Hyper-parameter Recommendation for Truth Discovery

Siying Chen, Xiaoou Ding, Zheng Liang, Yafeng Tang, Hongzhi Wang

Abstract: With the increase of data size, it has become particularly important to find the most trustworthy information among the diverse and contradictory data. The process of discerning claims consistent with the truth from various data sources is commonly referred to as truth discovery. While truth discovery has demonstrated commendable outcomes across diverse applications, most truth discovery algorithms are severely affected by hyper-parameters, thereby limiting their overall performance enhancement potential. Consequently, it is imperative to explore hyper-parameter recommendation for optimizing truth discovery. Initially, we advocate for data augmentation on the input dataset of the truth discovery. Given the limited availability of open-source datasets of truth discovery algorithms, employing data augmentation becomes crucial for enhancing data richness while upholding data quality. Subsequently, we propose a hyper-parameter recommendation method grounded in dataset similarity, model-agnostic meta learning and Bayesian optimization. The proposed method entails a multi-step process. First, a preliminary estimation of the hyper-parameter for the truth discovery algorithm is obtained through meta-learning. This initial estimation serves as input for the Bayesian optimization algorithm, which, in turn, predicts the hyper-parameter values tailored to each dataset. Leveraging similarity measures between datasets, the hyper-parameters for the target dataset are then computed. Following the hyper-parameter recommendation phase, the truth discovery algorithm attains optimal hyper-parameters, resulting in a noteworthy performance average improvement of 18.14% according to extensive experiments.

External IDs:dblp:conf/dasfaa/ChenDLTW24