When Peers Decide Who Trains: Norm-Driven Participation Feedback in Federated Learning

Federated learning trains one model across many devices or institutions without collecting their data: a server sends out the current model, each participant improves it on local data, and the server averages the results. Nearly all of this work assumes that who shows up in a given round is decided by hardware, such as a phone being charged or on Wi-Fi, or by payment. People and hospitals do not decide that way. Decades of technology-adoption research find that people join because a trusted authority endorsed the tool or because peers like them already use it. This is the "subjective norm," a person's sense of what the people who matter to them expect. Peers who influence each other also tend to hold similar data. So participation arrives in socially clustered waves, and the model is shaped first by whichever community adopted early.

We propose that this creates a feedback loop the field has not studied. Adoption driven by social norms over-represents early-adopting communities. The shared model then serves later, dissimilar communities poorly. Their disappointing experience makes them leave, and their departure deepens the bias. We will build a participation simulator whose social effect sizes are taken from published adoption studies. We will test whether this loop appears on standard federated benchmarks and whether existing debiasing methods break it. We will also propose a correction that estimates co-participating groups only from the server's own participation logs and protects the groups most likely to leave.

Core Claim

Correlated participation has already been shown to bias federated averaging. Debiasing Federated Learning with Correlated Client Participation models availability as an exogenous Markov chain with a fixed lockout of R rounds. It fixes the bias by having each client scale its learning rate by the inverse of its own sampling frequency. On economic incentives, Social Welfare Maximization for Federated Learning with Network Effects treats clients as utility maximizers who choose to participate, purchase, or abstain, and it solves for a static equilibrium.

We claim that neither framing covers the regime that adoption research describes, where participation is caused by other clients' participation and by the model's own performance. There are two reasons. First, a per-client frequency counter reweights clients who are present. It cannot give weight back to clients who have left, and it does not see that a whole group participated in the same burst. Second, a static equilibrium assumes perceived and realized benefit are the same thing. The adoption literature separates the drivers of joining from the drivers of staying (A Research Model of Consumers' Adoption and Continued Use of the PCEHR System).

The new claims are the following. The endogenous loop is measurable and harmful. It depends on homophily, meaning the tendency of socially linked clients to hold similar data. It can be corrected by reweighting at the level of co-participating groups and by accounting for the risk that a group leaves. The correction does not require the social graph.

Research Program

The first component is a participation simulator grounded in behavioral evidence. Each client $i$ decides in each round $t$ whether to participate according to

$$\text{logit}\, P(z_{i,t}=1) = a_i + \beta_v e_i + \beta_h \, s_{i,t} + \beta_u \, g_{i,t},$$

where $z_{i,t}$ indicates participation, $a_i$ is a baseline propensity, and $e_i$ indicates endorsement by an authority. This is the vertical channel, standing for a supervisor or clinician recommendation. The term $s_{i,t}$ is the fraction of $i$'s social neighbors who participated recently, the horizontal channel. The term $g_{i,t}$ is the realized gain of the current global model over local-only training on $i$'s data, the continued-use channel.

The vertical coefficient is not tuned by hand. We set it from reported odds ratios, taking log-odds between $\log 3.47 \approx 1.24$ and $\log 5.15 \approx 1.64$. These values come from the effect of supervisory priority on clinical guideline behavior in Factors influencing tobacco use treatment patterns among Vietnamese health care providers. We also use $\log 4.257 \approx 1.45$, the effect of provider advice on mHealth adoption in Willingness to Adopt mHealth Among Chinese Parents During the COVID-19 Outbreak.

These studies play a motivating role. They are cross-sectional surveys, a limitation the file flags, so the coefficients are plausible magnitudes and not causal estimates. For that reason we sweep $\beta_h$ and $\beta_u$ over wide ranges instead of claiming a single calibrated value. Xiaoyu Cheng reports that normative weight spikes under crisis, which motivates a regime-shift condition in which $\beta_h$ doubles partway through training. Beyond Virtual Bazaar describes adoption through same-background peers, and Reza Hadi Mogavi et al. find that older adults are persuaded by peers at the same life stage. Together these support making social ties homophilous. Both findings are culturally specific, so we treat homophily strength as a swept variable.

The second component is the correction. The server observes only participation timestamps, which it already has. From the co-participation matrix of the recent window it infers groups of clients whose participation moves together, using a Hawkes-style (self-exciting) contagion fit followed by spectral clustering. It then weights each update by

$$w_{i,t} \propto \frac{1}{\hat\pi_{c(i),t}\,|c(i)|}\,\big(1 + \lambda\,\hat r_{c(i),t}\big).$$

Here $c(i)$ is $i$'s inferred group and $|c(i)|$ its size. The term $\hat\pi_{c,t}$ is the group's time-decayed participation rate, so a burst from one group counts once rather than many times. The term $\hat r_{c,t}$ is the group's predicted departure risk, estimated from its falling participation and the realized-gain signal that clients report as a single scalar. The constant $\lambda$ trades average accuracy against retention.

