Social recommenders predict what a user will want next by mixing that user's own history with information passed along the user's social ties. They usually do this with a graph neural network, a model that repeatedly updates each user's representation by averaging in the representations of their neighbors. The social graph is almost always treated as fixed. We argue that this fixed graph, combined with repeated neighbor averaging, creates a specific and measurable error. A user's own past tastes travel out to a friend and then come back, so the model keeps hearing the user's old self as if it were independent social evidence. When tastes are stable this echo mostly double-counts. When a user's tastes drift, it actively anchors the model to who the user used to be.
The idea comes from network epidemiology. Spurious self-feedback of mean-field predictions inflates infection curves shows that standard mean-field epidemic models on static graphs let an infection bounce from A to B and straight back to A. That artifact inflates predicted outbreaks most strongly in sparse graphs. A second-order correction, known as TAP equations, removes the echo at linear cost. We propose to carry this diagnosis and its fix over to message-passing social recommenders. We also test the sharper claim that the echo is a hidden cause of the "preference inertia" the drift literature complains about.
Core Claim
In a two-layer social graph neural network, user $u$'s representation at the second layer depends on neighbor $v$'s first-layer representation. That first-layer representation already contains $u$'s own input, weighted by $w_{vu}$. Under a chronological split, $u$'s input summarizes $u$'s older interactions. So a term of size roughly $w_{uv} w_{vu}\, h_u^{(0)}$, where $w_{uv}$ is the normalized edge weight and $h_u^{(0)}$ is $u$'s own history embedding, returns to $u$ disguised as social signal.
We make three claims:
- The echo is a substantial and measurable share of what social aggregation contributes.
- Removing it improves accuracy mainly for users whose preferences have drifted between training and test windows.
- The gain is largest for low-degree users. With symmetric degree normalization, $w_{uv} w_{vu} \approx 1/(d_u d_v)$, so the echo weight grows as degrees $d_u$ and $d_v$ shrink. This mirrors the sparse-graph result of the Plefka-expansion paper.
If gains appear uniformly across drifting and stable users, or do not depend on degree, the echo story is wrong. In that case any improvement is ordinary regularization.
The fix is the cavity idea from message passing in physics. When $v$ sends a message to $u$, it sends a version of itself computed without $u$'s contribution:
$$ m_{v\to u}^{(\ell)} = \phi\Big(h_v^{(\ell-1)} - w_{vu}\, \psi(h_u^{(\ell-2)})\Big), $$
Here $m_{v\to u}^{(\ell)}$ is the layer-$\ell$ message from $v$ to $u$, and $\phi$, $\psi$ are the model's existing update and message maps. We will compare two versions. The exact edge-wise version costs O(number of edges) per layer and keeps one message per directed edge, as in belief propagation. The cheaper TAP-style version subtracts a single aggregated correction per node, $\sum_v w_{uv} w_{vu}\, \psi(h_u^{(\ell-2)})$, at O(number of nodes) extra cost.
Relation to Prior Work
Three social recommenders motivate the problem, and all keep the social graph fixed. GNN-DSR splits users into a short-term part (an LSTM over recent interactions) and a long-term part (graph attention). It aggregates friends over a static social graph. DyHGCN makes repost cascades time-sliced but explicitly keeps the follow graph static. GDMSR shows that most social ties are uninformative and prunes 10–40% of them. Its pruning is a one-off filter based on historical overlap.
None of these asks whether aggregating over the remaining ties feeds a user's stale self back into their representation. Pruning changes which edges exist. Echo cancellation changes what flows along each edge, so the two can be combined.
Two papers supply the method and constrain it:
- Spurious self-feedback of mean-field predictions inflates infection curves supplies the correction itself and the prediction that sparse graphs suffer most. It also leaves loops beyond back-and-forth exchanges, such as triangles, uncorrected. We therefore expect residual inertia in highly clustered neighborhoods and will measure it rather than claim to fix it.
- Learning Dynamic Preference Structure Embedding From Temporal Networks shows that per-node time decay beats a single global decay. That makes time-weighting a serious competing explanation for any drift gain.
On the drift side, NS-DPO handles drift by exponentially discounting old preference data. When preferences flip, standard DPO falls to roughly 50–55% reward accuracy while NS-DPO stays above 60–70%. NS-DPO treats each preference source in isolation. Our point is that in a social model, discounting a user's own old data does not stop that data from returning through neighbors whose representations were built from it.
COEVOLVE and CMMSB model the graph itself co-evolving with behavior. That is the more obvious way to connect these two directions. We deliberately do not build another co-evolution model. The question here is narrower and cheaper to settle: does the static-graph artifact already distort today's recommenders, even before the graph is allowed to change?
Research Program
The work has three parts.
Measuring the echo. For a trained social recommender, we compute how much of each user's final representation is attributable to that user's own input via paths that leave and return. This is the contribution of the user's input along walks of length two that return to the start, obtained from the gradient of the output with respect to the user's own input along those paths. We then compare that share to the contribution from genuinely other users.
