Published Paper Details:

Data mining and machine learning techniques extract valuable information from database to exploit problem-solving and decision-making. Dimensionality reduction plays a key role in decision-making. Dimensionality reduction methods effectively reduce data dimensionality for efficient decision-making process. Linear Discriminant Analysis (LDA) is the most popular dimensionality reduction algorithm that found a subspace which minimize within-class distance and maximize between-class distance. LDA used arithmetic mean to calculate between-class distance. However, it has some limitations which were overcome by Harmonic Linear Discriminant Analysis (HLDA) and HLDA pairwise (HLDAp). HLDA and HLDAp used the harmonic mean to compute the between-class distance. However, the time complexity is high for the initialization of HLDAp. In this paper, Improved HLDAp (IHLDAp) is introduced to reduce the time complexity of HLDAp. An initialization schema is used in IHLDAp which initializes the composition matrix in a consistent and robust way. In the initialization schema, a datum is generated based on the identity of data, uniqueness of the data and noise. Then, the error of the initialization process is reduced by using Alternating Direction Method of Multipliers (ADMM). By minimizing the error, a proper initialization of composition matrix is achieved which helps to reduce the dimensionality of data. Finally, the reduced data is given as input to Support Vector Machine (SVM). The propose IHLDAp can be applied for both single label classification and multi-label classification problem. The experiments are conducted to prove the effectiveness of IHLDAp in terms of accuracy, precision and recall.