Similarity = Value? Consultation Value Assessment and Alignment for Personalized Search

Chenglei Shen; Jianping Fan; Jun Xu; Ming He; Teng Shi; Weicong Qin; Weijie Yu; Xiao Zhang; Yi Xu

REVIEW 2 major objections 4 minor 1 cited by

Similarity = Value? Consultation Value Assessment and Alignment for Personalized Search

T0 review · 2 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Semantic similarity alone is a poor proxy for consultation value in personalized e-commerce search; a value-aware model that scores historical consultations by time decay, scenario scope, and posterior actions outperforms baselines in…

desk verdict VAPS is a plausible extension of MAPS with a value-gating idea, but the main experiments may be invalid because posterior-action value is computed from actions after the target search without any stated test-time safeguard. read the letter →

arxiv 2506.14437 v1 pith:REAICXWB submitted 2025-06-17 cs.IR

Weicong Qin , Yi Xu , Weijie Yu , Teng Shi , Chenglei Shen , Ming He , Jianping Fan , Xiao Zhang

show 1 more author

Jun Xu

This is my paper

classification cs.IR

keywords personalizedsearche-commerceconsultationvalueassessmentsemanticsimilaritytimedecayposterioractionscenarioscoperanking

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper targets personalized product search in e-commerce, where a user's past conversations with an AI assistant are used to personalize later searches. The authors argue that the standard practice of matching old consultations to the current query by semantic similarity confuses similarity with usefulness: a similar past consultation can be outdated, out of the product domain, or never acted on, and such consultations can even hurt ranking. To replace similarity, the paper defines consultation value as a weighted combination of three heuristic scores: time decay, scenario scope, and posterior action value, the last being the presence of later clicks, searches, or purchases. The proposed model, VAPS, filters each user's consultation history by this score and trains a consultation-action interaction module with an alignment objective, so that valuable consultations are selectively incorporated. The paper's claim, stated sympathetically, is that label-free consultation value assessment is a viable alternative to semantic-only alignment.

What carries the argument

The load-bearing object is the aggregated consultation value in Eq. (1): a label-free, interpretable scoring function that ranks and filters consultation history before the encoder sees it. The model side then re-encodes the surviving consultations with the Consultation-Action Interaction (CAI) module, a cross-attention block in which consultations act as queries and user actions (with action-type and discretized time embeddings) act as keys and values, so the model can see whether and how strongly a consultation was later supported by clicks, searches, or purchases. An explicit value-assessment alignment loss $L_{\mathrm{VA}}$ supervises those attention scores against the data-side consultation-action mapping. Together the three pieces, data-side scoring, CAI cross-attention, and alignment loss, let the model treat consultation value as something to be assessed and aligned rather than assumed from similarity.

What would settle it

A decisive test would hold the consultation set and sequence budget fixed and compare ranking quality when consultations are selected by the paper's value score versus by top semantic similarity; if the value-selected set does not beat the similarity-selected set under the same encoder, the central claim fails. A second, simpler check is to rerun the evaluation counting only actions that occurred before the target search as posterior actions; if the reported gains disappear, the improvement comes from chronological leakage rather than from value assessment.

Watch

Extended reading notes

Core claim

The central claim is that consultation value should be treated as a measurable property distinct from semantic relatedness. On the data side, the paper computes, for each consultation $c$ relative to a later search $s$, an aggregated score $O^{s,c}_u=(1-\lambda_1)O^{s,c}_{\mathrm{time}}+\lambda_1(\lambda_2 O^{c}_{\mathrm{scope}}+(1-\lambda_2)O^{c}_{\mathrm{action}})$, where the time component decays exponentially with the hour-level gap, the scenario component comes from the share of inverted-index product terms the consultation contains, and the action component is a scarcity-weighted count of buy, click, and search actions that the consultation can be linked to. Consultations are ranked by this score and filtered to the model's length budget. On the model side, VAPS enriches each consultation embedding with temporal and action-type information, uses cross-attention so consultations can draw posterior support from user actions, and adds an alignment loss that supervises the attention scores with the assessed consultation-action pairs. The paper reports that this combination outperforms all compared personalized search, retrieval, and multi-scenario baselines on both a commercial dataset and an Amazon-derived dataset, with roughly 2 percent gains on the commercial set and larger gains on Amazon.

