Pith. sign in

REVIEW 4 major objections 5 minor 33 references

SemRaD shows that structured LLM profiles plus hindsight distillation—reconciling pre- and post-conversion reasoning—improve cold-start LTV and CVR while using 9% of training data.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 19:07 UTC pith:A2CSGDBF

load-bearing objection Solid industrial recipe with a real online A/B, but the mechanism claim is under-supported: no dense-feature control, offline gains sit in the noise band, and the abstract overstates the data-efficiency result. the 4 major comments →

arxiv 2607.17070 v1 pith:A2CSGDBF submitted 2026-07-19 cs.AI

Bridging the Information Gap: Semantic Densification and Hindsight Distillation for Cold-Start Prediction

Hao Duong Le , Yifei Gao , Huan Li , Lun Jiang , Chen Bai , Ke Xing , Chen Zhang This is my paper
classification cs.AI
keywords cold-start predictionlifetime valueconversion rateknowledge distillationprivileged informationLLM semantic reasoningstructured reasoning schemadistillation experts
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper seeks to show that the cold-start problem in e-commerce—predicting lifetime value and conversion for users with only a few logged events—can be substantially improved by combining two ideas: densifying sparse logs with structured, schema-constrained LLM reasoning, and distilling privileged post-conversion knowledge through a reconciled hindsight target rather than raw future signals. A sympathetic reader would care because both levers are practical: the structured profile is servable at inference with a frozen embedder, and the hindsight target exists only at training. On a large industrial dataset the full method lifts LTV Gini by 1.9% and CVR AUROC by 1.0% over a production-grade base, matches production LTV with 9% of the training data, and a four-week online A/B confirms the gains. The claim rests on the mechanism that hindsight fusion narrows teacher-student disagreement for cold users—22% less LTV disagreement and 10% less CVR logit disagreement—making privileged knowledge reachable.

Core claim

SemRaD's central claim is that the information gap between a privileged teacher (which sees post-first-purchase behavior) and a sparse student (which sees only pre-purchase logs) is heterogeneous and can be bridged by two components: a Structured Semantic Reasoning Pipeline that turns free-form LLM rationales into a four-dimension schema, producing the servable Densified Semantic Profile and the training-only Hindsight Distillation Target; and a Hindsight-Aware Distillation Network that transfers the teacher's knowledge through Gumbel-routed Distillation Experts plus soft-target distillation. The paper argues that the hindsight target works because it explicitly reconciles pre- and post-conv

What carries the argument

The Hindsight Distillation Target is the central object: a per-user, per-dimension natural-language profile produced by an LLM fusion prompt that takes pre-conversion reasoning and post-conversion reasoning and explicitly states which pre-conversion signals were reliable and which were misleading, along with a teaching signal for the student. It carries the argument because it converts contradictory privileged supervision into self-consistent targets that a sparse student can learn from. Two supporting mechanisms are the Semantic-Gated Encoder, which gate-weights frozen text embeddings of the schema dimensions and doubles as a schema-audit utility score, and Distillation Experts, multiple ex

Load-bearing premise

The load-bearing assumption is that the LLM's structured rationales along the four schema dimensions (the servable Z_pre and the fused Z_fused) are faithful, informative summaries of latent user traits; the paper's own audit step (§4.1, Fig 5) validates dimensions only by gate weights, a usage measure rather than a correctness measure—if the profiles misrepresent sparse logs or drift across markets, the densification benefit and the 9%-data-efficiency claim collapse.

