Home / Research
One arc, four stages — from making overlapping community structure reliable, to making it scale, to deciding which structure is worth observing at all, to making structure inspectable.
Each project below asks one version of the same question: which structure is load-bearing, and how would you know? Triad answered it by optimizing all of the structure at once. Highway answered it by discarding most of it and keeping a backbone. Scout asks it before the structure is even visible, when looking costs. DSEG asks it of a model's internal representation.
Structural reliability
Overlapping community detection lets a node belong to several communities at once. Most existing methods are heuristics that lean on only part of the available structural signal, and two failure modes recur: mass collapse, where communities degenerate into empty or collapsed sets, and uniform fuzziness, where every membership is smeared into the same vague value.
Triad casts the problem as a single quadratic constrained program over continuous memberships — using all three structural dimensions (node, edge, community) and explicitly penalizing degeneracy in each, instead of patching heuristics together.
The limit it hitThe reliability was bought at a price that is intrinsic, not incidental: a quadratic constrained program does not scale to large networks. So the question had to change — not "how do I solve this faster," but "how much of this structure do I actually need?"
Scalability through sparse structure
Highway takes the reframed question seriously. Instead of solving the exact program faster, it asks how much structure overlapping-community inference really requires — and answers with a four-step method built on a deliberately sparse backbone: backbone construction, anchor-based initialization, neighbor-only propagation, and anchor-preserving pattern calibration.
What it opened upThe backbone is not only a speed-up — it is a hypothesis about which structure carries the inference. But that hypothesis assumes the graph is in front of you to begin with. The next stage removes the assumption: when the graph itself is stale and looking costs, which structure is even worth observing?
What is worth observing
Triad and Highway both assume the graph is there to be read. Scout drops that assumption. On an evolving network the maintained graph is always somewhat stale, and refreshing it costs observation and computation — so under a limited budget only part of it can be brought up to date, and the model has to decide which stale information to refresh.
The usual answer optimizes freshness. Scout's answer is task utility: it learns the value of querying each node from the maintained graph and the observation history, then spends the budget where the answer is worth the most downstream — because the part of the graph that changed the most is not necessarily the part the task needs.
What it opened upAcross all three, the operative claim is the same — this part of the structure carries the inference, that part does not. Making such a claim testable rather than merely asserted means being able to look at the structure and check it. The fourth stage asks that of a model's internal representation.
Structure you can read
The last stage turned the question into a model. In DSEG-Char every character is a node in a 5,000-node graph, and writing a line of verse is a walk over that graph — the semantic state selects the next node, the node rewrites the state, and the walk runs to the end of the line.
What makes it readable is that the three terms choosing the next character — state, edge and context — are kept apart instead of summed into one opaque activation. So every character comes with the term that actually pushed it up, and a browser visualiser colours the whole walk by that attribution. Interpretability here is a property of how the model is built, not a probe run afterwards.
Four stages, one arc: structure made reliable → structure made efficient → structure worth acquiring → structure made readable.
As artificial intelligence advances rapidly, both scalability and interpretability are becoming increasingly important. What I hope to investigate is how a model's decisions can be made easier to analyze, test, and trace — without giving up efficiency or the ability to scale — and how every piece of data can be put to fuller and more efficient use.