• We stress the overlooked role of feature extraction and processing phase in the standard multimodal recommendation pipeline. • Research often follows limited experimental settings, ignoring powerful new extraction models with careful selection their hyper-parameters, new datasets from uncommon domains, and usually-untested modalities. • We provide a new, end-to-end framework for standardized benchmarking which, unlike other recent benchmarking studies in multimodal recommendation, incorporates Ducho (a feature extraction framework), and Elliot / MMRec (two popular frameworks for reproducible multimodal recommendation); we highlight the implementative challenges to make all these separate frameworks interact under the same experimental pipeline. • We run ∼ 4,000 experiments spanning 8 datasets, 8 multimodal extractors, 15 (multimodal) recommender systems, under 5 different experimental settings. • Our comprehensive pipeline allows extensive and diversified benchmarks in multimodal recommendation. We observe (in most cases) performance improvements with more recent multimodal extractors, whose outcomes are validated under various domains, modalities, and extractors hyper-parameters. With the advent of deep learning and, more recently, large models, recommendation systems have greatly refined their capability of profiling users’ preferences and interests that, in most cases, are complex to disentangle. This is especially true for those recommendation algorithms that rely heavily on external side information, such as multimodal recommender systems. In specific domains like fashion, music, and movie recommendation, the multi-faceted features characterizing products and services may influence each customer on online platforms differently, paving the way to novel multimodal recommendation models that can learn from such multimodal content. According to the literature, the common multimodal recommendation pipeline involves (i) extracting multimodal features, (ii) refining their high-level representations to suit the recommendation task, (iii) optionally fusing all multimodal features, and (iv) predicting the user-item score. Although great effort has been put into designing optimal solutions for (ii-iv), to the best of our knowledge, very little attention has been devoted to exploring procedures for (i) in a rigorous way. In this respect, the existing literature outlines the large availability of multimodal datasets and the ever-growing number of large models accounting for multimodal-aware tasks, but (at the same time) an unjustified adoption of limited standardized solutions. As very recent works from the literature have begun to conduct empirical studies to assess the contribution of multimodality in recommendation, we decide to follow and complement this same research direction. To this end, this paper settles as the first attempt to offer a large-scale benchmarking for multimodal recommender systems, with a specific focus on multimodal extractors. Specifically, we take advantage of three popular and recent frameworks for multimodal feature extraction and reproducibility in recommendation, Ducho , and MMRec / Elliot , respectively, to offer a unified and ready-to-use experimental environment able to run extensive benchmarking analyses leveraging novel multimodal feature extractors. Results, largely validated under different extractors, hyper-parameters of the extractors, domains, and modalities, provide important insights on how to train and tune the next generation of multimodal recommendation algorithms.

Visual compatibility modeling is central to modern fashion recommendation systems. A key task is complementary item retrieval , where the goal is to identify a garment that harmonizes with a reference item, such as retrieving a compatible bottom for a given top. Recent generative approaches synthesize candidate garments to guide retrieval, but they either treat generation as an auxiliary signal or rely on computationally demanding architectures, limiting their practicality in large-scale deployments. In this work, we introduce GeCo , a generative–compositional framework that couples image synthesis and retrieval within an effective, two-stage design. In the first stage, a conditional GAN generates visually coherent bottom templates from top images; in the second stage, the generated template and top form a composed visual query for compatibility-based retrieval. This decoupled design avoids the heavy optimization pipelines used in prior generative approaches, resulting in stable training and low computational cost, while allowing the generated images to guide compatibility modeling. Experiments on three benchmarks, including the new FashionTaobaoTB dataset released with this work, show that GeCo offers competitive retrieval accuracy with a low memory footprint. Human evaluations further indicate that its generated items are perceived as realistic and stylistically compatible, supporting its suitability for practical, resource-constrained fashion recommendation scenarios. • GeCo : generative fashion compatibility model and complementary item retrieval. • GeCo uses a two-stage GAN under realistic hardware limits. • Human studies find its garments realistic and stylistically coherent. • FashionTaobaoTB and benchmarking codebase have been released.