Load-bearing premise

The load-bearing premise is that the three hand-built score functions in Eq. (1) capture what makes a consultation genuinely useful for a later search, and in particular that posterior actions can be counted from the user's later interaction history without leaking future labels into the evaluation.

Editorial extensions

If this is right

If the central claim is right, consultation logs can be pruned by value before encoding, cutting noise and shortening the sequence the ranking model must process, which directly addresses the length limit $L_{\mathrm{seq}}$.
The three-part value score provides weak supervision for any consultation-enhanced retrieval model, since it does not require human value labels.
The explicit posterior-action signal means a consultation's usefulness can be updated as the user acts, so the same consultation may carry different value for different later searches.
Because the action component is scarcity-weighted, the framework can absorb other action types such as favorites or add-to-cart by adding them to the action set, extending beyond buy, click, and search.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A natural extension the paper does not pursue is to make the value scores online and streaming, recomputing posterior-action value as new actions arrive rather than from a fixed offline inverted table, and testing whether refresh rate changes ranking quality.
The same value decomposition could transfer to recommendation or conversational systems where consultation logs exist, since nothing in the framework is specific to the search ranking loss.
A stricter test of the paper's causal story would compare, at equal sequence budget, consultations selected by the value score against consultations selected by top semantic similarity; the paper's Fig. 1 scatter plot is suggestive but correlational.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The core idea here is worth a moment: VAPS moves beyond MAPS's semantic-similarity alignment by scoring historical consultations on three value axes — time decay, scenario scope, and posterior action — and using that score to rank-and-filter which consultations feed a personalized search encoder. The packaging is new even if the individual ingredients are not, and the CAI cross-attention plus the LV_A alignment loss is a coherent way to make the model sensitive to post-consultation user actions. The ablations are reasonably informative, and the paper is transparent about its hyperparameters and limitations.

The soft spot is not the heuristics; it's the evaluation. The paper defines posterior-action value O^{c}_{action} from D(ts)_u, the interaction history after the target search s. It builds an offline inverted table mapping consultations to subsequent actions (App. A.1), and it never says that for validation or test queries this component is restricted to actions that occur before t_s. If the table covers the full log, then for a test query the model is filtering consultations based on items the user later searched, clicked, or bought. That is label leakage, and it would inflate every number in Tables 2 and 3. The paper does not address this at all. It is the load-bearing flaw.

Secondary issues: no error bars or confidence intervals beyond a vague t-test note; the introduction claims synthetic-data experiments that never appear; the code link is anonymous with no versioned release. These are fixable friction, not fatal on their own.

If the authors can confirm that posterior-action value is computed strictly from pre-query actions at evaluation time (or re-run with that constraint), the core contribution — that value-based consultation selection beats similarity-based selection — is plausible and worth a serious test. As written, the central evidence is compromised.

I'd send this to peer review because the question is important and the flaw is addressable, but I'd ask the referee to hold the authors to a clear statement of test-time data handling. The paper is for researchers in e-commerce search and consultation-based personalization; it does not reorganize the field, but it is a legitimate incremental step if the numbers hold.

Referee Report

2 major / 4 minor

Summary. The manuscript proposes a consultation-value assessment framework for personalized e-commerce search, replacing pure semantic similarity with three hand-designed value dimensions: time decay, scenario scope, and posterior action value. It introduces VAPS, which uses these scores to rank and filter historical consultations and then trains a model with a consultation-action interaction (CAI) module and a value-alignment loss. Experiments on a commercial dataset and the Amazon dataset compare VAPS against personalized search, multi-scenario, and retrieval baselines, reporting consistent gains in ranking and retrieval metrics. The core claim is that value-aware consultation selection improves personalized search beyond semantic-similarity alignment.

