Tony Idowu

Data Scientist | Python · XGBoost · NLP · Streamlit | Building ML solutions

[Tony Idowu @tonyds_xt]

Built a sentiment analysis tool for X posts, from raw scraped tweets to a live dashboard 🧵

[Tony Idowu @tonyds_xt]

@dlifeof_jay Pristine work chief

[Jay⚡️(The AI/ML Guy) @dlifeof_jay]

Marriott and Radisson Blu Hotels are direct competitors offering hospitality services in Ikeja Lagos

This project showcases how I used data science to audit their google map customer review, identify their strengths, weaknesses and also reveal their competitive adv..

A Thread🧵

[Tony Idowu @tonyds_xt]

Built a sentiment analysis tool for X posts, from raw scraped tweets to a live dashboard 🧵

[Tony Idowu @tonyds_xt]

Why it matters: brands and marketing teams can catch sentiment shifts in real time instead of finding out after the fact which is useful for campaign monitoring, brand tracking, and trend detection.

Live dashboard: https://t.co/NgaKVjsYeF

Code: https://t.co/BoTXt3K9fQ

[Tony Idowu @tonyds_xt]

Built a sentiment analysis tool for X posts, from raw scraped tweets to a live dashboard 🧵

[Tony Idowu @tonyds_xt]

Dashboard:
Built with Streamlit. KPIs, sentiment distribution, trend lines over time, word clouds, and a live text box to predict sentiment on anything you type in.

[Tony Idowu @tonyds_xt]

@dlifeof_jay 🫡

[Tony Idowu @tonyds_xt]

Live app: https://t.co/Ogh91n96lg

[Tony Idowu @tonyds_xt]

Month-to-month telecom customers churn 3x more than those on 2-year contracts.
High charges + low tenure is the deadliest combination.
I found this building a churn prediction model with XGBoost. The model hit 81% AUC and the feature importance rankings told the whole story 👇

[Tony Idowu @tonyds_xt]

Month-to-month telecom customers churn 3x more than those on 2-year contracts.
High charges + low tenure is the deadliest combination.
I found this building a churn prediction model with XGBoost. The model hit 81% AUC and the feature importance rankings told the whole story 👇

[Tony Idowu @tonyds_xt]

@iyandawaheed1 Built a telecom churn predictor with XGBoost that flags at risk customers before they leave. Churners were only 27% of the dataset so SMOTE was essential to get the recall right. 81% AUC and deployed live on Streamlit 👇

https://t.co/Ogh91n96lg

[Tony Idowu @tonyds_xt]

@naijadataprof Solid work chief

[Tony Idowu @tonyds_xt]

@CollinsKimotho Exactly, and that sequential correction is what makes it so dominant on structured data. The errors that matter most get the most attention in the next round. Random Forest has no memory of its mistakes

[Tony Idowu @tonyds_xt]

@Youraimldude Thank you 😊

[Tony Idowu @tonyds_xt]

@cdrophunter @X Connect

[Tony Idowu @tonyds_xt]

@ameenullahibr Same experience here. I used SMOTE on my churn dataset but scale_pos_weight is cleaner honestly. Less preprocessing, same result. Did you notice any difference in recall on the minority class between the two approaches?