Recommender systems learn from user–item interaction logs that are usually more than 99.9% empty. One recent response is to have the recommender write its own training data. The model imagines plausible continuations of each user's history, keeps the ones that pass a quality check, and retrains on real plus imagined sequences. RSIR shows this works for a few rounds, raising Recall@10 by 13.92% on Amazon-Sports after three iterations. It also shows the gains stall and then reverse after 5–8 rounds. The authors attribute this to an "irreducible noise floor" of compounding self-generated bias. This is a synthetic-to-real gap of the kind studied in simulation for robotics and perception: a model trained on its own approximations of reality learns the approximations' quirks.
We argue the floor is not irreducible. It comes from a specific design choice: the quality check is run by the same model being retrained. Recommenders trained on sparse data are biased toward popular items, as SimGCL and InvCF document. A popularity-biased judge therefore approves popular continuations, the student absorbs them, and the next judge is more biased still. Our proposal has two parts. First, freeze the judge as a model trained only on real data. Second, make its test popularity-stratified: it asks whether a user's true future items keep a high rank among items of similar popularity, rather than across the whole catalog. We predict this moves or removes the saturation point, raises long-tail accuracy without the usual head-item penalty, and lowers the cost of the check.
Core Claim
Self-generated interaction data drifts away from real behavior mainly because the filter and the student share their biases. Two changes break that loop. The first decouples the filter from the student by anchoring it to real data. The second removes the filter's popularity shortcut by comparing items only within popularity buckets. Together they should let recursive self-generation keep improving well past the point where RSIR degrades.
Concretely, consider user $u$ with current (real or partly synthetic) sequence $s$ and a candidate generated item $g$. Let $F_u$ be the user's held-out future items, taken from the training portion of their history and never from the test split. The anchored, stratified check accepts $g$ only if
$$\max_{v \in F_u} \; \mathrm{rank}_{J}\big(v \mid s \oplus g;\ B(v)\big) \le k .$$
Here $J$ is the frozen judge trained on real data, $\oplus$ means appending $g$ to $s$, $B(v)$ is the popularity bucket containing $v$, rank is computed only among items in $B(v)$, and $k$ is the threshold. RSIR's original check corresponds to $J$ equal to the current student and $B(v)$ equal to the whole catalog. Ranking within a bucket also shrinks the candidate set per check. This addresses the vocabulary-scale overhead RSIR lists as unsolved.
Relation to Prior Work
Three lines of work make sparse data denser with synthetic interactions, and they motivate this proposal.
RSIR generates sequences autoregressively and filters them with self-referential fidelity control. We keep its generator and bounded exploration and change only who judges and how. CF-SFL builds a "virtual user" with an adversarial reward estimator that unrolls synthetic feedback during collaborative filtering training, improving VAE NDCG@100 on MovieLens-20M from 0.426 to 0.435. Its learned discriminator is exactly the component that can latch onto synthetic artifacts. That is why we use a fixed ranking test rather than a trained real-versus-fake classifier. Sim2Rec and KuaiSim build user simulators for reinforcement learning. Sim2Rec's live A/B test (6.9% versus 0.1% for unadapted reinforcement learning) is strong evidence that the reality gap matters in recommendation. Both, however, target policy learning over small action spaces or short-video sessions, not densifying sparse catalogs, and Sim2Rec relies on hand-written "trend filters." Our anchored check plays the role of a trend filter but is derived from real data rather than written by hand.
Two papers supply the mechanism. InvCF shows that popularity acts as a spurious shortcut and separates it from preference using distance correlation. We borrow the diagnosis but not the architecture: stratifying by popularity removes the shortcut from the filter without retraining the backbone. SimGCL's uniformity analysis gives us a diagnostic for embedding collapse across rounds.
One paper constrains the design. ASM found in LiDAR adaptation that learned discriminators exploit "clean versus noisy" shortcuts. This is a warning against making the anchor a learned adversary. We use a discriminator only as an offline measuring instrument, never as a training signal.
Research Program
The work has three stages.
Stage 1: measure the drift. We reimplement RSIR on SASRec, CL4SRec and HSTU and run 12 rounds, well past the reported 5–8 round peak. After each round we record the following on the accepted synthetic data versus held-out real sequences:
- KL divergence between the item-popularity distributions;
- share of tail items;
- repeat and transition entropy;
- the uniformity loss of the student's embeddings, following SimGCL;
- the AUC of a held-out classifier distinguishing real from synthetic sequences.
The prediction to falsify is that popularity divergence and discriminator AUC rise in step with the performance downturn. If they stay flat while performance falls, our explanation is wrong and the noise floor lies elsewhere.
