<p>Cancer is a life-threatening disease where early detection is crucial for accurate diagnosis and effective treatment. Traditional machine learning methods for cancer diagnosis often depend on high-dimensional datasets and complex features, leading to slow analysis and sometimes unreliable results. Feature selection (FS) is a basic data pre-processing method in machine learning and classification that aims to reduce the vertical dimensionality of a dataset by selecting the lowest number of features that provides the highest accuracy for a classifier. As a result, FS can be framed as a multi-objective optimization problem with two conflicting.&#xa0;Objectives:&#xa0;minimizing the number of selected features while maximizing classification accuracy. Recently, the well-known Harris Hawks Optimization algorithm was converted to Binary HHO (BHHO) to address the FS problem. This paper introduces an Improved BHHO (IBHHO) algorithm to solve the multi-objective FS problems for cancer detection. The main contribution of the IBHHO algorithm is to improve the search process by enhancing diversity and search stability within a binary multi-objective optimization framework, thus balancing exploration and exploitation compared to current feature selection methods based on object swarms and evolutionary approaches. IBHHO involves three key modifications. First, IBHHO initializes its basic population of binary solutions using the Opposition-Based Learning (OBL) method. This modification ensures a more diverse initial population for the optimization process. Second, IBHHO employs a newly developed MLOBL technique, which combines Mixed Opposition-Based Learning (MOBL) and Lens Opposition-Based Learning (LOBL) to enhance search efficiency and solution quality during optimization. Third, it utilizes a non-dominated sorting algorithm to identify the Pareto front and guide the multi-objective selection process. IBHHO was evaluated across nine different and diverse cancer classification datasets and compared with eight algorithms recently used for multi-objective optimization problems. The comparison was based on five metrics usually used in multi-objective optimization: Pareto Front charts, Inverted Generational Distance (IGD), Hypervolume (HV), Generational Distance (GD), Pareto Front Diversity (PFD), and spread. Experimental results showed that the IBHHO algorithm achieved low IGD values across multiple datasets, outperforming most competing algorithms on the selected datasets in most trials and achieving the lowest IGD values in 6 out of 9 datasets. These values were 0.0011 for ALL-AML, 0.0004 for ALL-AML-4, and 0.0003 for SRBCT. It also achieved high HV values, the best in 6 out of 9 datasets, such as 0.9996 for the ovarian cancer dataset and 0.9994 for ALL-AML-4. These results indicate strong performance in terms of convergence and diversity compared to competing algorithms. IBHHO achieved an optimal Pareto front that is close to the true Pareto front. Furthermore, a comprehensive statistical evaluation was conducted to highlight the performance differences between algorithms. The Friedman test showed that IBHHO ranks first across all datasets, while the Wilcoxon signed-rank test validated IBHHO’s statistically superior performance in pairwise comparisons.</p>