acceptodds
Under review as a conference paper at ICLR 2027

Learning What to Recommend and Cache: Cost-Guided Multi-Agent Control at the Edge

Abstract

Recommendation-enabled edge caching does more than anticipate demand: it helps create the demand that the network must then serve. This coupling is powerful yet brittle, because recommendation response varies across users while every induced request must respect finite cache capacity and heterogeneous delivery routes. We cast the interaction as a sequential multi-agent control problem and introduce PA-MADDPG to connect behavior-conditioned demand with feasible edge decisions. Six descriptors distill each rating history into a behavioral profile, from which a Broad Learning System infers a Big Five representation. A cross-fitted response model blends this representation with list context, estimating how strongly recommendations may redirect individual demand without revealing the responses reserved for evaluation. The user model informs decentralized actors and centralized critics, while a cost-guided projection combines learned scores with marginal delivery savings to produce capacity-feasible caching and top-K recommendation decisions. Together, these elements align the demand that recommendations create with the resources available to serve it. On the synthesized Personality-INFOCOM setting, PA-MADDPG lowers cost across ten seeds against five adaptations, albeit with higher latency; ablations support projection but not a separable personality effect.

Primary area:
reinforcement learning
Keywords:
Edge Computing, Recommender System, Broad Learning, Federated Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_9GwvOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies joint recommendation and content caching in D2D-assisted edge networks, motivated by the observation that recommendations can influence which contents users request and, consequently, the network resources needed to serve them. The authors first introduce RITE, which extracts six behavioral descriptors from sparse rating histories and uses an incremental Broad Learning System to infer a user's Big Five personality traits. A separate response estimator combines the inferred traits with recommendation-list context to estimate user-specific responsiveness. The authors then propose DRAPE, a multi-agent controller based on centralized training and decentralized execution, to jointly select SBS cache placements and top-KK recommendation lists. A cost-guided projection converts continuous actor scores into decisions satisfying cache-capacity and recommendation constraints.

The evaluation uses a synthesized Personality–INFOCOM environment in which personality/rating records are paired with device-contact traces. DRAPE achieves a lower reported total cost than five scenario-adapted baselines over ten paired seeds, although its latency and some edge-service metrics are less favorable.

Critical strengths
  1. The paper explicitly models the interaction among recommendation decisions, recommendation-conditioned demand, SBS cache capacity, D2D availability, and content-delivery cost.
Critical weaknesses
  1. The central motivation in this paper is that inferred Big Five traits should capture user-specific differences in responses to recommendations and thereby improve downstream caching and recommendation decisions. However, the predictive evidence is weak.

  2. RITE uses an established incremental BLS architecture, while DRAPE relies on the standard MADDPG/CTDE framework with task-specific states, rewards, and a heuristic cost-guided projection. The projection is a plausible practical contribution, but the paper does not convincingly isolate its added value against simple cost-aware greedy caching or a matched MADDPG controller equipped with the same decoder.

  3. The paper does not name any specific public dataset. It refers only to "Personality" and "INFOCOM" sources, without providing the dataset names, versions, citations, download locations, licenses, or access procedures.

Questions for authors
  1. Could the authors separately ablate personality traits and response estimates, and compare directly using the six original behavioral features or a simpler learned response representation?

  2. What is the essential algorithmic contribution of DRAPE beyond standard MADDPG combined with cost-aware greedy projection? In particular, how does DRAPE compare with (i) a non-RL cost-aware caching heuristic using the same available information and (ii) a matched MADDPG controller with the same feasibility decoder and no personality input?

  3. Please provide the exact names, versions, citations, licenses, and access procedures for the Personality and INFOCOM datasets.

Reviewer_rbvkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies joint recommendation and edge caching in D2D-assisted edge networks. It proposes RITE, an incremental Broad Learning System that predicts Big Five personality traits from rating-derived behavioral descriptors, and combines the inferred traits with recommendation-list context to estimate user responses. Building on this representation, DRAPE uses centralized training and decentralized execution with multi-agent actor–critic learning. Its cost-guided projection converts continuous actor scores into capacity-feasible cache placements and top-K recommendation lists. Experiments in a trace-driven simulator compare DRAPE with five scenario-adapted baselines. Across ten paired seeds, DRAPE achieves lower total cost per request than all comparators, although it incurs higher latency and lower edge-service rates. The results support the overall cost advantage of DRAPE, while the contribution of personality-aware modeling remains unclear.

Critical strengths

1.Relevant joint optimization problem. The paper addresses the interaction between recommendation-induced demand, edge caching, and D2D delivery under storage and recommendation-list constraints. Jointly optimizing recommendation and caching is a meaningful direction because recommendation decisions can alter subsequent content demand and the associated delivery costs.

