🧠 Daily Study Log [2025-08-04]
From language self-reflection tools to ensemble weight tuning — today’s work was sharp, structured, and a step closer to submission excellence.

💡 Project Idea — Echo Mirror: AI Language Habit Reflector

Proposed an NLP-based tool that detects repetitive linguistic habits and gives feedback for improvement — inspired by self-awareness in communication.

Concept:

• Input: user text (e.g. emails, messages, transcripts)
• Analysis: frequent phrases, filler words, sentiment tone
• Output: heatmap-style feedback and suggestions for clarity

🔗 View full idea

🧪 Competition — Electricity Usage Forecasting (DACON)

24. Optuna-Tuned Voting Weights for Grouped Models

Kept the Time × Building Type segmented VotingRegressor structure,
but used Optuna to optimize the weights (XGB, LGBM, GBR) for each of the 24 subgroups.

📊 Results

• Local SMAPE: 11.7013
• Public LB: 9.46491 ✅ New best

25. Outlier Correction with IQR Filtering

Applied IQR-based outlier correction to VotingRegressor outputs for each group.
Helps reduce extreme prediction noise and increase overall stability.

📊 Results

• Local SMAPE: (Pending)
• Public LB: (Pending)

  Currently running…

📖 Paper Review — DoWhy: A Library for Causal Inference

Continued my review of the DoWhy framework, focusing on how it reframes causal inference as a transparent, testable process.
Rather than treating causality as just another statistical problem, DoWhy emphasizes that

“Your assumptions are your model.”

Key insights from today:

• Causal analysis is structured into four clear steps: Model → Identify → Estimate → Refute
• Explicit assumptions (e.g., ignorability) are not optional — they must be declared and tested
• Directed Acyclic Graphs (DAGs) help formalize what we believe, not just what we see
• Unlike traditional ML, DoWhy asks not “What predicts Y?” but “Did X cause Y?”

This mindset shift—from correlation to causation—feels essential for real-world decision-making tasks.

📂 View notes and code

📝 TOEIC Study

D-27 until the test. Staying focused with intensive practice on both Listening and Reading sections.
Currently working on an error log by part and identifying weak areas for targeted review.

🎧 Personal Project — Bass Seeker

Refined the direction for the Bass Seeker project.
Created a new module bass_selector to filter songs based on low-end characteristics.
First time working with ffmpeg and yt-dlp, which was tricky but rewarding.

📂 Project GitHub

✅ TL;DR

📍 Idea: Echo Mirror — NLP-based self-reflection tool for language habits
📍 Competition: VotingRegressor group-wise weight tuning via Optuna (SMAPE: 9.4649)
📍 Review: DoWhy – causal vs correlational thinking
📍 Study: Intensive TOEIC prep (D-27) — focusing on listening, reading, and error tracking
📍 Project: Developed bass_selector + first time using ffmpeg and yt-dlp