Various types of web applications have gained both higher customer satisfaction and more benefits since being successfully armed with personalized recommendation. However, the increasingly rampant shilling attackers apply biased rating profiles to systems to manipulate item recommendations, which not just lower the recommending precision and user satisfaction but also damage the trustworthiness of intermediated transaction platforms and participants. Many studies have offered methods against shilling attacks, especially user profile based-detection. However, this detection suffers from the extraction of the universal feature of attackers, which directly results in poor performance when facing the improved shilling attack types. This paper presents a novel dynamic time interval segmentation technique based item anomaly detection approach to address these problems. In particular, this study is inspired by the common attack features from the standpoint of the item profile, and can detect attacks regardless of the specific attack types. The proposed segmentation technique could confirm the size of the time interval dynamically to group as many consecutive attack ratings together as possible. In addition, apart from effectiveness metrics, little attention has been paid to the robustness of detection methods, which includes measuring both the accuracy and the stability of results. Hence, we introduced a stability metric as a complement for estimating the robustness. Thorough experiments on the MovieLens dataset illustrate the performance of the proposed approach, and justify the value of the proposed approach for online applications.