Personalized Art Recommendations

• Personalized artwork recommendation systems are algorithmic approaches that tailor art selections to individual preferences using multimodal data from visuals, metadata, and semantic cues.
• They integrate diverse representation modalities—such as deep visual embeddings, graph-based reasoning, and collaborative filtering—to optimize both retrieval-based and generative outputs.
• Recent advances blend diffusion models, LLM-based selection, and affect-aware feedback to deliver real-time, context-sensitive art recommendations, including applications in therapeutic settings.

Personalized artwork recommendation refers to algorithmic techniques, models, and evaluation practices for dynamically selecting or generating artworks tailored to the inferred or elicited preferences of individual users. This encompasses both retrieval-based systems operating over a finite item catalog (paintings, digital art, photographs) and generative systems synthesizing new art on demand. The field intersects recommender systems, computer vision, natural language processing, interaction design, and modern generative modeling.

1. Foundations and Representation Modalities

Personalized artwork recommendation systems can be categorized by the modalities used to represent users, items, and their interactions. The dominant strategies include:

• Metadata and explicit features: Early systems utilize structured metadata (e.g., color, subject, style, medium, mood), one-hot encoded as $m_i\in\{0,1\}^C$ for artwork $i$, and summary statistics such as brightness, entropy, and colorfulness—so-called “explicit attractiveness” features $a_i\in\mathbb{R}^7$ (Messina et al., 2017).
• Deep visual embeddings: Pre-trained convolutional neural networks (AlexNet (Messina et al., 2017), ResNet (Messina et al., 2020, Fosset et al., 2022)) are deployed to extract fixed-length embeddings (e.g., $v_i\in\mathbb{R}^{4096}$ for AlexNet fc7, $v_i\in\mathbb{R}^{2048}$ for ResNet-50). More recent works utilize BERT embeddings for textual metadata and advanced joint vision-language encoders (BLIP) (Yilma et al., 2024).
• Latent semantic representations: Topic models on textual descriptions, such as Latent Dirichlet Allocation (LDA) or BERTopic, enable high-level, interpretable, and explainable mapping of both user and item profiles into latent topic spaces, which facilitate semantic matching (Yilma et al., 2020, Yilma et al., 2023).
• Graph-based representations: Proximity and similarity graphs—between artworks, between artists, and hybrid graphs joining visual and contextual spaces—enable propagation of personalized signals and efficient community detection within collections (Fosset et al., 2022).
• Purely behavioral embeddings: User preferences can be captured via pooling and transformation of item embeddings corresponding to clicked, purchased, or liked items, resulting in compact user feature vectors suitable for real-time inference (Messina et al., 2020).

These representation modalities can be fused linearly, via ranking fusion (e.g., reciprocal-rank fusion), or by joint training in a shared latent space.

2. Algorithmic Approaches to Personalization

A variety of algorithmic paradigms underpin personalized artwork recommender systems, described below.

2.1 Content-Based Models

Content-based filtering exploits item-intrinsic descriptors. User profiles are constructed by aggregating features of prior interactions:

• Profile vectorization: $M_u = \frac{1}{|I_u|} \sum_{i\in I_u} m_i$, $V_u = \frac{1}{|I_u|} \sum_{i\in I_u} v_i$, $A_u = \frac{1}{|I_u|} \sum_{i\in I_u} a_i$ (Messina et al., 2017).
• Similarity-based scoring: Content match between a user profile and a candidate item is scored by cosine similarity, e.g., $$S_{\text{meta}}(u, i) = \frac{M_u \cdot m_i}{\|M_u\|\|m_i\|}$$.
• Hybrid fusion: Assigns weights $\alpha$, $\beta$, $\gamma$ to the different feature modalities, e.g., $$f(u, i) = \alpha S_{\text{meta}}(u, i) + \beta S_{\text{DNN}}(u, i) + \gamma S_{\text{EVF}}(u, i)$$, where the weights are learned via Bayesian Personalized Ranking (BPR) (Messina et al., 2017).

2.2 Collaborative Filtering and Hybrid Models

• Collaborative approaches associate each user and item with learned embeddings, typically trained via BPR or matrix factorization (Yilma et al., 2024). However, in the context of one-of-a-kind and sparse art datasets, pure collaborative methods often underperform.
• Hybridization fuses collaborative scores with content scores, weighted either globally or per-user, sometimes through multi-task learning (Fosset et al., 2022).

