Modeling Reference-dependent Choices with Graph Neural Networks

Liang Zhang, Guannan Liu, Junjie Wu, Yong TanarXiv 2024 · DOI · arXiv

Introduction

Most modern recommender systems operate on the assumption that a user’s preference for a product is based on its absolute utility. By analyzing past interactions, these models attempt to predict a score for a product in isolation. However, behavioral economics—most notably Prospect Theory—suggests that human decision-making is inherently comparative. Consumers do not evaluate products in a vacuum; instead, they judge them relative to "reference points," such as products they have previously owned or items currently popular within their social and market circles.

While the concept of reference-dependent choice is well-established in marketing science, integrating it into data-driven machine learning models presents significant challenges. Reference points are subjective, multi-faceted, and often latent. To bridge this gap, this paper introduces the Attributed Reference-dependent Choice Model for Recommendation (ArcRec), a Graph Neural Network (GNN) framework designed to capture how users evaluate products against both personal history and collective market behavior.

The ArcRec Framework

The ArcRec model moves beyond simple interaction matrices by incorporating the structural relationships between products and their attributes. The architecture is built upon several key innovations:

Attributed Reference Networks (ARNs)

To capture the "wisdom of the crowd," the authors construct global reference networks. These networks connect products that are frequently co-purchased or co-viewed. ArcRec further decomposes these into attribute-specific sub-networks (e.g., brand-based or category-based graphs). By applying GNNs to these ARNs, the model learns embeddings that reflect how a product is positioned relative to its competitors in the broader market.

Dual-Preference Utility Function

ArcRec evaluates a target product by comparing it to a user's personal reference points—their purchase history—across two distinct dimensions:

  1. Interest-inspired preference: This measures the alignment between the target product’s features and the user’s historical preferences.
  2. Price-inspired preference: This captures the "sticker shock" effect, evaluating how the product's price deviates from the user’s expected reference price.

Attribute-level Willingness-To-Pay (AWTP)

A critical contribution of this work is the AWTP module, an attention mechanism that recognizes that price sensitivity is not uniform across all product features. For example, a consumer might be highly price-sensitive regarding a product's size but willing to pay a premium for a specific brand. AWTP calculates a user's price tolerance for specific attributes by analyzing their historical deviations from average purchase prices, allowing for a more nuanced understanding of consumer priorities.

Experimental Evaluation

The authors validated ArcRec using both synthetic and real-world data:

  • Synthetic Dataset: Using multinomial logit models from marketing science, the authors simulated a retail environment to test the model against known ground-truth purchasing probabilities and price sensitivities.
  • Real-world Cosmetics Dataset: This dataset, sourced from a Taobao online store, included over 204,000 interactions among 17,590 consumers and 6,989 products. Attributes included brand, category, size, and price.

Key Results

ArcRec demonstrated substantial improvements over 14 state-of-the-art baselines, including LightGCN, xDeepFM, and price-aware models like PUP. On the Cosmetics dataset, ArcRec outperformed xDeepFM by approximately 44% in Hit Ratio (HR@5) and 45% in nDCG@5.

Beyond general accuracy, the model showed exceptional performance in two specific areas:

  1. Cold-Start Scenarios: In simulations where 1.5% of products were entirely new, ArcRec improved HR@5 by 16.16% over the strongest baseline. Because ArcRec leverages attribute-based reference networks, it can effectively position a new product in the graph based on its features (e.g., brand or category) even without a prior interaction history.
  2. Price Dynamics: In controlled tests where prices were fluctuated by 10%, ArcRec accurately predicted shifts in consumer behavior. It successfully identified price-sensitive groups and adjusted recommendations more aggressively for them compared to less sensitive segments.

Discussion and Future Work

The success of ArcRec highlights the importance of modeling the interaction between specific attributes and price tolerance rather than treating price as a standard feature. By grounding recommendation logic in Prospect Theory, the model provides a more psychologically accurate representation of consumer choice.

However, the model currently relies on implicit feedback (purchases). Future iterations could benefit from incorporating "negative" signals, such as products that were viewed but rejected, to further refine the understanding of reference points. Additionally, the interpretability of the AWTP metric offers significant potential for business intelligence, enabling companies to segment consumers and design targeted pricing strategies based on the specific attributes those consumers value most.

Summary updated

BibTeX
@misc{zhang2024modeling,
  author = {Liang Zhang and Guannan Liu and Junjie Wu and Yong Tan},
  title = {{Modeling Reference-dependent Choices with Graph Neural Networks}},
  year = {2024},
  eprint = {2408.11302},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2408.11302},
}
Figure