HelloFresh – AI-Driven Visitor Intent Prediction

Overview & Business Challenge:

For any online business, particularly subscription services, maximizing the conversion rate of website/app visitors into paying customers is crucial for ROI and sustainable growth. A common challenge is understanding and mitigating visitor abandonment before a desired action (e.g., subscription) is completed. This project aimed to address this by developing predictive models using clickstream data to identify, in real-time, a user’s likelihood to convert, thereby enabling targeted interventions to improve conversion rates and reduce wasted marketing spend.

Approach & AI/ML Methodology:

• Data Foundation: Analyzed 2.5 million rows of raw clickstream data (from BigQuery) representing 26,000 unique customer journeys on a leading subscription service platform.
• Extensive Data Engineering: Performed comprehensive data cleaning and feature engineering to transform raw user “hits” into meaningful features. This included:
  • Indexing unique customer journeys.
  • Dimensionality reduction for high-cardinality features (e.g., “screen_name”).
  • Creating aggregated statistical features (for shallow models) and sequential features (for LSTM) representing user behavior patterns.

Predictive Modeling (Dual Track):

• Shallow Machine Learning (Adjusted Statistical Behavior Predictive Method – ASBPM):
  • Utilized ensemble models like CatBoost on engineered statistical features.
  • Achieved a strong ROC AUC of 0.8715, demonstrating the ability to predict conversion intent without data leakage.

• Deep Learning (LSTM Method):

• • Developed a Bidirectional LSTM (BiLSTM) model to capture the nuances of sequential user behavior.
  • The BiLSTM model (using 55 hit steps) achieved an F1-macro score of 87% and ROC AUC of 0.875 (earlier tests showed 0.905 with 50 steps), indicating high predictive accuracy, especially when addressing data imbalance with SMOTE.

Rigorous Model Evaluation: Assessed models using ROC AUC, Precision-Recall AUC, and F1-Macro scores to ensure robustness.

Key Outcomes & Business Impact:

• High-Accuracy Intent Prediction: Developed models that accurately predicted a user’s likelihood to convert or abandon their journey.
• Real-Time Intervention Capability: The models provide probabilities of abandonment, enabling businesses to:
  • Identify at-risk users in real-time.
  • Deploy targeted interventions (e.g., personalized chatbot offers, support outreach) at optimal points in the user journey (e.g., “Checkout” page for non-converting users, as identified by the LSTM).

Significant Projected Revenue Uplift:

• The CatBoost model projected a potential 8% revenue increase with targeted interventions.
• The LSTM model projected a potential 12% revenue increase, demonstrating the value of more complex sequential modeling.

Optimized Marketing Efficiency:

Allows for more intelligent resource allocation by focusing retention efforts on users most likely to abandon.

Key AI Concepts & Technologies Applied:

• Machine Learning: CatBoost, Random Forest, Ensemble Methods.
• Deep Learning: Bidirectional LSTM (BiLSTM) for Sequence Classification.
• Clickstream Data Analysis: Processing large-scale user interaction data.
• Advanced Feature Engineering: For both statistical and sequential modeling.
• Big Data Technologies: Google BigQuery.
• Predictive Modeling: Forecasting user conversion intent.
• Data Imbalance Techniques: SMOTE.