Saloni Kwatra

Republic of India

Umeå University

Federated Learing: A collavorative and distributed machine learning framework with heterogeneous data


With the spread of AI, users/data subjects are becoming increasingly worried about how their personal information is being used or even exploited without their permission for commercial and political objectives. Because diverse individuals and organizations generate and own data, the old and naive method of transferring all data to a single location where powerful systems can train machine learning models is no longer desirable, as it is not efficient and violates data privacy. Federated Learning (FL) is an effective answer to this problem in this situation. The aim is to train a model at each device where a data source is located, then let the devices communicate their models to establish a global model consensus.

The original definition of FL is about building deep learning models. It is well known that deep learning models are in, most contexts, the most effective and with the best accuracy among other alternative data-driven models. Nevertheless, they are not so good with respect to their explainability. In contrast, other models and, specially, decision trees, provide explainability. Because of that, in our research project, we propose a framework for federated learning to build decision trees. Our approach will also incorporate privacy guarantees. According to how the data is partitioned among multiple parties in the feature and sample spaces, researchers classified FL into Horizontal Federated Learning (HFL), Vertical Federated Learning (VFL), and Federated Transfer Learning (FTL). HFL is a sample-partitioned FL, which means that datasets on various devices share the overlapping feature space but have separate sample space. VFL is a feature-partitioned FL, which is just the opposite of HFL. FTL can be used in situations when neither the features nor the data samples overlap. We propose a framework for implementing HFL with decision trees.

In this work, each device trains a decision tree classifier. Devices share the root node of their trees with the aggregator. The aggregator merges the trees by choosing the most common split attribute and grows the branches based on the split values of the chosen split attribute. This recursive process stops when all the nodes to be merged are leaf nodes. After the merging operation, the aggregator sends the merged decision tree to the distributed devices. Therefore, we aim to build a joint machine learning model based on the data from multiple devices while offering k-anonymity to the participants.