NIMS: MagIC - Documentos de conferências internacionais
URI permanente para esta coleção:
Navegar
Entradas recentes
- Spatiotemporal Statistical Analysis of Octopus Fishing Activity in the Algarve Based on Geographically Weighted Regression [poster]Publication . Oliveira, Beatriz; Costa, Ana Cristina; Information Management Research Center (MagIC) - NOVA Information Management School; NOVA Information Management School (NOVA IMS)
- Assessing BERTopic stability on a specialized scientific corpusPublication . Vasconcelos, Carolina; Mendonça, Sandro; Damásio, Bruno; Information Management Research Center (MagIC) - NOVA Information Management School; NOVA Information Management School (NOVA IMS)Neural topic models such as BERTopic have gained popularity for analyzing large text corpora, yet their stability properties remain underexplored. We evaluate BERTopic robustness along three dimensions using a corpus of 20,565 peer-reviewed abstracts published by banking institutions between 1980 and 2023. First, we measure seed sensitivity by fitting the same model specification across 11 random seeds and comparing document assignments, topic-level consistency, and word-level overlap. Document-level agreement ranges from 51% to 94%, while top-word Jaccard similarity is more stable (57–86%). Second, we compare five outlier reduction strategies and document a coherence–coverage trade-off: the baseline model achieves cv = 0:72 with 43% outliers, whereas the best post-hoc strategy reduces outliers to near zero at cv =0:66. Third, we assess temporal semantic stability by re-estimating topic representations within four decade windows and computing pairwise cosine similarities, which range from 0.76 to 0.99. These results show that BERTopic produces topics whose semantic content is more robust than their document assignments suggest, and that outlier handling and seed choice are first-order methodological decisions.
- Examining the Contradictions Between Centrality Measures and Self-Identified Influencers in Online Social NetworksPublication . Garcia, Ângela; António, Nuno; Rita, Paulo; NOVA Information Management School (NOVA IMS); Information Management Research Center (MagIC) - NOVA Information Management SchoolThe rapid growth of social media has significantly impacted how brands promote their products and interact with consumers. Consumers increasingly use the internet to gather information about products and brands. This fact has led brands to invest heavily in influencer marketing to boost brand awareness. Therefore, identifying influential figures who can help spread brand messages is crucial, and one effective way to achieve this is by calculating social network analysis’ centrality measures. This study explores the alignment between those identified through centrality measures in online social networks (OSNs) and self-proclaimed influencers. To validate the proposed methodology, this exploratory study uses Instagram data from the Portuguese brand Oliva Store as a case study. The analysis revealed a significant misalignment between self-identified influencers and those identified through network centrality measures. Among the various centrality measures, PageRank Centrality was the most effective, accurately identifying around 23% of self-proclaimed influencers. These findings challenge the notion that self-proclaimed influencers hold the highest influence. They highlight the complex dynamics of OSNs, where organizational entities can also play significant roles. This study provides critical insights for marketers and social media professionals, emphasizing the need for a nuanced approach to identifying and leveraging influencers to optimize marketing strategies.
- When Are Online Reviews Admissible Indicators of Perceived Tourism Pressure for Governance? [abstract]Publication . Arruda, Carlos; António, Nuno; NOVA Information Management School (NOVA IMS); Information Management Research Center (MagIC) - NOVA Information Management SchoolOvertourism increases the need for monitoring instruments that register when tourism intensity is experienced as pressure, rather than merely when activity volumes rise. Overtourism is often framed as a perceptual condition shaped by lived experience and tolerance thresholds, creating a measurement mismatch when monitoring relies on aggregated, low-frequency indicators such as arrivals, capacity, or economic contribution (Koens et al., 2018). Online reviews offer high temporal frequency and narrative detail that surface place-based frictions such as crowding, congestion, discomfort and disturbance, but participation is selective and platform-mediated (Yu & Egger, 2020). Automated text analysis can introduce distortions and validity risks that are rarely made explicit in governance applications (van Atteveldt et al., 2021). This study asks under what conditions online reviews can be interpreted as admissible indicators of perceived tourism pressure for cautious policy monitoring, separating interpretable signals from indeterminate outputs to reduce over-reaction (Peeters et al., 2018). This study adopts an admissibility-oriented design that treats review discourse as an experiential signal rather than a direct observation of pressure. The contribution is an interpretation and evaluation framework for review-based monitoring rather than a new natural language processing model or a definitive empirical account of overtourism dynamics. The objective is validation of interpretability rather than detection or prediction, prioritising specificity over sensitivity to reduce false-positive policy readings (Shmueli, 2010). Lisbon is used as a longitudinal demonstrator with TripAdvisor reviews for hotels, restaurants, and attractions from January 2023 to December 2025 analysed monthly. Multilingual content is treated as an interpretability constraint; extraction uses original-language review body text and excludes titles to avoid platform translation artefacts. Two semantic channels are derived from the same review corpus to support policy interpretation without over-committing to a single operationalisation. A conservative, definition-driven binary flag identifies overtourism-related expressions as an interpretable proxy