Ensemble methods are frequently used in recommender systems to improve accuracy by combining multiple models. Recent work reports sizable performance gains, but most studies still optimize primarily for accuracy and robustness rather than for energy efficiency. This paper measures accuracy energy trade offs of ensemble techniques relative to strong single models. We run 93 controlled experiments in two pipelines: 1. explicit rating prediction with Surprise (RMSE) and 2. implicit feedback ranking with LensKit (NDCG@10). We evaluate four datasets ranging from 100,000 to 7.8 million interactions (MovieLens 100K, MovieLens 1M, ModCloth, Anime). We compare four ensemble strategies (Average, Weighted, Stacking or Rank Fusion, Top Performers) against baselines and optimized single models. Whole system energy is measured with EMERS using a smart plug and converted to CO2 equivalents. Across settings, ensembles improve accuracy by 0.3% to 5.7% while increasing energy by 19% to 2,549%. On MovieLens 1M, a Top Performers ensemble improves RMSE by 0.96% at an 18.8% energy overhead over SVD++. On MovieLens 100K, an averaging ensemble improves NDCG@10 by 5.7% with 103% additional energy. On Anime, a Surprise Top Performers ensemble improves RMSE by 1.2% but consumes 2,005% more energy (0.21 vs. 0.01 Wh), increasing emissions from 2.6 to 53.8 mg CO2 equivalents, and LensKit ensembles fail due to memory limits. Overall, selective ensembles are more energy efficient than exhaustive averaging,
Figures & tables
Dataset
Users
Items
Interactions
Sparsity
Rating scale
MovieLens-100K
943
1,682
100,000
93.70%
1–5
MovieLens-1M
6,040
3,706
1,000,209
95.53%
1–5
ModCloth
47,958
1,378
82,790
99.87%
1–5
Anime
73,515
11,200
7,813,737
99.05%
1–10
TABLE I: Datasets used in the experiments (after preprocessing).
Dataset
Reference single model
RMSE
Energy (Wh)
Ensemble (best-efficiency / best-accuracy)
RMSE
Energy (Wh)
ML-100K
SVD++
0.9193
0.00182
Top Performers (eff.)
0.9154
0.00279
ML-1M
SVD++
0.8625
0.03886
Top Performers (eff.)
0.8542
0.04617
ModCloth
SVD++
0.9440
0.000117
Top Performers (eff.) / Average (acc.)
0.9384 / 0.9357
0.000817 / 0.00427
Anime
SVD
1.1335
0.00992
Top Performers (eff. & acc.)
1.1197
0.20842
TABLE II: Surprise results: best optimized single model vs. selected ensembles (energy includes training and evaluation).
Dataset
Reference single model
NDCG@10
Energy (Wh)
Ensemble (best-efficiency / best-accuracy)
NDCG@10
Energy (Wh)
ML-100K
User-KNN
0.2176
0.000087
Top Performers (eff.) / Average (acc.)
0.2238 / 0.2301
0.000148 / 0.000177
ML-1M
User-KNN
0.1959
0.00238
Average (eff.) / Rank Fusion (acc.)
0.2028 / 0.2071
0.00448 / 0.00889
ModCloth
Item-KNN
0.1595
0.00025
Top Performers (eff.) / Weighted (acc.)
0.1603 / 0.1622
0.000384 / 0.000441
Anime
User-KNN
0.2368
0.1777
(all ensembles failed due to memory limits)
–
–
TABLE III: LensKit results: best optimized single model vs. selected ensembles.
Recommendation recursive self-improvement (Rec-RSI) feeds recommender outputs into subsequent training. Evaluating each round solely through its latest model assumes that the successor consolidates the update, although pre- and post-update models may retain complementary ranking decisions. We term this distributed progress and quantify it using cross-generation advantage (CGA), a marginally matched contrast between cross- and within-generation model pairs. A rank-separation statistic, label-free at selection time, predicts which family to retain. Across four datasets and three sequential recommendation encoders, the preferred retention regime varies by architecture: cross-generation pairing benefits GRU4Rec and SASRec, whereas FMLP initially favors within-generation pairing and shifts toward cross-generation pairing after a second update. Rank separation selects the stronger family in 12/12 first-update and 5/6 second-update dataset-encoder settings; on held-out tests, the selected family outperforms the direct successor in 34/36 trajectories. Five transfer mechanisms do not consistently reproduce these gains in one model. These findings establish state retention as a distinct Rec-RSI problem: progress may reside in relations between generations as well as in the latest model. Code is available at https://github.com/Jinfeng-Xu/RecRSI.
Jinfeng Xu, Zheyu Chen, Ziyue Peng +5
The University of Hong Kong HongKong SAR, China · The Hong Kong Polytechnic University HongKong SAR, China · The Hong Kong University of Science and Technology HongKong SAR, China +1
Scaling recommendation models is a central challenge in recommender systems. Recently, RankMixer has emerged as an effective solution, operating on a unified token representation and alternating between token mixing and per-token feedforward networks (P-FFNs) to achieve scalable performance. However, RankMixer suffers from embedding collapse, where learned representations have low effective rank, limiting expressivity and underutilizing the expanded representation space. Through empirical analysis and theoretical insights, we identify rigid token mixing and P-FFN modules as the primary causes of this phenomenon, jointly inducing a damped oscillatory trajectory in effective-rank evolution across layers. To address it, we propose RankElastor, a novel architecture that produces spectrum-robust representations with provable collapse mitigation. RankElastor introduces two components: (i) parameterized full mixing, which enables expressive token mixing with improved spectral robustness; and (ii) GLU-improved P-FFNs, which stabilize representation spectra through GLU-style FFN modules. Extensive experiments on large-scale industrial datasets demonstrate that RankElastor consistently improves recommendation performance, mitigates embedding collapse, and exhibits robust scaling behavior. Code is available at this GitHub repository: https://github.com/vasile-paskardlgm/RankElastor
Guoming Li, Shangyu Zhang, Junwei Pan +7
The Hong Kong University of Science and Technology (Guangzhou) Guangzhou, China · Tencent Inc. Shenzhen, China
Early ranking stages in recommendation systems precompute item embeddings and cache them in-model for scoring within strict latency constraints. Because this cache exists only at serving time, outside the training loop, training and serving use different item representations, a structural discrepancy that limits quality and adds operational fragility. We show that co-designing the training and serving paths removes this representation discrepancy at its source. We introduce the memory layer, an in-model key-value embedding cache co-trained with the model: the item tower writes embeddings during training and the model reads them at serving, one source of truth for item representations by construction. Always-on embeddings cover items not yet cached, so every item receives a prediction, and the design consolidates three separate trainer-to-predictor update paths into a single self-contained pipeline. Deployed in production on Instagram Reels, the memory layer raises prediction coverage from 96% to 100%, improves embedding freshness from O(5 min) to O(20 s), and narrows the training-serving Normalized Entropy (NE) gap by up to 86%, yielding over 2× recall for the freshest content and a 5-6% cold start engagement lift. Because embeddings are produced during training, the system needs no separate bulk-evaluation or publish-time recomputation, cutting training-and-publish computational cost by 30% at neutral serving computational cost.