Customer Lifetime Value Prediction Using Embeddings (1703.02596v3)

Abstract: We describe the Customer LifeTime Value (CLTV) prediction system deployed at ASOS.com, a global online fashion retailer. CLTV prediction is an important problem in e-commerce where an accurate estimate of future value allows retailers to effectively allocate marketing spend, identify and nurture high value customers and mitigate exposure to losses. The system at ASOS provides daily estimates of the future value of every customer and is one of the cornerstones of the personalised shopping experience. The state of the art in this domain uses large numbers of handcrafted features and ensemble regressors to forecast value, predict churn and evaluate customer loyalty. Recently, domains including language, vision and speech have shown dramatic advances by replacing handcrafted features with features that are learned automatically from data. We detail the system deployed at ASOS and show that learning feature representations is a promising extension to the state of the art in CLTV modelling. We propose a novel way to generate embeddings of customers, which addresses the issue of the ever changing product catalogue and obtain a significant improvement over an exhaustive set of handcrafted features.

• The paper introduces a CLTV prediction system that integrates 132 handcrafted features with customer embeddings for improved forecasting.
• It employs a Random Forest model with Apache Spark and TensorFlow, achieving enhanced predictive accuracy measured by AUC and Spearman correlation.
• The system addresses zero-inflated, long-tailed customer value distributions and explores deep learning alternatives despite higher computational costs.

The paper "Customer Lifetime Value Prediction Using Embeddings" (1703.02596) discusses the development and deployment of a Customer Lifetime Value (CLTV) prediction system at ASOS.com, an online fashion retailer. The primary motivation for this system is to forecast the future value of customers to optimize marketing spend, identify high-value customers, and control churn.

Key Contributions:

1. System Architecture:
   • The paper describes a large-scale CLTV system integrated with operational systems at ASOS.
   • The system is based on a Random Forest (RF) regression model utilizing 132 handcrafted features derived from customer demographics, purchase history, returns history, and web/app session logs.
   • The architecture includes a feature generation process using Apache Spark and customer embedding generation using Tensorflow.
2. Feature Engineering:
   • Extensive use of handcrafted features is documented, with feature importance ranked using methods like random forest feature importance.
   • Features such as the number of orders, session frequency, and product views are highlighted as crucial for CLTV prediction.
3. Challenge of Prediction:
   • The paper identifies the challenge of predicting CLTV due to the zero-inflated and long-tailed nature of the distribution of customer values.
4. Embedding Techniques:
   • The authors explore customer embeddings generated from browsing session data. These embeddings are inspired by NLP techniques such as SkipGram with Negative Sampling (SGNS).
   • This method improves the system’s ability to predict future customer value by capturing richer patterns in the raw data.
5. Integration and Deployment Considerations:
   • The system is designed to retrain daily using up-to-date customer interaction data, with predictions calibrated for both churn and actual monetary spend.
   • Challenges in integrating learned features into live systems due to feature permutation issues are addressed, providing practical solutions for embedding continuity.
6. Improved Metrics:
   • The results demonstrate significant improvement in predictive accuracy when using embeddings, as measured by the Area Under Curve (AUC) and Spearman rank-order correlation.
7. Exploration of Deep Learning:
   • The paper also experiments with deep feedforward neural networks as a potential alternative or supplementary model to the current system but finds that the cost-benefit ratio currently favors the RF model given computational considerations.

Conclusions:

The authors conclude that while the baseline Random Forest model with handcrafted features performs well, integrating neural embeddings into the feature set provides further improvements and opportunities for enhancing CLTV predictions. The exploration towards using deep neural networks indicates possibilities but remains constrained by practical deployment costs.

Overall, the paper is a detailed account of how a machine learning approach was applied practically within a commercial setting, with insights into model development, deployment challenges, and enhancement through feature learning.