Generative modeling for 3D brain MRI is challenged by a trade-off between anatomical fidelity, sample diversity, and computational efficiency. Diffusion-based approaches achieve strong visual quality but typically require hundreds to thousands of sampling steps, while latent-space compression can introduce reconstruction artifacts and degrade fine-grained anatomy. We introduce FlowLet, a conditional generative framework that performs Flow Matching in an invertible 3D wavelet domain. This representation enables multi-scale generation without learned latent compression, while deterministic ODE sampling allows fast inference. Age conditioning is modeled through complementary feature-wise modulation and spatially adaptive cross-attention, enabling explicit control over age-related morphological variation. Across multi-site neuroimaging datasets, FlowLet achieves competitive and, in several settings, superior global fidelity compared to diffusion-based baselines using as few as 10 sampling steps. Region-based evaluation across 95 cortical and subcortical brain regions demonstrates improved local anatomical plausibility beyond what is captured by global similarity metrics alone. In a downstream brain age prediction study, models augmented with FlowLet-generated data consistently reduce prediction error relative to real-only training and other generative baselines. Rather than focusing on a single dominant metric improvement, these results highlight a consistent trade-off between efficiency, controllability, and anatomically meaningful 3D brain MRI generation. The proposed framework is released as open-source to support reproducibility.

Multimodal Recommender Systems aim to improve recommendation accuracy by integrating heterogeneous content, such as images and textual metadata. While effective, it remains unclear whether their gains stem from true multimodal understanding or increased model complexity. This work investigates the role of multimodal item embeddings, emphasizing the semantic informativeness of the representations. Initial experiments reveal that embeddings from standard extractors (e.g., ResNet50, Sentence-Bert) enhance performance, but rely on modality-specific encoders and ad hoc fusion strategies that lack control over cross-modal alignment. To overcome these limitations, we leverage Large Vision-Language Models (LVLMs) to generate multimodal-by-design embeddings via structured prompts. This approach yields semantically aligned representations without requiring any fusion. Experiments across multiple settings show notable performance improvements. Furthermore, LVLMs embeddings offer a distinctive advantage: they can be decoded into structured textual descriptions, enabling direct assessment of their multimodal comprehension. When such descriptions are incorporated as side content into recommender systems, they improve recommendation performance, empirically validating the semantic alignment encoded in LVLMs outputs. Our study highlights the importance of semantically rich representations and positions LVLMs as a compelling foundation to build robust and meaningful multimodal representations in recommendation tasks.

Pre-trained language models (PLMs) are widely used to derive semantic representations from item metadata in recommendation and search.In sequential recommendation, PLMs enhance ID-based embeddings through textual metadata, while in product search, they align item characteristics with user intent.Recent studies suggest task and domain-specific fine-tuning are needed to improve representational power.This paper challenges this assumption for e-commerce applications, showing that Generalist Text Embedding Models (GTEs), pre-trained on large-scale corpora, can guarantee strong zero-shot performance without specialized adaptation.Our experiments on popular e-commerce benchmarks demonstrate that GTEs outperform traditional and fine-tuned models in both sequential recommendation and product search.We attribute this to a superior representational power, as they distribute features more evenly across the embedding space.Finally, we show that compressing embedding dimensions by focusing on the most informative directions (e.g., via PCA) effectively reduces noise and improves the performance of specialized models.To ensure reproducibility, we provide our repository at https://github.com/sisinflab/GTE-Zero- Shot-Recsys.

Recommendation pipelines involve several stages that can critically affect performance and reproducibility. However, early pipeline stages remain under-standardized, limiting comparability and interoperability across studies. This tutorial addresses this gap by providing both theoretical insights and hands-on experience with tools and practices for standardized data processing in recommender systems. In the first part, we introduce DataRec, a Python library for reproducible and interoperable data management, and discuss data filtering, splitting, and topological analysis techniques. In the second part, we explore multimodal feature extraction in domains such as fashion, music, and movies, focusing on the challenges of meaningful multimodal integration. We introduce Ducho, a unified framework for extracting audio, visual, and textual features using modern backends, and demonstrate its integration with the evaluation framework Elliot. The tutorial targets researchers and practitioners with an interest in recommender systems, data preprocessing, and multimodal modeling. All materials, including slides, code, datasets, and recordings, will be openly available on a dedicated tutorial website: https://sites.google.com/view/dd4rec-tutorial/.