Pith. sign in

REVIEW 4 major objections 6 minor 48 references

A Text-Based Recommender System that Leverages Explicit Affective State Preferences

T0 review · 4 major / 6 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 A recommender that takes desired feelings as input ranks the user's actual next read first 67.6% of the time in a 20-book test.

desk verdict New task and a substantial dataset, but the evaluation leaks the target through the AC query, so the headline results do not support the central claim. read the letter →

arxiv 2505.20190 v2 pith:B6HPML4I submitted 2025-05-26 cs.IR cs.AI

Tonmoy Hasan , Razvan Bunescu This is my paper
classification cs.IRcs.AI
keywords recommendersystemsaffectivecomputingaffective-cognitivepreferencessequentialrecommendationtext-basedbookemotionwheelBayesianpersonalizedranking
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

Most recommenders infer what a user likes from ratings and clicks. This paper argues that users can instead tell the system how they want to feel—for example, entranced, cheerful, or pleasantly surprised by the conclusion—and that such explicit affective statements are a valuable input for finding the right book. The authors build ACRec, a model that combines the user's reading history with a text description of the desired affective-cognitive states, and train it on a dataset of book reviews mined for such statements. On a retrospective test, ACRec ranks the book the user actually read next at the top of a 20-book list 67.6% of the time, and reaches 15.6% hit rate at rank 10 across all 143,000 books, compared with 2.8% for the strongest text-only baseline. The claim is that letting users specify desired feelings, rather than just preferred item features, opens a new and much more expressive input channel for recommendation.

What carries the argument

The load-bearing object is the affective-cognitive (AC) description: a sentence or phrase, written by the user or selected from an ontology, that names the emotions, moods, or fine-grained states the user wants the book to produce, such as 'entranced' or 'pleasantly surprised by the conclusion.' The model embeds this text alongside the user's history—short-term preferences from a Transformer over the most recent books, long-term preferences from a time-weighted average of older ones—and a candidate book's original description plus review-based extended description, then scores the book with a small feed-forward network that also takes the cosine similarity between the AC text and the book description as an extra feature. The training objective is Bayesian personalized ranking with negative sampling. Supporting the selection mode, the authors organize mined affective statements into a 26-category emotion wheel adapted from the Geneva Emotion Wheel.

What would settle it

Run ACRec in a forward setting where users write or select their desired affective states before receiving recommendations, then measure whether books they actually choose rank at the top; if top-1 accuracy falls to the level of baselines that never see affective text, the reported gains come from the retrospective setup rather than from understanding feelings.

Watch

Extended reading notes

Core claim

The central claim is that a recommender can use an explicit textual description of the affective-cognitive states a user wants to experience—not just the user's rating history—as a first-class input, and that doing so substantially improves the system's ability to pick items that will induce those states. The paper states its result most directly in the evaluation: when the model is given the full affective and cognitive text, it achieves a top-1 accuracy of 67.6% in a 20-candidate ranking and an HR@10 of 15.6% over all 143,000 books, while the strongest baseline that uses item text but no affective input reaches 44.2% and 2.8% respectively. The authors also show that injecting the same affective statements into the prompts of LLM-based recommenders roughly doubles their accuracy, which they read as evidence that the affective-cognitive signal is broadly useful, not an artifact of their particular architecture.

Load-bearing premise

The result depends on treating the review a user wrote after finishing a book as a faithful proxy for the request the same user would have made before reading it, even though the two can differ in specificity, vocabulary, and timing.

Editorial extensions

If this is right

  • Users can express requests at a much finer grain than like/dislike, such as 'a book that makes me think about it for a long time' or 'thrilled, disturbed, and completely entranced.'
  • Adding the same affective statements to the prompts of LLM-based recommenders roughly doubles their top-1 accuracy, suggesting the input signal transfers beyond ACRec itself.
  • The gap between ACRec with full text (15.6% HR@10) and with only affective statements (2.9%) shows that objective cognitive content carries much of the ranking signal, so future systems should treat AC text as a mixture, not just emotion words.
  • The emotion-wheel selection interface offers a practical UI path for systems that require users to specify desires before any book is read.