2.3 Graph-Based Reasoning

• Random-walk with restart (RWR): Personalized PageRank-like propagation uses similarity graphs in the feature space for both artworks and artists, scoring unobserved items relative to a user’s liked set (Fosset et al., 2022). Parameters such as the restart probability ($\alpha$) and graph construction methods (Gaussian kernels vs. thresholded distances) are calibrated experimentally.
• Community and diversity control: Spectral clustering and Determinantal Point Process (DPP) re-ranking address overconcentration of recommendations and ensure diversity of style, medium, or theme (Fosset et al., 2022).

2.4 Preference Elicitation and Cold-Start Solutions

• Explicit feedback loops: Users rate, like, or select seed images to initialize profiles. Active learning (max-entropy methods) selects high-information queries to accelerate preference modeling (Parikh, 2020).
• Few-shot semantic initialization: For new users, systems offer the option to pick a handful of favorite items or attributes to bootstrap their latent representations (Fosset et al., 2022, Messina et al., 2017).

3. Generative and LLM-Based Approaches

Recent research advances have shifted from catalog-based retrieval to generative recommendation, synthesizing artwork personalized to user tastes.

3.1 Prompt-Based Generation and Model Recommendation

• Two-stage generative pipeline: (1) Retrieve the best generative model(s) given a prompt (via CLIP-based GRE-Score ensemble), then (2) generate candidate images and rank with user-in-the-loop pairwise or implicit feedback (Guo et al., 2023). The system operates over an “infinite” item space $I = \bigcup_{m \in M} m(P)$.
• Per-user ranking: Learning-to-rank losses are employed in the generated item space.

3.2 Personalized Diffusion Modeling

• Diffusion-based personal generation (REBECA): Assign users a learnable embedding $e_u\in\mathbb{R}^d$, sample image embeddings in CLIP-space via a conditional denoising diffusion prior $p_\theta(I_e|u, r)$, and decode into images with Stable Diffusion (Patron et al., 5 Feb 2025).
• Training: End-to-end by minimizing denoising score-matching loss and (optionally) rating prediction loss, with user feedback providing implicit supervision.
• Personalization assessment: Employs a verifier network estimating $\mathbb{P}(\text{user }u \text{ likes image }I)$ and permutation tests to evaluate statistical dependence between generated images and user profiles.

3.3 LLM-Based Personalization for Representational Selection

• LLM post-training for artwork selection: LLMs (e.g., Llama 3.1 8B) are post-trained with LoRA adapters to select the optimal artwork representation (e.g., cover image for a film) conditioned on user history and candidate captions (Nam et al., 6 Jan 2026).
• Prompt engineering: User histories and candidate artwork captions are embedded in a single prompt with custom delimiter tokens, and reasoning distillation from a stronger LLM provides justifications to improve interpretability and accuracy.
• Losses: Supervised fine-tuning, direct preference optimization (DPO), and reasoning-augmented objectives.

4. Special Contexts: Art Therapy and Cross-Domain Personalization

Artwork recommendation has recently been applied in affective and therapeutic contexts, prompting the development of affect- and preference-aware models.

4.1 Visual Art Therapy and Recommender Systems

• Therapeutic goals: Personalized art exposure is designed to shift affective states, alleviate anxiety, and support post-ICU recovery (Yilma et al., 2024). Pipelines elicit baseline preference via emotion-eliciting “sample paintings,” compute content-based similarities in visual, textual, or multimodal embedding spaces, and measure affective outcomes with validated scales (Pick-A-Mood, PANAS, PHQ-4).
• Content safety and human-in-the-loop: Visual and multimodal engines are preferred over unscreened text-based models, with explicit expert review to filter potentially detrimental content (Yilma et al., 2024).
• Personalization strategies: Adaptive loops are proposed in which user affective feedback re-weights model scoring.

4.2 Affect-Aware Cross-Domain Recommendation

• Cross-modal transfer with music elicitation: Music ratings and affective responses are used to drive art recommendations through contrastive alignment in joint affective/semantic spaces, leveraging deep autoencoders, affective distance metrics, and multimodal fusion with pre-trained music and vision LLMs (Yilma et al., 18 Jul 2025).
• Evaluation: User-centric metrics (Accuracy, Diversity, Novelty, Serendipity, Immersion, Engagement), as well as pre/post affective assessments, are used. User studies in psychiatric sequelae populations validate that music-driven CDR methods outperform or match traditional visual-only approaches.

5. Evaluation Protocols and Empirical Results

Evaluation in personalized artwork recommendation leverages both offline and user-centric methodologies:

• Offline metrics: Precision@K, Recall@K, nDCG@K, AUC, and FID/CMMD (for generative models) (Messina et al., 2017, Messina et al., 2020, Patron et al., 5 Feb 2025).
• User-centric studies: Likert-scale ratings of recommendation match, diversity, serendipity, novelty, interpretability, and affective impact (Yilma et al., 2020, Yilma et al., 2024, Yilma et al., 2023).
• Expert evaluation: Art professionals assess the “meaningfulness” of recommendations (75% precision@5 for graph-based contemporary art recommendation (Fosset et al., 2022)).
• Statistical testing: Wilcoxon signed-rank test, permutation tests for personalization dependence, and ablation studies quantify benefit from each system component.

