Soft set is a theory and tool for dealing with uncertain data. It is usually used in decision theory. The parameter reduction of soft set refers to the removal of redundant parameters that have little effect on decision-making. Since the 0-1 linear programming algorithm was proposed, the problem of parameter reduction of soft set has been basically solved, but the implementation of the 0-1 linear programming algorithm is complex and relies on the integer programming algorithm. Here, considering the practical application background of soft set, combining soft set with probability theory, a soft set parameter reduction method in the context of big data, called the variance toss algorithm, is designed. The time complexity of this algorithm is O(m2n), and 0-1 linear programming is usually regarded as a NP-hard problem. The variance toss algorithm is simple to implement. When the object set (or the complete set) is small and less than twice the size of the attribute set, the effect is poor. However, as the size of the object set (or the complete set) increases, the efficiency will gradually increase, and the final computing efficiency will be better than the 0-1 linear programming algorithm. It has higher efficiency for the soft set with high reduction density.