What would settle it

A direct falsifier: on a held-out set of cold-start users, replace the LLM-generated profiles with profiles generated from scrambled, truncated, or adversarially rephrased prompts; if LTV/CVR gains over the base persist unchanged, the semantic content is not what is carrying the effect. A sharper test: annotate a sample of pre-conversion rationale labels against outcomes from the post-conversion log; low agreement for users with five or fewer events would contradict the claim that the profiles are faithful enough to drive privileged distillation.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the claim is correct, cold-start LTV/CVR prediction improves without added serving cost: the deployed student consumes only cached frozen embeddings and a gated MLP, with no LLM, teacher, or fusion call at inference time.
  • The 9%-data-efficiency result implies that new markets or settings with expensive, slow-to-collect LTV labels can reach production-level quality with a fraction of the supervision.
  • Because the recipe generalizes across student backbones and profiler LLMs, the effect is attributable to the structured schema and hindsight fusion rather than to one upstream generator.
  • The measured student-teacher disagreement provides a concrete, inspectable 'distillability' signal that can be used to decide when privileged distillation will help.
  • The confirmed online A/B gains indicate the offline improvements are not an artifact of the experimental setup.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the paper's audit step measures dimension utility by the trained model's gate weights—a usage measure, not a correctness or consistency measure; a direct test of whether Z_pre and Z_fused labels agree with post-conversion outcomes under adversarial or shifted prompts would separate semantic fidelity from mere feature presence.
  • Editorial inference: the learned expert router splitting cleanly into converter and non-converter regimes suggests a simpler two-expert or regime-conditioned design might capture most of the distillation benefit.
  • Editorial inference: the 9%-data-efficiency result points toward a transfer recipe in which structured profiles generated in one market could reduce cold-start labeling cost in a new market, though the paper does not test cross-market transfer.
  • Editorial inference: because the fusion prompt explicitly labels which pre-conversion signals are reliable, the method implicitly teaches the student a sparse causal feature set; a sensitivity analysis on those identified signals would test whether the student actually learns to rely on them.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes SemRaD, a cold-start LTV/CVR prediction framework that combines structured LLM-generated semantic profiles with privileged-information distillation. A discover-curate-audit workflow defines a four-dimension behavioral schema; the student serves from pre-conversion logs plus a per-user Densified Semantic Profile Z_pre, while a training-only teacher consumes a Hindsight Distillation Target Z_fused reconciled from pre- and post-conversion reasoning. Knowledge is transferred via logit distillation and a per-user expert-routed reconstruction loss. Offline experiments on a 120k-user industrial dataset report +1.9% LTV Gini and +1.0% CVR AUROC over a production base, and a four-week online A/B reports +1.0% LTV and +0.43% CVR. The paper also claims production-level LTV with 9% of the training data.

Significance. If the reported results hold, SemRaD is practically significant: it shows a path from noisy LLM rationales to structured, servable semantic features, and it addresses the heterogeneous teacher-student gap with hindsight fusion and distillation experts. The strongest evidence is the online A/B with bootstrapped CIs and DeLong tests, the explicit component ablations, and the data-efficiency comparison. The main weaknesses are that the causal role of 'semantic reasoning' is not cleanly isolated from dense feature enrichment, and one robustness table contradicts the text's generalization claim. These issues are fixable but require additional experiments and careful rewriting.

