People in virtual worlds walk as if they were still in the physical world. In Second Life, avatars that can fly or teleport stay on the ground and follow paved paths around digital grass that has no reason to be avoided. Separately, studies of online networks keep finding that the most active users are not the most influential ones. These two facts collide whenever we try to measure who influences whom by how people move. If two avatars walk the same winding path a few seconds apart, the obvious reading is that one followed the other. Usually, though, both were obeying the same carried-over habit. Any influence measure built on movement will mistake shared habit for social following, and it will credit the people who move most with the most influence.
This proposal builds and tests a measure of movement influence that first learns where people walk when they are alone in a given space, and then counts as influence only the part of one person's movement that departs from that habit in step with another person's departure. We test it against interventional ground truth: scripted confederate avatars whose path choices we randomize. With a measure that is not fooled by habit, we can ask two questions the literature has not answered. Is spatial "activity" (how much someone explores or wanders) decoupled from spatial influence, as posting volume is decoupled from influence online? And does an avatar's influence over others change as they unlearn physical-world habits?
Core Claim
Movement-based influence estimates are inflated by shared environmental habit, and the inflation grows with how strongly a space channels people. The fix is to condition on a learned habit model, which plays the role of the "background distribution" in Copy or Coincidence?. That paper detects true links between records by weighing how similar two records are against how rare their shared feature is in the population. It proved that once noise is large enough, similarity alone is no better than chance (AUC near 0.5 at displacement noise $t = 1/\sqrt{2}$). We bring that logic to trajectories. Co-movement along a popular path is common under habit and should count for little. Co-movement across the grass, or a matched detour, is rare under habit and should count for a lot.
Concretely, let $s$ be the scene (geometry plus surface labels such as path or grass), $b_{t+1}$ person B's next displacement, $b_{\le t}$ and $a_{\le t}$ the recent histories of B and A, and $h_\theta(b_{t+1}\mid b_{\le t}, s)$ a habit model trained only on solo trajectories in comparable scenes. Habit-residual influence of A on B is
$$ I_{A\to B} = \mathbb{E}\left[\log \frac{p_\phi(b_{t+1}\mid b_{\le t}, a_{\le t}, s)}{h_\theta(b_{t+1}\mid b_{\le t}, s)}\right], $$
where $p_\phi$ is a pairwise model that additionally sees A's history. Formally this is a scene-conditioned transfer entropy, meaning how much A's past reduces uncertainty about B's next move. What is new is not the formula. It is the requirement that the denominator be a habit model fit on solo data, so that shared obedience to the environment cannot leak into the numerator. Without that, both terms are fit on the same mixed data, and the pairwise model can "explain" B with A simply because both track the path.
Research Program
The work has three parts. The first is the estimator. We train $h_\theta$ on single-person trajectories, which we get from solo stretches of LocoVR recordings and from solo sessions in our own study, and we train $p_\phi$ on pairs. Both use the same backbone, so the only difference is access to A. Because the input directions warn about synergy, we also report a second version. ITENE showed that ordinary transfer entropy mixes direct influence with "synergistic" information that exists only when several sources are seen together. In a tight doorway, B's detour may depend jointly on A and the walls rather than on A alone. We therefore estimate an ITENE-style intrinsic term on the bottleneck subsets of LocoVR to check whether the habit-residual score is picking up joint geometry–peer constraints rather than A's pull.
The second part is ground truth, which is where earlier work on movement influence is thinnest. The switching-parent model in Learning Influence among Interacting Markov Chains recovered a hardcoded leader in a four-agent synthetic grid, with the learned influence weight $\alpha_A \to 1.0$. Counterfactual influence in Intrinsic Social Motivation via Causal Influence showed that movement can become a signal among reinforcement-learning agents. Neither touches humans carrying physical habits. We build two sources of ground truth. One is simulation: agents placed in LocoVR's 131 home scenes and in a park-like scene with paths and grass, with a tunable preference for paths and a planted follow relation between some pairs. The other is a human study. Participants navigate a shared park scene in VR or a desktop virtual world alongside scripted confederate avatars. On each trial we randomize whether a confederate stays on the path, cuts across the grass, or stops at an off-path destination. The confederate's effect on participants is then a randomized causal effect, not an inference. We collect self-reported virtual-world experience directly, which avoids the "alt account" confound that limited tenure analyses in Please keep off the grass.
The third part is the science. We first ask whether movement activity predicts influence. We compute the place and step entropies of Something In The Way We Move, which measure how widely someone explores and how unpredictably they step, and correlate them with habit-residual influence. The online analogues are follower count versus retweets in the Million Follower Fallacy, and posting volume versus transfer entropy in Information Transfer in Social Media. Those motivate the hypothesis but do not settle it, because high-entropy movers in space might be genuine scouts rather than noise. We then ask whether unlearning habits changes influence: does a confederate who cuts the grass draw experienced participants off the path more than novices? Samperi et al. observed path adherence falling from 58% in new avatars to 39% in veterans, but they could not ask whether veterans' departures move anyone else.
