Pith. sign in

REVIEW 4 major objections 6 minor 59 references

Affect-aware Cross-Domain Recommendation for Art Therapy via Music Preference Elicitation

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that music-driven preference elicitation can serve as an alternative to visual browsing in personalized art therapy, and reports a 200-user study in which three cross-domain engines matched a visual-only baseline on…

desk verdict A genuine first study of music-driven cross-domain recommendation for art therapy, but the abstract's headline claim is contradicted by the paper's own significance tests. read the letter →

arxiv 2507.21120 v1 pith:LVRUTFBP submitted 2025-07-18 cs.IR cs.AI

Bereket A. Yilma , Luis A. Leiva This is my paper
classification cs.IRcs.AI
keywords cross-domainrecommendationarttherapymusicpreferenceelicitationvalence-arousalrecommendersystemspersonalizationaffectivecomputinguserstudy
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

The paper's goal is to show that music-driven preference elicitation can replace or complement the usual visual-only browsing step in personalized art therapy, by transferring affective knowledge from music to visual art. It proposes three cross-domain recommendation engines—Mozart, Haydn, and Salieri—that turn a user's ratings of short music clips into ranked painting recommendations. In a 200-user study with people who had psychiatric sequelae (diagnosed mental-health conditions), these engines were compared against a visual-only baseline on user-rated quality and mood change. The abstract claims music-driven elicitation outperforms the visual-only approach, but the paper's own significance tests found no statistically significant differences between engines on the quality ratings, except that the visual baseline scored higher than Haydn on diversity. The data do show a strong mood improvement from before to after the guided art-therapy session across all conditions, which supports music-driven engines as a viable alternative rather than a demonstrated improvement.

What carries the argument

The load-bearing object is a shared valence-arousal (V-A) coordinate system. The paper takes continuous V-A annotations from a music emotion database and translates the emotion labels of a painting emotion dataset into the same scale using a word-level lexicon of affective norms. All three engines measure emotional similarity as a distance or contrastive loss in this shared space: Mozart applies a Gaussian-kernel soft similarity to weight a contrastive alignment loss; Haydn retrieves by raw Euclidean distance in the V-A plane; Salieri fuses acoustic and image features with semantic descriptions generated by an LLM/VLM before computing cross-modal similarity.

What would settle it

Ask a separate group of participants to rate the emotional similarity of music-clip and painting pairs, then compare those human ratings with the V-A distances used by Haydn and with the joint-embedding distances used by Mozart and Salieri; if human similarity ratings do not track the model distances, the shared affective-space assumption fails and all three engines lose their retrieval basis.

Watch

Extended reading notes

Core claim

The central claim is that the affective content of music and paintings can be aligned in one valence-arousal space, and that this alignment makes a patient's music preferences a usable signal for retrieving therapeutic paintings. Mozart learns a shared embedding with a soft contrastive loss over continuous V-A similarity; Haydn retrieves by plain Euclidean distance in the V-A plane; Salieri fuses acoustic and image features with semantic text embeddings before scoring cross-modal similarity. The authors argue that music-driven elicitation captures affective cues missed by visual-only browsing and, in their 200-user study, matches the state-of-the-art visual engine's recommendation quality while producing the same mood improvement in a guided art-therapy session. They also state in the abstract that it outperforms visual-only elicitation, although their own significance tests found no statistically significant differences between engines.

Load-bearing premise

The load-bearing premise is that a painting's emotion labels and a music clip's continuous ratings can be meaningfully mapped onto the same valence-arousal scale, so a short V-A distance between a song and a painting really indicates emotional similarity.

