Algorithms that track groups in a changing network, such as friend circles in a social graph or co-active neuron modules in a brain recording, are usually judged by one similarity score against ground-truth groups and one wall-clock time on a server. Behavioral neuroscience learned long ago that a single readout misleads. A mouse that stops swimming may be despairing or just tired, so the field runs batteries of paired assays, where each test has a control that rules out one alternative explanation. We propose to evaluate dynamic community trackers the same way. One assay asks whether a tracker reports group births, merges, and splits promptly. A paired control asks whether it invents such events when only random churn occurs. A long-horizon assay measures drift. A cost assay counts processor cycles per graph update on a lean in-order RISC-V core and splits them into useful work and pipeline stalls, which are clock cycles where the processor sits idle waiting on data or memory.
The bet is that the single-score protocol hides two failure modes: inertia scoring as accuracy, and jitter scoring as responsiveness. A battery with controls will reorder which trackers look best, and the reordered ranking will better predict performance on held-out real tasks. The cost assay targets a deployment regime the field ignores. Embedded hardware next to a home cage or a neural recording rig cannot hide the irregular memory access of incremental graph updates. We test whether a cheap software change, batching update events so that dependent ones are not issued back to back, removes a measurable share of those stalls.
Core Claim
Dynamic community detection has a measurement problem that resembles a confound, not a modeling problem. Running a static clustering method independently on each time slice is known to produce arbitrary label swaps after tiny perturbations (Tracking evolving communities in large linked networks; Detecting communities and their evolutions in dynamic social networks - a Bayesian approach). The standard fixes all add inertia. One is a transition cost that penalizes nodes for switching groups, as in SwordPlots. Another is local force-balance moves that only fire when a neighbor's pull dominates, as in Quick Community Adaptation, QCA (Dynamic Social Community Detection and Its Applications). A third is cross-batch centroid calibration, as in Deep Temporal Graph Clustering.
Inertia raises agreement with ground truth whenever ground truth changes slowly, whether or not the tracker detects the changes that matter. This is the forced-swim confound in graph form: a high score can mean good tracking or merely a reluctance to move. The Temporal Graph Benchmark motivates the concern. It found that persistence-style heuristics beat a learned temporal model on three of four node-affinity datasets, and that the ordering inverts on "high-surprise" graphs where test interactions were unseen in training. Our claim is that a confound-controlled battery separates these regimes within a single evaluation. It should yield rankings that differ from single-score rankings, and those rankings should predict held-out task success better.
How the Directions Fit
The three directions carry unequal weight. Dynamic community detection is the object of study. Comprehensive behavioral assessment supplies the method. We borrow the logic of paired assays from work that pairs forced swimming with spontaneous locomotion to separate motivation from motor capacity (Possible Therapeutic Effect of T-794). We also borrow from broad standardized batteries like SHIRPA, which map one mutation's effects across many dimensions (Analysis of the mouse mutant Cloth-ears).
Pipeline stall overhead is the cost axis and the deployment constraint. Lean in-order RISC-V cores have no hardware that reorders instructions around a hazard, so their throughput depends directly on data dependencies and memory latency. This is documented in Ramping Up Open-Source RISC-V Cores and Specially-Designed Out-of-Order Processor Architecture for Microcontrollers. Graph accelerators so far attack stalls only for static graphs. Graphitron detects read-after-write conflicts and splits operations for algorithms like PageRank and BFS. Sparse Stream Semantic Registers speeds up sparse linear algebra. Neither evaluates streams of node and edge insertions and deletions, or community migration.
We deliberately do not propose a custom accelerator. The graphs that matter for live behavioral or neural monitoring are often small. The functional connectome study has 189 neurons. Building hardware before measuring where the cycles go would be premature.
Research Program
The battery has four assays, each paired with a control that a trivial tracker should pass or fail in a known way.
