it on every re-binning), and negative weights are refused with a clear error.
• New --maximum-strength (config maximum_strength, the C++ -maximumStrength) and --strength-intervals custom with --strength-intervals-file (config strength_intervals_file, the C++ 4.0.0 -strengthsIntervalsFile) (#20).
• Invalid configuration (from the YAML file or the flags), an unusable --strength-intervals-file and inputs the decomposition refuses (negative weights, --decomp dcores on an undirected graph) are reported as usage errors instead of tracebacks.
• K-dense decomposition (#19) now implements the C++ triangle-pair peeling (the k-truss decomposition): every edge starts with its triangle count and removing an edge lowers the count of the two other sides of each triangle it closed. The previous code took a plain vertex k-core of the edge/triangle dual graph, which over-estimates the index of edges whose triangles share a side, and its triangle enumeration assumed sorted adjacency, so it silently missed triangles on real edge lists. Parallel edges and self-loops are ignored. Per-edge indices are exposed as result.metadata["edge_indices"] (JSON export writes them as [u, v, index] triples). Verified against a brute-force k-truss on random graphs. The CAIDA 20251001 example now spans k-dense 2–105 (was 2–55); its image is regenerated.
• The color legend is titled after the decomposition (k-dense, d-core) instead of always k-core (#10).
• --directed --decomp dcores crashed with a Pydantic ValidationError while building components (Component was constructed with id/index instead of component_id/ shell_index). The d-core path through Network.decompose() and the CLI works again and is now covered by tests (#21).
• Edge-list reader (#22): columns may be separated by any whitespace (tabs and repeated spaces used to abort with invalid literal for int()); an unused third column is ignored when the graph is not weighted; --weighted on a two-column file uses weight 1.0 instead of NaN; self-loops are dropped with a warning instead of crashing the decomposition; non-integer node ids and single-column files give a clear error. All of this matches the C++ reader.
• --multigraph no longer crashes k-cores: parallel edges count towards the degree (the C++ behavior) and, for weighted multigraphs, their weights are summed into the strength.
• write_edge_list (and therefore lanet-vi generate) now defaults to space-separated output; its files were tab-separated and could not be read back by lanet-vi visualize.
• Node names may contain spaces (the name is the rest of the line, quotes stripped) and the colors file accepts tabs.
• lanet-vi config wrote YAML with !!python/object/apply tags that --config could not load; enums are now written as plain strings (#23).
• --config no longer discards the other command-line flags. Precedence is now the C++ one: defaults < YAML file < flags given explicitly on the command line (#23).
• --names now draws the labels (it loaded them and drew nothing); use --no-node-labels to load names without drawing them. A node named 0 is no longer skipped (#23).
• --show-degree-scale toggled the color legend instead of the degree legend. It now controls the degree (node size) legend, as the C++ -showDegreeScale; the color legend has its own --show-color-legend/--no-show-color-legend (config show_color_legend). show_size_legend is kept as a deprecated alias that folds into show_degree_scale (#23).
• Default-true boolean flags (--show-degree-scale, --gradient-edges, --color-by-community, --draw-community-boundaries) now have --no-... forms (#23).
• Options that the current implementation does not use (--from-layer, --use-spiral-layout and the --spiral-* settings, --detect-communities and the other community flags) say so in their help text, with the tracking issue (#18, #23). (--delta, --gamma, --alpha, --beta and --coord-distribution, inert at the time, act since the layout ports below.)
• write_decomposition_json crashed when components were present (it read non-existent Component.id / Component.index attributes).
• lanet_vi.community.base failed to import on Python 3.9 because of X | None return annotations evaluated at runtime.
• get_community_colors returned RGBA quadruples and used the removed matplotlib.cm.get_cmap API; it now returns RGB triples via pyplot.get_cmap.
• --community-algorithm now rejects unknown values with a usage error instead of a Pydantic validation traceback.
• D-core decomposition (--directed --decomp dcores) crashed because DecompositionResult had no metadata field; the field now exists and holds the (k_in, k_out) pairs. min_index is now derived from the data instead of hard-coded to 1.
• lanet_vi.community.base used X | None annotations that fail to import on Python 3.9.
• lanet_vi.__version__ reported 4.0.0; it now reads the installed package version (5.0.0) from package metadata, so pyproject.toml is the single source of truth.