major comments (4)
  1. [§6.3, Table 2b] The text states that 'gains remain under alternative profiler LLMs,' but Table 2b shows LTV Gini of 0.2834 for Qwen3.5-9B and 0.2820 for Qwen3.5-35B-A3B, both below the Base LTV of 0.2847 in Table 1, row (A). For LTV, the alternative profilers therefore do not preserve the claimed improvement; the CVR gains are small (+0.0039 and +0.0025) and are comparable to or smaller than the reported standard deviations. Please report significance tests for these differences, or revise the robustness claim so that it is consistent with the table.
  2. [Table 1 vs. §5.1, Fig. 4] The SemRaD row in Table 1 lists Z_pre as absent (✗), but §5.1 and Fig. 4 state that the deployed student consumes Z_pre while the teacher consumes Z_fused. If the Z_pre column is meant to denote something other than student input (e.g., whether the profile is used as an additional privileged augmentation), the labeling is confusing and prevents the reader from determining exactly what distinguishes SemRaD from PFD/HAPFD. This also makes the ablation rows ambiguous. Please correct the table or define the columns precisely.
  3. [§6.1–6.2, Table 1] The central mechanism claim is that structured LLM rationales provide semantic information that closes the cold-start gap. However, all LLM-based variants add dense text embeddings, and there is no control that adds a similarly dense, hand-computed feature summary derived from the same logs (e.g., event counts, session statistics, price-tier shares, entropy) without LLM reasoning. As a result, the +0.5pp LTV improvement of Sem. Des. over Base could be attributed to feature enrichment rather than to the semantic content of the rationales. The audit step in Fig. 5 measures only gate utilization, not the factual correctness or consistency of the rationales. Please add such a feature-engineering control, or explicitly reframe the claims from 'semantic reasoning' to 'dense semantic profiles.'
  4. [Table 1, lower block] The ablation labeled 'w/o Semantic Gating' reports LTV Gini 0.2834, which is below the Base value of 0.2847. It is unclear whether this condition removes the semantic view entirely or removes only the gating mechanism while retaining concatenated semantic embeddings. If the latter, a below-Base result requires explanation; if the former, the row does not isolate the gating mechanism. Please clarify the configuration and add a condition that removes gating but keeps the semantic embeddings (simple concatenation) so the effect of gating is identifiable.
minor comments (5)
  1. [Table 1] Report the number of seeds or resamples used for the mean±std values. As written, the reader cannot tell whether differences of 0.0005–0.002 are within seed noise.
  2. [§6.3, Table 2b] Even after correcting the over-claim, the profiler-LLM comparison lacks significance tests. Given the reported stds, several entries are close to overlapping; a small table of confidence intervals or p-values would strengthen the generalization claim.
  3. [§6.4, Table 3] Specify the statistical unit for the A/B (users, sessions, or events) and the meaning of '10,000 users / 6,791 conversions.' Also define 'log-Gini' in the significance sentence; it is not introduced in the metrics section.
  4. [§S3.2] The vocabulary for categorical sequence features is constructed from the union of all data splits. Although this may be standard for ID embeddings, it is a potential leakage path for categorical IDs; please justify or restrict vocabulary construction to the training split.
  5. [Fig. 7] The disagreement-reduction numbers (−22% LTV, −10% CVR logit) are presented without confidence intervals or per-bucket sample sizes. Please add these so the reader can judge whether the reduction is stable across the N≤5 subset.

Circularity Check

0 steps flagged

No significant circularity; central claims rest on external benchmarks and held-out A/B, not on self-citation or definitional fitting.

full rationale

The paper's headline LTV/CVR gains are empirical measurements against a production base, external prior-art baselines (PFD, HAPFD, unstructured LLM augmentation), and a four-week online A/B at Keeta. The load-bearing components—structured semantic profiles (Z_pre) and the hindsight target (Z_fused)—are generated by an LLM from pre/post-conversion logs and used as inputs; the student at serving time never sees Z_fused, so the prediction target is not an input to the student. The distillation objective in Eq. 9 does directly minimize student–teacher prediction/logit differences, and Fig. 7's 'distillability' measure is close to that training objective; however, Fig. 7 is a mechanism analysis comparing two distillation methods, not the paper's primary evidence, and the relative reduction is not forced because the comparison baseline also trains with distillation losses. The schema audit via gate weights is a usage measure, not a correctness proof, which weakens the mechanism claim but does not make the derivation circular. No load-bearing self-citations appear; all cited prior work is external. The cross-LLM robustness table (Table 2b) shows Qwen-9B LTV below Base, an internal inconsistency and a risk to the robustness claim, but this is a correctness concern, not circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 3 invented entities