Significance. The paper addresses a relevant and under-explored problem: how to exploit user-AI consultation logs for personalized product search. The proposed decomposition of consultation value into time, scope, and posterior-action components is intuitive, and the CAI module is a reasonable way to inject these signals into a search model. The authors provide code and use a public benchmark in addition to a commercial dataset, and the ablations indicate that each component contributes. The complexity of VAPS matches prior state-of-the-art methods. However, the central empirical claim is currently undermined by a potentially serious temporal-leakage issue in the computation of posterior-action value, so the significance is conditional on a corrected evaluation protocol.

major comments (2)

[Sec. 4.1.3, App. A.1, App. A.2] The posterior-action value is computed from interactions after the target search, which creates a risk of label leakage in the evaluation. In Sec. 4.1, D^{(t_s)}_u is defined as the item interaction history occurring after search s, and O^c_{action} is computed from this set. Appendix A.1 builds the offline inverted table from consumer actions and links them to consultations without any temporal restriction relative to the query, and Appendix A.2 uses this table to assign posterior-action frequencies. The paper nowhere states that for validation/test queries the posterior-action component is zeroed or restricted to actions strictly before t_s, nor that the CAI module receives a D_u truncated at the query timestamp. Since test queries are drawn from the same chronological logs, future clicks, searches, and buys can determine which consultations are selected and how they attend to actions, potentially inflating all gains in Tables 2-4. Please re-run the experiments with a strict temporal cutoff (actions before the query only, or posterior-action value explicitly zeroed at test time) and report results under that protocol.
[Sec. 4.2.2 and Sec. 4.2.4] Even if the value-based filter were fixed, the model itself may still consume future actions at inference. The CAI module uses all a_i in D_u with no time restriction, and the cascade encoder receives E_items built from item interactions; to compute p(v'|H_u, s^{(N+1)}_u, V'_u), the paper does not specify that D_u is cut off before the current query. At deployment, only interactions before the query are available, so the experimental protocol must state exactly which timestamp boundary is applied to D_u for each test example. Without this, the ranking results do not reflect a deployable system.

minor comments (4)

[Introduction, Contribution bullet, and Sec. 5] The text repeatedly claims experiments on 'real and synthetic data' and on 'real-world commercial and synthetic datasets,' but Section 5 reports only the Commercial and Amazon datasets, both real. There is no synthetic dataset experiment. Please remove or correct this claim, or add the missing synthetic experiments.
[Tables 2-4] The tables report only point estimates. The paired t-test significance is asserted with a dagger, but no standard deviations, confidence intervals, or exact p-values are provided. In addition, several metrics are not marked as significant (for example, Commercial HR@20 and Amazon HR@50), which makes the abstract's 'consistently outperforms' claim stronger than the reported evidence. Please report variance or significance details for all metrics.
[Sec. B.2] Several hyperparameters are only described as tuned over a range, without reporting the final selected values. In particular, the optimal τ1, τ2, λ3, and λ4 are not given. This makes reproduction difficult. Please list the final configuration for each dataset.
[Sec. 4.2.3 and Sec. 5.3] The L_VA loss is supervised with pseudo-labels generated by the same hand-designed heuristics that define consultation value. This is not circular for the final external ranking metrics, but the value assessment itself is not independently validated. A useful control would be to compare against a model trained with random or purely semantic value labels, to show that the specific three-dimensional definition, rather than any extra supervision signal, drives the gains.

