Dario Di Palma, Alessandro De Bellis, Giovanni Servedio, Vito Walter Anelli, Fedelucio Narducci, Tommaso Di Noia. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

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.

Unit testing is an essential but resource-intensive step in software development, ensuring individual code units function correctly. This paper introduces AgoneTest, an automated evaluation framework for Large Language Model-generated (LLM) unit tests in Java. AgoneTest does not aim to propose a novel test generation algorithm; rather, it supports researchers and developers in comparing different LLMs and prompting strategies through a standardized end-to-end evaluation pipeline under realistic conditions. We introduce the CLASSES2TEST dataset, which maps Java classes under test to their corresponding test classes, and a framework that integrates advanced evaluation metrics, such as mutation score and test smells, for a comprehensive assessment. Experimental results show that, for the subset of tests that compile, LLM-generated tests can match or exceed human-written tests in terms of coverage and defect detection. Our findings also demonstrate that enhanced prompting strategies contribute to test quality. AgoneTest clarifies the potential of LLMs in software testing and offers insights for future improvements in model design, prompt engineering, and testing practices.

Conversational Information Access systems have experienced widespread diffusion thanks to the natural and effortless interactions they enable with the user. In particular, they represent an effective interaction interface for conversational search (CS) and conversational recommendation (CR) scenarios. Despite their commonalities, CR and CS systems are often devised, developed, and evaluated as isolated components. Integrating these two elements would allow for handling complex information access scenarios, such as exploring unfamiliar recommended product aspects, enabling richer dialogues, and improving user satisfaction. As of today, the scarce availability of integrated datasets - focused exclusively on either of the tasks - limits the possibilities for evaluating by-design integrated CS and CR systems. To address this gap, we propose CoSRec, the first dataset for joint Conversational Search and Recommendation (CSR) evaluation. The CoSRec test set includes 20 high-quality conversations, with human-made annotations for the quality of conversations, and manually crafted relevance judgments for products and documents. Additionally, we provide supplementary training data comprising partially annotated dialogues and raw conversations to support diverse learning paradigms. CoSRec is the first resource to model CR and CS tasks in a unified framework, enabling the training and evaluation of systems that must shift between answering queries and making suggestions dynamically.

Large Language Models (LLMs) have become increasingly central to recommendation scenarios due to their remarkable natural language understanding and generation capabilities. Although significant research has explored the use of LLMs for various recommendation tasks, little effort has been dedicated to verifying whether they have memorized public recommendation dataset as part of their training data. This is undesirable because memorization reduces the generalizability of research findings, as benchmarking on memorized datasets does not guarantee generalization to unseen datasets. Furthermore, memorization can amplify biases, for example, some popular items may be recommended more frequently than others.

High Flow Nasal Cannula (HFNC) therapy is commonly used in infants hospitalized with bronchiolitis, but its early indication remains challenging. This study proposes an explainable machine learning framework to predict the need for HFNC at admission using real-world clinical data from 109 pediatric patients. After preprocessing, class balancing with SMOTE, and feature selection via RFECV, four classifiers were trained and compared. The Multi-Layer Perceptron (MLP) achieved the best performance, significantly outperforming other models in terms of F1-score. To ensure interpretability, SHAP values were used to identify key predictive features such as SpO2, CRP levels, and hospital stay duration while counterfactual explanations highlighted actionable decision boundaries. The proposed model combines predictive accuracy with clinical transparency, offering a promising tool to support early respiratory management in pediatric bronchiolitis

Maintaining a consistent brand identity across a global network of retail stores while adhering to local constraints has long challenged Visual Merchandisers.Legacy processes, often reliant on subjective "by-eye" adjustments, can drive up operational costs and lead to inconsistent in-store execution.We formalize a user modeling framework implementing a multi-criteria utility function that balances brand identity and operational overhead.We integrated our framework in a 3D virtual tour design platform, deploying it in the ecosystem of OVS, a global fashion firm.Through a preliminary user study, we showcase that our solution enables lower iteration cycles and decreases store-to-store discrepancies.