Editorial extensions

If this is right

  • Art-therapy systems could draw on a patient's existing music listening history or playlists as preference input, removing the need to browse visual stimuli during elicitation, a direction the paper's discussion explicitly raises.
  • Music-driven elicitation provides a cold-start path: a new patient can receive personalized painting recommendations from music ratings alone, before any painting preferences are known.
  • The three engines form a reusable family of baselines for future affect-aware cross-modal recommenders in therapeutic settings, with Salieri showing the strongest focus on Hope in the thematic analysis.
  • If affective transfer between music and visual art works, the same V-A alignment machinery could extend to other pairs of therapeutic media, such as film clips or literature, whenever equivalent affective annotations exist.
  • The pre-to-post mood improvement across all conditions suggests that, even when quality ratings do not separate the systems, any of these elicitation modalities can support the therapeutic goal of moving patients toward a more positive mood.

Reading between the lines

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

  • The cross-modal V-A mapping could be validated directly by collecting human similarity judgments between music clips and paintings and checking whether they correlate with the engines' distances; the paper does not report such a check, and this is the assumption on which all three engines rest.
  • Because the safety filtering deliberately removed negative and neutral stimuli, the absence of significant differences between engines may reflect restricted stimulus variance rather than true parity; a study with a wider affective range would be needed to distinguish these possibilities.
  • Music-driven elicitation may have an accessibility benefit for patients who find visual exposure overwhelming or triggering, since the elicitation phase requires no imagery; the paper does not test this, but its therapeutic framing points to it.
  • The soft-contrastive kernel, which converts continuous label distances into similarity weights, is a generic tool that could be transferred to other continuous-label cross-modal alignment problems beyond art therapy.
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 three cross-domain recommendation (CDR) engines—Mozart, Haydn, and Salieri—that use music preference elicitation to recommend therapeutic paintings in digital art therapy. Mozart learns a contrastive joint embedding of music and paintings using valence-arousal (V-A) labels as soft supervision; Haydn retrieves paintings by direct Euclidean distance in V-A space; Salieri concatenates MERT/ResNet and LLM/VLM text embeddings. The authors evaluate all three against a visual-only ResNet-50 baseline in a between-subjects study with 200 participants with psychiatric sequelae, measuring user-centric recommendation quality, mood change, and thematic reflections. The abstract and conclusion claim that music-driven elicitation outperforms the visual-only baseline and significantly improves mood.

Significance. If the central claim were supported, the paper would open a useful new direction: using music as a preference elicitation modality for visual art recommendation in mental-health contexts. The study has notable strengths: a relatively large clinical population (N=200), three methodologically distinct CDR baselines, a real guided art-therapy procedure, and public release of source code, data, and models. The thematic analysis and the reported mood shifts are also relevant to the RecSys community. However, the core empirical assertion—that music-driven methods outperform the visual baseline—is contradicted by the paper's own inferential statistics, and the methodological construction of Mozart makes the affective-transfer interpretation difficult to separate from the V-A labels themselves. As submitted, the contribution is better described as a first exploration of CDR for art therapy with comparable performance to a visual baseline, not as a demonstrated improvement.

