Krish

Most telecom companies lose customers before they know it's happening. I built an ML model that predicts churn before it costs revenue. Here's the business case + what I found 🧵 1/4

Day 7 of NLP OVERVIEW Project Update :- Build TF-IDF + Multinomial Naive Bayes pipeline Train/test split, fit, predict Evaluate: accuracy, precision, recall, confusion matrix Target: ~97% accuracy (achievable with this dataset) Save the model

Day 6 of NLP OVERVIEW Today I start building my spam classifier. Just setup. Downloaded the UCI SMS dataset, did EDA, found the first interesting thing: spam messages are 2x longer than ham on average. Pipeline sketch is done, I'll going to implement that tomorrow.

Day 5 of NLP OVERVIEW Core learning today - Why Transformers Won(Part 2) It's architecture : 6 encoders + 6 decoders Each encoder = Self-Attention layer + Feed-Forward layer Each decoder = Masked self-attention + Cross-attention + FFN Positional Encoding BERT vs GPT vs T5

@PlacementDiary Bro's time of posting day by day📈😂

Day 41📆 English: Spoken practice Not a good mental day. Still showed up. 😫 #Day41 #LearningInPublic #Consistency

Day 4 of NLP OVERVIEW Core learning today - Why Transformers Won(Part 1) 1. Classical NLP (TF-IDF, Bag of Words) - No word order → "dog bite man" = "man bite dog" 2. RNN/LSTM - No Parallelisation, Not remember early context Transformer solved all ! (Post this tomorrow)

Day 3 of NLP OVERVIEW Main tasks: Classification → label text (spam detection) NER → extract entities from text Summarization → shorten long text Question answering → retrieve answers from documents Most modern NLP uses transformers.

Day 2 of NLP OVERVIEW TF-IDF TF = word frequency in doc. IDF = rarity across all docs. Score = TF × IDF Word2Vec Maps words to dense vectors (~300 numbers). Similar words → similar vectors king - man + woman ≈ queen (vector math works) Day 3: NLP Tasks🔜

Started 10 Days of NLP Day 1: Tokenization 🧵 Before any LLM reads text, it breaks it into tokens. "I love machine learning" → ["I", "love", "machine", "learning"] Then removes noise (stop words), normalizes words (lemmatization), and feeds numbers into the model. #NLP

🚀 Tech stack completed so far (with projects): ✅Python ✅Data Science Stack • NumPy • Pandas • Matplotlib / Seaborn (visualizations) ✅Maths ✅ML FUNDAMENTALS 🔥 Next up: NLP(Overview) github.com/codewith-krishh

@SAHAJEEN @YouTubeIndia This guy is actually awesome, especially his 1 hour videos(numpy in 1 hour)

@YouTubeIndia No one can never beat this guy 🗿🗿🗿🗿

mention a 10/10 YouTube channel in the comments 👇

Stack: Python · Scikit-learn Pipelines · GridSearchCV · Cross Validation No data leakage. Consistent preprocessing. Production-ready structure. Full project on GitHub 👇 github.com/codewith-krish… 🧵4/4

What actually drives churn? 📈 Month-to-month contracts→highest churn risk 📈 Fiber optic plans→elevated risk 📉 Long tenure→strong retention signal 📉 2-year contracts→customers stay Businesses can use this to design smarter pricing, contracts, and loyalty programs. 🧵3/4

Most telecom companies lose customers before they know it's happening. I built an ML model that predicts churn before it costs revenue. Here's the business case + what I found 🧵 1/4

Use a Pipeline. train_test_split first, then: Pipeline ├─ scale numeric features ├─ encode categorical features └─ train model Preprocessing is now learned only from training data. Improved my Customer Churn model through pipeline Source code soon.... 🧵 2/2

✅Day 30/35 (ML Fundamentals) Most beginner ML projects preprocess data the wrong way. They do this: get_dummies() → StandardScaler() → train_test_split() → model. fit() Looks fine. But it causes data leakage. The fix? Check this thread ➡️ 🧵 1/2

✅Day 29/35 (ML Fundamentals) I'm just going to improve this project this week - Today I add : 1. numeric_features 2. binary_features 3. categorical_features Working on some other stuffs too, which I'm going to share in future (basically after learning tech stack)

Accuracy can lie. Balanced LR dropped accuracy (0.81→0.74) but recall jumped (0.56→0.78). For business: early churn detection > a pretty score. Next (Week 17): Pipelines · GridSearchCV · ROC-AUC · Feature importance Repo👇 github.com/codewith-krish… 🧵 4/4

Workflow: → Clean → EDA → Encode → Train/test split → Logistic Regression baseline → Cross-validation → Random Forest comparison Results: • LR → Acc: 0.81 | Recall: 0.56 • Balanced LR → Acc: 0.74 | Recall: 0.78✅ • Random Forest → Acc: 0.79 | Recall: 0.50 🧵 3/4

✅Day 27-28/35 (ML Fundamentals) Just shipped a Customer Churn Prediction model. 🧠 Goal: Predict which telecom customers leave - before they do. Here's the build, the results & the metric that matters most 👇 🧵 1/4 #MachineLearning #DataScience #Python