<p>The remarkable progress of federated learning (FL) has enhanced a range of tasks within computer vision, graph learning, and natural language processing fields. Existing frameworks such as Federated Graph Learning (FGL), TensorFlow Federated (TFF), and Federated AI Technology Enabler (FATE) facilitate its application in real-world scenarios. Formal Concept Analysis (FCA) is among the best-known unsupervised knowledge discovery techniques with a sound mathematical foundation. FCA has demonstrated its utility in addressing practical challenges within fields such as software engineering, knowledge engineering, and data mining. However, the adoption of federated FCA learning (FedFCA) remains limited, as its integration faces hurdles owing to the characteristics of lattice structures and formal context representation, which are central to FCA foundations, even though these elements are widely prevalent. Inspired by the evident demand for such a framework, we begin by identifying the hurdles in crafting an intuitive FedFCA framework. Following this exploration, in this paper, we unveil our contribution: FedFCA, a federated framework designed to provide Federated Formal Concept Analysis, which is a cutting-edge approach that uses differential privacy to ensure data protection for local parties. First, each party generates and encrypts the noisy local lattice, which contains the local data. Afterwards, a global server combines the encrypted noisy lattices from multiple untrusted parties, preserving the privacy of each party’s data. We conceptualize the comprehensive framework (FedFCA) as an autonomic microservice network of loosely integrated services, harmonizing with the principles of autonomy and decentralized management intrinsic to self-* systems. This approach facilitates synchronous or asynchronous communication strategies between servers and participants, encompassing both local lattice and global lattice mining processes. Extensive evaluation on real-world datasets validates that FedFCA achieves: (i) shorter runtimes compared to centralized FCA baselines; (ii) superior scalability to 250+ distributed providers; (iii) high global stability under different privacy budgets; and (iv) significant improvements in privacy-utility trade-offs over state-of-the-art federated mining of frequent itemsets (FedFIM and FedFPM). FedFCA is made publicly available as part of the FedFCA Framework suite on GitHub (<a href="https://github.com/msellamiTN/fedfca-framework.git">https://github.com/msellamiTN/fedfca-framework.git</a>), encouraging further FCA research and the expansion of applications previously hindered by the absence of specialized tools.</p>

FedFCA: Federated Formal Concept Analysis Framework

期刊论文
发表日期：

2026年1月6日

查看全文

Saida Sfaxi,
Mokhtar Sellami,
Mohamed Mohsen Gammoudi

摘要

The remarkable progress of federated learning (FL) has enhanced a range of tasks within computer vision, graph learning, and natural language processing fields. Existing frameworks such as Federated Graph Learning (FGL), TensorFlow Federated (TFF), and Federated AI Technology Enabler (FATE) facilitate its application in real-world scenarios. Formal Concept Analysis (FCA) is among the best-known unsupervised knowledge discovery techniques with a sound mathematical foundation. FCA has demonstrated its utility in addressing practical challenges within fields such as software engineering, knowledge engineering, and data mining. However, the adoption of federated FCA learning (FedFCA) remains limited, as its integration faces hurdles owing to the characteristics of lattice structures and formal context representation, which are central to FCA foundations, even though these elements are widely prevalent. Inspired by the evident demand for such a framework, we begin by identifying the hurdles in crafting an intuitive FedFCA framework. Following this exploration, in this paper, we unveil our contribution: FedFCA, a federated framework designed to provide Federated Formal Concept Analysis, which is a cutting-edge approach that uses differential privacy to ensure data protection for local parties. First, each party generates and encrypts the noisy local lattice, which contains the local data. Afterwards, a global server combines the encrypted noisy lattices from multiple untrusted parties, preserving the privacy of each party’s data. We conceptualize the comprehensive framework (FedFCA) as an autonomic microservice network of loosely integrated services, harmonizing with the principles of autonomy and decentralized management intrinsic to self-* systems. This approach facilitates synchronous or asynchronous communication strategies between servers and participants, encompassing both local lattice and global lattice mining processes. Extensive evaluation on real-world datasets validates that FedFCA achieves: (i) shorter runtimes compared to centralized FCA baselines; (ii) superior scalability to 250+ distributed providers; (iii) high global stability under different privacy budgets; and (iv) significant improvements in privacy-utility trade-offs over state-of-the-art federated mining of frequent itemsets (FedFIM and FedFPM). FedFCA is made publicly available as part of the FedFCA Framework suite on GitHub (https://github.com/msellamiTN/fedfca-framework.git), encouraging further FCA research and the expansion of applications previously hindered by the absence of specialized tools.