Anomaly detection in data streams requires a signal of an unusual event, and an actionable response requires diagnostics. Furthermore, monitoring for process control is often concerned with one or more target (controlled) attributes. Consequently, it is necessary to separate anomalies (and their contributing attributes) that could influence the controlled target strongly, and this becomes more important with the increased number of monitored attributes in modern processes. This task leads to a difficult problem not addressed directly by the machine learning/process control community. We introduce the target-aware anomaly detection problem and present a solution for process control in modern systems (with nonlinear dependencies, high dimensional noisy data, missing data, and so on). The main objective is to identify and rank outliers and also diagnose their contributing attributes with respect to the possible effect on the response. The method is different from traditional linear and/or univariate approaches, as it can deal with local data structure in the neighborhood of an outlier, and can handle complex interactions via the use of an appropriate learner. In addition, the method can be computed quickly and does not require time consuming matrix operations. Comparisons are made to traditional contribution plots computed from partial least squares.