CLARA: A Modular Framework for Unsupervised Transit Detection Using TESS Light Curves

Abstract

We present CLARA, a modular framework for unsupervised transit detection in TESS light curves, leveraging Unsupervised Random Forests (URFs) trained on synthetic datasets and guided by morphological similarity analysis. This work addresses two core questions: (a) How does the design of synthetic training sets affect the performance and generalization of URFs across independent TESS sectors? (b) Do URF anomaly scores correlate with genuine astrophysical phenomena, enabling effective identification and clustering of transit-like signals? We investigate these questions through a two-part study focused on (1) detection performance optimization, and (2) the physical interpretability of anomalies. In Part I, we introduce three URF model variants tuned via alpha-controlled scoring objectives, and evaluate their generalization across five TESS sectors. This large-scale test involved scoring 384,000 individual light curves (128,000 light curves per alpha variant), revealing stable, interpretable differences between recall-optimized, precision-optimized, and balanced models. In Part II, our optimized clustering (DPMM Cluster 2) yields a 14.04% detection rate (16 confirmed transits among 114 candidates) from the first five TESS SPOC sectors. This reflects a substantial enrichment over baseline rates: 0.4569% for the full TESS-SPOC project candidate set (7658 candidates across 1.68 million light curves), and 0.2650% for the FFI-based SPOC sample (7658 candidates across 2.89 million light curves;). Additionally, we perform Monte Carlo injection-recovery tests to assess feasibility, stability, justification and quantification of methods outlined. CLARA processed over 87,000 TESS SPOC light curves (Sectors 1-5).