Gene expression of social actions in Drosophilae has been attracting wide interest from biologists, medical scientists and psychologists. Gene-edited Drosophilae have been used as a test platform for experimental investigation. For example, Parkinson's genes can be embedded into a group of newly bred Drosophilae for research purpose. However, human observation of numerous tiny Drosophilae for a long term is an arduous work, and the dependence on human's acute perception is highly unreliable. As a result, an automated system of social action detection using machine learning has been highly demanded. In this study, we propose to automate the detection and classification of two innate aggressive actions demonstrated by Drosophilae. Robust keypoint detection is achieved using selective spatio-temporal interest points (sSTIP) which are then described using the 3D Scale Invariant Feature Transform (3D-SIFT) descriptors. Dimensionality reduction is performed using Spectral Regression Kernel Discriminant Analysis (SR-KDA) and classification is done using the nearest centre rule. The classification accuracy shown demonstrates the feasibility of the proposed system.