Addisu-Taye/Customer-Experience-Analytics-for-Mobile-Banking-Apps

Analyze user satisfaction and feedback on top Ethiopian mobile banking apps (CBE, BOA, Dashen Bank) using Google Play Store reviews. Includes web scraping, NLP, outlier detection, and EDA to uncover pain points, feature requests, and improvement opportunities. ## Methodology My approach combines **web scraping**, **natural language processing (NLP)**, **data engineering**, and **exploratory data analysis (EDA)** to derive meaningful insights from unstructured review data. ### Data Collection Strategy - Reviews, ratings, dates, and app names for **CBE**, **BOA**, and **Dashen Bank** applications are collected from the Google Play Store. - A **minimum target of 400 reviews per bank** (1,200 total) ensures robust data for analysis. - **The initial scraping methodology utilizes the google-play-scraper library.** The scrape_reviews function iterates through a dictionary of bank names and their corresponding Google Play Store application IDs. For each app, it fetches a specified number of the newest reviews, extracting the review content, rating, and date. This data is then compiled into a Pandas DataFrame for further processing. ### Data Preprocessing & Cleaning Pipeline Implemented via `eda_reviews.py`: - **Initial Data Loading**: Raw data loaded from CSV. - **Duplicate Removal**: Duplicates identified and removed. Logged separately. - **Missing Data Handling**: Rows missing `review` text or `rating` values are dropped. - **Date Normalization**: Dates converted to `YYYY-MM-DD` format. - **Language Detection** (Two-stage process using FastText): 1. **Pre-trained Model** (`lid.176.bin`): High-confidence English classification (threshold = 0.25). 2. **Custom Model** (`custom_lid_ethiopian.bin`): Classifies remaining reviews into Amharic, Tigrigna, Somali, or Afan Oromo using script detection, keyword matching, and confidence thresholds. - Reviews are split into: - `English` dataset - `Non-English` dataset (Amharic, Tigrigna, Somali, Afan Oromo) ### Outlier Analysis Module integrated in `eda_reviews.py` for anomaly detection: - **Emoji Analysis**: - Detects and saves reviews containing emojis. - Extracts top emojis. - **Review Length Outliers**: - Identifies extremely short or long reviews using the IQR method. ### Ex …