The rapid digitalization of the fashion industry has transformed marketing strategies, emphasizing the need for personalized and adaptive advertising content.This paper presents a case study on fine-tuning Large Language Models (LLMs) for fashion advertising, focusing on OVS, a major Italian fashion retailer.By leveraging real-world marketing data from OVS's newsletters and social media campaigns, we developed a fine-tuned model capable of generating engaging and stylistically coherent promotional content.To evaluate the effectiveness of this approach, we introduced a novel brand compliance index, measuring the alignment of AI-generated text with key branding requirements, such as audience targeting, event specificity, and platform appropriateness.Experimental results show that the fine-tuned model achieved a compliance score of 0.82, significantly outperforming the baseline model (0.63).Although this approach introduces a minor increase in generation latency, the enhanced alignment with brand identity justifies its use in marketing automation.Our findings highlight the potential of fine-tuned LLMs to streamline advertising content generation while maintaining brand consistency, offering valuable insights for the future of AI-driven digital marketing.

The investigation of neural correlates of mental imagery and perception has been a pivotal area of research in cognitive neuroscience, offering significant insights into how the brain represents both imagined and perceived experiences. While previous studies have successfully identified electrophysiological markers associated with motor and perceptual imagery, the neural signatures of motivational imagery—encompassing desires, needs, and cravings—remain underexplored. This study employs different machine learning classifiers applied to EEG data to classify and compare neural representations of twelve distinct motivational states under perception and imagery conditions. We conducted experiments using 14-channel and 18-channel EEG configurations to capture and analyze the neural responses of participants exposed to various stimuli. Our primary aims were to evaluate classification performance, measured by accuracy, and assess the impact of electrode density on performance. The results indicate that perception conditions generally yield higher accuracy in distinguishing motivational states than imagery conditions. Specifically, primary needs and somatosensory states exhibited strong and clear neural patterns in perception, with a peak accuracy of 88% in the 18-channel setup, while the accuracy for imagined states was more variable. Comparisons between 14-channel and 18-channel configurations revealed that higher electrode density slightly improved performance but was not significantly superior.

Software correctness is crucial, with unit testing playing an indispensable role in the software development lifecycle. However, creating unit tests is time-consuming and costly, underlining the need for automation. Leveraging Large Language Models (LLMs) for unit test generation is a promising solution, but existing studies focus on simple, small-scale scenarios, leaving a gap in understanding LLMs' performance in real-world applications, particularly regarding integration and assessment efficacy at scale. Here, we present AgoneTest, a system focused on automatically generating and evaluating complex class-level test suites. Our contributions include a scalable automated system, a newly developed dataset for rigorous evaluation, and a detailed methodology for test quality assessment.

Recommender systems, though widely used, often struggle to engage users effectively. While deep learning methods have enhanced connections between users and items, they often neglect the user’s perspective. Knowledge-based approaches, utilizing knowledge graphs, offer semantic insights and address issues like knowledge graph embeddings, hybrid recommendation, and interpretable recommendation. More recently, neural-symbolic systems, combining data-driven and symbolic techniques, show promise in recommendation systems, especially when used with knowledge graphs. Moreover, content features become vital in conversational recommender systems, which demand multi-turn dialogues. Recent literature highlights increasing interest in this area, particularly with the emergence of Large Language Models (LLMs), which excel in understanding user queries and generating recommendations in natural language. Sixth Knowledge-aware and Conversational Recommender Systems (KaRS) Workshop aims to disseminate advancements and discuss about challenges and opportunities.

Cardiovascular disease (CVD) is a general term referring to several heart or blood vessels abnormality. Heart failure (HF), directly associated to (CVD), is a significant global health problem as well as the leading cause of morbidity and mortality. The early detection of this condition is crucial for patient health. Traditional diagnostic methods for HF, such as history taking and physical examination, are often insufficient and require the use of advanced techniques such as Electrocardiogram (ECG). This study aims to extract temporal and morphological features from (ECG) signals and compare different Machine Learning (ML) classification models to enable rapid diagnosis and provide interpretable predictions. Specifically, we propose a Light Gradient Boosting (LGBM) model that can discriminate Normal Sinus Rhythm (NSR) and Arrhythmia (ARR) with a high accuracy of 0.99, achieving a Precision of 1.00, Recall of 0.99, and f1-score of 0.99 in the (NSR) class, and Precision of 0.99, Recall of 1.00, and f1-score of 0.99 in the (ARR) class, respectively. In addition, eXplainable Artificial Intelligence (XAI) analysis is performed to explain the model predictions.

