Silence Is Not Distrust: Evidential Trust States That Decay Toward Ignorance

When two users in a trust network stop interacting, almost every dynamic trust model quietly lowers the trust score between them. Two examples are the aging factor in Beta reputation systems and the recency weighting in neighbor-based trust predictors. Both treat six months of silence as weak evidence of distrust. This note argues that this is the wrong default. Silence means we know less, not that the relationship got worse. The right model should let old evidence fade so that its prediction slides back toward what this particular observer usually believes about strangers, while reporting that it is now uncertain. That baseline is the observer's personal "trustingness," which earlier work found to be the strongest single driver of trust ratings.

We propose an evidential trust state for each directed pair. It is a small set of quantities: accumulated positive evidence, accumulated negative evidence, and an observer-specific prior. These map directly onto subjective logic, the belief calculus underlying evidential deep learning, as belief, disbelief, and "vacuity" (the share of the opinion that is pure ignorance). Evidence decays over time, and the decay drives the opinion toward vacuity rather than toward distrust. Negative evidence is allowed to arrive heavier and fade slower than positive evidence. The claim can be checked directly on timestamped signed networks such as Bitcoin-OTC, Bitcoin-Alpha, and Wiki-RfA. After a long gap, do new ratings revert toward the rater's own baseline, or do they skew negative?

Core Claim

Following Evidential Deep Learning to Quantify Classification Uncertainty, write an opinion about whether observer $u$ trusts target $v$ in terms of positive evidence $r$, negative evidence $s$, a prior weight $W$, and a base rate $a_u$. The base rate is $u$'s default expected trust in someone they know nothing about. The expected trust and the vacuity are then

$$E_{uv}(t)=\frac{r_{uv}(t)+W a_u}{r_{uv}(t)+s_{uv}(t)+W},\qquad \text{vac}_{uv}(t)=\frac{W}{r_{uv}(t)+s_{uv}(t)+W},$$

with each evidence count decaying between interactions as $r(t)=r(t_0)e^{-\lambda_+ \Delta t}$ and $s(t)=s(t_0)e^{-\lambda_- \Delta t}$. Here $\Delta t$ is the time since the last observed evidence and $\lambda_\pm$ are separate decay rates.

As the evidence fades, $E_{uv}\to a_u$ and vacuity rises to 1. Contrast this with a conventional aging rule, under which $E_{uv}$ drifts toward a fixed untrusting value.

The proposal joins three findings that each paper reached alone. Trust Propagation with Mixed-Effects Models fit per-user random intercepts on CouchSurfing triads and showed that a source's trustingness predicts ratings better than the target's reputation. We reuse that intercept as $a_u$. Improving Trust Estimation in Human-Robot Collaboration Using Beta Reputation at Fine-grained Timescales updated a Beta reputation with an aging factor and found by maximum likelihood that failures must weigh more than successes. We keep the asymmetry but change where the decay goes. Evidential Stochastic Differential Equations for Time-Aware Sequential Recommendation enforced that epistemic uncertainty rises monotonically with elapsed time in user-item recommendation. We carry that principle into directed, signed interpersonal trust. There, the competing interpretation, that silence means distrust, is the standing assumption the principle has to beat.

What is new is not "evidential plus temporal." It is a specific, falsifiable modeling choice: silence decays toward the observer's own baseline under rising vacuity. We test that choice against its alternatives on real data.

Research Program

The model has three parts, and we keep each part as simple as the evidence allows.

The first part is an evidence encoder. It is a small learned function that maps each observed rating (for example, a Bitcoin-OTC score from -10 to +10) to positive and negative evidence amounts. It conditions on the rater's calibration, since a "3" from a harsh rater means more than a "3" from a lenient one. This replaces the hand-set thresholds of classical Beta reputation.

The second part is the observer base rate $a_u$. For users with history, it is estimated by a mixed-effects intercept. For cold-start users, a lightweight evidential probe on a frozen signed-graph encoder predicts it, in the style of Evidential Uncertainty Probes for Graph Neural Networks. That probe also reports its own uncertainty, so a new user's base rate is not falsely confident.

The third part handles pairs with no direct history. Opinions are composed over paths of at most two hops using subjective-logic discounting, so uncertainty grows along the chain. We limit chains to two hops because TrustGNN reports that path length 2 is optimal and that longer chains add noise.

The decay rates $\lambda_\pm$, the prior weight $W$, and the encoder are fit by maximizing the likelihood of each next observed rating under the Beta-predictive distribution implied by the current opinion.

Before any modeling, there is an analysis step that can redirect the project. We will measure, per dataset, how often a directed pair is rated more than once. If repeated pair ratings are rare, as they may be in rating-once networks, "silence" must instead be defined at the target level. In that case it means the time since $v$ last received any rating within $u$'s two-hop neighborhood. The hypothesis is then tested on that signal. We will report which regime each dataset falls in rather than force one definition.

