RecNet: A Resource-Constraint Aware Neural Network for Used Car Recommendation

Haihua Shi, Jianjun Qian, Nengjun Zhu, Tong Zhang, Zhen Cui, Qianliang Wu, Shanshan FengInt. J. Comput. Intell. Syst 2022 · OpenReview

Abstract

Resource constraints, such as limited product inventory or financial strength, significantly influence consumer choices but are often overlooked in traditional recommendation systems. This paper introduces RecNet, a framework designed to mine user preferences within resource-constrained environments. By utilizing a large-scale used car transaction dataset, the authors propose a network that captures dynamic user-item connections through mutually recursive recurrent neural networks (MRRNNs) and a dedicated resource-constraint branch. Experimental results on used car and Tmall datasets demonstrate that RecNet effectively incorporates these constraints to improve recommendation accuracy.

Introduction: The Challenge of Finite Resources

Most recommendation algorithms operate under the assumption of "infinite supply and infinite demand," assuming that past preferences directly translate to future ones. However, in high-value "bulk commodity" markets like used cars, two critical constraints disrupt this logic:

  1. Limited Supply: Unlike digital goods, a specific used car is unique; once sold, it is removed from inventory.
  2. Financial Strength: A major purchase, like a car, significantly impacts a user's immediate financial capacity and subsequent purchasing behavior.

Existing models often fail to track how a user's state or the market's state changes post-transaction. RecNet addresses this by treating recommendation as a resource-aware sequential task.

Methodology: The RecNet Architecture

RecNet employs a multi-branched architecture to capture the complex interplay between users, items, and their respective constraints.

The RecNet architecture, featuring mutually-recursive RNNs and a dedicated resource-constraint branch.

Mutually-Recursive RNNs (MRRNNs)

Instead of modeling users and items in isolation, RecNet uses two RNNs that take each other’s output as input. This captures the co-evolving relationship between a user’s shifting tastes and an item’s interaction history.

The Resource-Constraint Branch

This specialized module processes factors such as purchase history, time elapsed since the last transaction, item price, and initial inventory levels. It learns a "resource-aware" embedding that reflects the user's current financial and situational constraints.

Personal Mutual Information (MI) and Joint Discriminator

To capture non-linear dependencies, the model uses Noise-Contrastive Estimation (NCE) to maximize the mutual information between historical context and future actions. The final ranking is determined by a joint discriminator that combines supervised Mean Squared Error (MSE) loss with unsupervised MI estimation.

Experimental Evaluation

The model was tested on two datasets: the UsedCars dataset (57,000 users, 18.7 million actions) and the Tmall e-commerce benchmark. The UsedCars dataset served as the primary test for resource-constraint modeling due to its high prices and low purchase ratios.

Key Results

RecNet significantly outperformed state-of-the-art baselines in the resource-constrained environment:

  • Performance: RecNet achieved a Recall@10 of 0.0856 and an NDCG@10 of 0.0660, compared to SASRec (Recall@10: 0.0414) and JODIE (Recall@10: 0.0659).
  • Ablation Study: Removing the resource-constraint branch led to a performance drop across all metrics, confirming the necessity of modeling financial and inventory limits.
  • Unsupervised Efficacy: The NCE-only variant performed competitively, highlighting the strength of mutual information in representation learning.

Conclusion and Future Directions

RecNet proves that explicitly modeling resource constraints is vital for high-value recommendations. Future work will focus on disentangling latent factors—separating merchant inventory from user capital—and exploring more advanced architectures like Transformers to capture even longer-term dependencies.

Summary updated

Abstract

Resource constraints, e.g., limited product inventory or financial strength, may affect consumers’ choices or preferences in some recommendation tasks but are usually ignored in previous recommendation methods. In this paper, we aim to mine the cue of user preferences or intention in the case of resource-constraint recommendation tasks. For this purpose, we specifically build a largely used car transaction dataset possessing resource-constraint characteristics. Accordingly, we propose a resource-constraint-aware network to predict the user’s future interaction based on dynamic connections between users and items. To describe the user-item connection dynamically, mutually recursive recurrent neural networks (MRRNNs) are introduced to capture long-term interactive dependencies, and effective representations of users and items are obtained. To further consider the resource constraint, a resource-constraint branch is built to explore resource variation’s influence on user preferences. Finally, mutual information is introduced to measure the similarity between the future user action and fused historical behavior features to predict future interaction. The fused features come from both MRRNNs and resource-constraint branches. We test the performance on the built used car transaction dataset and the Tmall dataset, and the experimental results verify the effectiveness of our framework.
BibTeX
@article{shi2022recnet,
  author = {Haihua Shi and Jianjun Qian and Nengjun Zhu and Tong Zhang and Zhen Cui and Qianliang Wu and Shanshan Feng},
  title = {{RecNet: A Resource-Constraint Aware Neural Network for Used Car Recommendation}},
  journal = {Int. J. Comput. Intell. Syst.},
  year = {2022},
  url = {https://openreview.net/forum?id=haeUIvXNDH},
}
Figure