Veri Bilimci - Senaryo Soruları

1) Data quality assessment: Missing values, imbalance, noise. 2) Feature engineering: Behavioral proxies (website activity, email engagement), external data enrichment. 3) Baseline model: Logistic regression. 4) Advanced: Random Forest, XGBoost. 5) Interpretability: SHAP analysis. 6) Business validation: A/B test ile churn reduction campaign effectiveness ölçerim.

Phased approach: 1) Basit başlarım: Popularity-based, item-item collaborative filtering. 2) Cold-start problem için: Content-based filtering (product attributes). 3) Advanced: Matrix factorization (SVD), deep learning (Neural Collaborative Filtering). 4) Evaluation: Offline (precision@k, recall@k), Online (A/B test, CTR). 5) Scalability için Spark MLlib.

Investigation plan: 1) Data drift detection (feature distribution comparison), 2) Model degradation (concept drift), 3) Infrastructure issues (latency, memory), 4) Data pipeline issues (quality, freshness). Solutions: Retraining pipeline, feature update, model monitoring, alerting setup. Automated retrigger mechanism implement ederim.

Extreme imbalance challenge (fraud rate %0.1). Solutions: 1) Resampling (SMOTE, ADASYN), 2) Cost-sensitive learning, 3) Anomaly detection (Isolation Forest, Autoencoder), 4) Ensemble methods, 5) Human-in-the-loop verification. Metric optimization: Precision-recall tradeoff, business cost analysis.

Data-first approach: 1) Data strategy develop ederim (internal data, external sources, synthetic data), 2) MVP scope (minimum viable model), 3) Transfer learning (pretrained models), 4) Human labeling setup, 5) Iterative improvement. Eğer data acquisition cost > benefit ise, alternative solution öneririm.

Explainability approach: 1) Global interpretability: Feature importance, partial dependence plots. 2) Local interpretability: LIME, SHAP values for individual predictions. 3) Rule extraction: Decision tree surrogate model. 4) Model comparison: Simple vs complex model performance tradeoff. Interactive explanation dashboard build ederim.

Real-time architecture: 1) Offline: Batch model training, feature engineering. 2) Online: Precomputed recommendations, caching (Redis), model serving (TensorFlow Serving, TorchServe). 3) Fallback: Popular items. 4) Monitoring: Latency, cache hit rate. 5) Optimization: Model compression (quantization, pruning), approximate nearest neighbors.

Healthcare AI considerations: 1) Regulatory compliance (HIPAA, FDA), 2) Model validation (external dataset, clinical trials), 3) Explainability (doctor decision support, not replacement), 4) Bias detection (demographic parity), 5) Human oversight (clinician-in-the-loop). Legal review ve ethics board approval gerekir.

Attribution modeling: 1) Single-touch: Last-click, first-click (basit ama naive). 2) Multi-touch: Linear, time-decay, position-based. 3) Data-driven: Markov chains, Shapley values. 4) Incrementality testing: Randomized controlled trials. Marketing team'i education ederim ve pilot implementation öneririm.

Resource optimization: 1) Transfer learning (pretrained models fine-tuning), 2) Distributed training (data parallelism), 3) Mixed precision training (FP16), 4) Model compression (pruning, quantization), 5) Cloud spot instances. Ayrıca: Colab, Kaggle Kernels, academic partnerships. Cost-benefit analysis yaparım.