Stage 2: separate the two ingredients. We cross judge identity (current student, frozen real-only anchor, anchor refreshed every few rounds) with test type (full-catalog rank, popularity-stratified rank). The stratified test is also run with the bucket count varied, since one bucket recovers the full-catalog test.
Stage 3: cost. We report wall-clock time of the fidelity check under each test.
Baselines
The naive reference is the backbone trained on real data alone, with no synthetic sequences. Every method must beat it to justify itself.
The strongest existing method is RSIR, run at its best round and at matched round counts. On Amazon-Sports it reports a 7.71% Recall@10 gain for HSTU, growing from 8.02% to 13.92% cumulatively over three iterations. The file gives only these relative gains, not absolute values, so our reimplementation's real-only baseline sets the absolute scale. For fairness, RSIR and our variants share the generator, exploration pool, synthetic-to-real mixing ratio, number of accepted sequences per round, and a single hyperparameter budget. We tune $k$ for both on validation data. Beating RSIR means a higher peak Recall@10 and NDCG@10 and a later or absent downturn, with tail recall improved and head recall within noise.
CL4SRec stands in for heuristic augmentation, since RSIR already reports outperforming it. CF-SFL represents the adversarial virtual-user approach, run on its non-sequential setting for MovieLens-20M, where its reported VAE NDCG@100 is 0.435.
The key ablations isolate each ingredient: the frozen anchor with a full-catalog test, and the self-judge with a stratified test. If only one of the two carries the gain, the claim narrows to that ingredient.
Why not just…?
Why not just stop at the best round? Early stopping caps the damage but leaves the improvement ceiling where RSIR found it. The claim here is that the ceiling itself moves. The 12-round curves settle this directly. If anchored judging peaks at the same height and round as RSIR, the proposal adds nothing, and we will report that.
Why not just resample synthetic items to match real popularity? This is the most serious competitor. If drift were purely in the item-popularity marginal, reweighting accepted sequences would fix it cheaply. However, marginal matching keeps the self-judge, which still accepts sequences whose transitions reflect its bias. It can also overweight implausible tail items, recreating the accuracy-versus-tail trade-off that HID documented (on Tmall with STAMP, HR@20 rose from 26.10 to 28.26 and tail coverage from 69.46 to 73.65 only by adding explicit intent constraints). We include popularity-matched resampling as a full baseline. If it equals our method, the honest conclusion is that drift is a marginal-distribution problem and the simpler fix wins.
Why not just debias the backbone with InvCF and run RSIR unchanged? A debiased self-judge still co-evolves with the student, so the feedback loop remains even if its starting point is less skewed. InvCF's reported gains (10.2% average NDCG@20 with an MF backbone on Tencent) concern a single training run, not recursive rounds. We run InvCF-style debiasing plus self-judging as an additional arm. The judge-identity factor in Stage 2 shows whether decoupling matters once the bias is removed.
Proposed Experiments
Data. Amazon-Beauty, Amazon-Sports, Amazon-Toys and Yelp follow RSIR's sparse sequential setup and splits. KuaiRand-Pure, described as an unbiased log, serves as a check that gains are not just better exploitation of exposure bias.
Metrics. We report Recall@10 and NDCG@10 split by head and tail item buckets, tail coverage, the Stage 1 drift measures, and check cost. Results average five seeds with confidence intervals.
Decisive test. We plot performance and popularity drift by round for self-judging versus anchored stratified judging. The proposal succeeds if anchored judging keeps improving or holds steady beyond round 8 while self-judging declines, if drift tracks the decline, and if popularity-matched resampling does not reproduce the effect. It fails cleanly if all arms saturate at the same round regardless of the judge.
References
- Can Recommender Systems Teach Themselves? A Recursive Self-Improving Framework with Fidelity Control
- Are Graph Augmentations Necessary?: Simple Graph Contrastive Learning for Recommendation
- Invariant Collaborative Filtering to Popularity Distribution Shift
- Collaborative Filtering with A Synthetic Feedback Loop
- Sim2Rec: A Simulator-based Decision-making Approach to Optimize Real-World Long-term User Engagement in Sequential Recommender Systems
- KuaiSim: A Comprehensive Simulator for Recommender Systems
- Adversarially Masking Synthetic to Mimic Real: Adaptive Noise Injection for Point Cloud Segmentation Adaptation
- Bid Farewell to Seesaw: Towards Accurate Long-Tail Session-Based Recommendation via Dual Constraints of Hybrid Intents