Circularity Check

1 steps flagged · score 6.0 of 10

Posterior-action value is computed from interactions after the target search, so the value labels used to filter consultations and supervise attention can be built from the same future actions the ranking metrics evaluate; the paper never masks these at test time.

fitted input called prediction [Sec. 4.1.3, Sec. 4.1.4 (Eq. 1), and App. A.1/A.2; evaluation in Sec. 5.1/5.2]
"For ∀s∈ S_u, there exist a consultation set C^(t_s)_u ⊆ C_u that occurred before s and an item interaction history D^(t_s)_u ⊆ D_u that occurred after s, where t_s is the timestamp of s. ... O^c_action = Σ_a γ_a R^{T_a}_c ... T_a ⊆ D^(t_s)_u is the subset of interaction histories in D^(t_s)_u with action type a. ... C'_u = Rank&Filter O^{s,c}_u(C_u, L_seq, s, u)."

The consultation value for a search s is computed from item interactions that occur after s. Eq. (1) aggregates this posterior action value into O^{s,c}_u, which then selects the consultations used by the model and supervises CAI through L_VA. For the held-out queries in Sec. 5, those post-s interactions are the same future click/buy/search events used to define the ground-truth ranking and retrieval labels. The paper never states that the offline inverted table in App. A.1 is restricted to actions strictly before each evaluation query's timestamp, so the value signal can encode the target labels by construction. The reported gains in Tables 2-3 are therefore not a clean test of consultation value independent of the labels being predicted.

full rationale

The only formal circularity I can exhibit from the paper's own equations is the leakage loop above: D^(t_s)_u is defined as occurring after s, and O^c_action is computed from it, so the 'value' input can contain the future actions whose prediction is being evaluated. Apart from this, the paper is not circular in the usual self-citation sense: the three value heuristics are stated as empirical functions with tunable hyperparameters, the model is tested against external baselines on standard ranking and retrieval metrics, and no uniqueness theorem or ansatz is smuggled in through the same authors' MAPS citation. The reliance on MAPS for MoAE and the cascade encoder is architectural reuse and is benchmarked, not load-bearing. If the intended protocol was to compute posterior-action values only from actions strictly before each evaluation query, then the leakage loop disappears; but as written, the derivation uses post-search actions as both input and label. This is a single central circular/leakage step, so score 6 rather than higher.

Assumptions & free parameters 8 free parameters · 6 assumptions · 0 invented entities

The framework's central quantity, consultation value, is constructed from hand-chosen components and tuned weights rather than derived from first principles or external labels. The free parameters are tuned on the evaluation datasets, and the axioms are domain assumptions inherited from psychology, consumer behavior, and inverted-index practice. No new physical or abstract entities are introduced beyond the value score itself, which is computed from existing interaction data.

free parameters (8)

alpha (temporal decay factor) = 0.99 (tuned over {0.98, 0.99, 0.999, 0.9999})
Controls O_time in Sec. 4.1.1; chosen from empirical experience and validation tuning in App. B.2.
lambda1 (aggregation weight) = 0.5 (tuned over [0, 1])
Balances time-decay value against scope/action value in Eq. (1); optimal value reported in Sec. 5.6.
lambda2 (scope versus action weight) = 0.3 (tuned over [0, 1])
Balances scenario scope value against posterior action value in Eq. (1); optimal value reported in Sec. 5.6.
lambda_thresh (scope threshold) = 4
Threshold for f_scope in Sec. 4.1.2; set in App. B.2 after empirical selection.
tau1 and tau2 (softmax temperatures) = tuned over {0:0.1:1}
Control sharpness of contrastive losses L_VA and L_search in Sec. 4.2.3 and Sec. 4.2.4.
lambda3 (attention residual and loss weight) = tuned over {0:0.1:1}
Used in the CAI skip connection and in L_final; the paper reuses the same symbol for both roles in Sec. 4.2.2 and Sec. 4.2.4.
lambda4 (L2 regularization) = tuned over {0:0.1:1}
Weight decay in the final loss L_final in Sec. 4.2.4.
Action time window for inverted retrieval = 14 days
Chosen in App. A.1 because of dataset sparsity; affects which actions are linked to each consultation.

