FineSID: Progressive CTR learning from content through relevance to personalization

简体中文

Reference entry only. No executable local adapter or result; complete experiments not reproduced due to cost.

Paper information

Field	Details
Paper	Progressive Learning from Content to Relevance to Personalization for CTR Prediction
Affiliations	First author Zhiguang Yang: Alibaba Group
First public release	2026-10-21 03:19:02 UTC
Author code	No author source was found in the paper or arXiv as of 2026-10-22. Referenced VAR is not FineSID source.
Local key	finesid; no adapter
Local code	No executable local implementation.
Local evidence	No local experiment, metrics or GPU receipt.

What the paper changes

Preserve fine visual/text detail, train the whole shared predictor for query–item relevance, then transfer all shared parameters and normalization state to CTR. Stage I uses fixed VAR grids 1×1, 2×2, 3×3 and 4×4 (30 positions), codebook 4,096, preserving scale, location and fields. Code zero is valid; absence has a separate mask. Train randomly initialized CTR embeddings rather than copying VAR vectors. Fielded BPE/N-gram text, query-conditioned attention and native IDs feed ordered fusion.

Original paper figure

Original figure from the reviewed paper, reused without redrawing. See figure provenance.

Core equations

\[ \mathcal L_{rel}=-\sum_{c=1}^{4} t_c(x)\log \pi_c(x). \]

\[ \mathcal L_{CTR}=-y\log\hat p-(1-y)\log(1-\hat p). \]

Stage II learns four soft classes from a Qwen3.5 teacher with the complete shared predictor. Stage III retains all dense/embedding/normalization state (197 shared dense/normalization entries), replaces the head, resets Adagrad accumulators and fine-tunes on click BCE. The backbone uses two HiFormer layers, width 768, and MLP 768→256→64. Copying embeddings alone or optimizer state is not equivalent.

Results reported by the authors

Section 4.8/Table 5 reports seven-day Taobao main-search production A/B: clicks (IPV) +0.21%, GMV +0.85% relative to production baseline. Offline AUC/GAUC point changes differ from online relative lifts. Public MMCTR is not a substitute for private logs or live traffic.

Local reproduction and limits

Literature, original figure and author evidence only; no small fixture, local metrics, GPU receipt or Evolve.

Full experiment: not reproduced

64 H20s, batch 4,096/card and 262,144 global. 51.2B impressions, \(2.51\times10^8\) users, 200M queries, \(4.18\times10^8\) products and \(4.53\times10^8\) test impressions plus expert/private relevance labels. No full run, data download or automatic retry.