Changed

• The degree legend stacks its rows so that a sample never overlaps the previous one (needed in the pow / log modes, whose radii reach a fifth of the unit disc) and keeps the C++ text size instead of growing with the row pitch; unchanged on large classic pictures, where the C++ pitch is the larger.
• Layout (#18): the actual LaNet-vi placement is back. visualization/lanet_layout.py ports kcores_component.cpp / graph_kcores_components.cpp (classic mode): nested components (connected pieces of the inner core, recursively) with their own center, radius and scale (formulas (3)-(5) of NIPS 2005), rings one unit apart with the top core as a disc whose radius follows its log-degrees, node radius from the depth of its higher-index neighbors (formula (1)), angle from the circular average of their angles, top cores split into cliques laid along U-shaped paths in angular sectors, and the --no-cliques sector formula (2). The previous code placed every shell on a fixed 80-unit ring at a random radius and measured neighbor angles from the origin. K-dense and d-core results use the same classic placement, building the component tree with their own edge index (the edge dense index, as kdenses_component.cpp walks it); the rest of that file's variant (>= neighbors, tau, sibling circle packing, ratioConstant radii) belongs to the C++ pow/log mode, ported afterwards (see "Added" above). epsilon, delta, gamma, unit_length, seed, --no-cliques and --draw-circles now do what the C++ flags did; epsilon defaults to the C++ 0.18 again. The picture is framed as the C++ viewport (1.6 x 1.2 times the network radius), leaving the margin the legends sit in.
• Node radii follow the C++ computeHostRatio (0.4 (log(1+d)/log(dmax))^0.7 layout units, strength-based for weighted graphs) and are drawn in layout units for graphs of any size (an EllipseCollection replaces the point-sized scatter), floored at one pixel so peripheral nodes stay visible. node_size_scale is a plain multiplier and defaults to 1.0.
• CAIDA examples, README and docs now use the 20251001 AS-relationships snapshot (78,370 ASes, 489,407 relationships, k-cores 1-149); example images regenerated.

Infrastructure

• main is protected by a GitHub ruleset: changes land through pull requests with green CI and an automatically requested Copilot code review; force-pushes and deletions are blocked.
• mypy is now a blocking CI check (disallow_untyped_defs); the whole package type-checks cleanly. Pydantic models use ConfigDict instead of the deprecated inner Config class.
• CI split into lint (ruff check, ruff format, mypy), a test matrix on Python 3.9–3.13 (Ubuntu) plus macOS 3.12, and build (uv build + twine check). Coverage is enforced with a minimum threshold and uploaded as an artifact.
• Ruff formatter adopted; UP (pyupgrade) and B (bugbear) rule sets enabled and the codebase migrated to PEP 585 built-in generics.
• pre-commit hooks (ruff lint/format, file hygiene) via .pre-commit-config.yaml.
• Dependabot for GitHub Actions and Python (uv) dependencies.
• PyPI releases use trusted publishing (OIDC) through the pypi environment instead of a long-lived API token.
• Added CONTRIBUTING.md, SECURITY.md, CODE_OF_CONDUCT.md, CODEOWNERS, issue and pull request templates, and AGENTS.md (agent/contributor instructions; CLAUDE.md now includes it).
• New tests for the CLI, rendering pipeline, d-cores, writers and version metadata.
• Python 3.13 added to the supported versions.

5.0.0 - 2025-10-18

Overview

Note (September 2026): the parity claims in this entry were later found to be overstated. Several features listed as ported are missing, inert or behave differently; see issues #18–#26 for the current state.

Complete Python refactor of LaNet-vi 3.x (C++) with all legacy features included. This version brings all functionality from the latest C++ codebase (previously in legacy/Source/) into a modern Python implementation using NetworkX, pandas, and matplotlib.

Core Features (Ported from C++ LaNet-vi 3.x)

