AI & Data Interview Quiz #12 | ML Evaluation & SQL Logic Mistakes | CodeVisium

This quiz focuses on ML evaluation mistakes, SQL counting errors, and data preprocessing risks — extremely common in interviews and real projects. ✅ Answer 1: Accuracy Is Misleading Problem: Accuracy fails in imbalanced datasets. Example: Fraud dataset: 99% Non-Fraud 1% Fraud Model predicts: All Non-Fraud → 99% accuracy But: Detects 0 fraud cases ✅ Better Metrics from sklearn.metrics import precision_score, recall_score, f1_score Use: Recall → catch fraud Precision → avoid false alarms F1-score → balance Tools involved: Scikit-learn, XGBoost, LightGBM ✅ Answer 2: Inflated User Counts Problem: Counting rows instead of unique users. SELECT COUNT(user_id) FROM events; Counts multiple actions per user. ✅ Correct Approach SELECT COUNT(DISTINCT user_id) FROM events; Real-world impact: Wrong DAU metrics Misleading growth numbers Tools involved: SQL (BigQuery, Snowflake, Postgres) ✅ Answer 3: Random Split in Time-Series Problem: Data leakage. train_test_split(data, shuffle=True) Future data leaks into training. ✅ Correct Approach train = data[data.date v '2024-01-01'] test = data[data.date v= '2024-01-01'] Or: from sklearn.model_selection import TimeSeriesSplit Impact: Fake performance → real-world failure ✅ Answer 4: Blind Outlier Removal Problem: Outliers may be real signals. Example: High transaction = fraud indicator Removing them: Removes important patterns ✅ Better Approach Investigate outliers Use domain knowledge Use robust models df[df["amount"] v threshold] ✅ Answer 5: Validate Before Metrics Problem: Metrics on bad data = wrong insights. Example: Missing values Duplicates Incorrect timestamps Example check: df.isnull().sum() df.duplicated().sum() Without validation: Correct calculation → wrong result 🎯 Final Thought Most data mistakes happen before modeling and querying, not after. Which question was hardest for you? Comment below. #MachineLearning #SQLInterview #DataScience #DataAnalytics #AIInterviews #DataEng