Qualitative Instead of Quantitative: Towards Practical Data Analysis Under Differential Privacy

Differential privacy (DP) has become the de facto standard in the academic and industrial communities. Although DP can provide strong privacy guarantee, it also brings a major of performance loss for data mining systems. Recently there has been a flood of research into the quantitative mining of DP based algorithms, which are designed to improve the performance of data mining systems. However, industrial applications demand accurate quantitative mining results. Results containing noise are actually difficult to use. This paper rethinks to apply DP in industrial big data from another perspective: qualitative analysis, which aims to dig the data about rank, pattern, important set, etc. It does not require accurate results and naturally has a greater ability to accommodate noise. We design a framework about DP data publication based attribute importance rank to support the qualitative analysis of DP, which assists data buyers to perform qualitative analysis tasks and to know the credibility of their results. We show the realization of this framework using two typical qualitative tasks. Experimental results on public data and industrial data show that making use of this framework, qualitative analysis tasks can be completed with a high confidence support even when privacy budget \(\epsilon \) is very small (e.g., 0.05). Our observations suggest that qualitative analysis of DP has the potential ability to realize applying DP in industrial applications.