Implement InducedSubgraphView

[IanChenUIUC]

Implementation for issue #39

[IanChenUIUC]

Remaining items to implement are:

• option for compact IDs (so that it provides the same options as the GraphTools::SubgraphFromNodes)
• a few trailing test cases, namely the edge iterator and checking constructors
• adding cython bindings for InducedSubgraphView

Something to note is that I had to make the hasNode and hasEdge functions virtual, so that my subclass could override them. This may have some performance costs for the rest of the codebase, but I have not benchmarked anything.

I think we should save the base GraphView class, which InducedGraphView and CoarsenedGraphView inherit from, for a different PR, as this one is already becoming large.

[IanChenUIUC]

I'm looking at the GraphView interface that Arun proposed above, and I think I like the idea to not compute the degrees eagerly, and instead have the user explicitly call an invalidateCache (e.g. a synchronization function). Any other ideas?

[adsharma]

Or do it on the first request to access the degrees

[IanChenUIUC]

Or do it on the first request to access the degrees

Sounds good. I'll start working on adding that behavior.

[adsharma]

This keeps m from double-counting, but makes the public in-degree wrong for directed self-loops. Fixing this likely needs separate edge-count delta logic on removal too, since networkit/cpp/graph/InducedSubgraphView.cpp:119 currently subtract out-degree plus cached in-degree.

[IanChenUIUC]

I'm a bit confused -- I thought a directed self-loop from v to v would contribute +1 to both inDegree[v] and outDegree[v]. That contribution happens in the above loop, where we iterate over out-neighbors.

[adsharma]

performance: Full edge traversal of a small view scans the entire base graph. networkit/cpp/graph/InducedSubgraphView.cpp:198 and networkit/cpp/graph/InducedSubgraphView.cpp:225 call originalGraph.forEdges/parallelSumForEdges and filter with hasEdge, making view traversal O(E_base) plus set lookups. For induced views, this should iterate nodeSubset and filter each selected node’s neighbors, with undirected deduplication.

Also as an integration test, please run the abm14 graph and see if any runtime regressed due to the virtual keyword for some of the methods made them expensive. Another area to watch out is that some of the algorithms use virtual methods and at other times templated methods to access the graph.

[adsharma]

Details on the virtual vs templated. Only one problematic instance was found in the review.

The nuance is that C++ templates are only “polymorphic” through the operations they call inside the instantiated base template. For example, Graph::forNodes is a template, so if code has:

const Graph &g = view;
g.forNodes(...)

it instantiates/runs Graph::forNodes, not InducedSubgraphView::forNodes. The branch partially mitigates that by changing many internal checks from exists[v] to virtual hasNode(v), so those base-template paths become correct but potentially slow.

[IanChenUIUC]

Closing to defer until the refactor (#69) is implemented.