Knowing the activity type is important for many types of animal research. One example is to know how often a horse is standing still, walking, trotting or cantering. To automate this process, most scientific investigations rely on supervised neural networks, necessitating a resource-intensive and time-consuming manual labeling phase. This paper investigates the potential of unsupervised deep learning for automatic equine gait classification, employing the Deep Embedded Clustering (DEC) approach. Our evaluation compares it with alternative supervised and unsupervised algorithms, along with various pre-processing techniques. Findings indicate that probabilistic-based clustering algorithms, like Gaussian Mixture Models, are more effective for scenarios with significant variance in cluster densities, as frequently encountered in the domain of animal activity recognition. The paper establishes that transforming data from the time domain to the frequency domain using the Discrete Fourier Transform (DFT) contributes to improved performance. The integration of the DEC-based clustering algorithm with DFT achieves an overall accuracy approaching 83%. The research shows that while supervised methods still outperform unsupervised ones, with proper algorithm selection and data processing, unsupervised methods yield satisfactory results without the requisite for labeled data.