<p>This paper addresses the Wi-Fi fingerprinting-based indoor localization problem within a federated learning framework, where some unlabeled data are crowdsourced to the server. Wi-Fi fingerprinting involves building and maintaining a comprehensive database, which is expensive and time-consuming. Recent frameworks leveraging federated learning and crowdsourcing aim to address this issue, where clients collaboratively train a global model without sharing private data. However, the performance of the global model can degrade when clients have non-independent and identically distributed (non-IID) data or when they are incongruent: They assign different labels for the same fingerprint. The model may become biased toward distributions with more data or clients, requiring the training of multiple global models. We propose a new approach combining semi-supervised and clustered federated learning to determine the optimal number of global models using unlabeled data at the server without prior knowledge of the number of global models. Semi-supervised learning helps to improve the generalization performance or domain adaptation of a model by leveraging unlabeled data. We use a pseudo-labeling approach for semi-supervised learning, which assigns pseudo-labels with high-confidence predictions to the unlabeled data using the trained model on the labeled data. Further, in our methodology, each client trains a local model based on its non-IID data, uploads the model weights to the server, and the server clusters the models based on cosine similarity, reflecting client data distributions. We evaluate the proposed method using the UJIIndoor and Tampere University datasets, and experimental results show that it significantly outperforms existing approaches.</p>