Reading between the lines

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

  • A live pre-reading test might show smaller gains: the retrospective setup uses the ground-truth book's own review as the query, so ACRec is in part matching a book against its own review text rather than against a request written beforehand.
  • Because cognitive (C) statements outnumber affective ones in the dataset, the 'affective' framing may understate the role of ordinary preference text; a control experiment using non-affective expectations (e.g., 'set in the Arctic') would separate the affective contribution.
  • If pre-reading elicitation proves harder than mining reviews, the selection mode—choosing from an emotion wheel—becomes the practical bottleneck; its usability, not the model, may decide whether the approach works in the real world.
  • The same AC-statement mining and emotion-wheel mapping could be reused to build affective recommendation benchmarks for other media, such as movies or music, which the authors list as future work.
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.

Referee Report

4 major / 6 minor

Summary. The paper introduces ACRec, a text-based recommender system that takes as input a user's reading/rating/review history together with an affective-cognitive (AC) description of the states the user wants to experience, and scores candidate books by a Transformer+MLP architecture. The authors construct a Goodreads-based dataset for 1,000 users, use GPT-4o to extract affective, cognitive, and affective-cognitive statements from reviews, and report large gains over sequential and LLM-based baselines, e.g., HR@10 of 15.6% versus 2.8% for RecFormer in all-item ranking (Table 4) and 67.6% top-1 accuracy versus 44.2% for RecFormer in 20-item ranking (Table 5). The central empirical claim is that explicitly modeling user-specified affective-cognitive preferences substantially improves recommendation accuracy.

Significance. The proposed task—recommending items that will induce fine-grained affective states specified by the user—is novel and potentially valuable, and the paper makes useful contributions in the form of an emotion-wheel taxonomy, a large mined dataset of affective statements, and a reproducible architecture with released prompts. However, the evaluation protocol invalidates the central claim. As described in Section 4.2.1, the AC description used at test time is extracted from the review that the user wrote for the ground-truth book read at time t+1, and that same book is the positive item to be retrieved. This makes the task approximate text retrieval from the target item's own review, not prospective matching of affective preferences. The validation results in Appendix B, where a pure Cosine model with no user history reaches 81.6% HR@10, confirm that the AC text alone nearly identifies the target book. Consequently, the large margins over baselines in Tables 4 and 5 reflect an information leak rather than the practical value of explicit affective preferences. The paper's own acknowledgment that the retrospective use is not possible in practice further supports this reading.

