Hey everyone! 👋 I wanted to share something I've been working on recently. I've been exploring feature selection in high-dimensional datasets, particularly a problem where a feature can be genuinely important to the target but still have little or even zero marginal correlation with it. This becomes especially interesting when predictors are highly correlated. While exploring this, I came across a research paper by Tanaka and Matsui (2023) on Truncated Preconditioned Profiled Independence Screening (TPPIS). I decided to implement the method as an open-source Python library so it would be easier to experiment with and use in actual ML workflows. The package includes TPPIS along with baseline methods such as SIS, FPSIS, FPSIS-BIC, and PPIS, with scikit-learn compatibility. One interesting result from reproducing the paper's first simulation example was that TPPIS recovered the hidden predictor in all 50 runs, while SIS failed to recover it in every run under that specific setup. Obviously, that's just one simulation setting, and I'd like to see how the implementation behaves on other datasets too. If you're interested in feature selection, ML research, statistics, or working with high-dimensional data, I'd genuinely love to hear your thoughts. Here are the links if you'd like to explore it: 📄 Original paper: https://arxiv.org/abs/2306.05702 📦 GitHub: https://github.com/Mohammed-Aswath/tppis 📚 Documentation: https://mohammed-aswath.github.io/tppis/ 🐍 PyPI: https://pypi.org/project/tppis/ I've also shared a post about the project on LinkedIn, with a visual explanation of the problem and the simulation results. Here is the linked post link : https://lnkd.in/p/drWPGSpP I'm sharing this here because I'd like to learn from people who might approach the problem differently, try it on their own datasets, or spot limitations I've missed. Feedback and constructive criticism are very welcome!

Hi everyone! 👋 I'm Mohammed Aswath, an AIML undergraduate . I've been building AI projects around industrial document intelligence (my previous project was AI-based P&ID digitization), and recently I've been exploring AI security. I'm considering my next project around a problem that I believe is becoming increasingly important: Title : Semantic Data Loss Prevention (DLP) for AI Copilots and RAG Systems Traditional DLP protects files, emails, and databases. Modern AI systems, however, convert enterprise documents into embeddings stored in vector databases and retrieve knowledge through semantic search. This creates a new risk where AI can unintentionally expose sensitive knowledge through retrieval or generated responses, even when file-level permissions are enforced. Microsoft researchers have also highlighted this challenge in enterprise AI: https://arxiv.org/abs/2509.14608 https://arxiv.org/pdf/2509.14608 The idea is to build a security layer that enforces access control at the embedding, retrieval, and response generation stages, enabling enterprises to adopt AI more securely. I'd really appreciate your thoughts: Is this a meaningful problem statement? Are there existing approaches or research I should explore? If this problem interests you and you'd like to collaborate, feel free to comment or DM me. Looking forward to your feedback!