The event assay plants lifecycle events in synthetic streams: births, merges, splits, and dissolutions, following the lifecycle vocabulary of Blind Men and the Elephant. Streams are built on weighted LFR-style graphs, the benchmark family used to test the nested stochastic block model in the connectome study. The assay scores detection recall and latency in time steps.
Its paired null-churn control replays the same stream with events removed. Edges are rewired at the same rate while preserving degrees, and the control counts how many events the tracker reports anyway. A tracker that freezes its labels gets zero false events and zero recall. A tracker that reclusters each snapshot independently gets high recall and many false events. These two serve as the "sedated" and "hyperactive" reference animals.
The drift assay runs each tracker for long horizons on a stationary planted partition and measures the decay in partition quality against a fresh global reclustering. QCA's authors flag this kind of slow drift as a limitation of local incremental moves.
The cost assay compiles each tracker in C for RV64 and runs it in cycle-accurate RTL simulation on three open cores. These are the single-issue in-order CVA6, the dual-issue in-order CVA6S+ with renaming, and the out-of-order C910, all from the Cheshire setup in Ramping Up Open-Source RISC-V Cores. The assay reports cycles per graph event, broken down into data-hazard stalls, control stalls, and memory stalls.
On top of the battery we test one intervention, called hazard-aware event batching. Consecutive update events that touch the same community create read-after-write chains on that community's internal and external edge counts. A software scheduler can interleave events that touch disjoint neighborhoods, in the spirit of Graphitron's conflict splitting but applied to dynamic updates. We only reorder events whose affected neighborhoods are disjoint. We then check empirically that the output partition is unchanged, since QCA-style updates are order-dependent in general.
Proposed Experiments
Real streams come from the datasets named in the file. These are the QCA datasets (Enron email, arXiv citations, Facebook New Orleans with over 60,000 nodes and 1.5M links) and the BenchTGC sets DBLP, Brain, Patent, and arXivCS. For the largest out-of-memory stress case we use arXivLarge. We also use the NeuroPAL C. elegans recordings, where known stimulus onset times for NaCl, 2-butanone, and 2,3-pentanedione give an external anchor for when module reorganization should occur.
The decisive check is predictive validity. We rank trackers twice, once by mean normalized mutual information (NMI) and runtime and once by the battery profile. We then ask which ranking better predicts two held-out real tasks. The first task is recovery of stimulus-window module changes in NeuroPAL. The second is the boundary-node worm-containment simulation on Facebook New Orleans that QCA used, where QCA reported a 5–10% infection reduction over prior methods. If the two rankings agree, or if the single score predicts equally well, the battery adds nothing. We would report that result.
The cost hypothesis is that per-event updates on CVA6 are dominated by memory and data-hazard stalls rather than arithmetic. We further expect the server-measured speed advantage of incremental tracking over windowed reclustering to change size on in-order cores. If the cycle breakdown and rankings are the same on CVA6 and C910, the hardware axis can be dropped from the battery.
Baselines
The naive reference points are the two control trackers. The "frozen" tracker keeps its first partition forever. The "independent snapshot" tracker runs Louvain per time window and matches communities across windows by Jaccard overlap, the two-stage design surveyed in A Comparative Study of Different Approaches for Tracking Communities in Evolving Social Networks. Every real tracker must beat the frozen tracker on event recall and beat the snapshot tracker on null-churn false events. A tracker that fails either test is not tracking in any useful sense, whatever its NMI.
The strongest incremental competitor is QCA, chosen because it targets exactly this setting with per-event local moves that are guaranteed to increase modularity. QCA reports 3 to 4.5 seconds per dynamic update on Facebook New Orleans. We found no reported NMI value on Lacuna for QCA on these datasets, only the qualitative statement that it was superior. We therefore re-run it ourselves under identical event streams.
The strongest learned competitor is temporal graph clustering with cross-batch calibration from Deep Temporal Graph Clustering. It reports a 42.15% Adjusted Rand Index improvement on arXivCS and runs within 10GB of GPU memory on arXivLarge, where static methods ran out of memory. Because it needs a GPU, it enters the cost assay only on the server side, and its absence from the embedded column is itself a result.
