FedFlex: Federated TV Series Recommender

• FedFlex is a federated recommender system that uses on-device fine-tuning of matrix-factorization models to personalize TV series recommendations.
• It integrates client-side Maximal Marginal Relevance (MMR) to re-rank and diversify the top-N recommendations while balancing relevance and novelty.
• The system preserves privacy by processing raw user data on-device and transmitting only differentially private model updates during aggregation.

Searching arXiv for “FedFlex” and closely related papers to ground the article. FedFlex most directly denotes a federated recommender system for Netflix-style TV series recommendations that combines on-device federated fine-tuning of matrix-factorization models with client-side diversity re-ranking via Maximal Marginal Relevance (MMR). In the reported implementation, the system is built on SyftBox, initializes from a public “what’s-on-Netflix” snapshot, preserves privacy by keeping raw viewing data on device, and exchanges only model updates protected by differential privacy. The designation is not entirely unique in recent federated-systems literature, so identification by arXiv id is important when comparing works (Lankester et al., 15 Jul 2025).

1. System architecture and federated workflow

FedFlex is implemented on SyftBox, described as an open-source federated-learning platform that orchestrates secure communication between a central aggregator and many client devices (Lankester et al., 15 Jul 2025). The workflow is explicitly divided into initialization, local adaptation, aggregation, recommendation, and re-ranking.

At initialization, the aggregator holds a global matrix-factorization model, either SVD or BPR, pretrained on a public “what’s-on-Netflix” snapshot. On the client side, each user imports her Netflix viewing history, which is locally converted into an explicit rating vector, exemplified as 1–5 stars per TV series. Local fine-tuning is then performed on the device. Before updates are sent back, the client adds noise for differential privacy. The server collects noisy gradients or parameter deltas from all clients and computes a weighted average to produce a new global model. Using the updated global model together with her own ratings, each client generates Top-$N$ recommendations locally, after which MMR is applied as a post-processing step to produce a second list with higher diversity.

The user-study interface operationalizes this pipeline through a local web app displaying two side-by-side recommendation lists of five Netflix-series thumbnails. List A is the Top-5 by predicted rating. List B is the Top-5 after MMR re-ranking. This separation is methodologically important because it isolates the contribution of post-hoc diversity control from the contribution of the underlying federated recommender.

A plausible implication is that FedFlex should be understood less as a new federated optimizer than as an end-to-end recommendation stack in which federated fine-tuning and diversity-aware retrieval are deliberately coupled. The paper’s emphasis is therefore split between privacy-preserving training and ranking behavior at serving time, rather than on convergence theory or communication-efficiency analysis.

2. Latent-factor models used for local fine-tuning

FedFlex supports two interchangeable matrix-factorization algorithms for personalized fine-tuning: SVD for explicit feedback and BPR for implicit feedback (Lankester et al., 15 Jul 2025). The system keeps the federated orchestration fixed and swaps only the local objective.

For SVD, the model learns user factors $p_u \in \mathbb{R}^k$ and item factors $q_i \in \mathbb{R}^k$ and predicts ratings through

$\hat r_{ui} = p_u^T q_i.$

The stated objective is the regularized squared-error problem

$$\min_{P,Q}\; \sum_{(u,i)\in\mathcal{K}} \bigl(r_{u i}-p_{u}^{T}q_{i}\bigr)^{2} \;+\;\lambda\sum_{u}\lVert p_{u}\rVert^{2} \;+\;\lambda\sum_{i}\lVert q_{i}\rVert^{2}.$$

For BPR, the same latent-factor parameterization is used, but the learning criterion is pairwise ranking. Training encourages a user $u$ to rank observed items $i$ above unobserved items $j$, with objective

$$\min_{\Theta}\; -\sum_{(u,i,j)\in \mathcal{D}} \ln\bigl(\sigma\bigl(p_{u}^{T}(q_{i}-q_{j})\bigr)\bigr) \;+\;\lambda\lVert\Theta\rVert^{2},$$

where $\sigma$ is the sigmoid, $p_u \in \mathbb{R}^k$0, and $p_u \in \mathbb{R}^k$1 is the set of positive-negative pairs for each $p_u \in \mathbb{R}^k$2.

Within FedFlex, clients “simply swap out one loss for the other during local fine-tuning,” while SyftBox abstracts the federated averaging. This design means that the system can compare explicit-feedback and implicit-feedback personalization under an otherwise stable federated execution environment. That makes the SVD-versus-BPR comparison a property of model choice rather than of communication protocol.

