Dynamic job shop scheduling (DJSS) is an NP-hard optimisation problem requiring real-time responses to newly arrived jobs. Scheduling heuristics generated by genetic programming can make high-quality decisions in DJSS, but evaluating these heuristics through simulations is very time-consuming. To speed up evaluations, surrogate models based on machine learning have been developed to predict fitness values. However, existing surrogates are primarily simplistic machine learning models with overly simplified input features, which overlook important characteristics of scheduling heuristics. To enhance prediction accuracy, we propose a new feature representation that comprehensively represents the behaviour of a scheduling heuristic. Additionally, a neural network binary classifier is employed as the surrogate model to learn the complex patterns in the proposed feature representation. Experimental results indicate that the proposed algorithm can find better scheduling heuristics and converge faster compared to the existing algorithms. Further analysis reveals both the new feature representation and the neural network binary classification-based surrogate model enhance the prediction accuracy and contribute to the performance improvement.