major comments (4)
  1. [Abstract / Section 4.6.1 / Section 7] The abstract and conclusion state that music-driven preference elicitation 'outperforms the classic visual-only elicitation approach' and 'significantly improves mood,' but Section 4.6.1 reports that a linear mixed-effects model with Bonferroni-Holm-corrected pairwise contrasts found no statistically significant differences between any engines on any of the six user-centric metrics. The only exception is Visual outperforming Haydn on diversity (p=.007, r=0.2), which is the opposite direction of the claimed advantage. The paper's own evidence therefore does not support the headline claim; at best it supports a claim of comparable performance.
  2. [Section 4.6.2 / Section 7] The claim that the systems 'significantly improve mood' is not supported. The pre/post mood improvement has no control condition without art therapy or without recommendations, so it cannot be attributed to the recommendation engine. Moreover, between-group differences in mood change were not significant (chi-square p=.780 for Pick-A-Mood; PANAS differences were also not significant). The significant mood improvement mentioned in the conclusion is therefore an overreach relative to the design and the reported tests.
  3. [Section 3.2, Eqs. (2)-(5) and Eq. (7)] Mozart uses the raw V-A values both as input features and as the supervision signal. Equation (2) concatenates V-A values into the learned embeddings (z* = [z, v]), while Eqs. (3)-(4) define the soft similarity S_ij used as the contrastive target from exactly those V-A values. Thus the Euclidean distance in Eq. (7) is computed in a space whose input already contains the same coordinates that generated the training labels. This makes it difficult to interpret the model as learning an independent affective transfer: an ablation without the concatenated V-A values, or a comparison against a V-A-only oracle, would be needed to establish that the learned alignment contributes beyond the raw labels.
  4. [Section 3.1.3, Eqs. (3) and (8)] The cross-modal label commensurability assumption is load-bearing but unvalidated. The paper translates WikiArt emotion labels into V-A scores using the NRC VAD lexicon and treats them as directly comparable to DEAM's continuous per-song V-A annotations. Equations (3) and (8) then compute Euclidean distances between these two sets of coordinates as if they live in the same psychological space. No evidence is provided that word-based V-A norms for emotion labels are on the same scale as listener-based continuous music annotations. A concrete validation—for example, collecting human V-A ratings for a sample of paintings and music under the same protocol—is necessary before Haydn's retrieval or the soft labels in Mozart can be interpreted as measuring cross-modal emotional similarity.
minor comments (6)
  1. [Section 4.6.3] The final sentence says the music-based approaches 'even outperform SOTA in most user-centric metrics,' but this contradicts Section 4.6.1, which found no significant differences except Visual beating Haydn on diversity. The wording should be aligned with the statistical results.
  2. [Figure 1 caption] The caption contains an incomplete phrase: 'Before, the similarity matrix shows a rather effective alignment.' This appears to be a typo or a leftover fragment and should be corrected.
  3. [Section 4.4] The text says 'based on the elicited user preferences, see 2)' but should refer to 'Algorithm 2'.
  4. [Section 6] The attribution of the null results to 'stimulus homogeneity' is speculative and untested; the paper should either provide evidence for this explanation or explicitly label it as a hypothesis for future work.
  5. [Section 4.1-4.2] The user study with 200 participants with psychiatric sequelae does not report ethics approval, informed consent, or data-protection procedures. This is a standard reporting requirement for studies involving a clinical population.
  6. [References] Reference [44] is given as 'J. Tan et al.' with incomplete author information; the full citation should be provided.

Circularity Check

1 steps flagged · score 6.0 of 10

Mozart's 'affect-aware alignment' is partially circular: the contrastive target is computed from V-A coordinates that Eq. (2) concatenates directly into the input embeddings, so the affective component of the retrieval distance is constructed from its own labels rather than independently learned.

  1. self definitional [Section 3.2, Eqs. (2)-(5), retrieval Eq. (7)]
    "Finally we concatenate the raw V-A values with the learned embeddings: z∗m=[zm, va], z∗p=[zp, va]. (2) ... we first compute the Euclidean distance between the V-A vectors of two samples: dij=∥vi−vj∥2=√((vi−vj)^2+(ai−aj)^2), (3) ... To convert this distance into a soft, continuous similarity measure, we apply a Gaussian kernel: Sij=exp(−d2ij/2σ2), (4) ... the continuous similarity score Sij is integrated into the contrastive loss: L=... (5)"

    The supervision signal S_ij in Eq. (4) is a Gaussian kernel over dij, the Euclidean distance between V-A vectors v_i and v_j. Those exact V-A vectors are concatenated into the representations z*_i=[z_i,v_i] and z*_j=[z_j,v_j] in Eq. (2), and Eq. (5) trains the projection of z* to make its distance large or small according to S_ij. Thus the 'affect-aware alignment' is optimized to reproduce pairwise distances that are already fixed by coordinates fed directly into the model; a trivial solution is for the projection to read out the V-A dimensions. The retrieval distance of Eq.