3. Diversity control through Maximal Marginal Relevance

The diversity mechanism in FedFlex is a client-side re-ranking stage based on Maximal Marginal Relevance. The paper presents MMR as balancing relevance, represented by predicted rating, against novelty or diversity (Lankester et al., 15 Jul 2025): $p_u \in \mathbb{R}^k$3 Here, $p_u \in \mathbb{R}^k$4 is the candidate set, exemplified as Top-20 by rating, and $p_u \in \mathbb{R}^k$5 is the set of already selected items.

In the reported system, $p_u \in \mathbb{R}^k$6 is approximated by predicted rating, while $p_u \in \mathbb{R}^k$7 is computed as cosine similarity between 384-dimensional title embeddings generated with sentence-transformer all-MiniLM-L6-v2 on the series title. The trade-off parameter was set to $p_u \in \mathbb{R}^k$8, corresponding to 30% weight on relevance and 70% on the diversity penalty, explicitly to produce aggressively diverse results in a small-scale live study.

This architecture makes diversity a post-hoc control layer rather than an intrinsic term in the federated training objective. A plausible implication is that FedFlex separates personalization and diversification so that each can be modified independently: matrix factorization handles user-specific scoring, and MMR perturbs the final list to expand semantic spread. The paper’s qualitative description of newly introduced genres, such as documentaries or comedies, reflects this separation between score estimation and list construction.

4. Experimental protocol and evaluation criteria

The empirical study used a static catalog and initial ratings from whatsonnetflix.com, specifically a Dec 2024 snapshot (Lankester et al., 15 Jul 2025). The live user study involved $p_u \in \mathbb{R}^k$9 volunteers, of whom 10 completed the study, over a two-week period. Each participant ran two versions of FedFlex, one using SVD fine-tuning and one using BPR fine-tuning.

On each run, the local web app displayed two lists: List A as Top-5 raw model scores and List B as Top-5 after MMR re-ranking. Participants clicked on thumbnails they would watch, and the system recorded clicks, timestamps, and the two displayed lists. The evaluation combines behavioral, ranking, diversity, and qualitative measures.

The reported accuracy metrics are Click-through rate (CTR), Precision@5, nDCG@5, and Mean Reciprocal Rank (MRR). The diversity metrics are Intra-List Diversity (ILD) and Coverage ratio, defined as the fraction of unique items clicked versus served. The study also includes a post-study questionnaire concerning personalization perception and list preference.

Because the study couples online interaction with local on-device recommendation, its measurements reflect actual user exposure rather than purely offline replay. At the same time, the paper explicitly identifies scale limitations, which is important when interpreting the quantitative differences between raw and re-ranked lists.

5. Quantitative and qualitative findings

The paper reports averaged metrics across participants for four conditions: SVD, SVD+MMR, BPR, and BPR+MMR (Lankester et al., 15 Jul 2025). For SVD, the reported values are CTR $q_i \in \mathbb{R}^k$0, Precision@5 $q_i \in \mathbb{R}^k$1, nDCG@5 $q_i \in \mathbb{R}^k$2, MRR $q_i \in \mathbb{R}^k$3, ILD $q_i \in \mathbb{R}^k$4, and Coverage $q_i \in \mathbb{R}^k$5. For SVD+MMR, the values are CTR $q_i \in \mathbb{R}^k$6, Precision@5 $q_i \in \mathbb{R}^k$7, nDCG@5 $q_i \in \mathbb{R}^k$8, MRR $q_i \in \mathbb{R}^k$9, ILD $\hat r_{ui} = p_u^T q_i.$0, and Coverage $\hat r_{ui} = p_u^T q_i.$1.

