Spectral Clustering of Customer Transaction Data With a Two-Level Subspace Weighting Method

Area of Science:

• Data Mining
• Machine Learning
• E-commerce Analytics

Background:

• Customer segmentation from transaction data is crucial for retail and e-commerce.
• Existing Purchase Tree (PurTree) methods lack nuanced weighting for nodes and levels.
• This limitation hinders accurate clustering of customer transaction data.

Purpose of the Study:

• To propose a novel two-level subspace weighting spectral clustering (TSW) algorithm.
• To enhance customer transaction data clustering by introducing adaptive weighting mechanisms.
• To improve the accuracy and effectiveness of customer group discovery.

Main Methods:

• Developed a PurTree subspace metric incorporating level and sparse node weights.
• Implemented an iterative optimization algorithm for model training.
• Introduced an efficient method for calculating a regularization parameter.

Main Results:

• TSW adaptively learns a similarity matrix from local distances.
• The algorithm effectively distinguishes the importance of different tree levels and nodes.
• Comparative experiments on ten datasets demonstrated TSW's superiority over six other clustering algorithms.

Conclusions:

• The proposed TSW algorithm offers a significant advancement in customer transaction data clustering.
• Weighted spectral clustering using PurTree subspace metrics enhances customer segmentation accuracy.
• TSW provides a more effective approach for uncovering hidden cluster structures in transaction data.