within discourse while treating non-flagged text as non-evidence rather than absence of pressure. A complementary aspect-based framework extracts discrete criticism statements and maps them to a closed taxonomy designed for longitudinal comparability (Hu & Liu, 2004). Evidence snippets are retained to support auditability and reduce opaque aggregation in downstream reporting (Danilevsky et al., 2020). The closed taxonomy is treated as a governance-oriented validation decision that constrains interpretation to recurring, actionable dimensions of perceived strain rather than maximising descriptive coverage through unstable latent topics (Grimmer et al., 2022). Each extracted criticism statement is annotated with responsibility attribution, recording whether the narrative frames the friction as municipal, private, shared, or out of scope based only on textual support. Responsibility attribution is treated as a discourse framing label rather than a factual assignment of accountability and is only assigned when the text contains explicit institutional referents; otherwise the statement remains unassigned or out of scope (Foronda-Robles et al., 2025). The core contribution is a typology of interpretability states and failure modes operationalised through explicit guardrails. Outputs are classified as signal presence, interpretable non-expression when coverage is adequate and no admissible signal is observed, or non-interpretation when outputs are indeterminate due to missing coverage, sparsity, or extraction failure. Admissible signals are restricted to cases that satisfy guardrails designed to reduce volatility and selective expression artefacts, including temporal stability across multiple periods, recurrence across contributors, and semantic consistency at the aspect level (Cronbach & Meehl, 1955). Alignment checks against independent activity indicators are used for contextualisation rather than groundtruth validation, tightening plausibility when patterns co-move and constraining interpretation when they diverge because activity proxies may be internally valid yet not track perceived strain (OECD, 2025). Expected findings are methodological rather than substantive. The study delivers an admissibility framework that clarifies when review-derived outputs can be read as credible indicators of perceived pressure and when they must remain indeterminate, and it demonstrates how responsibility framing changes the governance reading of similar criticism types without making causal or accountability claims. A minimal longitudinal demonstrator illustrates how apparent signals in raw counts can become indeterminate under the guardrails while other signals become admissible when they recur across months and contributors, supporting cautious monitoring rather than definitive diagnosis (Kirilenko et al., 2023). Overall, the study proposes a conservative pathway for integrating online reviews into tourism policy interpretation as a monitoring layer, keeping claims proportionate to the mediated and selective nature of review discourse.
- Forecast Combination for Asset Classes ETFsPublication . Alvarez, Rodrigo Baggi Prieto; Bravo, Jorge Miguel; NOVA Information Management School (NOVA IMS); Information Management Research Center (MagIC) - NOVA Information Management SchoolThe Exchange-Traded Funds (ETFs) have transformed asset allocation, allowing investors to gain exposure to diverse asset classes with a single instrument. In turn, forecast combination models have emerged as advantageous methods for improving prediction accuracy. While the Efficient Market Hypothesis (EMH) posits that prices fluctuate randomly, making abnormal returns unattainable, empirical evidence reveals autocorrelation in stock returns, challenging the EMH's strict interpretation. This raises the question of whether new econometric and machine learning methods can predict asset values more effectively. We investigate the effectiveness of forecast combination in predicting financial asset classes. Analyzing ETFs across equity, fixed income, commodities, and cryptocurrency markets, we test the predictive accuracy of ten econometric and machine learning models, and preliminary results suggest that ensemble methods can indeed outperform simple models. Comparisons between forecasts and futures market prices reveal potential inefficiencies, suggesting opportunities for spot-futures index arbitrage. These findings contribute to discussions on market efficiency and highlight the role of forecast combinations in improving asset predictability and portfolio management.
- Revealing Process Structure in Urban Mobility NetworksPublication . Filonchik, Khristina; Pinto, José Pedro Diogo; Pinheiro, Flávio L.; Bação, Fernando; NOVA Information Management School (NOVA IMS); Information Management Research Center (MagIC) - NOVA Information Management SchoolUrban mobility is a multi-entity system that involves travelers, transport modes, and infrastructure. Beyond conventional origin/destination analysis, this paper investigates how process mining can structure and interpret mobility behavior from event data. Using Call Detail Records (CDRs) from Oeiras in the Lisbon metropolitan area (Portugal), we construct both case-centric and object-centric event logs and discover models that summarize flows and typical durations. Results show that most trips are intra-municipal, while inter-municipal flows connect strongly to neighboring areas, with typical inter-parish travel times of about 20 min. The object-centric perspective explicitly links trips and transport modes, revealing mode-specific duration differences (e.g., bus vs. car) that inform multimodal planning. Our contributions are: (i) a reproducible pipeline to transform CDRs into process mining artifacts, (ii) empirical evidence that mobility data exhibit a process-like structure, and (iii) the added value of object-centric models for multimodal analysis. Limitations include the low spatial precision of CDRs (tower-sector level) and heuristic transport-mode labels. Future work will integrate transport-network context (e.g., stations and routes) and model object-centric logs as heterogeneous graphs to enable richer and more reliable analysis.