major comments (4)
  1. [Section 4.2.1] The evaluation protocol leaks the target item into the query. For each training, validation, and test step, the AC description is formed by aggregating all A, AC, and C statements extracted from the review that the user wrote for the book read at time t+1, and the same book is used as the ground-truth positive sample. This means the model is not asked to match preferences expressed before reading; it is asked to retrieve a book from a paraphrase of its own review. The paper explicitly notes that 'in practice, this retrospective use is not possible,' yet the abstract and conclusion claim that ACRec is 'effective at identifying books that match user's expressed affective preferences.' This discrepancy invalidates the comparisons in Tables 4 and 5, since all baselines are evaluated without access to the AC text derived from the gold book's review.
  2. [Appendix B] The validation experiments show that the leakage alone, without any user modeling, nearly solves the task: the Cosine model, which ranks books by cosine similarity between the AC embedding and the concatenated original+extended book description embedding, achieves 81.6% HR@10 on validation. This is evidence that the AC description, extracted from the target book's own review, contains sufficient information to identify the target book by text matching. Moreover, the final ACRec model incorporates cos(ac, d_b) as an additional input to the scoring network (Appendix B, modifying Equation (2)), so the model is explicitly given a direct text-similarity signal between the query and each candidate book. The reported superiority of ACRec is therefore largely attributable to this retrieval-style signal, not to affective preference modeling.
  3. [Section 3.3 and Tables 4-5] The training objective is also retrofitted to the leak. Equation (3) optimizes the model to rank the book whose review generated the AC description above random negatives. As a result, the learned AC projection W_a is optimized to map a book's own review text to that book's description, rather than to map pre-reading affective preferences to future items. The comparison with baselines such as SASRec, BERT4Rec, UniSRec, and RecFormer is therefore not a fair test of the proposed input modality: those baselines never see AC text, while ACRec is given text that is derived from the item to be retrieved. The conclusion that 'the best results are obtained by models that can utilize textual descriptions of items and user affective preferences' is not supported by this experimental design.
  4. [Section 5] The statement that 'all results are conservative in that it is likely that in reality the ground truth book is not the only book satisfying the AC description' is misleading in the context of this evaluation. Because the AC description is generated from the ground-truth book's own review, the test setting is optimistic, not conservative: the query is engineered to match the positive item. The error analysis in Section 5 also inadvertently confirms the leak, since it notes that when the AC description contains a proper noun such as 'Australia,' the model tends to recommend books where that term appears prominently in their descriptions—exactly the behavior expected from a retrieval system, not from a system that matches affective preferences.
minor comments (6)
  1. [Table 4] The A + AC row reports results for only 554 users, whereas the A + AC + C row uses 1,000 users, but this subset difference is not explained in the text; if the smaller set is due to some users having no affective statements, this should be stated explicitly, and the comparison across rows should be made on the same user set.
  2. [Section 3.3] The phrase 'Bayesian personalized raking loss' should read 'Bayesian personalized ranking loss.'
  3. [Appendix B] The word 'dinmensional' should be 'dimensional.'
  4. [Appendix C] The LLaRA-AC prompt contains the placeholder '[CansHere]' twice in the same prompt; one of these should likely be a candidate list placeholder and the other should not be duplicated.
  5. [Table 2] The quote 'brakes my heart' should presumably be 'breaks my heart.'
  6. [Section 4.2.1] The phrase 'second-to-last acceptable time step' uses 'acceptable' where 'useful' is used elsewhere; this terminology should be made consistent.

Circularity Check

3 steps flagged · score 8.0 of 10

Evaluation is circular: the AC query is the ground-truth book's own review and the scorer includes cos(ac, db), so reported gains measure retrieval from the target item, not prospective affective matching.

  1. self definitional [Section 4.2.1 (Training, Validation, and Testing)]
    "For each training, validation, and test time steps, we utilize as input an AC description that aggregates all A, AC, and C statements that were extracted (with the approach described in Section 4.2) from the review that the user wrote for the book at that time step, whereas the book itself is used as the ground truth positive sample. ... AC statements are extracted from the review of the book that was read at time t + 1 and are used as input ... The book that the user actually read at time t+1 (whose review was used to extract AC statements) is used as the ground truth book for evaluation."

    The AC query is generated from the gold book's own review, so the input and the target are two encodings of the same document. The evaluation measures whether the model can retrieve a book from a paraphrase of its own review, not whether it can match affective preferences expressed before reading. The paper concedes this: 'in practice, this retrospective use is not possible because the review of a book does not exist before recommending and reading the book.' Yet the abstract and conclusion claim effectiveness at identifying books matching users' expressed affective preferences, which the retrospective protocol cannot test.

  2. self definitional [Appendix B (Validation Experiments) and ACRecCos score equation]
    "In the Cosine model, we create a combined book description by concatenating the books' original and extended descriptions. For each test sample, we calculate the cosine similarity between the Jina embedding of the AC description ac and the Jina embeddings of the combined book description db. ... Given that incorporating Cosine improved the performance of ACRec, in the rest of the paper we use ACRec to refer to the recommendation model that utilizes cosine as a feature."

    The final scoring function includes the direct term cos(ac, db), where db is the candidate book's extended description built from reviews and ac is extracted from the target book's own review. Validation shows the cosine-only model, with no user history, reaches 81.6% HR@10, so the AC text alone almost identifies the target. The large margins over SASRec, RecFormer, and the LLM baselines in Tables 4 and 5 therefore follow from a text-similarity shortcut whose inputs are the target's own review, not from learned affective-preference matching.