The central result rests on three load-bearing inputs the reader does not pay for upstream: (1) the LUPI premise that a teacher reading the same 30-day post-conversion window that defines the labels can transfer usable knowledge to a pre-conversion-only student; (2) the faithfulness of GPT-OSS 120B's structured rationales, which is asserted rather than audited; (3) the gate-weighted semantic embedding being predictive beyond raw features. No wholly new physical/formal entities are postulated; the new artifacts (Z_pre, Z_fused, the schema) carry only internal evidence. Hand-set hyperparameters (distillation weights, M=10, temperatures, schema thresholds) are tuned on validation with sensitivity unreported.

free parameters (4)
  • λ_kd, λ_DE (distillation loss weights, Eq. 9) = not stated in main text (supplementary §S1)
    Hand-tuned weights controlling the logit-KD and expert-reconstruction losses; central to the training objective; no sensitivity analysis reported.
  • Number of distillation experts M and Gumbel temperature = M=10; temperature not stated
    Figure 6 shows only 2 of 10 experts receiving >1% of assignments, so M=10 is over-provisioned; no ablation on M or router temperature is reported.
  • Schema K=4 dimensions and curation thresholds = K=4; embedding-cluster cosine ≥0.85; 6–8 LLM finalists
    The deployed schema depends on outcome-stratified LLM discovery (batches balanced by high/low LTV), embedding clustering, LLM consolidation, and an expert rubric; these heuristics shape the entire semantic signal.
  • Preprocessing thresholds = τ_session=300,000 ms; L_max=5,600; cardinality τ=50
    Session segmentation, sequence truncation, and categorical/numerical partitioning constants (§S3) that shape the dynamic-feature distribution; no sensitivity check.
axioms (5)
  • domain assumption The teacher's post-conversion input S_post over [t0, t0+30d] shares the same window that defines the labels y_LTV and y_CVR, and knowledge from this future-reading teacher transfers to a student restricted to pre-conversion inputs.
    Eqs (2)–(3) in §3 define both labels on exactly the window the privileged teacher reads; the LUPI transfer premise is the entire basis of the distillation component.
  • domain assumption GPT-OSS 120B's structured rationales faithfully represent user latent traits along the four schema dimensions.
    §4.3–4.4 and prompts in S2; the paper's own motivation cites hallucination and prompt sensitivity of LLM rationales ([6],[15]) for unstructured output, yet no correctness/consistency audit of the structured rationales is performed — Fig 5 audits gate usage, not rationale accuracy.
  • domain assumption The gate-weighted semantic embedding adds predictive signal beyond raw event sequences and static features.
    §5.2; the support is component D's gain over Base in Table 1 (+0.8% LTV Gini) plus the full-system online A/B — no independent evidence isolates the semantic view.
  • domain assumption The strictly temporal split by first-conversion timestamp prevents target or privilege leakage between train/val/test.
    §S3.4 Eq (21); vocabularies are built from the union of all splits (§S3.2), and the paper assumes this does not leak label information.
  • standard math Standard ML background — Transformer attention-pooling, Gumbel-Softmax reparameterization, Huber/BCE losses, temperature-scaled KL distillation — performs as intended.
    §5, §S1; unproved background the implementation and training rely on.
invented entities (3)
  • Densified Semantic Profile Z_pre (per-user 4-dimension LLM rationale set) no independent evidence
    purpose: Serving-time input that densifies sparse pre-conversion logs into interpretable latent-trait descriptions (§4.3)
    Supported only by internal ablations (Table 1 rows D and SemRaD) and two illustrative examples (§S4.3); no external falsifiable handle, no released correctness audit.
  • Hindsight Distillation Target Z_fused no independent evidence
    purpose: Training-only reconciled supervision replacing raw concatenation of pre/post profiles (§4.4)
    Justified only by internal comparison to PFD/HAPFD (Table 1) and the student-teacher disagreement analysis (Fig 7); its value rests entirely on the LLM fusion call being correct.
  • Four-dimension behavioral schema (Temporal Resolution, Transaction Magnitude, Navigation Intent Specificity, Domain Entropy) no independent evidence
    purpose: Structures LLM reasoning and enables gate-weighted semantic encoding (§4.1)
    Constructed via outcome-stratified LLM discovery and embedding clustering; validated only by gate weights from the same trained model (Fig 5).