We also include the multilayer transition-cost method from SwordPlots and DyPerm, which maximizes permanence rather than modularity. On NeuroPAL we add the offline nested stochastic block model pipeline from the connectome study as an upper reference. It outperformed Louvain and Combo in 7 of 9 weighted-LFR categories and produced the experimentally validated AWB gustation prediction.
For fairness, all methods consume the same event streams and window boundaries. Each method gets the same tuning budget for its inertia or smoothing parameter, a fixed number of configurations selected on separate synthetic streams that are never used for reported numbers. Everything runs single-threaded in the cost assay. The key ablation removes hazard-aware batching from an otherwise identical QCA build. The batching beats plain QCA only if it lowers cycles per event on CVA6 while leaving the output partition identical, or within seed-to-seed NMI variation where exact identity is not achievable.
Why not just…?
Why not just report NMI and runtime on a server, as everyone does? That protocol is exactly what we suspect is confounded. TGB already shows that a heuristic can lead or trail depending on how surprising the graph is (Temporal Graph Benchmark), which means a pooled score blends different regimes. The predictive-validity experiment settles this directly: if the NMI ranking predicts held-out task success as well as the battery does, the conventional protocol wins and we will say so.
Why not just use many metrics on planted-event benchmarks, without the paired controls? This is the most serious threat, because a battery could collapse into "more metrics." The paired controls are what make the battery more than that. Recall without the null-churn false-event rate rewards jitter. Stability without event recall rewards freezing. We will report the multi-metric-without-controls ranking as its own condition. If it predicts the held-out tasks as well as the full battery, the controls are unnecessary.
Why not just profile on an x86 or ARM out-of-order machine with hardware counters? For server deployment that would suffice. Out-of-order cores hide much of the hazard cost that in-order embedded cores pay, though: C910 gains 119.5% IPC over CVA6 at much larger area, while CVA6S+ gains 34.4% with 6% area (Ramping Up Open-Source RISC-V Cores). Including C910 in the cost assay tests this objection directly. If cost rankings are the same across all three cores, out-of-order profiling is enough and the in-order column is dropped.
Why not just build a Graphitron-style accelerator for dynamic updates? Graphitron's reported speedups are for static graphs. We do not yet know whether dynamic community updates are stall-bound at all at the graph sizes relevant to behavioral monitoring. The cost assay answers that question first, and only a large measured stall share would justify hardware.
References
- Tracking evolving communities in large linked networks
- Detecting communities and their evolutions in dynamic social networks - a Bayesian approach
- SwordPlots: Exploring Neuron Behavior within Dynamic Communities of Brain Networks
- Dynamic Social Community Detection and Its Applications
- Deep Temporal Graph Clustering: A Comprehensive Benchmark and Datasets
- Temporal Graph Benchmark for Machine Learning on Temporal Graphs
- Possible Therapeutic Effect of T-794, a Novel Reversible Inhibitor of Monoamine Oxidase-A, on Post-Stroke Emotional Disturbances, Assessed in Animal Models of Depression.
- Analysis of the mouse mutant Cloth‐ears shows a role for the voltage‐gated sodium channel Scn8a in peripheral neural hearing loss
- Ramping Up Open-Source RISC-V Cores: Assessing the Energy Efficiency of Superscalar, Out-of-Order Execution
- Specially-Designed Out-of-Order Processor Architecture for Microcontrollers
- Graphitron: A Domain Specific Language for FPGA-based Graph Processing Accelerator Generation
- Sparse Stream Semantic Registers: A Lightweight ISA Extension Accelerating General Sparse Linear Algebra
- Learning dynamic representations of the functional connectome in neurobiological networks
- Blind Men and the Elephant: Detecting Evolving Groups in Social News
- A Comparative Study of Different Approaches for Tracking Communities in Evolving Social Networks
- DyPerm: Maximizing Permanence for Dynamic Community Detection