SPAR: Prefix-conditioned recommendation and frozen-codebook load repair¶
L1 RQ-PIB repair only; diagnostic only. No PATH decoder, quantizer training or preference optimization.
Paper information¶
| Field | Details |
|---|---|
| Paper | SPAR: Stepwise Personalized Autoregressive Recommendation with Prefix-Balanced Semantic IDs |
| Affiliations | First author Chunrui Zhang: University of Electronic Science and Technology of China; collaborators include Alibaba HUJING and Youku. |
| First public release | 2026-10-20 09:17:14 UTC |
| Author code | No author implementation found in paper/arXiv. Youku logs are private. |
| Local key | spar-prefix-balance; no adapter |
| Local code | src/auto_research/reproductions/spar_prefix_balance.py |
| Local evidence | Local mechanism metrics |
What the paper changes¶
Relevant history depends on generated SID prefix; popular prefixes can dominate interaction load. SPAR uses PATH for coarse-to-fine prefix-conditioned memory, RQ-PIB to rebalance a frozen residual codebook, and SFT/DPO for preferences. The local scope is only the repair stage.
Original paper figure¶
Original figure from the reviewed paper, reused without redrawing.
Core equations¶
For a parent prefix, compute mean branch load and a feasible cap using total weight, largest item and sibling count. From an overloaded branch, choose a positive-weight item and feasible destination minimizing increase in squared residual-to-center distance. The greedy rule is not globally optimal or exactly equalizing. Never update centers, move an item more than once, move zero-load groups or change parent prefix.
Results reported by the authors¶
Paper reports mean relative HR@10/20/50/100 +2.66% on Youku logs and seven-day, 10%-traffic A/B with watch time +3.95%, play conversion +3.38%. Full model outcomes cannot be attributed to local greedy moves.
Local reproduction and limits¶
Provide frozen centers, residuals, assignments, train-only weights and parents. Tests for Eq.13–15/Algorithm 1 check smallest feasible move, empty destination, heavy-item cap, conservation, single move and unrepaired overload. Three CPU synthetic seeds do not establish recommendation performance or exposure fairness. No quantizer/PATH/beam/DPO or Evolve.
PYTHONPATH=src python scripts/run_oct21_ranking_components.py \
--component spar-prefix-balance --device cpu --output runs/spar-prefix-fresh.json
Full experiment: not reproduced¶
Qwen3-0.6B–4B, months of production history, daily preference data and live traffic. Full training/A/B not reproduced; no retries.