1 more flagged steps
  1. fitted input called prediction [Section 3.3 (Training Objective) with Section 4.2.1]
    "as positive sample we use the book know to match the AC description for the user, whereas negative samples will be drawn at random at each gradient update step from books that the user has not read. ... For each training, validation, and test time steps, we utilize as input an AC description that aggregates all A, AC, and C statements that were extracted ... from the review that the user wrote for the book at that time step, whereas the book itself is used as the ground truth positive sample."

    The BPR objective is optimized on pairs where the positive book is exactly the book whose review generated the AC description. The learned projections Wa, Wd, and We are therefore fit to map review-derived AC text to that same book's own description. At test time the same retrospective pairing is reused, so the 'prediction' is the training target re-encountered through a learned retrieval function. Baselines that never see the AC text are compared against a model trained and evaluated on this self-paired input, making the comparison structurally unfair.

full rationale

The central experimental claim reduces to retrieval by construction. The AC descriptions used as test queries are extracted from the ground-truth book's own review, and the final scoring function includes a direct cosine term between that AC text and candidate book descriptions built from reviews; the validation-only Cosine model reaches 81.6% HR@10 without any user history, demonstrating that the target is nearly identified by the query alone. Training uses the same retrospective pairing, so the learned projections are fitted to map a book's own review text to that book's description. This is not a minor caveat: the reported margins in Tables 4 and 5 reflect an information leak rather than the value of prospective explicit affective preferences, and the paper's own statement that retrospective use is impossible in practice concedes that the deployed setting was not evaluated. No self-citation circularity is load-bearing; the Hasan and Bunescu (2023) citation and the GPT-4o extraction quality checks are independent of the main claim. The score is 8 because the main comparison is forced by input construction, though the dataset and interface contributions retain some independent content.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the assumption that post-hoc review statements, rewritten by GPT-4o as if uttered before reading, faithfully capture preferences a user would express in advance. This assumption is introduced in Section 4.2 and is not validated against real user queries. The model also depends on standard ML hyperparameters chosen by grid search, and on the semantic quality of Jina embeddings for affective text. No new physical entities are postulated.

free parameters (7)
  • Transformer hidden dimension = 128
    Selected by grid search over 64, 128, 256, 512 on validation data (Appendix D).
  • Number of transformer blocks = 4
    Grid search over 1 to 6 encoder blocks (Appendix D).
  • Short-term history length m = 30
    Grid search over 15 to 60 latest items; Table 7 shows m=30 gives best HR@10 on validation.
  • Dropout probability = 0.2
    Grid search over 0.1, 0.2, 0.3, 0.4, 0.5 (Appendix D).
  • Learning rate = 0.0001
    Grid search over 0.1, 0.01, 0.001, 0.0001, 0.00001 (Appendix D).
  • Useful time step filters = rating >=4; review tokens >=20; description tokens >=250
    Section 4.1 applies these ad hoc criteria, restricting data to high-rated, verbose interactions.
  • Max useful steps per user = 20
    Section 4.1 caps the number of selected time steps, spread evenly, to balance user history lengths.
assumptions (4)
  • domain assumption GPT-4o extraction yields accurate AC statements
    Manual checks on 100 statements show 99% guideline adherence and 88% affective/objective precision, but no validation on the full set or against real user preferences.
  • ad hoc to paper Rewritten post-hoc review statements represent pre-reading user preferences
    Section 4.2 instructs the LLM to phrase statements as if written before reading the book; there is no evidence that users would choose those affective states a priori, and this is the core evaluation assumption.
  • domain assumption Jina embeddings capture affective and cognitive text similarity
    The model and the cosine feature operate entirely on Jina embedding similarities; no analysis of embedding quality for affective language is provided.
  • domain assumption Extended book descriptions from other users' reviews are uncontaminated
    The paper excludes train/test users' reviews from extended descriptions, but other reviews of the same book naturally share content with the test user's review, which is the main matching signal.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Text-Based Recommender System that Leverages Explicit Affective State Preferences." pith.science (2026). https://pith.science/paper/B6HPML4I

