Introduction
Recommender Systems (RS) primarily rely on user behavior data, typically represented as a high-dimensional user-item rating matrix. However, real-world data is characterized by extreme sparsity and heterogeneity, where different users exhibit vastly different rating patterns. While Deep Neural Network (DNN) models like AutoRec have advanced rating prediction, most existing approaches are "single-metric," utilizing a fixed loss function and regularization term. This rigidity often fails to capture the diverse underlying properties of complex user data.
This paper introduces the Multi-Metric AutoRec (MMA) framework, which adopts a "multi-lens" approach to recommendation by integrating multiple mathematical norms to better represent sparse and varied user behaviors.
The Multi-Metric AutoRec (MMA) Framework
The MMA framework extends the standard AutoRec architecture by training four distinct base model variants. Each variant utilizes a unique combination of $L_1$ and $L_2$ norms for its loss function $l(\Delta_{j,k})$ and regularization term $r(w_i)$. These norms act as different mathematical "lenses": the $L_1$ norm provides robustness against outliers and promotes feature selection, while the $L_2$ norm ensures mathematical stability and fine-grained data representation.
Beyond simply training multiple models, MMA employs a self-adaptive ensemble strategy. During the training phase, the framework monitors the accumulative loss of each variant on a validation set. It dynamically adjusts the weights of each model's predictions based on their real-time performance, allowing the system to automatically favor the metric combination that best suits the specific characteristics of the dataset.
Experimental Evaluation
The performance of MMA was evaluated against five benchmark datasets: MovieLens (1M, 100k, and HetRec), Yahoo (music ratings), and Douban (social networking ratings). These datasets represent a wide range of sparsity levels, with the Douban dataset (D5) being the most challenging at a density of only 1.52%.
MMA was compared against seven state-of-the-art baselines, including traditional Matrix Factorization (MF), DNN-based models (AutoRec, NRR, SparseFC, GLocal-K), and Graph Neural Network-based models (IGMC).
Results and Analysis
Experimental results demonstrate that MMA consistently outperforms existing state-of-the-art models across nearly all tested benchmarks. As shown in the performance table below, MMA achieved the lowest Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) in the majority of scenarios.
Key observations from the results include:
- Superior Accuracy: MMA provides more precise rating predictions than single-metric models by leveraging the strengths of both $L_1$ and $L_2$ norms.
- Robustness to Sparsity: On the highly sparse Douban dataset, MMA maintained competitive performance, highlighting the effectiveness of its multi-metric approach in data-starved environments.
- Efficiency: Despite training four variants, the autoencoder-based architecture of MMA remains computationally efficient compared to more complex graph-based models like IGMC, while delivering superior or comparable accuracy.
Conclusion
The Multi-Metric AutoRec framework proves that the choice of metric space—how we measure error and constrain model weights—is as critical as the neural network architecture itself. By dynamically weighting multiple base models, MMA adapts to the inherent heterogeneity of user behavior data. Future research may explore applying this multi-metric ensemble approach to more complex architectures, such as Transformers or specialized Graph Neural Networks, to further push the boundaries of recommendation accuracy.
Summary updated