full rationale

The central Mozart mechanism is partially circular: the contrastive loss's target similarity is a monotone function of the V-A Euclidean distance, and those same V-A values are concatenated into the input embeddings that the loss reshapes. The subsequent retrieval therefore inherits the label coordinates by construction, which weakens the interpretation of the 'affective alignment' as independently learned cross-modal transfer. Haydn is explicitly a V-A nearest-neighbor search, so it is a direct use of labels rather than a hidden circularity, and Salieri feeds V-A values into its text prompts, which is a similar leakage but less cleanly reducible to a specific equation. The paper's headline claim that music-driven elicitation 'outperforms' visual-only elicitation is not a circularity problem but is contradicted by the paper's own statistics: Section 4.6.1 reports no significant differences between engines except Visual outperforming Haydn on diversity, and Section 4.6.2 reports non-significant mood differences; those are correctness and overclaim concerns. The self-citations to the authors' prior VA RecSys work [49, 50] are used as a baseline and are not load-bearing circularity, because the comparison is an external user study. Overall, one constructional component of the proposed method warrants a score of 6: a partial circularity rather than a fully forced derivation.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a chain of domain assumptions: the NRC VAD lexicon provides a valid cross-modal bridge between WikiArt emotion labels and DEAM continuous V-A ratings; fixed pre-trained feature extractors (MERT, ResNet-50) encode affective content; the Gaussian-kernel soft contrastive target is an appropriate learning signal; and the curated positive/non-neutral stimulus filters preserve therapeutic safety without distorting the comparison. Several hyperparameters (sigma, margin, dimension, filtering thresholds) are chosen from data or grid search. No new physical or conceptual entities are invented.

free parameters (5)
  • Gaussian kernel bandwidth (sigma) = 0.5
    Selected via grid search in Section 3.2.3; controls decay of soft similarity S_ij in Eq. (4), affecting the contrastive training target.
  • Contrastive margin m = 0.5
    Selected via grid search in Section 3.2.3; sets minimum separation for dissimilar pairs in Eq. (5).
  • DEAM filtering thresholds = valence SD > 1.75; arousal SD > 1.0
    Chosen based on label distributions (Section 3.1.1); removes 893 of 1802 songs, shaping the music domain.
  • Curation valence/arousal thresholds = valence > 0.1; arousal <= -0.1 or >= 0.1
    Used in Section 4.1 to keep positive, non-neutral stimuli for the user study; a post hoc safety filter.
  • Embedding dimension 256D = 256
    Chosen to balance expressiveness and efficiency; an offline ranking-overlap check showed no gains with 512D or 1024D (Section 3.2 footnote 4).
assumptions (5)
  • domain assumption NRC VAD Lexicon maps emotion words to valence-arousal coordinates that are commensurable with DEAM's continuous V-A ratings.
    Used in Section 3.1.3 to translate WikiArt emotion labels into V-A space; if the mapping is not cross-modally valid, the shared affective space is invalid.
  • domain assumption Pre-trained MERT and ResNet-50 features encode the affective content of music and paintings sufficiently for cross-modal alignment.
    Both feature extractors are used as fixed backbones (Sections 3.2 and 3.4); no fine-tuning or validation of affective encoding is provided.
  • ad hoc to paper Gaussian kernel soft similarity is an appropriate target for continuous-valued contrastive learning.
    Equation (4) defines the training target; the choice is motivated by a citation but the bandwidth is a fitted hyperparameter.
  • domain assumption Participants' self-reported psychiatric diagnoses via Prolific are accurate enough to constitute a psychiatric sequelae population.
    Recruitment in Section 4.2 relies on Prolific screening; PHQ-4 is collected but all participants are said to show anxiety and depression symptoms, which is not verified clinically.
  • domain assumption Guided art therapy with three paintings produces measurable mood change via Pick-A-Mood and PANAS in a single session.
    The pre/post design in Section 4.5 relies on this; no validation of sensitivity to a 35-minute online session is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Affect-aware Cross-Domain Recommendation for Art Therapy via Music Preference Elicitation." pith.science (2026). https://pith.science/paper/LVRUTFBP

@misc{pith2026250721120,
  author       = {Pith},
  title        = {Pith review of: Affect-aware Cross-Domain Recommendation for Art Therapy via Music Preference Elicitation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LVRUTFBP}},
  note         = {Machine review of arXiv:2507.21120}
}
read the original abstract

