AAI_2025_Capstone_Chronicles_Combined

9

cluster centroids and iteratively assigning data points to the nearest centroid while adjusting the centroids towards the mean of these assigned data points. Agha et al. (2023) used k-means clustering to create groupings based on academic performance, helping to identify at-risk students who might otherwise go unnoticed. For large-scale AI conversation analysis, Anthropic’s Clio platform combines k-means and large language models (LLMs) to drive analysis on hierarchical “facets” from interactions with their Claude models, providing insight into the most frequent conversation topics across hundreds of thousands of conversations (Tamkin et al., 2024). Other frequently used clustering methods in educational data science include hierarchical clustering and fuzzy c-means (see Figure 3). Hierarchical clustering creates nested sets of clusters in a tree data structure, allowing for more effective exploration of hierarchical data. For example, Yotaman et al. (2020) used this approach to assign students to hierarchical learning paths. Howlin and Dziuban (2019) used the degree of cluster membership from fuzzy clustering to identify students who displayed outlier behaviors in their learning progression. Our project utilizes Deep Embedded Clustering (DEC), an unsupervised deep learning method designed to simultaneously learn optimal feature representations and cluster data points. DEC achieves this by first training a deep autoencoder to map input data to a lower-dimensional latent space. It then iteratively optimizes both the encoder’s parameters and a clustering objective. This objective encourages similar data points to be mapped close together in the latent space and dissimilar points farther apart while continuously refining cluster assignments. This dual optimization allows DEC to discover more meaningful and intricate clusters by learning a feature space that is specifically tailored for the data being clustered, moving beyond the limitations of traditional clustering methods like k-means. An additional aspect of our approach involves a facet extraction process designed to guide unsupervised deep clustering models toward predefined thematic patterns, mirroring Clio’s pipeline to discover broad usage patterns in conversations with the AI assistant Claude (Tamkin et al., 2024). Instead of feeding the entire dialogue directly to the DEC model, this process leverages an LLM to extract specific attributes - such as the task, user’s mood, or language - from

233