Implementation of novel loss and activation function-based deep network in big data analytics for pattern discovery across multiple data sources

Big data analytics utilizes specialized technologies and advanced analytical methods to manage large, complex datasets and uncover insights, which are inaccessible through traditional processing. A core component of data mining is pattern discovery that identify structures like associations, clusters, classifications, and sequential patterns. Deep learning (DL) has gained significant attention due to its ability to autonomously learn multi-level feature representations from large-scale datasets, in contrast to conventional machine learning techniques that rely heavily on handcrafted features. This capability enables DL models to efficiently manage the volume, velocity, and variety characteristics of big data. Big Data analytics for pattern discovery holds immense potential for driving innovation and transforming industries by unlocking actionable insights that were once hidden within vast amounts of data. Although numerous methods have been introduced for high-utility item mining, most of them are limited to small-scale data and function primarily within standalone or independent computing environments. Therefore, a novel big data analytics model for pattern discovery is developed in this work. Initially, the required data for the experiment are sourced from variety of large-scale datasets. The proposed Spatial Attention-based Residual Bidirectional Long Short-Term Memory with Novel Loss and Activation Function (SARBiLSTM-NLAF) is used to form patterns from this dataset. The proposed model utilizes a BiLSTM network to analyze sequential big data by learning information from both preceding and succeeding states, thereby capturing comprehensive temporal context and improving pattern recognition performance. An integrated spatial attention mechanism dynamically focuses on identifying most relevant data points across space and time for effectively filtering noise and improving detection accuracy. The residual structure mitigates the vanishing gradient problem and allows for more efficient training process to enhance the model’s ability. The effectiveness of model is validated through detailed experimental analysis across multiple data sources using significant performance measures, such as accuracy for demonstrating robust and reliable performance. The proposed model achieves 90.49%accuracy, 93.16%precision, and 88.96%recall by significantly outperforming existing approaches. This minimizes false positives while ensuring comprehensive detection, and ultimately provides a more trustworthy framework for real-world deployment.