assumptions (6)

domain assumption Exponential forgetting curve models consultation usefulness: O_time = alpha^(ts-tc).
Sec. 4.1.1 assumes recent consultations are more valuable, following Rubin and Wenzel (1996).
domain assumption Textual overlap between actions and consultations is sufficient to link them.
App. A.1 builds the action-consultation inverted table by checking whether action text appears in the consultation text; this may miss semantically related but lexically different actions.
domain assumption Scarcer consumer actions are more valuable than common ones.
Sec. 4.1.3 sets gamma_a inversely proportional to action frequency, citing Verhallen (1982) and Teubner and Graul (2020).
domain assumption Inverted-index term counts identify scenario scope.
Sec. 4.1.2 assumes consultations with more product-term hits are in-scope and therefore valuable.
ad hoc to paper Aggregate consultation value is a weighted sum of time, scope, and action value.
Eq. (1) is proposed without external value labels or a derivation; the weighting is validated only through downstream ranking performance.
domain assumption Semantic similarity alone is an inadequate proxy for consultation value.
Motivated by Fig. 1, which shows a scatter plot without a reported correlation or significance test.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Similarity = Value? Consultation Value Assessment and Alignment for Personalized Search." pith.science (2026). https://pith.science/paper/REAICXWB

@misc{pith2026250614437,
  author       = {Pith},
  title        = {Pith review of: Similarity = Value? Consultation Value Assessment and Alignment for Personalized Search},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/REAICXWB}},
  note         = {Machine review of arXiv:2506.14437}
}

read the original abstract

Personalized search systems in e-commerce platforms increasingly involve user interactions with AI assistants, where users consult about products, usage scenarios, and more. Leveraging consultation to personalize search services is trending. Existing methods typically rely on semantic similarity to align historical consultations with current queries due to the absence of 'value' labels, but we observe that semantic similarity alone often fails to capture the true value of consultation for personalization. To address this, we propose a consultation value assessment framework that evaluates historical consultations from three novel perspectives: (1) Scenario Scope Value, (2) Posterior Action Value, and (3) Time Decay Value. Based on this, we introduce VAPS, a value-aware personalized search model that selectively incorporates high-value consultations through a consultation-user action interaction module and an explicit objective that aligns consultations with user actions. Experiments on both public and commercial datasets show that VAPS consistently outperforms baselines in both retrieval and ranking tasks.

Figures

Figures reproduced from arXiv: 2506.14437 by the authors.

**Figure 1.** NDCG@20 difference with/without consultations vs. normalized semantic similarity, using MAPS (Qin et al., 2025) on Amazon dataset. Semantic similarity does not consistently reflect the value of consultations for improving search. Recent studies can be broadly divided into two directions. The first mainly extracts user-specific preferences from user interactions (across multiple scenarios) to personalize search re… view at source ↗

**Figure 2.** Illustration of consumer consultation patterns and action verification in E-commerce platform. Yellow [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Overview of VAPS. 4.1.2 Scenario Scope Value An observation is that out-of-scope consultations (e.g., discussions about politics) provide little value to e-commerce platforms, while consultations with scenario-specific terms are more likely to accurately represent user needs. To operationalize this insight, we quantify scenario scope value Oc scope with scenario-specific terms. Terms can be directly obtained from t… view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Value distribution visualization. Fig. 4a reveals that the commercial dataset exhibits lower normalized scores overall, attributable to its sparser user interactions, consistent with the fact that its primary product categories are nondaily necessities. In Fig. 4b, w…

**Figure 6.** Figure 6: Impact of different hyperparameters on Ama [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

**Figure 8.** Figure 8: Examples of consultations on the Commercial dataset. Time Decay Value. While these design choices are empirically justified, some aspects may invite scrutiny—particularly the heuristic nature of value functions, hyperparameter sensitivity, and dataset constraints. Her…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Bridging Search and Recommendation through Latent Cross Reasoning
cs.IR 2025-08 conditional novelty 5.0 of 10

A latent cross reasoning model with contrastive learning and GRPO reinforcement learning improves search-enhanced recommendation on Qilin and KuaiSAR.