- Inference of Firm-Firm Competing Networks in the Portuguese Public Procurement MarketPublication . Semedo, Luís P.; Pinto, Carolina; Monteiro, Beatriz; Sturm, Niclas Frederic; Damásio, Bruno; Pinheiro, Flávio L.; NOVA Information Management School (NOVA IMS); Information Management Research Center (MagIC) - NOVA Information Management SchoolThis work explores the Portuguese public procurement landscape by developing two Firm-Firm networks − one for open tenders (competitive procedure) and other for direct awards (non-competitive procedure). Using data from the official Portuguese public procurement platform, it explores and access these two different procedures and accounts for their differences when building and analyzing the networks. Results show that half of the firms involved in direct awards also participate in open tenders and their level of product and geographical specialization is slightly higher.
- Augmenting Firm Diversification Behavior Prediction with Graph EmbeddingsPublication . Sturm, Niclas F.; Candia, Cristian; Damásio, Bruno; Pinheiro, Flávio L.; NOVA Information Management School (NOVA IMS); Information Management Research Center (MagIC) - NOVA Information Management SchoolPublic procurement plays a crucial role in modern economies, serving not only as a means to procure the necessary services, works, and goods by public administrations but also as a policy tool to foster innovation and development. Here, we utilize machine learning models to predict the competitiveness of firms in securing contracts by activity sector, thereby enhancing our understanding of market dynamics, specifically by predicting firm capabilities and their evolution over time. To that end, we utilize a large dataset of public procurement contracts from 2009 to 2024, and we extend the existing literature on activity and technological diversification by applying these methodologies to a new domain, thereby contributing methodologically to the study of firm-level behavior. We develop a machine learning-based approach that enables a higher degree of explainability in the drivers of diversification outcomes, a common requirement in policy-relevant applications. We develop feature sets derived from a Node2Vec approach, inferring an embedding space using a network of firm activities. Our experiments show that machine learning models outperform heuristic baselines, including autocorrelation models of firm behavior, and achieve a performance comparable to feature sets derived from the relatedness paradigm. These findings suggest that embedding-space features may serve as substitutes for established measures of firm capabilities. Using predictive models has additional potential for decision makers in firms to identify future opportunities for diversification and gather market intelligence, as well as to estimate whether the firm can remain competitive in its activities.
- Mapping Mobility Networks Between Government Roles in PortugalPublication . Shaul, Carolina; Damásio, Bruno; Pinheiro, Flávio L.; NOVA Information Management School (NOVA IMS); Information Management Research Center (MagIC) - NOVA Information Management SchoolThis paper examines career mobility within Portuguese governments between 2011 and 2025, focusing on both formal government members (Ministers, Secretaries of State) and their private office staff. Using data collected from official government records, we construct networks where nodes correspond to either governmental positions or policy domains, and links capture shared careers. A descriptive analysis of network and centrality measures allowed for the identification of patterns on cross-portfolio mobility and career progression. Results show that advisory staff occupy central roles in connecting otherwise fragmented networks, while top political offices such as Ministers and Prime-Ministers are structurally peripheral. Moreover, policy domains such as Presidency and Modernization exhibit strong interconnections, suggesting their bridging role across government portfolios, while highly specialized policy domains rarely trade members with other portfolios.
- Revisiting SLIMPublication . Silva, Gorka; Stewart, Lachlan; Bakurov, Illya; Castelli, Mauro; Farinati, Davide; Muñoz Contreras, Jose Manuel; Trujillo, Leonardo; Vanneschi, Leonardo; NOVA Information Management School (NOVA IMS); Information Management Research Center (MagIC) - NOVA Information Management SchoolGeometric Semantic Genetic Programming (GSGP) induces unimodal error surfaces for supervised learning problems, enabling efficient search within semantic space. However, its conventional operators cause rapid model growth, limiting interpretability. The Semantic Learning algorithm based on Inflate and deflate Mutations (SLIM) addresses this limitation through size-reducing deflate mutations, while preserving GSGP’s theoretical properties. This work presents a systematic enhancement of SLIM through two sets of improvements: (i) theoretically grounded refinements, including optimal mutation step and linear scaling; and (ii) heuristic extensions, including Pareto-based tournament selection, multi-objective model identification, bounded mutation steps for implicit regularization and automatic algebraic simplification. A comprehensive evaluation on regression tasks demonstrates that all of the proposed SLIM enhancements, when considered separately, match or exceed baseline performance on test data, while achieving a significant reduction of model size. However, the best results were achieved when all enhancements are employed together. These results position the enhanced SLIM as a promising step toward more accurate, efficient, and interpretable symbolic regression for real-world data modeling.