Dimensionality reduction is an important technique in both machine learning and data mining. Many varieties of techniques are already available for removing the least important attributes from the training datasets. In this paper, a new dimensionality reduction technique is proposed based on the relationship between input attributes and the output class label attribute in the training datasets. The proposed dimensionality reduction technique is very simple to analyze, compute, and implement. The proposed dimensionality reduction technique is verified experimentally using WEKA as well as NETBEANS Java tools. UCI machine learning datasets are employed for experimentation. Output results are carefully analyzed and observed. Expected results and experimental results are matching. It shows that the proposed dimensionality reduction technique is correct. The proposed technique is true for both categorical as well as numerical attributes in the training datasets. The main fundamental and very useful characteristic of the proposed dimensionality reduction technique is that it is very easy to compute and apply. The challenges in the dimensionality reduction technique are attribute selection, computational overhead, high storage cost, complexity of algorithms, and difficulty of data preservation. Recent techniques that are proposed for dimensionality reduction are—New hamming distance, minimizing the reconstruction error, non-linear discrimination analysis techniques, Principle Component Analysis (PCA), and Linear Discriminant Analysis (LDA).