Current AI regulations require discarding sensitive features (e.g., gender, race, religion) in the algorithm’s decision-making process to prevent unfair outcomes. However, even without sensitive features in the training set, algorithms can persist in discrimination. Indeed, when sensitive features are omitted (fairness under unawareness), they could be inferred through non-linear relations with the so-called proxy features. In this work, we propose a way to reveal the potential hidden bias of a machine learning model that can persist even when sensitive features are discarded. This study shows that it is possible to unveil whether the black-box predictor is still biased by exploiting counterfactual reasoning. In detail, when the predictor provides a negative classification outcome, our approach first builds counterfactual examples for a discriminated user category to obtain a positive outcome. Then, the same counterfactual samples feed an external classifier (that targets a sensitive feature) that reveals if the modifications to the user characteristics needed for a positive outcome moved the individual to the non-discriminated group. When this occurs, it could be a warning sign for discriminatory behavior in the decision process. Furthermore, we leverage the deviation of counterfactuals from the original sample to determine which features are proxies of specific sensitive information. Our experiments show that, even if the model is trained without sensitive features, it often suffers discriminatory biases.

Recommender systems have become ubiquitous in daily life, but their limitations in interacting with human users have become evident. Deep learning approaches have led to the development of data-driven algorithms that identify connections between users and items, but they often miss a critical actor in the loop - the end-user. Knowledge-based approaches are gaining attention due to the availability of knowledge-graphs, such as DBpedia and Wikidata, which provide semantics-aware information on different knowledge domains. These approaches are being used for recommendation and challenges such as knowledge graph embeddings, hybrid recommendation, and interpretable recommendation. Moreover, the emergence of neural-symbolic systems, which combine data-driven and symbolic methods, can significantly improve recommendation systems. A growing number of research papers on such topics demonstrate the growing interest and research potential of these systems. Furthermore, content features become crucial when interaction requires it. The development of conversational recommender systems presents new challenges, as they require multi-turn dialogues between users and systems, blurring the line between recommendation and retrieval. Evaluation of these systems goes beyond simple accuracy metrics and is hampered by the limited availability of datasets. While research and development into conversational recommender systems has been less prominent in the past, recent literature shows growing interest and potential for these systems.

Brain-computer interfaces are widely used to control machines using Electroencephalography (EEG) signals. Several low-cost electroencephalographs are available on the market that achieves good-quality EEG signals. One of the most intriguing issues for developing biofeedback systems is classifying users' emotional states using EEG signals and Machine Learning (ML) methods. In our study, we propose a novel ML-based biofeedback tool using a BCI to detect two different users' mental states: Focus, and Relaxation. We compared several ML algorithms achieving an average accuracy on the Test Set of 0.90 by using SVM. Finally, we propose a prototype for music generation according to the classification output that could be adopted in neurorehabilitation scenarios.

Mild Cognitive Impairment (MCI) is a syndrome charac-terized by cognitive impairment that is greater than expected for a subject's age and level of education. Nevertheless, it does not interfere with daily activity. Prevalence in epidemiological and population-based studies ranges from 3% to 19% in adults older than 65 years. A very interesting approach in this area is related to the identification of an Artificial Intelligence (Al)-based model and a subset of relevant features to predict the MCI clinical outcome. In our study, we propose a Pareto-optimality-based approach to identify the best model for predicting MCI. In fact, the best model achieves an Accuracy and Recall on Yes MCI of 71 % and 80% respectively. With this approach, it is possible to select the best model in order to predict Yes MCI (highest risk class). Our study presents a new best model selection approach that can be applied in identifying the best model that can be applied in various disease classification problems.

