With the growing interest in Multimodal Recommender Systems (MRSs), collecting high-quality datasets provided with multimedia side information (text, images, audio, video) has become a fundamental step. However, most of the current literature in the field relies on small- or medium-scale datasets that are either not publicly released or built using undocumented processes. In this paper, we aim to fill this gap by releasing M3L-10M and M3L-20M, two large-scale, reproducible, multimodal datasets for the movie domain, obtained by enriching with multimodal features the popular MovieLens-10M and MovieLens-20M, respectively. By following a fully documented pipeline, we collect movie plots, posters, and trailers, from which textual, visual, acoustic, and video features are extracted using several state-of-the-art encoders. We publicly release mappings to download the original raw data, the extracted features, and the complete datasets in multiple formats, fostering reproducibility and advancing the field of MRSs. In addition, we conduct qualitative and quantitative analyses that showcase our datasets across several perspectives. This work represents a foundational step to ensure reproducibility and replicability in the large-scale, multimodal movie recommendation domain. Our resource can be fully accessed at the following link: https://zenodo.org/records/18499145, while the source code is accessible at https://github.com/giuspillo/M3L_10M_20M.

The evaluation of Conversational Recommender Systems necessitates robust protocols to measure utility and user satisfaction. While human-in-the-loop testing remains the gold standard, scalability and reproducibility constraints have driven the field toward User Simulators. However, current simulation paradigms predominantly utilize rigid templates or closed-source Large Language Models that exhibit idealized behaviors. These approaches fail to capture user ambiguity, resulting in benchmarks that overestimate system proficiency by assuming crystallized user intent. To address this limitation, we introduce a family of open-weight user simulation models capable of generalizing across diverse e-commerce domains. Leveraging Teacher-Student distillation, we operationalize three distinct behavioral stereotypes: Direct, Vague-Proactive, and Vague-Reactive. Our evaluation of state-of-the-art Agentic Generative Conversational Recommender Systems reveals a critical Robustness Gap: while agents perform proficiently with decisive users, performance collapses when facing passivity and ambiguity. These findings underscore the necessity of our scalable framework for rigorously stress-testing the next generation of conversational agents against realistic, non-cooperative user behaviors.

The integration of Large Language Models into recommender systems has introduced a new paradigm in which models leverage their pre-trained knowledge to generate recommendations. A prevailing assumption is that an LLM’s inherent knowledge is sufficient to support high-quality recommendations across diverse domains. This article challenges that assumption, positing that in specialized domains, recommendation efficacy is limited by the textual nature of an LLM’s knowledge. To address this limitation, we propose REKALM, a comprehensive integration framework for enhancing LLM-based recommenders through knowledge integration. Central to our approach is the extension of established text-conversion techniques to non-traditional data modalities. We utilize knowledge lexicalization , a process that translates heterogeneous data sources into a unified natural language format. This lexicalized corpus is then used in a knowledge-aware instruction-tuning pipeline to explicitly align the LLM’s internal representations with four distinct types of domain-specific information. We conduct experiments across four distinct domains to validate our framework. Our findings provide consistent empirical evidence that while an LLM’s inherent knowledge may suffice for universally familiar domains like movies, recommendation quality in more specialized areas is significantly improved through knowledge integration. The proposed approach demonstrates that augmenting LLMs with lexicalized, domain-specific knowledge is an effective system-level strategy for advancing the next generation of recommender systems.

The last few years have seen an increasing interest of the RecSys community in the multimodal recommendation research field, as shown by the numerous contributions proposed in the literature. Our paper falls in this research line, as we released a multimodal extension of three state-of-the-art datasets (MovieLens-1M, DBbook, Last.fm-2K) in the movie, book, music recommendation domains, respectively. Although these datasets have been widely adopted for classical recommendation tasks (e.g., collaborative filtering), their use in multimodal recommendation has been hindered by the absence of multimodal information. To fill this gap, we have manually collected multimodal item raw files from different modalities (text, images, audio, and video, when available) for each dataset. Specifically, we have collected, for MovieLens-1M, movie plots (textual information), movie posters (images) and movie trailers (audio and video); for Last.fm-2K, we have collected, for each artist, the tags provided by users (textual information), the most popular album covers (images), and the most popular songs (audio); finally, for DBbook we have collected book abstracts (textual information) and book covers (image). We encoded all this information using state-of-the-art feature encoders, and we released the extended datasets, which include the mappings to the raw multimodal information and the encoded features. Finally, we conduct a benchmark analysis of various recommendation models using MMRec as a multimodal recommendation framework. Our results show that multimodal information can further enhance the quality of recommendations in these domains compared to single collaborative filtering. We release the multimodal version of such datasets to foster this research line, including links to download the raw multimodal files and the encoded item features.

