Laser powder bed fusion (LPBF) is one of the additive manufacturing (AM) techniques and the most studied laser-based AM process for metals and alloys. The optimization of the laser process parameters of LPBF and the prediction of defects, for example, keyholes, cracks, and lack of fusion (LOF), are important for improving the quality of products made with LPBF. Deep learning (DL) is powerful in analyzing complex processes and predicting anomalies; however, much data is generally required for training a DL model. Experimental studies on AM (e.g., LPBF) habitually employ the design of experiments to decrease the number of experiments and save time and costs. Hence, the experimental data are not prepared for DL model creation in most situations. This paper studies the creation of a DL model on a small experimental dataset with unbalanced data and the prediction of the LOF defect of LPBF utilizing the created DL model. Data analytics is mainly conducted based on four DL methods, including Elman neural networks, Jordan neural networks, deep neural networks (DNN) with weights initialized by the deep belief network, and the regular DNN based on four algorithms: “rprop+”, “rprop−”, “sag,” and “slr.” It is shown that the regular DNN after the z-score standardization of the small dataset helps create a more accurate DL model and achieve better analytics and prediction results than the three other DL methods in this paper. The three other DL methods do not work well in the prediction of LOF based on the small dataset (with unbalanced data).