2.Careful end-to-end evaluation. The authors use ten paired seeds, share the stochastic request environment across methods, and apply paired statistical tests with Holm correction. The paper also reports cost decomposition, operational trade-offs, mapping sensitivity, robustness tests, and component diagnostics rather than relying solely on an aggregate performance metric.

Critical weaknesses

1.The contribution of personality-aware modeling is not supported by the downstream results. RITE improves macro MAE over the training-mean predictor only slightly (0.1873 versus 0.1888), while macro RMSE remains unchanged. More importantly, adding inferred personality traits worsens recommendation-response prediction on the 78 mapped evaluation users: MAE increases from 0.1359 to 0.1373, and Pearson correlation decreases from 0.044 to 0.008. The five-seed diagnostic ablation also finds no resolved cost improvement from mean, shuffled, or oracle personality/response information. These results do not establish a measurable benefit from personality-aware modeling in the current evaluation setting. Since personality-aware response modeling is a central motivation of the paper, its contribution to the reported end-to-end cost reduction remains unverified.

2.The independent algorithmic contribution and evidence for cost-guided projection need clearer justification. DRAPE combines a standard MADDPG-style centralized-training/decentralized-execution framework with deterministic ranking and feasibility projection. The main experimental configuration sets (\kappa_x=0.5) and (\kappa_y=0), so cache-side cost guidance is active while the explicit recommendation-side cost-guidance term is disabled. The reported NoGuide ablation therefore primarily tests the removal of cache-side guidance; it does not independently establish the contribution of both projection components. Moreover, the five-seed diagnostic reports a cost increase of 0.0452 after removing guidance, but the paired Wilcoxon test yields (p=0.0625). The direction is consistent, but the evidence is not conventionally statistically significant. These results support the potential value of cache-side cost guidance, but do not independently establish the contribution of the recommendation-side guidance term or the novelty of the overall projection mechanism relative to prior constrained-action methods. A clearer comparison with existing approaches and component-specific evidence would strengthen the methodological claims.

Questions for authors

1.Could the authors clarify the role of inferred personality in the final control policy? Given that personality-conditioned response prediction performs worse on the mapped evaluation users and that the oracle diagnostic does not show a resolved cost improvement, what evidence supports retaining the personality-aware component in the proposed framework?

2.The formal configuration sets (\kappa_y=0), while the method description presents cost guidance for both caching and recommendation. Could the authors provide a controlled comparison that separately evaluates cache-side guidance and recommendation-side guidance, including a configuration in which both terms are enabled?

3.Could the authors better distinguish the novelty of DRAPE's cost-guided projection from prior action-constrained reinforcement learning and projection methods? In particular, what properties of the proposed ranking-and-feasibility procedure are specific to joint recommendation and caching, and what measurable benefit do they provide beyond a standard centralized-training/decentralized-execution controller with a feasible-action decoder?

4.The five baselines are scenario-adapted implementations rather than reproductions of the original authors' code. Could the authors clarify which design elements and hyperparameters were retained from each original method, and provide sufficient details to assess whether the adaptations preserve the main strengths of the respective baselines?

Reviewer_wP7GOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

Joint recommendation and caching in D2D-assisted edge networks. The authors first propose RITE: six behavioral descriptors computed from a sparse rating history (rating mean, variance, popularity-stratum entropy and breadth, log activity, extremity share), mapped to a Big Five vector through an incremental Broad Learning System. The predicted traits are combined with recommendation-list context in a cross-fitted ridge model to estimate each user's response to a list. Joint recommendation and caching is then cast as a constrained sequential decision problem, trained with MADDPG-style CTDE, and decoded into capacity-feasible binary decisions by a cost-guided projection. Evaluation runs on a simulator stitched together from a real rating dataset and a real contact trace: 78 users, 100 contents, 5 base stations.

I'd welcome a response to the questions above. The experimental discipline and the honesty of the reporting are both above average for this kind of work, and there are two ways this could become a much better paper: replace the personality pathway with a representation that actually moves the downstream cost, or shrink the claim to "cost-guided projection yields a 1% cost improvement under a stated weight vector." The problem here isn't sloppy execution — it's that the claims are much larger than the experiments can carry.

Critical strengths

The statistical discipline here is better than most systems papers of this kind. Ten paired seeds, Holm correction, paired bootstrap, leave-one-seed-out, Friedman plus Kendall's W — and, importantly, a clear statement that the independent unit is a complete training/environment seed rather than an episode or a slot. Appendices C and D add an information-boundary table that spells out which quantities are used for fitting, which reach the policy, and which live only in the hidden simulator. Table 3 turns "we prevented leakage" from a claim into a checklist. Appendix H.1 applies a one-standard-error rule for the projection weights and explicitly rejects a candidate that looked better on validation. That's more careful than picking the minimum.