Art Therapy (AT) is an established practice that facilitates emotional processing and recovery through creative expression. Recently, Visual Art Recommender Systems (VA RecSys) have emerged to support AT, demonstrating their potential by personalizing therapeutic artwork recommendations. Nonetheless, current VA RecSys rely on visual stimuli for user modeling, limiting their ability to capture the full spectrum of emotional responses during preference elicitation. Previous studies have shown that music stimuli elicit unique affective reflections, presenting an opportunity for cross-domain recommendation (CDR) to enhance personalization in AT. Since CDR has not yet been explored in this context, we propose a family of CDR methods for AT based on music-driven preference elicitation. A large-scale study with 200 users demonstrates the efficacy of music-driven preference elicitation, outperforming the classic visual-only elicitation approach. Our source code, data, and models are available at https://github.com/ArtAICare/Affect-aware-CDR

Figures

Figures reproduced from arXiv: 2507.21120 by the authors.

Figure 1
Figure 1. Proposed system architectures. From left to right: Affect-aware contrastive alignment (Mozart), Affective Space [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Screenshots (in chronological order) of the web application we developed for the user study. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Distribution of ratings of our user-centric recommendation quality metrics. Dots denote mean values. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Mood improvement comparison before and after art therapy. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Emotion score changes according to the Positive Affect Negative Affect Schedule (PANAS) scale. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Heatmap of healing concepts identified by recom [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

59 extracted references · 45 canonical work pages

  1. [1]

    Annemarie Abbing, Anne Ponstein, Susan van Hooren, Leo de Sonneville, Hanna Swaab, and Erik Baars. 2018. The effectiveness of art therapy for anxiety in adults: A systematic review of randomised and non-randomised controlled trials. PloS one 13, 12 (2018), e0208716

    work page 2018

  2. [2]

    Sinem Aslan, Giovanna Castellano, Vincenzo Digeno, Giuseppe Migailo, Raffaele Scaringi, and Gennaro Vessio. 2022. Recognizing the emotions evoked by art- works through visual features and knowledge graph-embeddings. InInternational Conference on Image Analysis and Processing . Springer, 129–140

    work page 2022

  3. [3]

    Kelly Sarah Barnett and Fabian Vasiu. 2024. How the arts heal: a review of the neural mechanisms behind the therapeutic effects of creative arts on mental and physical health. Frontiers in behavioral neuroscience 18 (2024), 1422361

    work page 2024

  4. [4]

    David M Blei, Andrew Y Ng, and Michael I Jordan. 2003. Latent dirichlet allocation. Journal of machine Learning research 3, Jan (2003), 993–1022

    2003

  5. [5]

    Rebecca Bokoch, Noah Hass-Cohen, April Espinoza, Tyler O’Reilly, and Elad Levi

    work page

  6. [6]

    Olivia Brancatisano, Amee Baird, and William Forde Thompson. 2020. Why is music therapeutic for neurological disorders? The Therapeutic Music Capacities Model. Neuroscience & Biobehavioral Reviews 112 (2020), 600–615

    work page 2020

  7. [7]

    José J Campos-Bueno, Octavio DeJuan-Ayala, Pedro Montoya, and Niels Bir- baumer. 2015. Emotional dimensions of music and painting and their interaction. The Spanish journal of psychology 18 (2015), E54

    work page 2015

  8. [8]

    Suvin Choi, Jong-Ik Park, Cheol-Ho Hong, Sang-Gue Park, and Sang-Cheol Park

    work page

Show all 59 references
  1. [9]

    Shreyan Chowdhury, Andreu Vall, Verena Haunschmid, and Gerhard Widmer

  2. [10]

    Pieter MA Desmet, Martijn H Vastenburg, and Natalia Romero. 2016. Mood measurement with Pick-A-Mood: review of current methods and design of a pictorial self-report scale. Journal of Design Research 14, 3 (2016), 241–279

