A key challenge in video anomaly detection is the identification of rare abnormal patterns in the positive instances as they exhibit only a small variation compared to normal patterns, and they are largely biased by the dominant negative instances. To address this issue, we propose a weakly supervised video anomaly detection model called NTCN-ML - Novel Temporal Convolutional Network Multi-Instance Learning Model. The NTCN-ML model extracts temporal representations of video data to construct a time-series pattern to optimize the multi-instance learning process. The model examines the correlation between positive and negative samples in the multi-instance learning process to balance the feature association between rare positive and negative instances. The video anomaly detection with the NTCN-ML model achieved 95.3\% and 85.1\% accuracy for UCF-Crime and ShanghaiTech datasets, respectively, and outperformed the baseline models.