pith-pipeline@v1.3.0-alltime-deepseek · 17497 in / 25404 out tokens · 238843 ms · 2026-08-01T19:07:49.564917+00:00 · methodology

0 comments
read the original abstract

New-user cold-start is a critical bottleneck for e-commerce platforms: predicting user lifetime value (LTV) and conversion rate (CVR) for users with sparse interaction history. Two prior directions -- LLM-based semantic augmentation and learning using privileged information (LUPI) -- each face a key limitation. First, LLM augmentation produces unstructured rationales that are noisy and hard to operationalize in production. Second, naive student-teacher distillation can be brittle due to an information gap between the privileged teacher and the sparse student; moreover, this gap is heterogeneous across users. We propose SemRaD, a Semantic Reasoning-aware Distillation framework addressing both limitations. First, a Structured Semantic Reasoning Pipeline replaces free-form rationales with a structured schema built via a discover-curate-audit workflow, producing per user a Densified Semantic Profile (consumed by the deployed student via a Semantic-Gated Encoder that focuses on the most informative dimensions) and a Hindsight Distillation Target reconciled from pre- and post-conversion reasoning (used only at training). Second, to bridge this gap and handle its heterogeneity, a Hindsight-Aware Distillation Network transfers privileged knowledge via the hindsight target, with Distillation Experts improving transfer under per-user variability. On a large-scale industrial dataset, SemRaD lifts +1.9% LTV (Gini) and +1.0% CVR (AUROC) over a production-grade base; a four-week online A/B at Keeta confirms +1.0% LTV / +0.43% CVR. SemRaD also matches the production system's LTV using only 9% of the training data while improving CVR by 0.8%.

Figures

Figures reproduced from arXiv: 2607.17070 by Chen Bai, Chen Zhang, Hao Duong Le, Huan Li, Ke Xing, Lun Jiang, Yifei Gao.

Figure 1
Figure 1. Figure 1: Performance and data efficiency of SemRaD. (A–B) [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Semantic encoding and distillation paradigms. (A, C) prior unstructured encoding and large-gap distillation vs. (B, D) [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Structured Semantic Reasoning Pipeline (Stages 0–3): schema construction via [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Hindsight-Aware Distillation Network. Shared student/teacher encoders (Static, Dynamic, Semantic-Gated); student [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Semantic gating weights per dimension, by LTV [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Post-hoc routing of Distillation Experts. Of [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Student-Teacher Predictive Disagreement. SemRaD [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

33 extracted references · 1 canonical work pages · 1 internal anchor

  1. [1]

    Ke Bao, Jizhi Zhang, Yang Zhang, Ruobing Xie, and Xinyu Zhang. 2023. TALLRec: An Effective and Efficient Tuning Framework to Align Large Language Model with Recommendation. arXiv preprint arXiv:2305.12300

    Pith/arXiv arXiv 2023

  2. [2]

    Chunfeng Gao, Yongfeng Zhang, Fuzhen Zhuang, Yanchi Liu, Haoyuan Liu, Xiangnan He, and Qing Li. 2023. Chat-REC: Towards Interactive and Explainable Recommendation via Large Language Models. arXiv preprint arXiv:2305.14524

    Pith/arXiv arXiv 2023

  3. [3]

    Shijie Geng, Shuchang Liu, Zuohui Fu, Ye Yuan, Zhaochun Ren, and Gerard de Melo. 2022. Personalized Prompt for Recommendation. arXiv preprint arXiv:2203.13366

    Pith/arXiv arXiv 2022

  4. [4]

