Federated recommenders keep each user's clicks on their own phone. A central server never sees raw interactions; it only combines model updates sent by devices. To cope with users who differ a lot from one another, the strongest recent systems group devices by behavioral similarity and train a separate model per group. The filter-bubble literature, meanwhile, finds that grouping users by past behavior and showing each group what that group already likes is how echo chambers form. This proposal treats the federated server's choice of which model each device receives as the main lever for bursting bubbles. That lever has no direct counterpart in centralized systems, and it can be pulled without seeing anyone's data.
Our claim has two parts. First, cluster-based federated personalization measurably speeds up narrowing of exposure over repeated train-recommend rounds, compared with a single shared model. Second, a small change at the server repairs most of that narrowing while keeping most of the accuracy gain from clustering. We call the change bridge aggregation. Each device scores items with a blend of its own cluster's model and one deliberately chosen "bridge" cluster, a group whose members are similar to it but whose consumption covers different categories. The device itself tunes how much weight the bridge gets, based on whether its user actually engages with the out-of-cluster items. We will test both parts in closed-loop simulation on a fully observed dataset. There, every user's true response to every item is known, so feedback loops can be replayed without inventing a user model.
Core Claim
ReFRS is the clearest instance of the mechanism we worry about. The server treats each device's model parameters as a "semantic fingerprint," clusters the fingerprints with k-means, and sends each device a model trained only on its cluster. This raises HR@10 by over 25% relative to plain federated averaging on a MovieLens subset. ReFRS reports only accuracy and runtime and never measures diversity, despite "diversified preferences" in its title. The same structure appears in Taobao logs: Understanding Echo Chambers in E-commerce Recommender Systems found that users who follow recommendations show more stable click clusters (adjusted Rand index 0.0986 versus 0.0765) and declining content diversity (1.0945 to 1.0882).
Our hypothesis is that federated clustering imposes this segregation by design. The server is the right place to undo it, because it already holds the fingerprints and the per-cluster models.
The closest connecting work does not address this. FedDistRec splits a diffusion-based preference model into shared and on-device parts and evaluates single-round accuracy only. The federated recommender survey lists bias and echo chambers as open problems. Our work differs from all three in two ways: it measures multi-round narrowing, and it intervenes in the aggregation topology rather than in the local loss.
Research Program
The audit. We build a closed-loop federated simulator on the KuaiRec small matrix, which covers 1,411 users and 3,327 short videos at 99.6% density. Each user is one client. Each round proceeds in three steps. Clients receive their model and recommend a top-k slate. Simulated feedback is read from the ground-truth watch ratios. Clients then train locally on the items they were shown and send updates back to the server.
Because the matrix is almost fully observed, we can evaluate any slate against true preferences. We can also measure narrowing directly, using three quantities: the categories each client is exposed to over rounds, the stability of client cluster assignments (the adjusted-Rand-index protocol of Ge et al.), and intra-list diversity. We compare plain federated averaging, ReFRS-style hard clustering, and centralized training under the same loop.
The intervention. Let $c(i)$ be client $i$'s cluster and $h_k$ the normalized category histogram of items recommended to cluster $k$. The server computes $h_k$ from slates it already routes, or through secure aggregation of client counts, so no individual history is revealed. The server picks a bridge cluster
$$b(i) = \arg\max_{k \neq c(i)} \; \mathrm{sim}(f_i, \mu_k)\cdot \mathrm{JS}(h_k, h_{c(i)}),$$
where $f_i$ is the client's fingerprint, $\mu_k$ is cluster $k$'s centroid, and JS is the Jensen–Shannon divergence. The bridge is thus near the client in taste but different in what it consumes. On device, item $j$ is scored by
$$s_i(j) = (1-\alpha_i)\, s_{c(i)}(j) + \alpha_i\, s_{b(i)}(j).$$
The weight $\alpha_i$ is chosen per client from a small grid, for example {0, 0.1, 0.2, 0.4}, by a local bandit. Its reward is engagement with items outside the client's dominant categories. This borrows the explore/exploit logic of hierarchical contextual bandits but over four arms instead of an item tree. It also echoes the user steering of User-controllable Recommendation Against Filter Bubbles, except that $\alpha$ is learned and private rather than set by an explicit command. Blending scores rather than parameters avoids any problem with averaging models from different clusters; parameter interpolation is a secondary variant.
Robustness. We test the dynamics under the boredom-and-quit behavior of KuaiEnv, used by CIRS. We also run the behavior-aware Bubble Escape Potential protocol of Quantifying the Potential to Escape Filter Bubbles on MovieLens-1M. That protocol compares how much diversity active, exploring agents receive relative to passive ones, so it checks whether gains reach users who actually want to escape.