• K-core decomposition: Classic k-core algorithm using NetworkX
• K-dense decomposition: Triangle-based decomposition (m-cores)
• D-core decomposition: Directed graphs with (k_in, k_out) pairs per node
• Not a port of legacy/Source/graph_dcores.cpp, which computed the per-l (k, l)-core table and drew nothing (see --dcore-table under Unreleased); the independent in/out cores are the older graph_dcores_old.cpp (#21)
• CLI: --directed --decomp dcores
• Spiral/semicircular layout: Mathematical spiral placement using Newton-Raphson solver
• Inspired by legacy/Source/espiral.cpp, which was never linked into a C++ release; the option is accepted but inert (#18)
• CLI: --use-spiral-layout
• Community detection: Louvain and greedy modularity algorithms
• A NetworkX-based replacement, not a port: the C++ community code (solution_lanci*.cpp, solution_submodular.cpp) was research code never linked into a LaNet-vi binary, and its local-growth and submodular methods are not in Python (#26)
• CLI: --detect-communities
• Community visualization: Color-coded nodes with boundary overlays
• Random graph generation: Testing and benchmarking utilities
• Erdős-Rényi, Barabási-Albert, Watts-Strogatz, Powerlaw cluster
• NetworkX wrappers; the C++ erdos_renyi.cpp was a 12-line unlinked G(n, p) loop (#26)
• CLI: lanet-vi generate

New Python-Specific Features

• Type-safe configuration: Pydantic models for all settings
• Modern CLI: Typer + Rich with --double-hyphen flags (Unix/GNU standard)
• Comprehensive logging: DEBUG and INFO levels (--verbose, --quiet, --log-file)
• Enhanced JSON export: D3.js-compatible graph exports with full metadata
• Spatial indexing: KD-tree based indexing for O(log N) queries on large graphs
• Information theory metrics: MI, NMI, ARI, VI for partition comparison
• Pandas integration: Efficient data management for large networks
• YAML configuration: Easy-to-edit config files

Architecture Improvements

• NetworkX integration: Leverages battle-tested graph algorithms
• Modular design: Clean separation of concerns (io, decomposition, visualization, metrics)
• Better performance: NumPy vectorization, spatial indexing, optimized layouts
• No external renderers: Pure Python/Matplotlib (removed POV-Ray dependency)
• Comprehensive testing: Type checking with mypy, linting with ruff

Dependencies

Core libraries: - networkx>=3.0 - Graph algorithms - pandas>=2.0 - Data management - matplotlib>=3.7 - Visualization - scipy>=1.10 - Spatial indexing, convex hulls - scikit-learn>=1.3 - Clustering metrics - pydantic>=2.0 - Configuration validation - typer>=0.9 + rich>=13.0 - CLI interface

Usage Examples

# Basic k-core visualization
lanet-vi visualize --input network.txt --output viz.png

# D-core decomposition on directed networks
lanet-vi visualize --input network.txt --directed --decomp dcores --output dcores.png

# Community detection
lanet-vi visualize --input network.txt --detect-communities --output communities.png

# Spiral layout
lanet-vi visualize --input network.txt --use-spiral-layout --output spiral.png

# Generate random graph for testing
lanet-vi generate --output test.txt --model barabasi-albert --nodes 1000 --edges 3

# Verbose logging
lanet-vi visualize --input network.txt --verbose --log-file debug.log

---

Migration from C++ LaNet-vi 3.x

LaNet-vi 5.x is a complete Python rewrite of the C++ version. Parity was overstated in 5.0.0 and restored piece by piece in 5.1.0 (see the 5.1.0 entry and issues #18–#26):

Key Differences: - Language: C++ → Python 3.10+ - Rendering: POV-Ray / SVG → Matplotlib - Configuration: Custom format → YAML - CLI: Single-hyphen → Double-hyphen flags (Unix/GNU standard) - Dependencies: No external renderers, pure Python stack

Feature Parity (as of 5.2.0): - ✅ K-core decomposition, including the weighted peeling - ✅ K-dense (m-core) decomposition (triangle-pair peeling) - ✅ D-core decomposition (directed graphs), with the (k, l)-core table of the 4.0.0 tool - ✅ The LaNet-vi placement: classic, pow and log coordinate distributions - ✅ Color scale, grayscale, legends, edge sampling, -window, -fromlayer, -names - ✅ K-connectivity (-kconn, -kconntype) - ✅ SVG and PDF output (matplotlib, by the --output extension; not the C++ SVG writer) - ⚠️ Community detection: NetworkX Louvain / greedy modularity drawn on the picture; a 5.x addition, the C++ had no community rendering (#26) - ⚠️ Random graph generation: NetworkX wrappers (#26) - ❌ Not ported: Gomory-Hu connectivity and POV-Ray scenes (won't-fix, #25). The spiral layout of the development tree was never in a C++ release and was removed. - ➕ Enhanced JSON exports - ➕ Information theory metrics - ➕ Type-safe configuration

---