In the last few years, a renewed interest of the research community in conversational recommender systems (CRSs) has been emerging. This is likely due to the massive proliferation of Digital Assistants (DAs) such as Amazon Alexa, Siri, or Google Assistant that are revolutionizing the way users interact with machines. DAs allow users to execute a wide range of actions through an interaction mostly based on natural language utterances. However, although DAs are able to complete tasks such as sending texts, making phone calls, or playing songs, they still remain at an early stage in terms of their recommendation capabilities via a conversation. In addition, we have been witnessing the advent of increasingly precise and powerful recommendation algorithms and techniques able to effectively assess users’ tastes and predict information that may be of interest to them. Most of these approaches rely on the collaborative paradigm (often exploiting machine learning techniques) and neglect the huge amount of knowledge, both structured and unstructured, describing the domain of interest of a recommendation engine. Although very effective in predicting relevant items, collaborative approaches miss some very interesting features that go beyond the accuracy of results and move in the direction of providing novel and diverse results as well as generating explanations for recommended items. Knowledge-aware side information becomes crucial when a conversational interaction is implemented, in particular for preference elicitation, explanation, and critiquing steps.

Brain-Computer Interfaces allow controlling machines through signals coming from Electroencephalography (EEG) analysis. Nowadays, there are several cheap electroencephalographs available on the market that guarantee good quality EEG signals. A very interesting approach in this area is related to detecting the emotional states of a user through the analysis of her EEG signal. In our study, we tried to detect the emotional polarity (Valence), the state of emotional excitement (Arousal), and the level of emotion control (Dominance). Through metric interpolation and Russell’s circumplex model, it is possible to characterize and define the current emotional state of the user who wears the device. Our study presents a prototype of an EEG-based emotion recognizer that provides the user’s emotional state exploitable as bio-feedback.

Machine Learning could help the healthcare industry manage huge amounts of data and discover hidden trends and patterns that could help us better understand disease development and treatment. The goal is to define a Neural Network model (NN) to classify physical frailty in aging cohort to identify the frail food and clinical profile. In a 1, 929 older cohort from Southern Italy, the Food Frequency Questionnaire (FFQ) and clinical data were collected with blood tests. A NN was built with a hyperparameter tuning technique using accuracy as a performance parameter to select the best model. Confusion matrices, Garson and Olden’s variable importance were evaluated. Older age, female gender, high BMI, and high blood pressure were associated with physical frailty. In frail subjects, the lipid profile and RBC levels were significantly lower than their counterpart. On the contrary, serum levels of interleukin-6 and CRP were higher in the frail group. Frail subjects show higher consumption of spaghetti soup, pecorino cheese, fennel and chocolate, while a lower consumption of ham. The NN model has a respective training and testing accuracy of 86.49% and 85.77%. NN performs well on the train. The test dataset makes few mistakes and can predict healthy subjects with high specificity. According to Garson’s method, age, gender, foods rich in fats, and smoking habits are essential in predicting the frailty condition. In contrast, Olden’s method underlined the higher consumption of legumes and unrefined cereals.

Recommender systems (RSs) have widely grown thanks to the outstanding capability of providing users with accurate and tailored recommendations. Recently, public awareness and new regulations forced RS researchers and practitioners to study solutions to user privacy endangerment. This tutorial will guide the attendees through the possible threats and the solutions towards private RSs.

In the last few years, a renewed interest of the research community on conversational recommender systems (CRSs) is emerging. This is probably due to the great diffusion of Digital Assistants (DAs) such as Amazon Alexa, Siri, or Google Assistant that are revolutionizing the way users interact with machines. DAs allow users to execute a wide range of actions through an interaction mostly based on natural language messages. However, although DAs are able to complete tasks such as sending texts, making phone calls, or playing songs, they are still at an early stage on offering recommendation capabilities by using the conversational paradigm.