Methodology / Metric	Domain (Sample Paper)	Key Empirical Result
Content-based (visual, metadata, hybrid)	(Messina et al., 2017)	nDCG@5 DNN 0.0810, Hybrid 0.0841
Graph-based RWR (artwork + artist graphs)	(Fosset et al., 2022)	75% precision@5 (expert)
Triplet-based vision ranking	(Messina et al., 2020)	nDCG@20=0.0966 (CuratorNet)
Textual LDA/BERT-topic fusion	(Yilma et al., 2023, Yilma et al., 2020)	LDA outperforms ResNet for explainability
Generative diffusion-based personalization	(Patron et al., 5 Feb 2025)	FID=117.77 (lower than baselines), strong user-image dependence (p<1e-3)
LLM-based artwork selector	(Nam et al., 6 Jan 2026)	+5.21% IPS over baseline
Art therapy CDR (music-based)	(Yilma et al., 18 Jul 2025)	CDR outperforms visual baseline in several user-centric qualities

Best practices include temporal train/test splits (leave-one-out, session-based), statistical testing for significance, and direct user/expert involvement in subjective or affective settings. Interpretability is addressed by explicit topic disclosure or model-generated reasoning justifications.

6. Key Challenges, Design Principles, and Future Directions

Methodological Challenges

• Sparsity and cold start: Physical art markets and personalized collections are often one-of-a-kind and have few interactions; content-based and hybrid approaches remain critical.
• Interpretability: Deep visual models offer accuracy but little transparency; semantic models and reasoning distillation offer user trust and acceptance.
• Affective and safety concerns: In therapeutic settings, controlling for negative or inappropriate content is mandatory, requiring human-in-the-loop review and affective annotation (Yilma et al., 2024, Yilma et al., 18 Jul 2025).
• Diversity and novelty: Determinantal Point Process and submodular optimization help promote serendipity and avoid recommendation homogeneity (Fosset et al., 2022).

Emerging Strategies

• Prompting-free generative personalization: Combining semantic injection (LoRA/DiffLoRA) and genetic linguistic optimization enables aesthetic personalization in major art styles within minutes, using interactive user feedback (Zhou et al., 2024).
• Latent-space personalization for generation: REBECA and similar frameworks provide fully generative personalized artwork, relying on learned user embeddings and diffusion priors (Patron et al., 5 Feb 2025).
• LLM reasoning: Training LLMs on paired histories and artwork captions, with explicit justification generation, supports granular personalization of visual representations and explanations (Nam et al., 6 Jan 2026).

Open Research Questions

• Efficient representation of generative models for direct recommendation (Guo et al., 2023).
• Longitudinal measurement of affective and therapeutic outcomes (Yilma et al., 18 Jul 2025).
• Balancing diversity vs. preference fit in generative settings (Fosset et al., 2022, Guo et al., 2023).
• Multimodal and cross-domain transfer, e.g., music–art–text, for robust user modeling across creative domains (Yilma et al., 18 Jul 2025).

Deployment and Scalability

• Scalable feature extraction: Precompute embeddings (visual, textual, semantic) for catalog-scale rankings.
• Approximate nearest neighbor search: Accelerates inference for large candidate sets (e.g., FAISS for visual features).
• Incremental and online adaptation: Real-time profile updates, feedback loops, and periodic retraining are critical for live systems (Messina et al., 2017, Fosset et al., 2022).

7. Synthesis and Outlook

Personalized artwork recommendation has matured from classical metadata- and content-based filtering to graph-based, hybrid, and fully generative systems, now harnessing the modeling power of deep multimodal networks and LLMs. User studies and benchmarks demonstrate that coupling visual, textual, and affective representations is essential for both accuracy and user satisfaction, especially where subjective or therapeutic impact is central.

Current state-of-the-art models balance high-dimensional embedding spaces, fast approximate retrieval, dynamic user modeling, and robust evaluation protocols. The future trajectory points toward personalized artwork generation, cross-domain preference transfer (e.g., music to art), affect-aware frameworks for therapy and well-being, and seamless explainability, as research continues to push beyond the limitations of static catalogs and fixed-content recommendation (Patron et al., 5 Feb 2025, Nam et al., 6 Jan 2026, Yilma et al., 18 Jul 2025).