    2016

  3. [11]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding.arXiv preprint arXiv:1810.04805 (2018)

    2018 arXiv

  4. [12]

    Antonella Di Vita, Mario Augusto Procacci, Martina Bellagamba, Maria Jacomini, Roberta Massicci, and Maria Paola Ciurli. 2022. Psychotherapy and Art Therapy: A pilot study of group treatment for patients with traumatic brain injury.Journal of health psychology 27, 4 (2022), 836–846

    2022

  5. [13]

    Ali Mamdouh Elkahky, Yang Song, and Xiaodong He. 2015. A multi-view deep learning approach for cross domain user modeling in recommendation systems. In Proceedings of the 24th international conference on world wide web . 278–288

    2015

  6. [14]

    Yingjie Feng and Mingda Wang. 2025. Effect of music therapy on emotional re- silience, well-being, and employability: a quantitative investigation of mediation and moderation. BMC psychology 13, 1 (2025), 47

    2025

  7. [15]

    Ignacio Fernández-Tobías, Matthias Braunhofer, Mehdi Elahi, Francesco Ricci, and Iván Cantador. 2016. Alleviating the new user problem in collaborative filtering by exploiting personality information. User Modeling and User-Adapted Interaction 26 (2016), 221–255

    2016

  8. [16]

    Maarten Grootendorst. 2022. BERTopic: Neural topic modeling with a class-based TF-IDF procedure. arXiv preprint arXiv:2203.05794 (2022)

    2022 arXiv

  9. [17]

    Suzanne Haeyen and Merel Staal. 2021. Imagery rehearsal based art therapy: Treatment of post-traumatic nightmares in art therapy. Frontiers in psychology 11 (2021), 628717

    2021

  10. [18]

    Kathy Hathorn and Upali Nanda. 2008. A guide to evidence-based art. The Center for Health Design 1 (2008), 1–23

    2008

  11. [19]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition . 770–778

    2016

  12. [20]

    Taisei Hirakawa, Keisuke Maeda, Takahiro Ogawa, Satoshi Asamizu, and Miki Haseyama. 2021. Cross-domain recommendation method based on multi-layer graph analysis with visual information. In 2021 IEEE International Conference on Image Processing (ICIP). IEEE, 2688–2692

    2021

  13. [21]

    Jingxuan Hu, Jinhuan Zhang, Liyu Hu, Haibo Yu, and Jinping Xu. 2021. Art therapy: a complementary treatment for mental disorders.Frontiers in psychology 12 (2021), 686005

    2021

  14. [22]

    Papadopoulos Stefanos Iordanis. 2021. Emotion-aware music recommendation systems. (2021)

    2021

  15. [23]

    Dian Jiao. 2025. Advancing personalized digital therapeutics: integrating music therapy, brainwave entrainment methods, and AI-driven biofeedback. Frontiers in Digital Health 7 (2025), 1552396

    2025

  16. [24]

    Juslin and Daniel Västfjäll

    Patrik N. Juslin and Daniel Västfjäll. 2008. Emotional responses to music: The need to consider underlying mechanisms. Behavioral and Brain Sciences 31, 5 (2008), 559–575. https://doi.org/10.1017/S0140525X08005293

    2008 doi

  17. [25]

    Stefan Koelsch. 2014. Brain correlates of music-evoked emotions. Nature reviews neuroscience 15, 3 (2014), 170–180

    2014

  18. [26]

    AI Ladas, T Gravalas, C Katsoridou, and CA Frantzidis. 2024. Harmony in the brain: A narrative review on the shared neural substrates of emotion regulation and creativity. Brain Organoid and Systems Neuroscience Journal 2 (2024), 81–91

    2024

  19. [27]

    Cheng-Che Lee, Wan-Yi Lin, Yen-Ting Shih, Pei-Yi Kuo, and Li Su. 2020. Crossing you in style: Cross-modal style transfer from music to visual arts. In Proceedings of the 28th ACM international conference on multimedia . 3219–3227

    2020

  20. [28]