@misc{pith2026250520190,
  author       = {Pith},
  title        = {Pith review of: A Text-Based Recommender System that Leverages Explicit Affective State Preferences},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B6HPML4I}},
  note         = {Machine review of arXiv:2505.20190}
}
read the original abstract

The affective attitude of liking a recommended item reflects just one category in a wide spectrum of affective phenomena that also includes emotions such as entranced or intrigued, moods such as cheerful or buoyant, as well as more fine-grained affective states, such as "pleasantly surprised by the conclusion". In this paper, we introduce a novel recommendation task that can leverage a virtually unbounded range of affective states sought explicitly by the user in order to identify items that, upon consumption, are likely to induce those affective states. Correspondingly, we create a large dataset of user preferences containing expressions of fine-grained affective states that are mined from book reviews, and propose a Transformer-based architecture that leverages such affective expressions as input. We then use the resulting dataset of affective states preferences, together with the linked users and their histories of book readings, ratings, and reviews, to train and evaluate multiple recommendation models on the task of matching recommended items with affective preferences. Experiments show that the best results are obtained by models that can utilize textual descriptions of items and user affective preferences.

Figures

Figures reproduced from arXiv: 2505.20190 by the authors.

Figure 1
Figure 1. Selection of AC statements, starting from top emotion categories organized around the Emotion Wheel. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Overall architecture of the proposed ACRec model. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 20 canonical work pages

  1. [1]

    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 INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    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 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [4]

    Aminu Da’u and Naomie Salim. 2019. https://doi.org/10.1109/ACCESS.2019.2907729 Sentiment-aware deep recommender system with neural attention networks . IEEE Access, 7:45472--45484

    work page arXiv 2019

  4. [5]

    Zhe Fu, Xi Niu, and Mary Lou Maher. 2023. https://doi.org/10.1145/3605145 Deep learning models for serendipity recommendations: A survey and new perspectives . ACM Comput. Surv., 56(1)

    work page doi:10.1145/3605145 2023

  5. [6]

    Emotions in everyday life

    Peter Goldie. 2004. https://doi.org/10.1177/0539018404047705 The life of the mind: commentary on “ Emotions in everyday life” . Social Science Information, 43(4):591--598. Publisher: SAGE Publications Ltd

    work page doi:10.1177/0539018404047705 2004

  6. [7]

    Tonmoy Hasan and Razvan Bunescu. 2023. https://doi.org/10.1145/3604915.3608851 Topic-level bayesian surprise and serendipity for recommender systems . In Proceedings of the 17th ACM Conference on Recommender Systems, RecSys '23, page 933–939, New York, NY, USA. Association for Computing Machinery

    arXiv 2023

  7. [8]

    Shirley Anugrah Hayati, Dongyeop Kang, Qingxiaoyang Zhu, Weiyan Shi, and Zhou Yu. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.654 INSPIRED : Toward Sociable Recommendation Dialog Systems . In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing ( EMNLP ) , pages 8142--8152, Online. Association for Computational Linguistics

    doi:10.18653/v1/2020.emnlp-main.654 2020

  8. [9]

    Zhankui He, Zhouhang Xie, Rahul Jha, Harald Steck, Dawen Liang, Yesu Feng, Bodhisattwa Prasad Majumder, Nathan Kallus, and Julian Mcauley. 2023. https://doi.org/10.1145/3583780.3614949 Large Language Models as Zero - Shot Conversational Recommenders . In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management , CIKM '2...

    arXiv 2023

Show all 48 references
  1. [11]

    Leo Iaquinta, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro, Michele Filannino, and Piero Molino. 2008. https://doi.org/10.1109/HIS.2008.25 Introducing serendipity in a content-based recommender system . In 2008 Eighth International Conference on Hybrid Intelligent Systems...