Reference graph

Works this paper leans on

40 extracted references · 26 canonical work pages · cited by 1 Pith paper

[1]

Qingyao Ai, Daniel N Hill, SVN Vishwanathan, and W Bruce Croft. 2019 a . A zero attention model for personalized product search. In Proceedings of the 28th ACM International Conference on Information and Knowledge Management, pages 379--388

work page 2019
[2]

Qingyao Ai, Yongfeng Zhang, Keping Bi, Xu Chen, and W Bruce Croft. 2017. Learning a hierarchical embedding model for personalized product search. In Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 645--654

work page 2017
[3]

Qingyao Ai, Yongfeng Zhang, Keping Bi, and W Bruce Croft. 2019 b . Explainable product search with a dynamic relation embedding model. ACM Transactions on Information Systems (TOIS), 38(1):1--29

work page 2019
[4]

Keping Bi, Qingyao Ai, and W Bruce Croft. 2020. A transformer-based embedding model for personalized product search. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1521--1524

work page 2020
[5]

Hongru Cai, Yongqi Li, Wenjie Wang, Fengbin Zhu, Xiaoyu Shen, Wenjie Li, and Tat-Seng Chua. 2024. https://arxiv.org/abs/2410.17236 Large language models empowered personalized web agents . Preprint, arXiv:2410.17236

work page Pith review arXiv 2024
[6]

Jianneng Cao, Panagiotis Karras, Panos Kalnis, and Kian-Lee Tan. 2011. Sabre: a sensitive attribute bucketization and redistribution framework for t-closeness. The VLDB Journal, 20:59--81

work page 2011
[7]

Matteo Catena, Craig Macdonald, and Iadh Ounis. 2014. On inverted index compression for search engine efficiency. In Advances in Information Retrieval: 36th European Conference on IR Research, ECIR 2014, Amsterdam, The Netherlands, April 13-16, 2014. Proceedings 36, pages 359--371. Springer

work page 2014
[8]

Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. 2024. Bge m3-embedding: Multi-lingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation. arXiv preprint arXiv:2402.03216

arXiv 2024

Show all 40 references

[9]

Jiayi Chen, Wen Wu, Liye Shi, Wei Zheng, and Liang He. 2023. Long-tail session-based recommendation from calibration. Applied Intelligence, 53(4):4685--4702

2023
[10]

Shitong Dai, Jiongnan Liu, Zhicheng Dou, Haonan Wang, Lin Liu, Bo Long, and Ji - Rong Wen. 2023. Contrastive learning for user sequence representation in personalized product search. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD 2...

2023
[11]

Rafael Ferreira, David Semedo, and Jo \ a o Magalh \ a es. 2023. Rating prediction in conversational task assistants with behavioral and conversational-flow features. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retri...

2023
[12]

Huifeng Guo, Ruiming Tang, Yunming Ye, Zhenguo Li, and Xiuqiang He. 2017. Deepfm: a factorization-machine based neural network for ctr prediction. In Proceedings of the 26th International Joint Conference on Artificial Intelligence, pages 1725--1731

2017
[13]

Rebecca Hamilton and Linda L Price. 2019. Consumer journeys: Developing consumer-based strategy

2019
[14]

Henderi Henderi, Tri Wahyuningsih, and Efana Rahwanto. 2021. Comparison of min-max normalization and z-score normalization in the k-nearest neighbor (knn) algorithm to test the accuracy of types of breast cancer. International Journal of Informatics and Information Systems, 4(...