Baselines
Plain federated averaging with one shared model. This is the naive reference. It has no cluster segregation, so it may already be less bubbly. Its lower accuracy defines the cost side of the trade-off.
ReFRS-style hard clustering. This is the strongest existing federated recommender in the file and the system whose dynamics we audit. The strongest reported number is ReFRS-SASRec beating federated averaging by over 25% in HR@10 on a MovieLens 10% subset (ReFRS). No reported result was found on Lacuna for federated recommenders on KuaiRec, so we reproduce both. All federated methods share the same SASRec backbone, the same number of clusters (tuned once for hard clustering and then frozen), the same number of rounds, the same per-client local compute, and an equal hyperparameter search budget.
Local differentiable diversity training. This is the strongest diversity method. Each client adds the differentiable top-k diversity objective of Gradient-Based Diversity Optimization to its local loss under hard clustering. The centralized version reports up to a 78% relative diversity gain over matrix factorization on Yahoo-R2 with near-identical recall. We also run the centralized version as an upper reference.
Ablations isolating the key ingredient. First, we blend with the global federated-averaging model instead of a chosen bridge. Second, we use a randomly chosen other cluster as the bridge. Third, we fix $\alpha$ globally rather than learning it per client.
What counts as beating the baselines. Bridge aggregation wins if it dominates hard clustering and local diversity training on the accuracy–exposure-diversity frontier at matched NDCG@10, and if its Bubble Escape Potential is higher. The targeted-bridge ablation must also beat both the global-blend and random-bridge ablations.
Why not just…?
Why not just use one global federated model and skip clustering? This is a real threat. If the shared model sits on the same frontier as clustering, then segregation was never the problem. The audit tests exactly this. Comparing federated averaging with hard clustering across rounds tells us whether clustering adds narrowing beyond what personalization alone causes. The global-blend ablation tests whether a per-client interpolation toward the shared model recovers everything. If it does, we will report that the interpolation weight, not bridge selection, is what matters.
Why not just re-rank on the device with maximal marginal relevance or determinantal point processes? Re-ranking can only reorder candidates that the cluster model already scores highly, so it cannot surface what the cluster never learned. The DDT authors report that post-hoc re-ranking "often tanks accuracy" relative to in-training diversity (Gradient-Based Diversity Optimization). We will still include client-side re-ranking as a baseline at matched accuracy.
Why not just inject random items? The Bubble Escape Potential study found that up to about 30% randomness actually lowered the measure for LightGCN and TiCoSeRec, and that only 80–90% randomness helped, which destroys utility (Quantifying the Potential to Escape Filter Bubbles. We include epsilon-greedy at matched exploration rates. The random-bridge ablation tests the analogous claim one level up, at the level of clusters rather than items.
Why not rely on differential-privacy noise to add variety? ReFRS reports that differential privacy degraded both accuracy and clustering quality. Its noise is unguided, so it adds no relevant diversity. We will run a privacy-noise variant only to confirm this in the closed loop.
Decisive Experiments
The first check is the audit. Over 30 rounds on KuaiRec, does hard clustering shrink per-client category coverage and raise cluster stability faster than federated averaging? If it does not, the premise fails and the proposal reduces to ordinary federated diversity training.
The second check asks whether bridge aggregation keeps at least most of hard clustering's NDCG@10 gain over federated averaging while restoring exposure coverage to the federated-averaging level or above. The comparison is made against ground-truth watch ratios, at matched accuracy against local DDT training.
The third check is the bridge ablation. Targeted bridges must beat random bridges and the global blend. Otherwise the selection rule contributes nothing.
The fourth check covers the two robustness conditions. Under KuaiEnv's boredom-and-quit behavior, does session length improve? This is the condition in which CIRS fell back to the level of DeepFM for highly sensitive users. On MovieLens-1M, does Bubble Escape Potential rise rather than fall as accuracy is held fixed, reversing the accuracy–escape correlation that the original study found across centralized models?
References
- ReFRS: Resource-efficient Federated Recommender System for Dynamic and Diversified User Preferences
- Understanding Echo Chambers in E-commerce Recommender Systems
- Beyond Single Embedding: Modeling User Preferences as Distribution in Federated Recommendation
- A Scenario-Oriented Survey of Federated Recommender Systems: Techniques, Challenges, and Future Directions
- KuaiRec: A Fully-observed Dataset and Insights for Evaluating Recommender Systems
- Show Me the Whole World: Towards Entire Item Space Exploration for Interactive Personalized Recommendations
- User-controllable Recommendation Against Filter Bubbles
- CIRS: Bursting Filter Bubbles by Counterfactual Interactive Recommender System
- Quantifying the Potential to Escape Filter Bubbles: A Behavior-Aware Measure via Contrastive Simulation
- Gradient-Based Diversity Optimization with Differentiable Top-$k$ Objective