Semi-supervised learning (SSL) methods are renowned for their capacity to utilise unlabelled data. In most SSLs, the threshold settings are fixed. They ignored the fact that learning difficulty varies among different categories. Moreover, many SSLs only learn from high-confidence data, overlooking the valuable facial expression features present in low-confidence data. As a consequence, the utilisation of data is limited and incomplete. To address the above issues, we propose the Adaptive Threshold-Driven Semi-Supervised (ATD-SS) method. Through adaptively adjusting thresholds, our method can make full use of the data. ATD-SS involves three parts: 1) Adaptive Adjustment Threshold Module. This module adjusts the confidence threshold by the predicted probability distribution of the labelled data. 2) Pseudo-labelling Module. For high-confidence data, this module generates pseudo labels to constrain the image prediction through the loss function. 3) Contrastive Learning Module. For low-confidence data, the similarity between features is learnt through contrastive clustering. The module is able to pull together features that belong to the same facial expression category. Experiments show that our semi-supervised method outperforms existing FER approaches and even the fully-supervised baseline.