    Huifeng Guo, Ruiming Tang, Yunming Ye, Zhenguo Li, and Xiuqiang He. 2017. DeepFM: A Factorization-Machine based Neural Network for CTR Prediction. InProceedings of the 26th International Joint Conference on Artificial Intelligence (IJCAI 2017). 1725–1731

    2017

  5. [5]

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. 2015. Distilling the Knowledge in a Neural Network. arXiv preprint arXiv:1503.02531

    Pith/arXiv arXiv 2015

  6. [6]

    Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. 2023. Survey of Hallucination in Natural Language Generation.Comput. Surveys55, 12 (2023), 1–38. doi:10.1145/ 3571730

    2023

  7. [7]

    SeongKu Kang, Junyoung Hwang, Wonbin Kweon, and Hwanjo Yu. 2020. DE-RRD: A Knowledge Distillation Framework for Recommender System. InProceedings of the 29th ACM International Conference on Information and Knowledge Management (CIKM ’20). 605–614. doi:10.1145/3340531.3412005

    arXiv 2020

  8. [8]

    Wang-Cheng Kang and Julian McAuley. 2018. Self-Attentive Sequential Recom- mendation.arXiv preprint arXiv:1808.09781(2018)

    Pith/arXiv arXiv 2018

  9. [9]

    Jieyong Kim, Hyunseo Kim, Hyunjin Cho, SeongKu Kang, Buru Chang, Jiny- oung Yeo, and Dongha Lee. 2024. Review-driven Personalized Preference Rea- soning with Large Language Models for Recommendation. arXiv preprint arXiv:2408.06276

    Pith/arXiv arXiv 2024

  10. [10]

    Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2023. Large Language Models are Zero-Shot Reasoners. arXiv preprint arXiv:2205.11916. arXiv:2205.11916 [cs.CL]

    Pith/arXiv arXiv 2023

  11. [11]

    David Lopez-Paz, Léon Bottou, Bernhard Schölkopf, and Vladimir Vapnik. 2016. Unifying distillation and privileged information. InInternational Conference on Learning Representations (ICLR)

    2016

  12. [12]

    Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H. Chi. 2018. Entire Space Multi-Task Model: An Effective Approach for Estimating Post-Click Conversion Rate. arXiv preprint arXiv:1804.07931

    Pith/arXiv arXiv 2018

  13. [13]

    Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. 2015. FitNets: Hints for Thin Deep Nets. In International Conference on Learning Representations (ICLR)

    2015

  14. [14]

    Schein, Alexandrin Popescul, Lyle H

    Andrew I. Schein, Alexandrin Popescul, Lyle H. Ungar, and David M. Pennock

  15. [15]

    Melanie Sclar, Yejin Choi, Yulia Tsvetkov, and Alane Suhr. 2024. Quantifying Language Models’ Sensitivity to Spurious Features in Prompt Design or: How I learned to start worrying about prompt formatting. InInternational Conference on Learning Representations (ICLR)

    2024

  16. [16]

    Yunpeng Song. 2024. Predicting User Behavior in Smart Spaces with LLM- Enhanced Logs and Personalized Prompts (Data Description). arXiv preprint arXiv:2412.12653. doi:10.48550/arXiv.2412.12653

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2412.12653 2024

  17. [17]

    Chi, and Xinyang Yi

    Alicia Tsai, Adam Kraft, Long Jin, Chenwei Cai, Anahita Hosseini, Taibai Xu, Zemin Zhang, Lichan Hong, Ed H. Chi, and Xinyang Yi. 2024. Leveraging LLM Reasoning Enhances Personalized Recommender Systems. InFindings of the Association for Computational Linguistics: ACL 2024. 13176–13188. doi:10.18653/ v1/2024.findings-acl.780

    2024

  18. [18]

    Vladimir Vapnik and Akshay Vashist. 2009. A New Learning Paradigm: Learning Using Privileged Information.Neural Networks22, 5–6 (2009), 544–557. doi:10. 1016/j.neunet.2009.06.042