In this article, we present MyrrorBot , a personal digital assistant implementing a natural language interface that allows the users to: (i) access online services, such as music, video, news, and food recommendation s, in a personalized way, by exploiting a strategy for implicit user modeling called holistic user profiling ; (ii) query their own user models, to inspect the features encoded in their profiles and to increase their awareness of the personalization process. Basically, the system allows the users to formulate natural language requests related to their information needs. Such needs are roughly classified in two groups: quantified self-related needs (e.g., Did I sleep enough? Am I extrovert? ) and personalized access to online services (e.g., Play a song I like ). The intent recognition strategy implemented in the platform automatically identifies the intent expressed by the user and forwards the request to specific services and modules that generate an appropriate answer that fulfills the query. In the experimental evaluation, we evaluated both qualitative (users’ acceptance of the system, usability) as well as quantitative (time required to complete basic tasks, effectiveness of the personalization strategy) aspects of the system, and the results showed that MyrrorBot can improve the way people access online services and applications. This leads to a more effective interaction and paves the way for further development of our system.

More and more precise and powerful recommendation algorithms and techniques have been proposed over the last years able to effectively assess users' tastes and predict information that would probably be of interest for them. Most of these approaches rely on the collaborative paradigm (often exploiting machine learning techniques) and do not take into account the huge amount of knowledge, both structured and non-structured ones, describing the domain of interest for the recommendation engine. The aim of knowledge-aware and conversational recommender systems is to go beyond the traditional accuracy goal and to start a new generation of algorithms and interactive approaches which exploit the knowledge encoded in ontological and logic-based knowledge bases, knowledge graphs as well as the semantics emerging from the analysis and exploitation of semi-structured textual sources.

In this paper we present ExpLOD, a framework which exploits the information available in the Linked Open Data (LOD) cloud to generate a natural language explanation of the suggestions produced by a recommendation algorithm. The methodology is based on building a graph in which the items liked by a user are connected to the items recommended through the properties available in the LOD cloud. Next, given this graph, we implemented some techniques to rank those properties and we used the most relevant ones to feed a module for generating explanations in natural language. In the experimental evaluation we performed a user study with 308 subjects aiming to investigate to what extent our explanation framework can lead to more transparent, trustful and engaging recommendations. The preliminary results provided us with encouraging findings, since our algorithm performed better than both a non-personalized explanation baseline and a popularity-based one.

This paper presents T-RecS (Temporal analysis of Recommender Systems conference proceedings), a framework that supplies services to analyze the Recommender Systems Conference proceedings from the first edition, held in 2007, to the last one, held in 2015, under a temporal point of view. The idea behind T-RecS is to identify linguistic phenomena that reflect some interesting variations for the research community, such as topic drift, or how the correlation between two terms changed over time, or how similarity between two authors evolved over time.

In the last years, hundreds of social networks sites have been launched with both professional (e.g., LinkedIn) and non-professional (e.g., MySpace, Facebook) orientations. This resulted in a renewed information overload problem, but it also provided a new and unforeseen way of gathering useful, accurate and constantly updated information about user interests and tastes. Content-based recommender systems can leverage the wealth of data emerging by social networks for building user profiles in which representations of the user interests are maintained.

Basic content personalization consists in matching up the attributes of a user profile, in which preferences and interests are stored, against the attributes of a content object. This paper describes a content-based recommender system, called FIRSt, that integrates user generated content (UGC) with semantic analysis of content. The main contribution of FIRSt is an integrated strategy that enables a content-based recommender to infer user interests by applying machine learning techniques, both on official item descriptions provided by a publisher and on freely keywords which users adopt to annotate relevant items. Static content and dynamic content are preventively analyzed by advanced linguistic techniques in order to capture the semantics of the user interests, often hidden behind keywords. The proposed approach has been evaluated in the domain of cultural heritage personalization.

Recommender systems (RSs) proved to make easier the task of accessing relevant information in a broad range of domains. In content-based RSs, preferences on content items expressed by users turned out to be reliable indicators to suggest and filter interesting contents. Item representation plays a key role in content-based RSs, thus choosing proper facets to represent items is a fundamental task for deploying effective RSs. Contextual facets are often marginally relevant to predict user preferences, but in some domains disregarding contextual facets makes recommendations useless. This paper proposes a strategy to improve the effectiveness of a content-based RS that dynamically suggests tours within a museum by exploiting contextual facets such the physical layout of items and the interaction of users with the environment.