Abstract
CHISSL is a scalable interactive machine learning (IML) framework designed to help users organize large unlabeled datasets through intuitive drag-and-drop interactions. By leveraging a dendrogram-based approximation for label transduction, the system achieves $O(n)$ performance, enabling real-time feedback loops on the client side. This approach avoids the computational and cognitive bottlenecks of traditional semi-supervised learning and high-dimensional visualizations, allowing analysts to triage massive datasets efficiently.
Introduction
Data analysts are frequently overwhelmed by vast quantities of unlabeled data, such as log entries, social media posts, or sensor data. While machine learning offers a path toward categorization, it typically requires high-quality labeled data, which is expensive to produce. Existing "human-in-the-loop" systems attempt to bridge this gap but often fail due to two primary bottlenecks: computational scalability and cognitive scalability. Traditional algorithms are often too slow for millisecond-level human interaction, and common 2D projections (like t-SNE) create cluttered "star field" visualizations that overwhelm the user. CHISSL (Computer-Human Interaction for Semi-Supervised Learning) addresses these issues by rethinking the relationship between data representation and user interaction.
Methods and Architecture
CHISSL utilizes a client-server architecture to optimize the interactive labeling process.
Server-Side Heavy Lifting
The server performs representation learning and executes agglomerative clustering to construct a dendrogram—a tree-like structure representing data relationships. Instead of transmitting a massive high-dimensional feature matrix, the server sends only the dendrogram to the client as a memory-efficient "parent pointer array."
Client-Side Interaction and Transduction
The web-based client uses the dendrogram as a proxy for distance. It performs label transduction by identifying the closest labeled instance within the tree structure. This process requires only two passes over the pointer array, resulting in $O(n)$ time and space complexity. This deterministic responsiveness ensures that the UI remains snappy even as the user provides new labels.
The CHISSL Interface
Rather than displaying every data point simultaneously, CHISSL organizes data into a table where each row represents a user-defined group. The interface displays a user-labeled example alongside several "recommendations"—representative instances from sub-trees that help the user determine if a group is cohesive or requires adjustment.
Key Insights and Performance
The core innovation of CHISSL is the use of dendrograms as distance proxies. The shortest path distance within a pre-computed dendrogram serves as a "good enough" approximation for distances in high-dimensional space, allowing for instant predictions without heavy mathematical overhead. This shifts the focus from exhaustive visualization to relevant recommendation, scaling to the limits of human attention rather than screen resolution.
Evaluation Across Domains
The authors demonstrated the framework's versatility across several diverse datasets:
- Social Media: 24,000 images and 8,000 text messages.
- Cloud Computing: 42,000 sequences of resource utilization data.
- Cyber Security: 15,000 entity-day pairs from the CERT Insider Threat dataset.
- Geo-temporal Analysis: 500 vehicle-day pairs from the VAST Challenge 2014 dataset.
In all cases, the system maintained millisecond response times, proving its viability for truly interactive data exploration.
Conclusion and Future Work
CHISSL provides a robust framework for non-experts to organize data through intuitive "organizing by example" rather than complex parameter tuning. While the current system is transductive—focused on labeling existing data—future work aims to extend this into inductive modeling. This would allow users to export their groupings to train classifiers capable of automatically labeling new, unseen data. Additionally, the authors plan to conduct rigorous empirical evaluations against ground truth datasets to compare CHISSL's accuracy with other state-of-the-art IML frameworks.
Summary updated