    2009

  19. [19]

    Maksims Volkovs, Guangwei Yu, and Tomi Poutanen. 2017. DropoutNet: Ad- dressing Cold Start in Recommender Systems. InProceedings of the 11th ACM Conference on Recommender Systems (RecSys ’17). 88–96

    2017

  20. [20]

    Ruoxi Wang, Bin Fu, Gang Fu, and Mingliang Wang. 2017. Deep & Cross Network for Ad Click Predictions. InProceedings of the 1st Workshop on Deep Learning for Recommender Systems (DLRS at RecSys 2017)

    2017

  21. [21]

    Le Wu, Xiangnan He, Yunshan Ma, Liwei Chen, Shuqin Li, and Tat-Seng Chua

  22. [22]

    Chen Xu, Quan Li, Junfeng Ge, Jinyang Gao, Xiaoyong Yang, Changhua Pei, Fei Sun, Jian Wu, Hanxiao Sun, and Wenwu Ou. 2020. Privileged Features Distillation at Taobao Recommendations. InProceedings of the 26th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD ’20). 2590–2598. doi:10.1145/ 3394486.3403309

    arXiv 2020

  23. [23]

    Huining Yuan, Wenpeng Zhang, Zijie Hao, and Zengde Deng. 2025. Hardness- aware Privileged Features Distillation with Latent Alignment for CVR Prediction. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD ’25). 5182–5193. doi:10.1145/3711896.3737231

    arXiv 2025

  24. [24]

    Weiqi Yue, Yuyu Yin, Xin Zhang, Binbin Shi, Tingting Liang, and Jian Wan. 2025. CoT4Rec: Revealing User Preferences Through Chain of Thought for Recom- mender Systems. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 13142–13151. doi:10.1609/aaai.v39i12.33434

    doi:10.1609/aaai.v39i12.33434 2025

  25. [25]

    Guorui Zhou, Chengru Song, Xiaoqiang Zhu, Ying Fan, Han Zhu, Xiao Ma, Yanghui Yan, Junqi Jin, Han Li, and Kun Gai. 2018. Deep Interest Network for Click-Through Rate Prediction. InProceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (KDD ’18). 1059–

    2018

  26. [26]

    Section X

    Guorui Zhou, Xiaoqiang Zhu, Chenru Song, Ying Fan, Han Zhu, Xiao Ma, Yanghui Yan, Junqi Jin, Han Li, and Kun Gai. 2019. Deep Interest Evolution Network for Click-Through Rate Prediction. InProceedings of the AAAI Conference on Artificial Intelligence. GenAI Usage Disclosure We disclose two forms of generative AI usage in this work.(i) As a core methodolog...

    2019

  27. [30]

    Temporal Resolution: Analyze the latency between discovery and action

  28. [31]

    Transaction Magnitude: Evaluate the relative value tier of interacted items

  29. [32]

    Navigation Intent Specificity: Determine the linearity of the user journey

  30. [33]

    Ground Truth

    Domain Entropy: Measure the distribution variance of interacted categories. ... For each dimension, provide a specificReasoning Rationalebased on the evidence. Explain WHY you assigned the score. S2.2 Teacher Prompt (Hindsight Fusion) The following prompt is used to generate the “Ground Truth” rea- soning by fusing the pre-conversion hypothesis with the p...

    2018

  31. [1068]

    doi:10.1145/3219819.3219823

    arXiv

  32. [2002]

    InProceedings of the 25th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’02)

    Methods and Metrics for Cold-Start Recommendations. InProceedings of the 25th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’02). 253–260. doi:10.1145/564376.564421

    arXiv

  33. [2023]

    arXiv preprint arXiv:2305.19860

    A Survey on Large Language Models for Recommender Systems. arXiv preprint arXiv:2305.19860

    Pith/arXiv arXiv