Abstract
Traditional recommender systems rely on negative sampling to handle implicit feedback, a process that is often biased and computationally sensitive. Furthermore, many models fail to exploit the rich, heterogeneous nature of user behaviors, such as viewing or adding items to a cart before a purchase. This paper introduces Efficient Heterogeneous Collaborative Filtering (EHCF), a framework that eliminates negative sampling by optimizing over the entire dataset with linear complexity. By incorporating a transfer-based prediction layer, EHCF models the sequential dependencies between different user behaviors as translations in a vector space. Experimental results on large-scale datasets like Taobao and Beibei demonstrate that EHCF significantly outperforms state-of-the-art methods in both accuracy and training efficiency, achieving up to a 23x speedup.
Introduction
Most modern recommendation algorithms utilize implicit feedback, where only positive interactions are observed. To train these models, researchers typically employ Negative Sampling (NS), which treats a subset of unobserved items as negative signals. However, NS is inherently biased and its performance is highly sensitive to the sampling ratio.
Beyond the challenges of sampling, users leave a trail of diverse interactions—views, clicks, and cart additions—before making a purchase. These heterogeneous behaviors are often ignored or treated as independent tasks. EHCF addresses these gaps by providing an efficient way to learn from the entire dataset while capturing the natural funnel of user intent.
Methodology
Fast Whole-Data Optimization
EHCF's primary innovation is a mathematical rearrangement of the loss function that allows the model to consider every unobserved item as a negative signal without the $O(Users \times Items)$ computational burden. By leveraging the sparsity of positive interactions and rearranging nested sums, the authors reduce the complexity to a linear scale relative to the number of users and items. This ensures stable convergence and removes the randomness associated with sampling.
Transfer-Based Prediction Layer
Rather than treating behaviors like "viewing" and "buying" as isolated events, EHCF models them as a sequence of related tasks. Inspired by knowledge graph embeddings, the model uses a "transfer scheme" where high-level behaviors are viewed as translations of lower-level ones in a shared vector space.
This architecture allows the model to use abundant auxiliary data (e.g., views) to bolster the prediction of target behaviors (e.g., purchases), which are often much sparser.
Experimental Results
Accuracy and Convergence
EHCF was evaluated against state-of-the-art models including Neural Collaborative Filtering (NCF) and Neural Multi-Task Recommendation (NMTR). On the Taobao and Beibei datasets, EHCF showed average improvements of 57.1% and 47.5% respectively over the best baselines.
As illustrated in the figure above, EHCF reaches its peak performance rapidly. In practical terms, on the Beibei dataset, EHCF converged in just 24 minutes, while the previous state-of-the-art (NMTR) required 550 minutes—a 23x improvement in training efficiency.
Robustness to Sparsity
A key strength of EHCF is its performance for "cold-start" users. By effectively transferring patterns from a user's viewing history, the model can make accurate purchase recommendations even for users with very few historical transactions. This robustness is a direct result of the joint optimization of heterogeneous behaviors.
Conclusion
EHCF provides a robust and efficient alternative to traditional negative sampling-based recommendation models. By combining whole-data optimization with a behavior-aware transfer scheme, it sets a new benchmark for both accuracy and speed in heterogeneous collaborative filtering. Future work will focus on extending this non-sampling optimization to non-linear deep neural network architectures.
Summary updated