DMLA: A Dynamic Model-Based Lambda Architecture for Learning and Recognition of Features in Big Data
Real-time event modeling and recognition is one of the major research areas that is yet to reach its fullest potential. In the exploration of a system to fit in the tremendous challenges posed by data growth, several big data ecosystems have evolved. Big Data Ecosystems are currently dealing with various architectural models, each one aimed to solve a real-time problem with ease. There is an increasing demand for building a dynamic architecture using the powers of real-time and computational intelligence under a single workflow to effectively handle fast-changing business environments. To the best of our knowledge, there is no attempt at supporting a distributed machine-learning paradigm by separating learning and recognition tasks using Big Data Ecosystems. The focus of our study is to design a distributed machine learning model by evaluating the various machine-learning algorithms for event detection learning and predictive analysis with different features in audio domains. We propose an integrated architectural model, called DMLA, to handle real-time problems that can enhance the richness in the information level and at the same time reduce the overhead of dealing with diverse architectural constraints. The DMLA architecture is the variant of a Lambda Architecture that combines the power of Apache Spark, Apache Storm (Heron), and Apache Kafka to handle massive amounts of data using both streaming and batch processing techniques. The primary dimension of this study is to demonstrate how DMLA recognizes real-time, real-world events (e.g., fire alarm alerts, babies needing immediate attention, etc.) that would require a quick response by the users. Detection of contextual information and utilizing the appropriate model dynamically has been distributed among the components of the DMLA architecture. In the DMLA framework, a dynamic predictive model, learned from the training data in Spark, is loaded from the context information into a Storm topology to recognize/predict the possible events. The event-based context aware solution was designed for real-time, real-world events. The Spark based learning had the highest accuracy of over 80% among several machine-learning models and the Storm topology model achieved a recognition rate of 75% in the best performance. We verify the effectiveness of the proposed architecture is effective in real-time event-based recognition in audio domains.
Table of Contents
Introduction -- Background and related work -- Proposed framework -- Results and evaluation -- Conclusion and future work