This design reuses the inverse-frequency idea from Debiasing Federated Learning with Correlated Client Participation. It lifts that idea from individual clients to groups and makes it non-stationary. Unlike the leave-one-out influence of Influence-oriented Personalized Federated Learning or the Hessian-based cascades of DICE, it needs no extra model evaluations. Those methods define influence purely through gradients and assume a fixed topology. We use their notion only as a baseline.

Baselines

The naive reference is plain federated averaging under the same simulated participation. It shows how large the feedback bias is when nothing is done about it. A second reference is federated averaging with participation drawn uniformly at random, at the same average rate. It isolates the harm that comes from social correlation rather than from low participation.

The strongest existing competitor for correcting participation bias is Debiasing FedAvg from Debiasing Federated Learning with Correlated Client Participation. That paper reports 73.8% CIFAR-10 test accuracy under minimum separation R = 5, against 67.6% for standard FedAvg and 68.0% for FedVARP. It is the right competitor because it is the only method in the file with a guarantee under correlated participation. We will first reproduce its CIFAR-10 R = 5 setting to check our implementation against those numbers. We then run it in our simulator with the same model, rounds, and learning-rate search budget as our method. Our method beats it only if it achieves higher accuracy on the whole eventual population and a smaller early-versus-late group gap, with neither degrading, averaged over five seeds and across the homophily sweep.

The strongest personalization competitor is FedC²I from Influence-oriented Personalized Federated Learning, which reports 85.50% test accuracy on Digit-5 against 82.88% for FedAvg. It uses the same tuning budget but is allowed its extra leave-one-out computation. We report that compute separately.

The economic competitor is the participation subsidy from SEMTS in Social Welfare Maximization for Federated Learning with Network Effects, which reports welfare gains of 148.86% on MNIST and 300% on CIFAR-10. It enters our simulator as a transfer that raises $a_i$ for subsidized clients, under a budget equal to what our method spends in retention weight. There is no reported result on Lacuna for any of these methods on LEAF Reddit, FEMNIST, or Sentiment140 under socially correlated participation, so those comparisons are new.

Three ablations isolate the key ingredients. The first uses group reweighting without the departure-risk term. The second gives the method the true social graph instead of inferred groups. The third uses per-client rather than per-group reweighting while keeping the risk term.

Why not just…?

Why not just use Debiasing FedAvg, since inverse frequency corrects bias whatever its cause? This is the most serious threat. When participation is stationary and no one leaves permanently, inverse-frequency weighting may well suffice. The failure we predict comes from departures, where a counter cannot weight a client who never returns, and from within-window group bursts. Our homophily and $\beta_u$ sweeps settle the question directly. If Debiasing FedAvg matches our method whenever $\beta_u = 0$, that confirms our account. If it also matches when $\beta_u > 0$, the correction is unnecessary, and we would report the loop as a diagnostic finding only.

Why not just personalize with clustered or influence-based aggregation, so each community gets its own model? Personalization could weaken the loop, because a late group's realized gain depends less on the shared model. But late groups are small when they arrive, and the Digit-5 gains of FedC²I come with a leave-one-out cost that grows with the number of clients. We therefore compare retention and late-group accuracy at matched compute. We also test combining personalization with our weighting, since the two may be complementary rather than competing.

Why not just pay the groups at risk of leaving, as SEMTS does? Money acts on the baseline propensity, but it does not change the realized-gain term that drives departure. The PCEHR model separates perceived from realized benefit for exactly this reason. The matched-budget comparison will show whether subsidies retain clients who then still receive a poor model. We will measure that outcome as retained-but-underserved clients.

Proposed Experiments

The primary testbed is the Reddit dataset from LEAF. Clients are users, and a social tie connects users who share a subreddit, which gives natural homophily without synthetic graphs. We add FEMNIST and Sentiment140, building ties by k-nearest-neighbor matching on label histograms at controlled homophily levels. Following LEAF's own emphasis that the 25th percentile of accuracy can drop dramatically for data-poor users, we report the 10th, 25th, and 50th percentile of accuracy over all clients who would ever join. We also report the early-versus-late group accuracy gap, the attrition rate by group, and the bias relative to the uniform-participation reference.

The cross-silo check uses FLamby Fed-ISIC2019, with six acquisition systems, and Fed-Heart-Disease, with four countries. Endorsement comes from one "prestige" silo. FLamby reports about 65% balanced accuracy for pooled training, 50–55% for federated strategies on Fed-ISIC2019, and as low as 35% for some local-only clients. We ask whether silos whose realized gain is small leave under norm-driven entry.

The decisive checks are three. First, under plain averaging, the early-versus-late gap and the attrition of late groups must grow with both homophily and $\beta_u$. There must be a null effect at zero homophily, where social correlation carries no data signal. Without that pattern, the loop is not real. Second, our method must beat Debiasing FedAvg specifically in the $\beta_u > 0$, high-homophily cells. Third, groups inferred from logs must recover most of the benefit of the true graph. If they do not, the method needs social data that servers rarely hold.