Predict Seller Intent From Subscription Data

You are given a take-home dataset, `seller_intent_take_home_dataset.csv`, containing about 5,000 new subscription records from a website-building platform. The business goal is to understand which factors are associated with a new user becoming an active seller and to build an interpretable prediction model for seller intent. The dataset includes `subscription_id`, `subscription_start`, nullable `first_sale`, `days_in_trial`, `subscription_plan`, `subscription_period`, `discount_amount`, `country`, `site_topic`, `site_need`, and raw `user_agent`. Important issues include missing seller labels, inconsistent timestamps, missing and long-tail onboarding fields, and raw user-agent strings. ### Constraints & Assumptions - Define the target from `first_sale` and `subscription_start`; do not use future information as features. - State how you handle rows where `first_sale` occurs before `subscription_start`. - Prioritize practical data reasoning, feature engineering, leakage prevention, interpretability, and presentation clarity over complex model tuning. - Distinguish predictive associations from causal claims. ### Clarifying Questions to Ask - At what time will the model score a subscriber: at subscription start, after trial, or after onboarding? - What prediction horizon matters: any future sale, 7-day seller, 30-day seller, or another window? - What is the data extraction date so right-censoring can be handled? - What business action will use the prediction: lifecycle messaging, sales outreach, onboarding personalization, or product analysis? - What cost tradeoff matters more: missing potential sellers or contacting low-intent users? ### Part 1 - Define The Target And Handle Timestamp Edge Cases How would you define the seller-intent label and handle timestamp anomalies? #### What This Part Should Cover - A fixed-horizon target such as seller within 30 days after subscription start. - Exclusion or flagging of records with `first_sale < subscription_start`, with a clear rule for minor timezone-like differences if used. - Right-censoring treatment for subscriptions without a full observation window. - Avoidance of leakage from `first_sale` or `days_to_first_sale` as features. ### Part 2 - Perform Focused EDA What exploratory analysis would you perform? #### What This Part Should Cover - Data quality checks, uniqueness, timestamp parsing, missingness, class balance, and outliers. - Conversion by plan, period, country, trial length, discount, topic, need, and device/browser features. - Long-tail category analysis and high-signal tables or plots that support modeling decisions. ### Part 3 - Engineer Features What features would you create? #### What This Part Should Cover - Timestamp, trial, discount, plan, billing period, country, topic, need, and missingness features. - Rare-category grouping and one-hot encoding. - User-agent-derived device type, operating system, and browser. - Train-only preprocessing to prevent validation leakage. ### Part 4 - Train And Evaluate An Interpretable Model Which model and metrics would you use? #### What This Part Should Cover - Baseline model, regularized logistic regression, and optional comparison to a tree-based model. - Stratified or temporal validation based on the business use case. - ROC-AUC, PR-AUC, log loss or Brier score, calibration, precision/recall, and lift or precision at top K. ### Part 5 - Explain Drivers And Limitations How would you explain the main drivers of seller intent? #### What This Part Should Cover - Coefficient or odds-ratio interpretation for logistic regression. - Clear language that features are predictive associations, not proven causes. - Limitations around right-censoring, timestamp quality, small sample size, missing fields, and observational data. ### Part 6 - Prepare The Final Presentation What should the final notebook or presentation include? #### What This Part Should Cover - Goal, target definition, EDA findings, feature engineering, model choice, metrics, drivers, business recommendations, limitations, and next steps. - Concise visuals and an executive summary. - Recommendations for improved tracking, experiments, monitoring, and productionization. ### What a Strong Answer Covers - Defines the prediction problem precisely before modeling. - Prevents leakage and handles censoring and timestamp anomalies honestly. - Uses interpretable modeling and business-relevant metrics. - Explains what the model can and cannot claim. ### Follow-up Questions - What if the seller rate is only 5%? - How would you choose a threshold for outreach? - What if annual plans are highly predictive but not causal? - How would you monitor the model after launch? - How would you improve the data collection process?