Semi-greedy heuristics for feature selection with test cost constraints

In real-world applications, the test cost of data collection should not exceed a given budget. The problem of selecting an informative feature subset under this budget is referred to as feature selection with test cost constraints. Greedy heuristics are a natural and efficient method for this kind of combinatorial optimization problem. However, the recursive selection of locally optimal choices means that the global optimum is often missed. In this paper, we present a three-step semi-greedy heuristic method that directly forms a population of candidate solutions to obtain better results. In the first step, we design the heuristic function. The second step involves the random selection of a feature from the current best k features at each iteration. This is the major difference from conventional greedy heuristics. In the third step, we obtain p candidate solutions and select the best one. Through a series of experiments on four datasets, we compare our algorithm with a classic greedy heuristic approach and an information gain-based \(\lambda \)-weighted greedy heuristic method. The results show that the new approach is more likely to obtain optimal solutions.