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. 2022. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. arXiv preprint arXiv:2201.12086 (2022)

    2022 arXiv

  21. [29]

    Yizhi Li, Ruibin Yuan, Ge Zhang, Yinghao Ma, Xingran Chen, Hanzhi Yin, Cheng- hao Xiao, Chenghua Lin, Anton Ragni, Emmanouil Benetos, et al. 2023. Mert: Acoustic music understanding model with large-scale self-supervised training. arXiv preprint arXiv:2306.00107 (2023)

    2023 arXiv

  22. [30]

    Bojun Liu and Bohong Liu. 2022. Human-Centric C ross-Domain T ransfer N etwork for M usic Recommendation. In International Advanced Computing Conference. Springer, 407–414

    2022

  23. [31]

    Bernd Löwe, Inka Wahl, Matthias Rose, Carsten Spitzer, Heide Glaesmer, Katja Wingenfeld, Antonius Schneider, and Elmar Brähler. 2010. A 4-item measure of depression and anxiety: validation and standardization of the Patient Health Questionnaire-4 (PHQ-4) in the general populat...

    2010

  24. [32]

    Cathy A Malchiodi. 2011. Handbook of art therapy . Guilford Press

    2011

  25. [33]

    Tong Man, Huawei Shen, Xiaolong Jin, and Xueqi Cheng. 2017. Cross-domain recommendation: An embedding and mapping approach. In IJCAI, Vol. 17. 2464– 2470

    2017

  26. [34]

    Saif M Mohammad. 2025. NRC VAD Lexicon v2: Norms for valence, arousal, and dominance for over 55k English terms. arXiv preprint arXiv:2503.23547 (2025)

    2025 arXiv

  27. [35]

    Sharaj Panwar, Paul Rad, Kim-Kwang Raymond Choo, and Mehdi Roopaei. 2019. Are you emotional or depressed? Learning about your emotional state from your music using machine learning. The Journal of Supercomputing 75 (2019), 2986–3009

    2019

  28. [36]

    Varesh Patel and Khyati Mehta. 2024. Emotion-Aware Music Recommendations: Evaluating Custom CNN vs. VGG16 and InceptionV3. In 2024 IEEE International Conference on Future Machine Learning and Data Science (FMLDS). IEEE, 499–504

    2024

  29. [37]

    Pearl Pu, Li Chen, and Rong Hu. 2011. A user-centric evaluation framework for recommender systems. InProceedings of the fifth ACM conference on Recommender systems. 157–164

    2011

  30. [38]

    Alfredo Raglio, Marcello Imbriani, Chiara Imbriani, Paola Baiardi, Sara Manzoni, Marta Gianotti, Mauro Castelli, Leonardo Vanneschi, Francisco Vico, and Luca Manzoni. 2020. Machine learning techniques to predict the effectiveness of music therapy: A randomized controlled trial...

    2020

  31. [39]

    Yiren Ren, Sophia Kaltsouni Mehdizadeh, Grace Leslie, and Thackery Brown

  32. [40]

    Mathieu Roy, Isabelle Peretz, and Pierre Rainville. 2008. Emotional valence contributes to music-induced analgesia. Pain 134, 1-2 (2008), 140–147

    2008

  33. [41]

    Gillian M Sandstrom and Frank A Russo. 2010. Music hath charms: The effects of valence and arousal on recovery following an acute stressor. Music and Medicine 2, 3 (2010), 137–143

    2010

  34. [42]

    Susanne Schweizer, Ian H Gotlib, and Sarah-Jayne Blakemore. 2020. The role of affective control in emotion regulation during adolescence. Emotion 20, 1 (2020), 80

    2020

  35. [43]

    Cognitive, Affective, & Behavioral Neuroscience 24, 5 (2024), 912–930

    Affective music during episodic memory recollection modulates subsequent false emotional memory traces: an fMRI study. Cognitive, Affective, & Behavioral Neuroscience 24, 5 (2024), 912–930

    2024

  36. [44]

