Comparative Market Analysis (CMA) is standard process for real estate professional while assessing the market value of a property. A CMA report usually lists the target property and multiple recently sold houses in the same neighborhood and compares their features, e.g., built age, living size, number of bedrooms, etc. Professional realtors can then determine the appropriate price of the target property. However, without professional training, it is difficult to determine the importance of individual features and thus it remained challenging for buyers/sellers to appropriately use CMA reports. This dissertation aims to address this challenge by machine learning approach and automating the CMA based house appraisal procedure. The key idea is a self-supervised learning framework which includes two training stages. In the first stage, a set of nearly sold houses are identified for each target house in the training set and a contrastive loss function is introduced to encourage similarities or dissimilarities between the sold houses and the target house. This loss does not access the true market value of training samples and is used to pre-train the machine learning model. In the second stage, a set of supervised house-value pairs are used to fine-tune the pre-trained model for regression purposes. Experiments with various regression models are evaluated on a newly collected housing dataset. Results suggested the proposed self-training framework can consistently improve the appraisal accuracies of various regression models.