2021
[15]

Geoffrey Hinton. 2015. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531

2015 arXiv
[16]

Michael Jahrer, Andreas T \"o scher, and Robert Legenstein. 2010. Combining predictions for accurate recommender systems. In Proceedings of the 16th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 693--702

2010
[17]

Chen Jin, Luyi Yang, and Kartik Hosanagar. 2023. To brush or not to brush: Product rankings, consumer search, and fake orders. Information Systems Research, 34(2):532--552

2023
[18]

Pooria Joulani, Andras Gyorgy, and Csaba Szepesv \'a ri. 2013. Online learning under delayed feedback. In International conference on machine learning, pages 1453--1461. PMLR

2013
[19]

Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980

2014 arXiv
[20]

Quoc Le and Tomas Mikolov. 2014. Distributed representations of sentences and documents. In International conference on machine learning, pages 1188--1196. PMLR

2014
[21]

Jiacheng Li, Yujie Wang, and Julian McAuley. 2020. https://doi.org/10.1145/3336191.3371786 Time interval aware self-attention for sequential recommendation . In Proceedings of the 13th International Conference on Web Search and Data Mining, WSDM '20, page 322–330, New York, NY...

2020
[22]

Hezheng Lin, Xing Cheng, Xiangyu Wu, and Dong Shen. 2022. Cat: Cross attention in vision transformer. In 2022 IEEE international conference on multimedia and expo (ICME), pages 1--6. IEEE

2022
[23]

Fengran Mo, Abbas Ghaddar, Kelong Mao, Mehdi Rezagholizadeh, Boxing Chen, Qun Liu, and Jian-Yun Nie. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.135 CHIQ : Contextual history enhancement for improving query rewriting in conversational search . In Proceedings of the 2024 ...

2024 doi
[24]

Jianmo Ni, Jiacheng Li, and Julian McAuley. 2019. Justifying recommendations using distantly-labeled reviews and fine-grained aspects. In EMNLP-IJCNLP

2019
[25]

Weicong Qin, Yi Xu, Weijie Yu, Chenglei Shen, Ming He, Jianping Fan, Xiao Zhang, and Jun Xu. 2025. https://arxiv.org/abs/2503.01711 Maps: Motivation-aware personalized search via llm-driven consultation alignment . Preprint, arXiv:2503.01711

2025 arXiv
[26]

Stephen Robertson, Hugo Zaragoza, et al. 2009. The probabilistic relevance framework: Bm25 and beyond. Foundations and Trends in Information Retrieval , 3(4):333--389

2009
[27]

David C Rubin and Amy E Wenzel. 1996. One hundred years of forgetting: A quantitative description of retention. Psychological review, 103(4):734

1996
[28]

Teng Shi, Zihua Si, Jun Xu, Xiao Zhang, Xiaoxue Zang, Kai Zheng, Dewei Leng, Yanan Niu, and Yang Song. 2024. Unisar: Modeling user transition behaviors between search and recommendation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development ...

2024
[29]

Teng Shi, Jun Xu, Xiao Zhang, Xiaoxue Zang, Kai Zheng, Yang Song, and Enyun Yu. 2025. Unified generative search and recommendation. arXiv preprint arXiv:2504.05730

2025 arXiv
[30]

Zihua Si, Zhongxiang Sun, Xiao Zhang, Jun Xu, Xiaoxue Zang, Yang Song, Kun Gai, and Ji - Rong Wen. 2023. When search meets recommendation: Learning disentangled search representation for recommendation. In Proceedings of the 46th International ACM SIGIR Conference on Research ...