    Tan et al

    J. Tan et al . 2024. Contrastive Learning Is Spectral Clustering On Similarity Graph. In Proceedings of CVPR 2024

    2024

  37. [45]

    Michael Thaut. 2013. Rhythm, music, and the brain: Scientific foundations and clinical applications. Routledge

    2013

  38. [46]

    Berit Marie Dykesteen Vik, Geir Olve Skeie, and Karsten Specht. 2019. Neuroplas- tic effects in patients with traumatic brain injury after music-supported therapy. Frontiers in human neuroscience 13 (2019), 177

    2019

  39. [47]

    Heather L Stuckey and Jeremy Nobel. 2010. The connection between art, healing, and public health: A review of current literature. American journal of public health 100, 2 (2010), 254–263

    2010

  40. [48]

    Barbara L Wheeler, Estate Sokhadze, Joshua Baruth, Gene Ann Behrens, and Carla F Quinn. 2011. Musically induced emotions: Subjective measures of arousal and valence. Music and Medicine (2011)

    2011

  41. [49]

    Yilma, Chan Mi Kim, Gerald C

    Bereket A. Yilma, Chan Mi Kim, Gerald C. Cupchik, and Luis A. Leiva. 2024. Artful Path to Healing: Using Machine Learning for Visual Art Recommendation to Prevent and Reduce Post-Intensive Care Syndrome (PICS). In Proceedings of the CHI Conference on Human Factors in Computing...

    2024

  42. [50]

    Bereket A Yilma, Chan Mi Kim, Geke Ludden, Thomas van Rompay, and Luis A Leiva. 2025. The AI-Therapist Duo: Exploring the Potential of Human-AI Col- laboration in Personalized Art Therapy for PICS Intervention. arXiv preprint arXiv:2502.09757 (2025)

    2025 arXiv

  43. [51]

    David Watson, Lee Anna Clark, and Auke Tellegen. 1988. Development and validation of brief measures of positive and negative affect: the PANAS scales. Journal of personality and social psychology 54, 6 (1988), 1063

    1988

  44. [52]

    Tianzi Zang, Yanmin Zhu, Haobing Liu, Ruohan Zhang, and Jiadi Yu. 2022. A survey on cross-domain recommendation: taxonomies, methods, and future directions. ACM Transactions on Information Systems 41, 2 (2022), 1–39

    2022

  45. [53]

    Qian Zhang, Wenhui Liao, Guangquan Zhang, Bo Yuan, and Jie Lu. 2021. A deep dual adversarial network for cross-domain recommendation. IEEE Transactions on Knowledge and Data Engineering 35, 4 (2021), 3266–3278

    2021

  46. [54]

    Feng Zhu, Yan Wang, Chaochao Chen, Jun Zhou, Longfei Li, and Guanfeng Liu

  47. [55]

    Saba Yousefian Jazi, Marjan Kaedi, and Afsaneh Fatemi. 2021. An emotion- aware music recommender system: bridging the user’s interaction and music recommendation. Multimedia Tools and Applications 80 (2021), 13559–13574

    2021

  48. [2019]

    Towards Explainable Music Emotion Recognition: The Route via Mid-level Features. In Proceedings of the 20th International Society for Music Information Retrieval Conference, ISMIR 2019, Delft, The Netherlands, November 4-8, 2019 , Arthur Flexer, Geoffroy Peeters, Julián Urbano...

    2019

  49. [2021]

    arXiv preprint arXiv:2103.01696 (2021)

    Cross-domain recommendation: challenges, progress, and prospects. arXiv preprint arXiv:2103.01696 (2021)

    2021 arXiv

  50. [2024]

    PloS one 19, 5 (2024), e0300607

    Accelerated construction of stress relief music datasets using CNN and the Mel-scaled spectrogram. PloS one 19, 5 (2024), e0300607

    2024

  51. [2025]

    Frontiers in Psychology 16 (2025), 1569609

    A scoping review of integrated arts therapies and neuroscience research. Frontiers in Psychology 16 (2025), 1569609

    2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.