Ranya Zeba

Key outcomes

• Cleaned the dataset by imputing missing values with column means and validating data types before analysis
• Found a strong correlation (r=0.72) between production budget and gross revenue, but a near-zero correlation (r=0.07) between audience ratings and likes
• Found that view count correlated strongly with both likes and dislikes (r=0.78), showing that visibility drives engagement in both directions, not just positive sentiment
• Found a weak correlation (r=0.29) between a film's social media following and its revenue, challenging the assumption that follower count predicts box office success
• Recommended prioritizing budget allocation and marketing reach over follower-count-driven casting or promotion decisions
• Deliverable: Python/pandas exploratory data analysis notebook with visualizations

Key outcomes

• Cleaned and merged five linked datasets (items, restaurants, orders, order details, summarized orders) and validated data types and categorical values before analysis
• Ran menu item frequency analysis to identify the top-selling dishes forming the core of the restaurant's sales
• Ran time-based analysis showing order volume peaking at lunch and dinner, with Fridays as the busiest day and summer as the busiest season
• Built a heatmap of the most popular main-and-side combinations and found delivery made up only 11.4% of orders, the lowest of all order types
• Recommended a new combo menu item based on the top-selling pairing and a 15%-off first delivery promotion to grow the underperforming delivery channel
• Deliverable: Python/pandas analysis notebook with visualizations (bar charts, pie chart, heatmap)

Key outcomes

• Cleaned and merged two datasets of roughly 45,000 fake and true news articles, removing duplicates and combining headline and body text into a single feature
• Engineered TF-IDF features from the combined text and trained a Logistic Regression classifier on a 70/30 train-test split
• Diagnosed a severe overfitting problem: the model scored near-perfect accuracy on training data but only 49.51% on an independent test set, barely above chance
• Attempted regularization tuning to correct the overfitting and documented why the fix did not resolve the underlying generalization failure
• Deliverable: Python/scikit-learn classification notebook with model evaluation

Key outcomes

• Calculated actual CLV against the $720 expected value (12 months at $60), finding every region fell far short: North America ($136.62), Rest of World ($96.84), Western Europe ($158.14)
• Identified Instagram, direct visits, and Facebook as the highest-CLV acquisition channels, with referral programs and Google Ads underperforming
• Diagnosed early customer drop-off as the core driver of low CLV, pointing to a retention problem rather than an acquisition problem
• Recommended shifting to tiered pricing (6-month and 12-month plans), regional pricing adjustments, and reallocating marketing spend toward high-CLV channels
• Deliverable: Python analysis with business recommendation report

Key outcomes

• Cleaned and explored a 35-column resume dataset, examining match score distributions and missing data patterns
• Built a linear regression model (matched_score ~ num_skills) that was statistically significant but explained less than 1% of the variance in match scores, showing skill count barely moves the needle
• Identified a bimodal distribution in match scores: resumes tended to be either a strong fit or a poor one, with little middle ground
• Translated the findings into practical guidance for both job seekers (prioritize relevance over keyword-stuffing) and recruiters (move beyond keyword filtering toward skill-application screening)
• Deliverable: R/ggplot2 analysis report with regression modeling

Key outcomes

• Built an initial Decision Tree that scored 91% training accuracy but only 85% on test data, correctly diagnosing overfitting from excessive tree depth
• Applied regularization (tuning max_depth, min_samples_split, min_samples_leaf) to simplify the tree while preserving classification performance
• Trained an XGBoost classifier that reached 91.29% test accuracy with strong generalization, outperforming the regularized decision tree
• Tested Logistic Regression and correctly explained its poor fit: the data is not linearly separable, so a linear decision boundary was the wrong tool
• Trained an MLP neural network as a fourth comparison point, reaching comparable performance (91.27% test accuracy) to XGBoost
• Deliverable: Python/scikit-learn notebook comparing 4 classification models