Baselines

The naive reference point is a global and per-rater mean. The rater mean predicts each rating as the rater's average past rating. It ignores the target and ignores time, and it isolates how much of the signal is pure trustingness. For orientation, the mixed-effects paper reports a simple-mean MSE of 0.7498 against 0.5327 for the mixed-effects model on CouchSurfing triads. We will run the same comparison on the temporal datasets rather than import that number.

The strongest classical competitor is a mixed-effects model with explicit time features, following Trust Propagation with Mixed-Effects Models. It adds random intercepts for raters and targets plus log-gap and exponentially decayed interaction counts as fixed effects. It is strongest because it already contains the observer baseline and can learn any average effect of time. It gets the same features and the same chronological splits. Beating it on point accuracy is not required. Beating it means matching its error overall while giving better-calibrated uncertainty, and lower error specifically on long-gap and cold-start pairs.

The strongest neural competitor is DynamiSE. It is a continuous-time model for signed networks that uses a differential equation with a graph network as the derivative. It reports an RMSE of 0.3246 on bitcoin-otc for the full model (taken from its ablation table) and an MAE of 0.6523 on Wiki-RfA, versus 0.6848 for M²DNE. We will rerun DynamiSE under its own protocol to confirm that we reproduce these numbers. We then evaluate both methods on identical chronological splits with equal hyperparameter-search budgets. DynamiSE produces no uncertainty, so we give it an uncertainty estimate by training a five-member deep ensemble. Beating it means coming within noise of its RMSE while outperforming the ensemble on selective-prediction risk.

The key ablations hold everything fixed and change one thing at a time. The first changes where decay goes: toward vacuity (proposed), toward a fixed distrust value (the conventional aging rule), or toward the global mean. The second removes the observer-specific base rate. The third replaces separate decay rates with a single symmetric one.

Why not just…?

Why not just add an exponential decay feature to the mixed-effects model? This is the most serious threat to the point-prediction claim. A well-specified mixed-effects model with a log-gap term could learn "revert to the rater's intercept after long gaps" on average. The difference is that it gives a fixed residual variance. It cannot say that one prediction is uncertain because evidence is old while another is uncertain because past ratings conflicted. If the mixed-effects model matches us on error and its prediction intervals stratified by gap are equally well calibrated, then the evidential state adds nothing beyond interpretability. We would report that plainly, and the contribution would shrink to the empirical finding about silence.

Why not just use classical Beta reputation with per-user base rates and hand-tuned aging? Our model is close to this, deliberately. Classical Beta reputation is a special case of ours with a fixed evidence mapping. The human-robot study found that replacing fixed end-of-task indicators with a learned continuous mapping cut its error from 18.17% to 0.18%. That was in a single human-robot pair, not a social network, so it suggests but does not prove that learning the mapping pays off here. We run the hand-set version as an ablation. If it ties the learned encoder, we will recommend the simpler system.

Why not just a dynamic graph network with an ensemble for uncertainty? Ensembles of DynamiSE-style models can produce uncertainty, at roughly five times the cost. Nothing in their structure, however, makes uncertainty grow with silence. Any such behavior has to be learned from sparse long-gap examples. Our gap-stratified calibration test measures exactly this. If the ensemble's uncertainty rises with gap length as reliably as ours, the inductive bias of decaying toward vacuity is unnecessary for uncertainty, though it may still matter for accuracy on cold pairs.

Proposed Experiments

We will use Bitcoin-OTC, Bitcoin-Alpha, and Wiki-RfA with strictly chronological splits. Each test rating is predicted using only earlier evidence.

The decisive empirical check needs no model. We bin test ratings by gap length, either pair-level or target-level depending on which regime the dataset falls in. For each bin we compute the mean deviation of the new rating from the rater's historical mean, and separately from the target's prior pair rating. Two outcomes are possible. If deviations from the prior rating grow with the gap while deviations from the rater's mean shrink, silence behaves as lost evidence. If ratings after long gaps are systematically more negative than the rater's mean, conventional aging is right, and our decay should target distrust instead.

Model comparisons report RMSE and MAE on rating value and AUC on sign. Uncertainty quality is reported three ways: expected calibration error of the Beta-predictive distribution, area under the risk-coverage curve when abstaining on high-vacuity predictions, and the correlation between vacuity and absolute error within each gap bin. All results are stratified by gap length, by cold-start status of the rater, and by direct versus two-hop pairs.

A secondary test uses the attacks from Towards Secrecy-Aware Attacks Against Trust Prediction in Signed Graphs. In those attacks, poisoned edges drop the specialized detector's AUC from 0.974 to 0.460. We will test whether bursts of fresh negative evidence on targets that were recently silent stand out as high-dissonance opinions. We treat this as exploratory, and a null result would not undercut the main claim.