    2008 doi

  2. [12]

    Marius Kaminskas and Derek Bridge. 2016. https://doi.org/10.1145/2926720 Diversity, serendipity, novelty, and coverage: A survey and empirical analysis of beyond-accuracy objectives in recommender systems . ACM Trans. Interact. Intell. Syst., 7(1)

    2016 doi

  3. [13]

    Wang-Cheng Kang and Julian McAuley. 2018. https://doi.org/10.1109/ICDM.2018.00035 Self-attentive sequential recommendation . In 2018 IEEE International Conference on Data Mining (ICDM), pages 197--206

    2018

  4. [14]

    Rahul Katarya and Om Prakash Verma. 2016. https://doi.org/10.1016/j.physa.2016.05.046 Recent developments in affective recommender systems . Physica A: Statistical Mechanics and its Applications, 461:182--190

    2016 doi

  5. [15]

    Yejin Kim, Kwangseob Kim, Chanyoung Park, and Hwanjo Yu. 2019. https://doi.org/10.24963/ijcai.2019/380 Sequential and diverse recommendation with long tail . In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI-19 , pages 2740--2...

    2019 doi

  6. [16]

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

    2014 arXiv

  7. [17]

    Xiaoyu Kong, Jiancan Wu, An Zhang, Leheng Sheng, Hui Lin, Xiang Wang, and Xiangnan He. 2024. https://proceedings.neurips.cc/paper_files/paper/2024/file/cd476d01692c508ddf1cb43c6279a704-Paper-Conference.pdf Customizing language models with instance-wise lora for sequential reco...

    2024

  8. [18]

    Denis Kotkov, Shuaiqiang Wang, and Jari Veijalainen. 2016. https://doi.org/10.1016/j.knosys.2016.08.014 A survey of serendipity in recommender systems . Knowledge-Based Systems, 111:180--192

    2016 doi

  9. [19]

    Wenqiang Lei, Xiangnan He, Yisong Miao, Qingyun Wu, Richang Hong, Min-Yen Kan, and Tat-Seng Chua. 2020. https://doi.org/10.1145/3336191.3371769 Estimation-action-reflection: Towards deep interaction between conversational and recommender systems . In Proceedings of the 13th In...

    2020

  10. [20]

    Jiacheng Li, Ming Wang, Jin Li, Jinmiao Fu, Xin Shen, Jingbo Shang, and Julian McAuley. 2023. https://doi.org/10.1145/3580305.3599519 Text is all you need: Learning language representations for sequential recommendation . In Proceedings of the 29th ACM SIGKDD Conference on Kno...

    2023

  11. [21]

    Pan Li and Alexander Tuzhilin. 2020. https://doi.org/10.1145/3404855 Latent unexpected recommendations . ACM Trans. Intell. Syst. Technol., 11(6)

    2020 doi

  12. [22]

    Raymond Li, Samira Ebrahimi Kahou, Hannes Schulz, Vincent Michalski, Laurent Charlin, and Chris Pal. 2018. https://papers.nips.cc/paper_files/paper/2018/hash/800de15c79c8d840f4e78d3af937d4d4-Abstract.html Towards Deep Conversational Recommendations . In Advances in Neural Info...

    2018

  13. [23]

    Jianxun Lian, Yuxuan Lei, Xu Huang, Jing Yao, Wei Xu, and Xing Xie. 2024. https://doi.org/10.1145/3589335.3651242 RecAI : Leveraging Large Language Models for Next - Generation Recommender Systems . In Companion Proceedings of the ACM Web Conference 2024 , WWW '24, pages 1031-...

    2024

  14. [24]

    Jiayi Liao, Sihang Li, Zhengyi Yang, Jiancan Wu, Yancheng Yuan, Xiang Wang, and Xiangnan He. 2024. https://doi.org/10.1145/3626772.3657690 Llara: Large language-recommendation assistant . In Proceedings of the 47th International ACM SIGIR Conference on Research and Development...