For BPR, the reported values are CTR $\hat r_{ui} = p_u^T q_i.$2, Precision@5 $\hat r_{ui} = p_u^T q_i.$3, nDCG@5 $\hat r_{ui} = p_u^T q_i.$4, MRR $\hat r_{ui} = p_u^T q_i.$5, ILD $\hat r_{ui} = p_u^T q_i.$6, and Coverage $\hat r_{ui} = p_u^T q_i.$7. For BPR+MMR, the corresponding values are CTR $\hat r_{ui} = p_u^T q_i.$8, Precision@5 $\hat r_{ui} = p_u^T q_i.$9, nDCG@5 $$\min_{P,Q}\; \sum_{(u,i)\in\mathcal{K}} \bigl(r_{u i}-p_{u}^{T}q_{i}\bigr)^{2} \;+\;\lambda\sum_{u}\lVert p_{u}\rVert^{2} \;+\;\lambda\sum_{i}\lVert q_{i}\rVert^{2}.$$0, MRR $$\min_{P,Q}\; \sum_{(u,i)\in\mathcal{K}} \bigl(r_{u i}-p_{u}^{T}q_{i}\bigr)^{2} \;+\;\lambda\sum_{u}\lVert p_{u}\rVert^{2} \;+\;\lambda\sum_{i}\lVert q_{i}\rVert^{2}.$$1, ILD $$\min_{P,Q}\; \sum_{(u,i)\in\mathcal{K}} \bigl(r_{u i}-p_{u}^{T}q_{i}\bigr)^{2} \;+\;\lambda\sum_{u}\lVert p_{u}\rVert^{2} \;+\;\lambda\sum_{i}\lVert q_{i}\rVert^{2}.$$2, and Coverage $$\min_{P,Q}\; \sum_{(u,i)\in\mathcal{K}} \bigl(r_{u i}-p_{u}^{T}q_{i}\bigr)^{2} \;+\;\lambda\sum_{u}\lVert p_{u}\rVert^{2} \;+\;\lambda\sum_{i}\lVert q_{i}\rVert^{2}.$$3.

The paper summarizes these outcomes by stating that accuracy saw a modest drop when applying MMR, as expected. Diversity improved slightly for BPR+MMR and stayed near-constant for SVD+MMR. Coverage, described as unique items clicked, rose from 35 to 43 for SVD and from 46 to 58 for BPR when using MMR. Figure 1 is reported to show that reranked lists introduced new genres while still reflecting core preferences, with BPR+MMR showing the strongest genre-diversification effect.

Qualitatively, most participants did not report a clear preference between List A and List B. Participants were uncertain whether either list was truly personalized, which the paper attributes likely to the small data volume. However, clicks on the reranked list remained high, suggesting that diversity was not penalized severely in user satisfaction. This suggests that the principal empirical claim of FedFlex is not that MMR improves every conventional accuracy metric, but that it can alter the semantic profile of recommendations without necessarily collapsing engagement.

6. Privacy properties, limitations, and terminological ambiguity

FedFlex’s privacy claim rests on two mechanisms: training entirely on-device and transmitting only model updates, with noise added for differential privacy (Lankester et al., 15 Jul 2025). The paper states that all raw viewing data stays on the device and that only model updates travel over the network. This positions the system within standard privacy-preserving federated recommendation practice, but the reported contribution is specifically the combination of such training with a diversity-oriented recommendation interface.

The limitations listed in the paper are also specific. The two-week, 10-user study is described as suffering from low training volume and potential selection bias due to a Linux/macOS requirement. Future directions explicitly proposed include learning a per-user MMR $$\min_{P,Q}\; \sum_{(u,i)\in\mathcal{K}} \bigl(r_{u i}-p_{u}^{T}q_{i}\bigr)^{2} \;+\;\lambda\sum_{u}\lVert p_{u}\rVert^{2} \;+\;\lambda\sum_{i}\lVert q_{i}\rVert^{2}.$$4, incorporating richer features such as genre tags, cast metadata, or scene summaries, and adding group-fairness constraints in a privacy-preserving federated setting alongside diversity.

A common misconception is to treat “FedFlex” as a uniquely identified method name across federated-systems research. The supplied literature indicates otherwise. The exact title “FedFlex: Federated Learning for Diverse Netflix Recommendations” refers to the recommender system described above (Lankester et al., 15 Jul 2025). By contrast, “Federated Learning with Flexible Control” introduces FlexFL, an FL algorithm for flexible computation and bidirectional communication control rather than a recommender system (Wang et al., 2022). “FLEX: FLEXible Federated Learning Framework” refers to a research framework rather than a single algorithm (Herrera et al., 2024). “Federated Aggregation of Demand Flexibility” is a privacy-preserving aggregation framework for demand-side resources whose summary also uses the shorthand “FedFlex,” but its primary title is different (Dong et al., 23 Sep 2025). A further adjacent case is “FLEX-MoE,” a federated mixture-of-experts framework whose supplied summary refers to “FedFlex” in describing the algorithmic procedure, again under a different primary title (Zhang et al., 28 Dec 2025).

For accurate scholarly usage, the term is therefore best resolved by context and arXiv id. In the strict bibliographic sense, FedFlex most directly denotes the Netflix-style federated recommender with SVD/BPR fine-tuning and MMR-based diversity control (Lankester et al., 15 Jul 2025).