In this paper, we propose a recommendation model that exploits a graph augmentation technique based on Large Language Models (LLMs) to enrich the information encoded in its underlying Knowledge Graph (KG). Our work relies on the assumption that the triples encoded in a KG can often be noisy or incomplete, and this may lead to sub-optimal modeling of both the characteristics of items and the users' preferences. In this setting, graph augmentation can be a suitable solution to improve the quality of the data model and provide users with high-quality recommendations. Accordingly, in this work, we align with this research line and propose GAL-KARS (GraphAugmentation with LLMs forKnowledgeAware Recommender Systems). In our framework, we start from a KG, and we design some prompts for querying an LLM and augmenting the graph by incorporating: (a) further features describing the items; (b) further nodes describing the preferences of the users, obtained by reasoning over the items they like. The resulting KG is then passed through a Knowledge Graph Encoder that learns users' and items' embeddings based on the augmented KG. These embeddings are finally used to train a recommendation model and provide users with personalized suggestions. As shown in the experimental session, graph augmentation based on LLMs can significantly improve the predictive accuracy of our recommendation model, thus confirming the effectiveness of the model and the validity of our intuitions.

Conversational Recommender Systems (CRSs) facilitate item discovery through multi-turn dialogues that elicit user preferences via natural language interaction. This field has gained significant attention following advancements in Natural Language Processing (NLP) enabled by Large Language Models (LLMs). However, current CRS research remains constrained by datasets with fundamental limitations. Human-generated datasets suffer from inconsistent dialogue quality, limited domain expertise, and insufficient scale for real-world application, while synthetic datasets created with proprietary LLMs ignore the diversity of real-world user behavior and present significant barriers to accessibility and reproducibility. The development of effective CRSs depends critically on addressing these deficiencies. To this end, we present DistillRecDial, a novel conversational recommendation dataset generated through a knowledge distillation pipeline that leverages smaller, more accessible open LLMs. Crucially, DistillRecDial simulates a range of user types with varying intentions, preference expression styles, and initiative levels, capturing behavioral diversity that is largely absent from prior work. Human evaluation demonstrates that our dataset significantly outperforms widely adopted CRS datasets in dialogue coherence and domain-specific expertise, indicating its potential to advance the development of more realistic and effective conversational recommender systems.

Recommender systems (RSs) have become increasingly versatile, finding applications across diverse domains. Large Language Models (LLMs) significantly contribute to this advancement since the vast amount of knowledge embedded in these models can be easily exploited to provide users with high-quality recommendations. However, current RSs based on LLMs have room for improvement. As an example, knowledge injection techniques can be used to finetune LLMs by incorporating additional data, thus improving their performance on downstream tasks. In a recommendation setting, these techniques can be exploited to incorporate further knowledge, which can result in a more accurate representation of the items. Accordingly, in this paper, we propose a pipeline for knowledge injection specifically designed for RS. First, we incorporate external knowledge by drawing on three sources: (a) knowledge graphs; (b) textual descriptions; (c) collaborative information about user interactions. Next, we lexicalize the knowledge, and we instruct and fine-tune an LLM, which can easily return a list of recommendations. Extensive experiments on movie, music, and book datasets validate our approach. Moreover, the experiments showed that knowledge injection is particularly needed in domains (i.e., music and books) where the encoded knowledge within LLMs may not be suitable for recommendation tasks, even if such content was used during the training of the model. This finding points to several promising future research directions.

In this paper, we present a methodology to provide users with knowledge-aware recommendations based on the fusion of multimodal item embeddings. Our approach relies on the intuition that each modality (i.e., graph, text, video, images, etc.) emphasizes different characteristics and nuances of the items, so it is necessary that a comprehensive knowledge-aware recommender system (KARS) encodes and exploits all the different data sources that are available in a specific domain. Accordingly, we design a multimodal KARS architecture based on a deep neural network that: (a) learns a representation of each uni-modal feature (i.e., description, trailers, covers, audio signals, and so on) through an appropriate encoder; (b) exploits self-attention and cross-attention to fuse the different sources and refine the embeddings; (c) returns a prediction score which represents user’s interest in the item, which is finally used to generate a top-k recommendation list. In the evaluation, we carried out experiments against two datasets, and the results showed that our approach overcame several baselines for multimodal and knowledge-aware recommendations, thus confirming the intuitions behind this work.