    2024

  15. [25]

    Jan Mizgajski and Miko aj Morzy. 2019. Affective recommender systems in online news industry: how emotions influence reading choices. User Modeling and User-Adapted Interaction, 29(2):345--379

    2019

  16. [26]

    Claudia Orellana-Rodriguez, Ernesto Diaz-Aviles, and Wolfgang Nejdl. 2015. https://doi.org/10.1145/2700171.2791042 Mining affective context in short films for emotion-aware recommendation . In Proceedings of the 26th ACM Conference on Hypertext & Social Media, HT '15, page 185...

    2015

  17. [27]

    Deuk Hee Park, Hyea Kyeong Kim, Il Young Choi, and Jae Kyeong Kim. 2012. https://doi.org/10.1016/j.eswa.2012.02.038 A literature review and classification of recommender systems research . Expert Systems with Applications, 39(11):10059--10072

    2012 doi

  18. [28]

    Gustavo Penha and Claudia Hauff. 2020. https://doi.org/10.1145/3383313.3412249 What does BERT know about books, movies and music? Probing BERT for Conversational Recommendation . In Proceedings of the 14th ACM Conference on Recommender Systems , RecSys '20, pages 388--397, New...

    2020

  19. [29]

    Alessandro Petruzzelli, Alessandro Francesco Maria Martina, Giuseppe Spillo, Cataldo Musto, Marco De Gemmis, Pasquale Lops, and Giovanni Semeraro. 2024. https://doi.org/10.1145/3627043.3659565 Improving transformer-based sequential conversational recommendations through knowle...

    2024

  20. [30]

    Marco Polignano, Fedelucio Narducci, Marco de Gemmis , and Giovanni Semeraro. 2021. https://doi.org/10.1016/j.eswa.2020.114382 Towards emotion-aware recommender systems: an affective coherence model based on emotion-driven behaviors . Expert Systems with Applications, 170:114382

    2021

  21. [31]

    Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt-Thieme. 2009. Bpr: Bayesian personalized ranking from implicit feedback. In Proceedings of the Twenty-Fifth Conference on Uncertainty in Artificial Intelligence, UAI '09, page 452–461, Arlington, Virginia,...

    2009

  22. [32]

    Paul Resnick and Hal R. Varian. 1997. https://doi.org/10.1145/245108.245121 Recommender systems . Communications of the ACM, 40(3):56–58

    1997

  23. [33]

    Sacharin, Schlegel K., and Scherer K

    V. Sacharin, Schlegel K., and Scherer K. R. 2012. https://www.unige.ch/cisa/files/4514/6720/4016/Geneva_Emotion_Wheel_Rating_Study_Report_2012_08_11_2.0.pdf Geneva Emotion Wheel rating study . Unpublished report. University of Geneva: Swiss Center for Affective Studies

    2012

  24. [34]

    Camilo Salazar, Jose Aguilar, Juli\' a n Monsalve-Pulido, and Edwin Montoya. 2021. https://doi.org/10.1016/j.cosrev.2021.100377 Affective recommender systems in the educational field. a systematic literature review . Comput. Sci. Rev., 40(C)

    2021

  25. [35]

    Klaus R. Scherer. 2005. https://doi.org/10.1177/0539018405058216 What are emotions? And how can they be measured? Social Science Information, 44(4):695--729. Publisher: SAGE Publications Ltd

    2005 doi

  26. [36]

    Saba Sturua, Isabelle Mohr, Mohammad Kalim Akram, Michael G \"u nther, Bo Wang, Markus Krimmel, Feng Wang, Georgios Mastrapas, Andreas Koukounas, Nan Wang, et al. 2024. jina-embeddings-v3: Multilingual embeddings with task lora. arXiv preprint arXiv:2409.10173

    2024 arXiv

  27. [37]

    Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang. 2019. https://doi.org/10.1145/3357384.3357895 Bert4rec: Sequential recommendation with bidirectional encoder representations from transformer . In Proceedings of the 28th ACM International Conference ...

