Satisfying users’ requests based on the service level agreements of network slices is one of the most basic and vital topics of network slicing in 6G networks, and anomaly detection is regarded as a key technique for locating the abnormal status of slices. However, current studies on slice anomaly detection mostly focused on real-time monitoring of slices and ignored the prediction of potential anomalies. Generally, when anomalies trigger, it is hard for slices to adjust the resources in time due to resource competition among physical/virtual nodes. Besides, the resource provisioning strategies can also be optimized when slices are running normally, which is seldom considered when performing slice anomaly detection. To cope with these challenges, in this paper, we are motivated to locate the potential slice anomalies and optimize the resource allocation strategies in a holistic view by learning users’ historical behaviors. Specifically, we design a general network architecture, model the process of slice resource provisioning, and formulate the problem as maximizing the long-term system net promoter score (NPS). To solve this problem, we propose a framework to locate the potential slice anomalies and decide the resource allocation strategies simultaneously by predicting the users’ future requests and positions. As a result, simulation results demonstrate that our proposed scheme outperforms other baselines in improving the long-term system NPS and reducing the average latency of users.