Shopper intent prediction from clickstream e-commerce data with minimal browsing information

Area of Science:

• Computational intelligence
• Data science
• E-commerce analytics

Background:

• User intent prediction from clickstream data is crucial for e-commerce personalization and optimization.
• Existing methods often require detailed data, which may not always be available or practical.
• Symbolic representation of clickstream data offers a way to simplify analysis while retaining key information.

Purpose of the Study:

• To develop and compare two distinct approaches for user intent prediction from coarse-grained e-commerce clickstream data.
• To evaluate the effectiveness of hand-crafted feature-based and deep learning-based classification methods.
• To assess the performance of these methods for both arbitrary and limited-length trajectory classifications, including imbalanced datasets.

Main Methods:

• Development of a proprietary, coarse-grained clickstream dataset to create symbolic trajectories.
• Implementation of a hand-crafted feature-based classification using k-gram statistics and visibility graph motifs.
• Benchmarking and improvement of deep learning models, specifically a proposed Long Short-Term Memory (LSTM) architecture, against state-of-the-art (SOTA) methods.

Main Results:

• K-gram statistics with visibility graph motifs demonstrated fast and accurate classification performance.
• Purchase prediction was found to be reliable even with extremely short observation windows using the feature-based approach.
• The proposed LSTM architecture achieved improved classification accuracy compared to SOTA models on the new dataset.

Conclusions:

• Both hand-crafted feature engineering and deep learning are viable for user intent prediction from simplified clickstream data.
• The feature-based approach offers efficiency and accuracy, particularly for early purchase prediction.
• The deep learning approach, specifically the LSTM model, provides enhanced accuracy, with a thorough analysis of trade-offs for industry applications.