Conversational Recommender Systems (CRS) have recently drawn attention due to their capacity of delivering personalized recommendations through multi-turn natural language interactions. In this paper, we fit into this research line and we introduce a Knowledge-Aware Sequential Conversational Recommender System (KASCRS) that exploits transformers and knowledge graph embeddings to provide users with recommendations in a conversational setting. In particular, KASCRS is able to predict a suitable recommendation based on the elements that are mentioned in a conversation between a user and a CRS. To do this, we design a model that: (i) encodes each conversation as a sequence of entities that are mentioned in the dialogue (i.e., items and properties), and (ii) is trained on a cloze task, that is to say, it learns to predict the final element in the sequence - that corresponds to the item to be recommended - based on the information it has previously seen. The model has two main hallmarks: first, we exploit Transformers and self-attention to capture the sequential dependencies that exist among the entities that are mentioned in the training dialogues, in a way similar to session-based recommender systems [25]. Next, we used knowledge graphs (KG) to improve the quality of the representation of the elements mentioned in each sequence. Indeed, we exploit knowledge graph embeddings techniques to pre-train the representation of items and properties, and we fed the input layer of our architecture with the resulting embeddings. In this way, KASCRS integrates both knowledge from the KGs as well as the dependencies and the co-occurrences emerging from conversational data, resulting in a more accurate representation of users and items. Our experiments confirmed this intuition, since KASCRS overcame several state-of-the-art baselines on two different datasets.

In this paper, we present a knowledge-aware recommendation framework based on neuro-symbolic graph embeddings that encode first-order logical (FOL) rules. In particular, our workflow starts from a knowledge graph (KG) encoding user preferences (based on explicit ratings [13]) and item properties. Next, knowledge-aware recommendation are obtained through the combination of three modules: (i) a rule learner, that extracts FOL rules from the KG; (ii) a graph embedding module, that learns the embeddings of users and items based on the triples of the KG and the FOL rules previously extracted; (iii) a recommendation module that uses the embeddings to feed a deep learning architecture. In the experimental session, we evaluate the effectiveness of our strategy on two datasets and the results show that the combination of KG embeddings and FOL rules led to an improvement in the accuracy and in the novelty of the recommendations.

In this paper, we present a strategy to provide users with explainable cross-domain recommendations (CDR) that exploits large language models (LLMs). Generally speaking, CDR is a task that is hard to tackle, mainly due to data sparsity issues. Indeed, CDR models require a large amount of data labeled in both source and target domains, which are not easy to collect. Accordingly, our approach relies on the intuition that the knowledge that is already encoded in LLMs can be used to more easily bridge the domains and seamlessly provide users with personalized cross-domain suggestions. To this end, we designed a pipeline to: (a) instruct a LLM to handle a CDR task; (b) design a personalized prompt, based on the preferences of the user in a source domain, and a list of items to be ranked in target domain; (c) feed the LLM with the prompt, in both zero-shot and one-shot settings, and process the answer in order to extract the recommendations and a natural language explanation. As shown in the experimental evaluation, our approach beats several established state-of-the-art baselines for CDR in most of the experimental settings, thus showing the effectiveness of LLMs also in this novel and scarcely investigated scenario.

In this paper, we introduce a Knowledge-aware Recommender System (KARS) based on Graph Neural Networks that exploit pre-trained content-based embeddings to improve the representation of users and items. Our approach relies on the intuition that textual features can describe the items in the catalog from a different point of view, so they are worth to be exploited to provide users with more accurate recommendations. Accordingly, we used encoding techniques to learn a pre-trained representation of the items in the catalogue based on textual content, and we used these embeddings to feed the input layer of a KARS based on GCNs. In this way, the GCN is able to encode both the knowledge coming from the unstructured content and the structured knowledge provided by the KG (ratings and item descriptive properties). As shown in our experiments, the exploitation of pre-trained embeddings improves the predictive accuracy of the KARS, which overcomes all the baselines we considered in several experimental settings.