2023
[31]

Timm Teubner and Antje Graul. 2020. Only one room left! how scarcity cues affect booking intentions on hospitality platforms. Electronic Commerce Research and Applications, 39:100910

2020
[32]

Theo MM Verhallen. 1982. Scarcity and consumer choice behavior. Journal of Economic Psychology, 2(4):299--322

1982
[33]

Jiayi Xie, Shang Liu, Gao Cong, and Zhenzhong Chen. 2023. Unifiedssr: A unified framework of sequential search and recommendation. arXiv preprint arXiv:2310.13921

2023 arXiv
[34]

Shunyu Yao, Noah Shinn, Pedram Razavi, and Karthik Narasimhan. 2024. https://arxiv.org/abs/2406.12045 -bench: A benchmark for tool-agent-user interaction in real-world domains . Preprint, arXiv:2406.12045

2024 arXiv
[35]

Jingying Zeng, Hui Liu, Zhenwei Dai, Xianfeng Tang, Chen Luo, Samarth Varshney, Zhen Li, and Qi He. 2025. Cite before you speak: Enhancing context-response grounding in e-commerce conversational llm-agents. arXiv preprint arXiv:2503.04830

2025 arXiv
[36]

Changshuo Zhang, Teng Shi, Xiao Zhang, Qi Liu, Ruobing Xie, Jun Xu, and Ji-Rong Wen. 2024 a . Modeling domain and feedback transitions for cross-domain sequential recommendation. arXiv preprint arXiv:2408.08209

2024 arXiv
[37]

Changshuo Zhang, Teng Shi, Xiao Zhang, Yanping Zheng, Ruobing Xie, Qi Liu, Jun Xu, and Ji-Rong Wen. 2024 b . Qagcf: Graph collaborative filtering for q&a recommendation. arXiv preprint arXiv:2406.04828

2024 arXiv
[38]

Kepu Zhang, Teng Shi, Sunhao Dai, Xiao Zhang, Yinfeng Li, Jing Lu, Xiaoxue Zang, Yang Song, and Jun Xu. 2024 c . Saqrec: Aligning recommender systems to user satisfaction via questionnaire feedback. In Proceedings of the 33rd ACM International Conference on Information and Kno...

2024
[39]

online" 'onlinestring :=

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...
[40]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Qingyao Ai, Daniel N Hill, SVN Vishwanathan, and W Bruce Croft. 2019 a . A zero attention model for personalized product search. In Proceedings of the 28th ACM International Conference on Information and Knowledge Management, pages 379--388

work page 2019

[2] [2]

Qingyao Ai, Yongfeng Zhang, Keping Bi, Xu Chen, and W Bruce Croft. 2017. Learning a hierarchical embedding model for personalized product search. In Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 645--654

work page 2017

[3] [3]

Qingyao Ai, Yongfeng Zhang, Keping Bi, and W Bruce Croft. 2019 b . Explainable product search with a dynamic relation embedding model. ACM Transactions on Information Systems (TOIS), 38(1):1--29

work page 2019

[4] [4]

Keping Bi, Qingyao Ai, and W Bruce Croft. 2020. A transformer-based embedding model for personalized product search. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1521--1524

work page 2020

[5] [5]

Hongru Cai, Yongqi Li, Wenjie Wang, Fengbin Zhu, Xiaoyu Shen, Wenjie Li, and Tat-Seng Chua. 2024. https://arxiv.org/abs/2410.17236 Large language models empowered personalized web agents . Preprint, arXiv:2410.17236

work page Pith review arXiv 2024

[6] [6]

Jianneng Cao, Panagiotis Karras, Panos Kalnis, and Kian-Lee Tan. 2011. Sabre: a sensitive attribute bucketization and redistribution framework for t-closeness. The VLDB Journal, 20:59--81

work page 2011

[7] [7]

Matteo Catena, Craig Macdonald, and Iadh Ounis. 2014. On inverted index compression for search engine efficiency. In Advances in Information Retrieval: 36th European Conference on IR Research, ECIR 2014, Amsterdam, The Netherlands, April 13-16, 2014. Proceedings 36, pages 359--371. Springer

work page 2014

[8] [8]

Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. 2024. Bge m3-embedding: Multi-lingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation. arXiv preprint arXiv:2402.03216

arXiv 2024