With the exponential growth of online videos, issues related to video copyright and content moderation have become increasingly pressing. Consequently, research fields addressing Partial Video Copy Detection (PVCD) have produced a variety of solutions. Previous methods first extract frame-level features and then perform temporal alignment, which may result in the loss of original temporal information. To address these limitations, this paper proposes a Shot-Based Partial Copy Detection (SPCD) method. This approach introduces shot segmentation into the PVCD problem, thereby eliminating the need for temporal alignment. Additionally, it employs 3D mixed convolution to extract spatio-temporal features within shots, helping to preserve dynamic features within each shot. Moreover, the proposed method utilizes a metric learning framework to train a Deep Neural Network (DNN) as a projection network, facilitating similarity calculation and effectively enhancing performance. The method is validated on the benchmark dataset VCDB, where it outperforms previous state-of-the-arts in terms of F1-score. Further robustness testing with distraction videos and ablation studies demonstrate the effectiveness of feature projection in this approach.