Federated deep active learning for attention-based transaction classification

Usman Ahmed

doi:10.1007/s10489-022-04388-1

2023-01-23

PRE

Usman Ahmed

Jerry Chun‐Wei Lin *

Philippe Fournier‐Viger

DOI：10.1007/s10489-022-04388-1

原文链接

原文求助

摘要

En 中文

Cyber-physical transactions can be clustered and classified using an effective vector representation. Traditionally, methods for finding patterns have relied on heuristics and pruning. Despite the high dimensionality of transactional data, approaches that use frequent item sets as features can be affected by dimensionality, sparsity, and privacy issues. As a result, FIs are not uniformly distributed over the data. In this paper, we propose an embedded architecture for transaction classification based on distributed learning. The model transforms transaction data into frequent sets and implements attention mechanisms based on encoder-decoder structures. Consequently, the model can learn continuous vectors in low dimensions while maintaining context and colocation. We analyzed a high-dimensional transaction dataset to test attention-based methods and federated learning. To improve decision limits while maintaining privacy and security, the proposed model lowers the global loss function. In the experiment, four datasets are used for comparison. The data are randomly selected and distributed to different clients for each dataset. We run each experiment with five different random partitions of the dataset to evaluate the active learning. The training set is the same size in each round, and the test set is not evaluated in each round. Our methods are compared to the best performing baseline method using the F1 score and the percentage of the dataset used. Compared to the baseline model, the proposed model performed better in terms of percentage increases and output classes, namely retail (1.7%, 2), cancer (17.38%, 3), food (0.74%, 2), and snippet (3.47%, 8).

Keyword:

Active learning

Classification

Federated learning

Attention