A heuristically self-organised Linguistic Attribute Deep Learning for edge intelligence

With the development of the internet of things (IoT), IoT intelligence is becoming a new technology. The “curse of dimensionality” is the barrier of data fusion in edge devices to the success of IoT intelligence. Deep Learning has attracted a lot of attention recently due to its successful application in various fields such as image processing and natural language processing. However, the success of Deep Learning benefits from the Graphics Processing Unit (GPU) computing. Similar to conventional Deep Learning, the computational complexity of optimising LAHs blocks the applications of LAHs. However, unlike conventional Deep Learning, an Linguistic Attribute Hierarchy (LAH) embedded with Linguistic Decision Trees (LDTs) could overcome the shortcoming of the lack of interpretation by providing transparent information propagation through the rules created by LDTs in the LAH. To address the challenge in the construction of high-performance LAHs, we propose a heuristic approach for constructing an LAH embedded with LDTs for decision making or classification by using the distance correlations between attributes and between attributes and the target variable. The set of attributes is divided into some attribute clusters and then heuristically organised into an LAH. The proposed approach has been validated on some benchmark decision or classification problems from the UCI machine learning repository. The experimental results show that the proposed self-organisation algorithm can create an effective and efficient LAH. Such a self-organised LAH embedded with LDTs can not only efficiently address “curse of dimensionality” in a single LDT for data fusion with massive attributes, but also achieve better or comparable performance in decision making or classification compared to a single LDT for the problem to be solved. The main contribution of this work is that (1) an LAH provides a new attribute Deep Learning approach with good transparency and extends the conventional Deep Learning, which usually denotes a black-box deep neural network; (2) the self-organisation algorithm for SOLAH is much more efficient than the genetic algorithm in the wrapper for optimising LAHs, and SOLAH can achieve comparable results to the LAH, which was optimised by the genetic algorithm in the wrapper (GAW). This is crucial for edge intelligence and makes it feasible to embed SOLAH and the self-organisation algorithm in edge devices for IoT applications. Hence, this research will take a step towards edge intelligence.