Federated load event recognition stops sharing semantics unconditionally: FedReST edges an architecture-matched FedAvg by 0.00524 Event F1 on REFIT 2014 at 10% labels
Synopsis
Federated non-intrusive load monitoring lets households collaborate without centralizing raw power traces, yet sparse local labels and household-specific operating patterns make cross-client semantic evidence unevenly reliable. FedReST splits collaboration into two admissions: the server uses reliability-gated partial optimal transport to accept only semantic mass that both sides support, and each receiving client filters the resulting update through function-preserving gradient projection so locally mastered states are not disturbed. On frozen REFIT 2014 rolling folds with 10% visible labels and matched two-epoch local budgets, corrected-core FedReST shows paired differences of +0.00524 in Event F1 and +0.00104 in Window F1 against an architecture-matched Head+BN FedAvg reference.
Interpretation
Federated load event recognition can now decline to share part of its semantic evidence instead of forcing every client's prototypes into one global correspondence. Parameter averaging previously treated local updates as commensurate estimates of a shared objective, prototype federation averaged prototypes unconditionally, and standard balanced optimal transport required all mass to participate in a match. On frozen REFIT 2014 rolling folds with 10% visible labels and matched two-epoch local budgets, corrected-core FedReST shows paired differences of +0.00524 in Event F1 and +0.00104 in Window F1 against the architecture-matched Head+BN FedAvg reference, with higher Event F1 in seven of nine fold-seed pairs.
Sharing is judged twice, once by the server on whether evidence is supported and once by the receiving client on whether the direction is compatible with its local function. Masked reconstruction and state-aware contrastive learning first build temporal representations that are mapped to simplex-valued semantic distributions, and each client uploads evidence-weighted task-state prototype measures. The server aligns these measures with reliability-gated partial optimal transport, leaving unsupported mass unmatched; the receiver applies function-preserving gradient projection onto a low-rank sensitive subspace derived from anchor activations, filtering only the admitted semantic residual while the local objective update stays unprojected.
Perspective
The demonstrated setting is the frozen REFIT 2014 rolling protocol over houses 2, 3, 5, 7, 15 and 20, four tasks (refrigeration, washing machine, dishwasher, tumble dryer), 10% visible labels and two local epochs, plus post-freeze robustness audits on REFIT 2015 and UK-DALE; the authors propose domain-adaptive reliability estimation as the next extension.
The descriptive 95% interval for the Event F1 paired difference is [−0.00780, 0.01828], and the authors note that repeated seeds within overlapping rolling folds are not independent population samples; Event F1 changes are negative at 1% and 5% visible labels, and the UK-DALE audit shows negative Event F1 and Window F1 changes, which the authors cite in proposing domain-adaptive reliability estimation.