The reporting honesty is worth naming as well. The unhappy results are not buried. RITE beats the training-mean predictor by 0.0015 MAE. Adding inferred personality makes response prediction worse. DRAPE-Oracle, given true traits and true response, does not beat DRAPE-Full. No paired test in the stress sweep survives Holm correction. All of this is stated plainly in the main text and appendix. A paper willing to build ablations that argue against itself is doing something right.

The trouble is that those honest numbers are also the reason to reject.

Critical weaknesses
  1. The authors' own ablations already kill the first two contributions. In Table 1, RITE's macro MAE is 0.1873 against 0.1888 for a training-mean predictor, and macro RMSE is 0.2276 for both, identical to four decimals. Of the five traits, RITE is worse on Emotional stability for both MAE (0.2136 vs 0.2060) and RMSE (0.2477 vs 0.2401). The Pearson correlations run 0.143, 0.153, 0.234, 0.088, −0.080, the negative one being Emotional stability. A predictor statistically indistinguishable from outputting the training mean is not "inferring personality from behavior." And the next row is the direct hit: adding RITE traits moves response MAE from 0.1359 to 0.1373 and Pearson from 0.044 down to 0.008. The authors write it themselves — "Adding RITE traits does not improve response prediction." The chain promised in the first two sentences of the abstract is broken in the experiments.

  2. That projection buys about one percent and pays for it in two other metrics. Table 6: DRAPE costs 4.0778 per request against 4.1168–4.1540 for the five comparators, a relative gain of 1.04% to 1.95%. Same table, latency is 92.60 ms versus 86.20 ms for OOA, and edge service rate 26.71% versus 33.45%, a 6.74 point drop. The response is that "the advantage is confined to the optimized cost." That's circular. The objective is assembled from hand-set weights (α_m = 1.0, α_e = 4.0, α_u = 0.2, λ_r = 0.03, λ_c = 0.05). Winning 1% under those weights while losing on latency and service rate doesn't show the method is better; it shows it was fitted to those weights. Move α_e to 2.0 and the ordering may well flip. There is no weight-sensitivity analysis anywhere.

  3. The evaluation environment is built by the authors, and "trace-driven" doesn't carry the weight. Appendix C.2 says the 78 Personality users and 78 INFOCOM identifiers are sampled and paired one-to-one under seed 2027, and states that "it does not imply that the two datasets describe the same real individuals." The response model is then built by the authors from the enjoy watching field, and the hidden simulator samples requests from it. So: traits from dataset A, contact traces from dataset B, welded together arbitrarily, with the connecting function defined by the authors, and the policy trained and evaluated on that same function. It's a closed loop with no point of contact with reality. Nor will 78 users, 100 contents and 5 base stations support any claim about edge networks.

  4. A load-bearing modeling assumption with no justification. Eq. (3) sets ξ_{u,f} = 4q(1−q) because it peaks at q = 0.5 and vanishes at both ends. But why should recommendation influence disappear for items a user hates, or items a user already loves? Being pushed something you dislike is a response too, and so is being recommended your favorite. The parabolic form is mathematical convenience dressed up as a behavioral claim. Eq. (4) and Eq. (25) have the same character — a functional form written first, an interpretation attached afterwards.

  5. The baselines are scenario adaptations, not author code. Appendix D.2 says so directly: "transparent scenario adaptations rather than exact author-code reproductions," and each one is adapted differently (OOA is greedy placement with trigger-based re-caching, uEDC-L is affine robust scoring). Under those conditions a 1% cost difference can't be attributed to anything.

Questions for authors
  1. Given that DRAPE-Oracle shows no resolved difference from DRAPE-Full, how much does the personality pathway actually contribute under the current objective? If the answer is negligible, how should the first two sentences of the abstract be rewritten?

  2. Does the cost ordering against OOA survive if α_e goes from 4.0 to 2.0, or λ_r from 0.03 to 0.1? A sweep with fixed trained weights answers this without retraining. If the ordering flips, what does the 1% mean?

  3. Latency goes from 86.20 to 92.60 ms and edge service rate from 33.45% to 26.71%. What cost equivalent do those two changes carry in your objective? Without that conversion the reader can't tell whether the 1.04% is cancelled out.

  4. Where is the boundary of "trace-driven"? Given that traits and contacts come from unrelated datasets and the response function is yours, in what sense does this simulator predict deployment behavior?

  5. What justifies ξ_{u,f} vanishing at q = 0 and q = 1? Can you give an alternative form that doesn't rely on the parabola, and say how sensitive the results are to it?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.