Baselines
The naive reference is activity volume: rank people by distance travelled or by place entropy and treat that ranking as influence. The analogous baseline in meetings, speaking length, scored a KL divergence of 0.226 against human influence judgments, versus 0.106 for the two-level influence model (Learning Influence among Interacting Markov Chains). We include it to measure how far volume and influence diverge in space. Beating it means a clearly higher AUC for recovering the randomized confederate effect.
The proximity baseline labels B as following A when B stays within a lagged distance threshold of A for several consecutive steps. We tune the threshold and lag on the same validation split as our method. This is what a practitioner would deploy. Copy or Coincidence predicts it will collapse as habit strength rises.
The plain transfer-entropy baseline uses the same $p_\phi$ backbone but replaces the solo-trained habit model with a single-history model fit on the same pooled data, following Ver Steeg and Galstyan. It is the ablation that isolates our key ingredient. If it matches our method across habit strengths, the solo habit prior adds nothing.
The strongest existing competitor is Shapley edge attribution, as in You Mostly Walk Alone: drop A from a trained scene-aware social trajectory predictor and measure the increase in B's prediction error. That paper found near-zero neighbor contribution on ETH-UCY and SDD, and a roughly 57% error increase on interactive SportVU basketball. We apply it to a predictor trained on LocoVR, whose reported 3-second average displacement error is 0.44 m versus 0.81 m for a GIMO-trained model (LocoVR). We first confirm that our reimplementation reaches comparable error, so the attribution is not run on a weak predictor. All methods share data, the predictor architecture where applicable, and an equal hyperparameter-search budget.
We found no result on Lacuna for any method recovering ground-truth follower relations from human trajectories. Beating Shapley attribution therefore means higher AUC on the randomized confederate effect, together with a smaller loss of AUC as simulated path preference increases.
Why not just…?
Why not just use proximity and lag to detect following? On a crowded path, everyone is close to everyone. The 58% path adherence among new Second Life avatars means unrelated pairs will look like followers, and Copy or Coincidence shows that distance alone falls to chance at moderate noise. Our habit-strength sweep in simulation measures this directly. If proximity holds up when path preference is weak, we will report that, and the habit correction then matters only in strongly channeled spaces.
Why not just run transfer entropy or Granger causality on positions? This is the most serious threat, because our estimator is a conditioned transfer entropy. The difference lies entirely in where the baseline model comes from. When pairs dominate the training data, a pooled single-history model absorbs some of the shared habit and some of the social signal, and the leftover is hard to interpret. The ablation above settles the question. If pooled transfer entropy matches the solo-habit version on confederate ground truth, we drop the claim of a new estimator and present the work as a measurement study.
Why not just use edge-dropping attributions from a good trajectory predictor? A scene-aware predictor already conditions on geometry, so it partly encodes habit. But attribution measures what the model relies on, not what the person responded to. A shared destination, such as both people heading to the stage, can make A useful for predicting B without A causing anything. The randomized confederate trials separate the two, because the confederate's path is assigned independently of where the participant is going. If Shapley attribution recovers the randomized effect as well as our method does, then the contribution reduces to the benchmark and the scientific findings, and we will say so.
Proposed Experiments and Decisive Checks
In simulation, we sweep path preference from none to strong and planted follow strength from zero upward. We report AUC for recovering the true pairs for every method. The decisive pattern is that the proximity and pooled transfer-entropy baselines lose AUC as path preference rises, while the habit-residual score stays flat. A second check is that our score stays near zero for unlinked pairs even under strong path preference.
In the confederate study, with roughly 60 participants split across self-reported experience levels, the target quantity is the randomized change in participants' off-path rate caused by a grass-cutting confederate. Each estimator should rank confederate–participant pairs by influence in line with that measured effect. The decisive check is agreement between the estimated influence and the randomized effect, compared across methods. The same data tests the habit-unlearning question: whether the confederate effect differs between experienced and novice participants.
On LocoVR, we compute habit-residual influence for all pairs and correlate it with place and step entropy. We also compare the plain and intrinsic (ITENE-style) estimates in doorway and hallway segments to see how much apparent yielding is joint geometry–peer synergy. A near-zero correlation with entropy would extend activity–influence decoupling to physical movement. A strong positive correlation would show that, unlike online posting, moving a lot does steer others.
One result would sink the core claim. If all three alternatives match the habit-residual score on confederate ground truth across habit strengths, then shared habit is not the confound we believe it is.
References
- Copy or Coincidence? A Model for Detecting Social Influence and Duplication Events
- ITENE: Intrinsic Transfer Entropy Neural Estimator
- Learning Influence among Interacting Markov Chains
- Intrinsic Social Motivation via Causal Influence in Multi-Agent RL
- Please keep off the grass: individual norms in virtual worlds
- Something In The Way We Move: Quantifying patterns of exploration in virtual spaces
- Measuring User Influence in Twitter: The Million Follower Fallacy
- Information Transfer in Social Media
- You Mostly Walk Alone: Analyzing Feature Attribution in Trajectory Prediction
- LocoVR: Multiuser Indoor Locomotion Dataset in Virtual Reality