Building the correction. We implement the edge-wise cavity variant and the TAP-style variant as drop-in changes to the message step of GNN-DSR and of a plain LightGCN-style social aggregator. Embedding size, optimizer and tuning budget stay unchanged.
A controlled simulator. We generate synthetic users with latent preferences on random graphs of controlled degree (mean degree 2 to 20) and clustering. Users switch preference regimes at known times, and the ground truth is known. This lets us vary drift magnitude, degree and clustering independently. Real data cannot do that.
All real-data evaluation uses chronological splits. We label a user as a drifter when the divergence between their item-category distributions in the training and test windows exceeds a threshold. We report results for several thresholds so the conclusion does not depend on one cutoff.
Baselines
Naive reference. A no-social model: GNN-DSR's LSTM-plus-item-graph components with the social graph removed. If social aggregation helps drifters less than it helps stable users, this reference shows it.
Strongest existing method. GNN-DSR, because it already separates dynamic from static preferences and so is the system most likely to absorb drift without our change. It reports RMSE 0.9444 on Ciao against a best prior baseline of 0.9750, and NDCG@10 of 0.2805 on Delicious against 0.2669. Those figures come from its own protocol. We will rerun it on our chronological split with the authors' hyperparameter ranges. Every variant gets the same tuning budget in trials and GPU hours, and we report the means and spread of five seeds. Beating it means a statistically reliable improvement in NDCG@10 and RMSE on the drifting-user subset, with no loss on stable users.
Pruning baseline. GDMSR reports a 10% Recall@1 gain on Ciao over Diffnet++ and MHCN. The file gives no absolute value. It is run both alone and combined with echo cancellation, to test whether the two effects add up.
Time-weighting baseline. A per-node time-decay weighting of interactions and edges in the style of DPS, applied to GNN-DSR.
Ablations isolating the key ingredient. Three ablations separate the echo correction from cruder fixes:
- removing explicit self-loops only;
- a one-layer model, which has no return paths but also loses two-hop information;
- the correction applied with a random partner instead of the true return path, which controls for the extra regularization the subtraction might add.
Why not just…?
Why not just use one aggregation layer? A single layer has no return path, so it has no echo. It also discards friends-of-friends, which GDMSR's noise analysis suggests are where much of the useful signal sits, since direct ties are often uninformative. The one-layer ablation settles this directly. If it matches the cavity model on drifters, the correction is unnecessary and we will say so.
Why not just prune noisy ties, as GDMSR does? Pruning decides which edges exist, using historical similarity. The echo flows along every edge that survives, and pruning by past similarity keeps exactly the ties most aligned with the user's old self. We therefore predict pruning alone leaves inertia for drifters intact. The pruning-plus-cancellation combination tests whether the effects add or overlap.
Why not just discount old interactions, as NS-DPO and DPS do? This is the real threat. Strong enough time decay shrinks the stale input before it can echo. Our claim is that decay and cancellation remove different things. Decay removes old evidence wholesale, including old evidence that is still valid. Cancellation removes only self-generated evidence. The simulator settles it: when drift is partial (only some preference dimensions change), cancellation should beat decay. If tuned per-node decay matches cancellation everywhere, the proposal reduces to "use decay" and we will report that.
Decisive Experiments
We run on Ciao, Epinions, Yelp and Douban with timestamps, re-split chronologically, plus Delicious for comparability with GNN-DSR's reported NDCG@10. Metrics are NDCG@10, Recall@10 and RMSE, each reported separately for drifting and stable users and by degree bucket.
First check: is the echo large enough to matter? On trained GNN-DSR, the echo share must be non-trivial, above a few percent of the social contribution, and must be larger for low-degree users. If it is negligible, we stop.
Second check: is the gain where the theory says it should be? In the simulator and on real data, cancellation's improvement over GNN-DSR must grow with drift magnitude. It must be concentrated in low-degree buckets and must vanish for stable users. The random-partner control must show no gain.
Third check: does it beat tuned decay? In the simulator's partial-drift regime, cancellation must beat tuned per-node decay at equal tuning budget.
Secondary analysis: loops beyond pairs. We regress the remaining drifter error on local clustering coefficient. A positive slope would indicate that triangle-sized loops, which neither our correction nor the TAP equations address, are the next source of inertia.
References
- Spurious self-feedback of mean-field predictions inflates infection curves
- Graph Neural Networks with Dynamic and Static Representations for Social Recommendation
- DyHGCN: A Dynamic Heterogeneous Graph Convolutional Network to Learn Users' Dynamic Preferences for Information Diffusion Prediction
- Robust Preference-Guided Denoising for Graph based Social Recommendation
- Learning Dynamic Preference Structure Embedding From Temporal Networks
- Right Now, Wrong Then: Non-Stationary Direct Preference Optimization under Preference Drift
- COEVOLVE: A Joint Point Process Model for Information Diffusion and Network Co-evolution
- Co-Evolution of Selection and Influence in Social Networks