Abstract
In e-commerce platforms, profit-aware recommender systems aim to improve the platform's profits while maintaining high overall accuracy by recommending items with high profits as top-ranked items. We explore two issues faced by existing model-based profit-aware approaches (MBAs) when training recommendation models for profit enhancement. First, current MBAs tend to inaccurately infer the item ranking by the profit-based weighting scheme; the ranking of observed items by a user is inferred without considering the user preference for each item, while all unobserved items are assumed to have an equally low ranking. Second, current MBAs train the model without employing the item ranking as ground truth; during training, the model is optimized for the preference score for each item independently rather than being directly optimized for the overall ranking of items. To tackle these issues, we propose a novel MBA that involves three key steps: (S1) defining the Current Preference incorporated with Profit (CPP) for items; (S2) classifying items through CPP; and (S3) training the model by list-wise learning-to-rank (LTR) based on CPP. Extensive experimental results using real-world platform datasets demonstrate that our approach improves accuracy by approximately 4% and profits by about 24% compared to the best-competing method.
Introduction
E-commerce platforms operate in a state of constant tension between two primary objectives: maximizing user satisfaction (accuracy) and maximizing platform revenue (profit). While traditional recommendation systems focus almost exclusively on predicting what a user is most likely to click or buy, profit-aware systems attempt to nudge users toward high-margin products.
However, existing Model-Based Approaches (MBAs) often struggle to find the right balance. They frequently suffer from a "profit-bias" that ignores actual user intent or rely on optimization techniques that do not reflect how users actually interact with recommendation lists. This paper introduces a novel framework designed to bridge this gap by integrating user preference, purchase recency, and item profit into a unified list-wise learning framework.
The Limitations of Current Profit-Aware Models
The researchers identify two fundamental flaws in current profit-aware recommendation strategies:
- Inaccurate Ranking Inference: Most models assume that a user’s preference for an item is directly proportional to its profit margin. In reality, users are generally unaware of margins; they purchase based on personal utility. Furthermore, these models often treat all unpurchased items as equally irrelevant, failing to identify "hidden gems"—items that are both highly profitable for the platform and highly relevant to the user's specific tastes.
- Point-wise vs. List-wise Optimization: Standard recommendation models often utilize "point-wise" learning, where the system attempts to predict an independent score for each item in isolation. However, recommendation is inherently a ranking problem. Users evaluate items relative to others in a list. Optimizing for individual scores does not necessarily result in an optimal overall sequence for the user.
The Proposed Framework: CPP and List-wise LTR
To address these challenges, the authors propose a three-step methodology that redefines how models are trained for profit maximization.
S1: Defining Current Preference incorporated with Profit (CPP)
Instead of relying solely on profit margins, the authors introduce the CPP metric. This metric serves as the new "ground truth" for training and is calculated by combining three distinct factors:
- Pre-trained Preference: An initial estimate of what the user likes based on historical data.
- Recency Score: A weight that prioritizes a user's most recent interactions, acknowledging that tastes evolve over time.
- Item Profit: The actual margin the platform earns from the sale.
S2: Classification of Unobserved Items
To solve the problem of "hidden gems," the model classifies unpurchased items. Rather than assuming all unobserved items are of zero interest, the framework identifies items that align with a user’s current interests and have high profit margins. These items are then sampled and treated as positive examples during the training phase, allowing the model to learn to "nudge" users toward profitable items they are likely to enjoy.
S3: List-wise Learning-to-Rank (LTR)
Moving away from point-wise scoring, the authors implement a list-wise loss function. This approach trains the model to optimize the entire sequence of recommendations simultaneously. By focusing on the relative order of items, the model ensures that items with a higher CPP (those that balance preference, recency, and profit) are consistently ranked above those with lower scores.
Experimental Evaluation
The researchers evaluated the proposed approach using three real-world datasets from Amazon: Beauty (ABeauty), Video Games (AGames), and Cellphones (ACellphones). Item prices were used as a proxy for profit margins, and the data was split chronologically (60% training, 20% validation, 20% testing) to simulate a real-world production environment.
The method was tested against five state-of-the-art baselines across three different collaborative filtering architectures: Matrix Factorization (MF), Collaborative Denoising Auto-Encoder (CDAE), and LightGCN.
Key Findings
- Substantial Profit Gains: The approach achieved a 24% increase in total profit compared to the best-performing baseline methods.
- Improved Accuracy: Unlike many profit-aware models that see a decline in user relevance, this method improved recommendation accuracy (measured by Recall@10) by approximately 4%.
- Efficiency and Scalability: The training time for the model scales linearly with the number of users, suggesting it is viable for large-scale commercial applications.
Conclusion and Future Work
By shifting the focus from independent item scores to a holistic, list-wise ranking based on current user preferences and profit, this research provides a path forward for e-commerce platforms to increase revenue without degrading the user experience. The "hidden gem" strategy proves that profit and preference do not have to be a zero-sum game.
Future research directions include adapting the model for dynamic profits, where margins change due to seasonal sales or supply chain shifts, and reducing the dependency on the initial pre-training phase to streamline the model's deployment.
Summary updated