A Redundancy Reduction Approach for Controllable Sequential Recommendations
Veronika Ivanova, Marina Munkhoeva, Ivan Razvorotnev, Evgeny Frolov
cs.IR
2026-09-21
A Barlow Twins decorrelation term with next-item-matched positive pairs lifts accuracy on five benchmarks and exposes a knob trading head accuracy for long-tail exposure.
Sequential recommendation lives with a long tail: a few popular items absorb most exposure, and models reinforce the concentration, because frequently updated item embeddings drift toward directions shared across user histories and pick up a scoring advantage unrelated to the specific user. Practitioners trade head-of-list accuracy against long-tail exposure with engineering compromises, but a training-time handle for dialing that trade-off continuously has been missing. A team from Yandex and HSE turns it into a representation-geometry problem, using the Barlow Twins objective from self-supervised learning.
The training objective is Lpred plus alpha times LBT. Lpred is a standard next-item loss. LBT is Barlow Twins redundancy reduction: embeddings from two views are L2-normalized and mean-centered (replacing the original projection head), their cross-correlation matrix is computed, the diagonal is pushed to 1 and the off-diagonals to 0.
The positive-pair construction is the cleanest part. Instead of synthetic corruptions such as masking, cropping or reordering, the method pairs two user histories that end with the same next item. The semantics hold up, since histories leading to the same item should embed similarly, and the labels are consistent by construction.
The geometric analysis supplies the mechanism. Decompose item embeddings against the direction of the mean user embedding, and frequent items carry a large component there, adding a user-agnostic bonus to their scores. Decorrelation pushes user representations toward zero mean and identity covariance, erasing the shared low-rank direction that gives popular items their global edge. A bucket-based Alignment Concentration metric quantifies how strongly each popularity bucket aligns with the shared subspace.
Five public benchmarks, three base losses (BCE, CE, SCE); adding BT improves all of them. On ML-1M with CE, NDCG@10 goes from 0.0494 to 0.0583. Gains over the full baseline suite (SASRec variants, CL4SRec, DuoRec, EC4SRec, two-stage BT pretraining):
| Dataset | HR@1 | NDCG@10 | cov@10 |
| ML-1M | +12.7% | +6.2% | -12.5% |
| Yelp | +8.9% | +2.7% | +9.5% |
| Gowalla | +4.6% | +5.7% | +7.6% |
| Beauty | +6.9% | +4.8% | -6.2% |
| Kindle | +5.8% | +4.1% | +13.9% |
(relative to the second-best method)
Accuracy is uncontroversial: nearly every dataset and metric improves with statistical significance (paired t-tests, p < 0.05). Exposure splits by dataset: catalog coverage rises on Yelp, Gowalla and Kindle, and falls on ML-1M and Beauty.
By bucket on ML-1M, HR@1 is slightly below baselines on the head bucket and clearly higher on mid and tail; at HR@10 even the head matches, so blockbuster accuracy does not pay for the tail gains. Alpha is the knob: on Yelp, raising it from 0.1 to 0.4 holds HR@10 while coverage climbs; on ML-1M the direction reverses and both degrade, which the authors attribute to timestamp collisions flooding the pair construction with frequent terminal items. The practical recommendation is alpha between 0.2 and 0.4.
Two directly portable pieces. One, a one-line training-time regularizer for any dot-product sequential recommender: no architecture change, no inference cost, consistent accuracy gains. Two, alpha is a post-deployment dial for exposure allocation, which matters more for platforms with ecosystem-diversity mandates (content distribution, long-tail marketplace supply) than reranking after the fact. The two-stage BT-PT comparison also settles that end-to-end joint training beats pretrain-then-finetune here.
Head-tail effects are strongly dataset-dependent: coverage drops on ML-1M and Beauty, and the temporal-structure explanation is an attribution rather than a verified mechanism. The causal chain from geometry to exposure is explicitly not established. There is no comparison against dedicated popularity-debiasing methods, which the authors frame as a different setting. Pair construction can be skewed by frequent terminal items on timestamp-dense data, and all five benchmarks are conventional sequential-recommendation scale, with no LLM-based recommenders.