    2019

  28. [38]

    Thompson, Mehmet H

    Cynthia A. Thompson, Mehmet H. Göker, and Pat Langley. 2004. A personalized system for conversational recommendations. J. Artif. Int. Res., 21(1):393--428

    2004

  29. [39]

    Abhishek Tripathi, T. S. Ashwin, and Ram Mohana Reddy Guddeti. 2019. https://doi.org/10.1109/ACCESS.2019.2911235 Emoware: A context-aware framework for personalized video recommendation using affective video sequences . IEEE Access, 7:51185--51200

    2019

  30. [40]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. 2017. https://proceedings.neurips.cc/paper_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf Attention is all you need . In Advances in Ne...

    2017

  31. [42]

    Mengting Wan and Julian McAuley. 2018 b . https://doi.org/10.1145/3240323.3240369 Item recommendation on monotonic behavior chains . In Proceedings of the 12th ACM Conference on Recommender Systems, RecSys '18, page 86–94, New York, NY, USA. Association for Computing Machinery

    2018

  32. [43]

    Dandan Wang and Xiaoming Zhao. 2022. https://doi.org/10.3389/fnins.2022.984404 Affective video recommender systems: A survey . Frontiers in Neuroscience, 16

    2022

  33. [44]

    Xiaolei Wang, Kun Zhou, Ji-Rong Wen, and Wayne Xin Zhao. 2022. https://doi.org/10.1145/3534678.3539382 Towards unified conversational recommender systems via knowledge-enhanced prompt learning . In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data M...

    2022

  34. [45]

    Xiaoyu Zhang, Ruobing Xie, Yougang Lyu, Xin Xin, Pengjie Ren, Mingfei Liang, Bo Zhang, Zhanhui Kang, Maarten de Rijke, and Zhaochun Ren. 2024. Towards empathetic conversational recommender systems. arXiv preprint arXiv:2409.10527

    2024 arXiv

  35. [46]

    Yuan Cao Zhang, Diarmuid \' O S\' e aghdha, Daniele Quercia, and Tamas Jambor. 2012. https://doi.org/10.1145/2124295.2124300 Auralist: introducing serendipity into music recommendation . In Proceedings of the Fifth ACM International Conference on Web Search and Data Mining, WS...

    2012

  36. [47]

    Zihuai Zhao, Wenqi Fan, Jiatong Li, Yunqing Liu, Xiaowei Mei, Yiqi Wang, Zhen Wen, Fei Wang, Xiangyu Zhao, Jiliang Tang, and Qing Li. 2024. https://doi.org/10.1109/TKDE.2024.3392335 Recommender Systems in the Era of Large Language Models ( LLMs ) . IEEE Transactions on Knowled...

    2024

  37. [48]

    Yong Zheng. 2017. https://doi.org/10.1145/3106426.3106535 Affective prediction by collaborative chains in movie recommendation . In Proceedings of the International Conference on Web Intelligence, WI '17, page 815–822, New York, NY, USA. Association for Computing Machinery

    2017

  38. [49]

    Kun Zhou, Wayne Xin Zhao, Shuqing Bian, Yuanhang Zhou, Ji-Rong Wen, and Jingsong Yu. 2020 a . https://doi.org/10.1145/3394486.3403143 Improving conversational recommender systems via knowledge graph based semantic fusion . In Proceedings of the 26th ACM SIGKDD International Co...

    2020

  39. [50]

    Kun Zhou, Wayne Xin Zhao, Hui Wang, Sirui Wang, Fuzheng Zhang, Zhongyuan Wang, and Ji-Rong Wen. 2020 b . https://doi.org/10.1145/3340531.3412098 Leveraging historical interaction data for improving conversational recommender system . In Proceedings of the 29th ACM Internationa...

    2020

  40. [51]

    Jie Zou, Aixin Sun, Cheng Long, and Evangelos Kanoulas. 2024. https://doi.org/10.1145/3677376 Knowledge-enhanced conversational recommendation via transformer-based sequential modelling . ACM Trans. Inf. Syst. Just Accepted

    2024 doi

Pith tools

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