Recomendação Jurídica Interativa sob Alta Esparsidade: Uma Avaliação Empírica com Algoritmos Bandit

• Mateus Macorini Alves UEM
• Matheus Leme UEM
• Marcos Aurelio Domingues UEM

DOI: https://doi.org/10.5753/webmedia.2026.30974

Resumo

Recommender systems based on Multi-Armed Bandit (MAB) algorithms have been widely investigated in domains such as entertainment and e-commerce, where user-item interactions are relatively recurrent and collaborative signals are more abundant. However, their applicability to the legal domain remains underexplored. This work empirically evaluates the feasibility of bandit-based interactive recommendation for legal documents, a scenario characterized by high sparsity, large item catalogs, and scarce positive feedback. Using the iRec framework, we conducted offline experiments on a real legal interaction dataset, named JurisLog, comparing classical algorithms such as 𝜀-Greedy, Upper Confidence Bound (UCB), and Thompson Sampling against a non-personalized popularity baseline. To isolate the effect of sparsity and item recurrence, we constructed three versions of the dataset under different density constraints and compared them with reference datasets commonly used in the literature. The results showed that the legal domain imposes substantial structural challenges, with all evaluated algorithms achieving considerably lower performance than in traditional recommendation domains. Moreover, the global popularity baseline consistently outperformed the bandit-based approaches in the legal datasets, suggesting that sparse collaborative feedback limits the effectiveness of direct interactive learning. These findings indicate that traditional Multi-Armed Bandit methods are not directly sufficient for legal recommendation scenarios and highlight the need for hybrid approaches that incorporate semantic representations and domain-specific information.

Palavras-chave: Recommender Systems, Legal Recommendation, Bandit Algorithms, High Sparsity

Referências

Olivier Chapelle and Lihong Li. 2011. An empirical evaluation of thompson sampling. Advances in neural information processing systems 24 (2011).

Xiaocong Chen, Lina Yao, Julian McAuley, Guanglin Zhou, and Xianzhi Wang. 2021. A survey of deep reinforcement learning in recommender systems: A systematic review and future directions. arXiv preprint arXiv:2109.03540 (2021).

Keunho Choi, Donghee Yoo, Gunwoo Kim, and Yongmoo Suh. 2012. A hybrid online-product recommendation system: Combining implicit rating-based collaborative filtering and sequential pattern analysis. electronic commerce research and applications 11, 4 (2012), 309–317.

Leandro Carísio Fernandes, Leandro dos Santos Ribeiro, Marcos Vinícius Borela de Castro, Leonardo Augusto da Silva Pacheco, and Edans Flávius de Oliveira Sandes. 2026. JurisTCU: A Brazilian Portuguese Information Retrieval Dataset with Query Relevance Judgments. Language Resources and Evaluation 60 (2026), 23. DOI: 10.1007/s10579-025-09881-w

Tor Lattimore and Csaba Szepesvári. 2020. Bandit algorithms. Cambridge University Press.

Yixiao Ma, Yueyue Wu, Qingyao Ai, Yiqun Liu, Yunqiu Shao, Min Zhang, and Shaoping Ma. 2023. Incorporating Structural Information into Legal Case Retrieval. ACM Transactions on Information Systems 42, 2, Article 40 (2023), 28 pages. DOI: 10.1145/3609796

Miguel Martín, Antonio Jiménez-Martín, and Alfonso Mateos. 2019. A numerical analysis of allocation strategies for the multi-armed bandit problem under delayed rewards conditions in digital campaign management. Neurocomputing 363 (2019), 99–113.

James McInerney, Benjamin Lacker, Samantha Hansen, Karl Higley, Hugues Bouchard, Alois Gruson, and Rishabh Mehrotra. 2018. Explore, exploit, and explain: personalizing explainable recommendations with bandits. In Proceedings of the 12th ACM conference on recommender systems. 31–39.

Dattaraj Rao. 2020. Contextual bandits for adapting to changing user preferences over time. arXiv preprint arXiv:2009.10073 (2020).

Yuta Saito, Shunsuke Aihara, Megumi Matsutani, and Yusuke Narita. 2020. Open bandit dataset and pipeline: Towards realistic and reproducible off-policy evaluation. arXiv preprint arXiv:2008.07146 (2020).

Javier Sanz-Cruzado, Pablo Castells, and Esther López. 2019. A simple multiarmed nearest-neighbor bandit for interactive recommendation. In Proceedings of the 13th ACM conference on recommender systems. 358–362.

Sulthana Shams, Daron Anderson, and Douglas Leith. 2021. Cluster-based bandits: Fast cold-start for recommender system new users. In Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval. 1613–1616.

Thiago Silva, Nícollas Silva, Heitor Werneck, Carlos Mito, Adriano CM Pereira, and Leonardo Rocha. 2022. Irec: An interactive recommendation framework. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval. 3165–3175.

Thiago Silva, Nícollas Silva, Heitor Werneck, Adriano CM Pereira, and Leonardo Rocha. 2020. The impact of first recommendations based on exploration or exploitation approaches in recommender systems’ learning. In Proceedings of the Brazilian Symposium on Multimedia and the Web. 173–180.

Qing Wang, Chunqiu Zeng, Wubai Zhou, Tao Li, S Sitharama Iyengar, Larisa Shwartz, and Genady Ya Grabarnik. 2018. Online interactive collaborative filtering using multi-armed bandit with dependent arms. IEEE Transactions on Knowledge and Data Engineering 31, 8 (2018), 1569–1580.

Chunyi Zhou, Yuanyuan Jin, Xiaoling Wang, and Yingjie Zhang. 2020. Conversational music recommendation based on bandits. In 2020 IEEE international conference on knowledge graph (ICKG). IEEE, 41–48.