Abstract
Recently, Deep Neural Networks (DNNs) have been widely introduced into Collaborative Filtering (CF) to produce more accurate recommendation results due to their capability of capturing the complex nonlinear relationships between items and users. However, the DNNs-based models usually suffer from high computational complexity, i.e., consuming very long training time and storing huge amount of trainable parameters. To address these problems, we propose a new broad recommender system called Broad Collaborative Filtering (BroadCF), which is an efficient nonlinear collaborative filtering approach. Instead of DNNs, Broad Learning System (BLS) is used as a mapping function to learn the complex nonlinear relationships between users and items, which can avoid the above issues while achieving very satisfactory recommendation performance. However, it is not feasible to directly feed the original rating data into BLS. To this end, we propose a user-item rating collaborative vector preprocessing procedure to generate low-dimensional user-item input data, which is able to harness quality judgments of the most similar users/items. Extensive experiments conducted on seven benchmark datasets have confirmed the effectiveness of the proposed BroadCF algorithm.
Introduction
Collaborative Filtering (CF) is the industry standard for predicting user preferences based on historical behavior. While traditional Matrix Factorization (MF) is efficient, it often struggles to capture the complex, nonlinear relationships between users and items. Deep Neural Networks (DNNs) were introduced to solve this, but they come with significant drawbacks: high computational complexity requiring numerous training epochs, memory bottlenecks on large-scale datasets due to millions of parameters, and a tendency to struggle with "long-tail" data distributions.
BroadCF aims to provide a model that captures nonlinearities as effectively as a DNN but trains as fast as a shallow model by utilizing a Broad Learning System (BLS) architecture.
Methodology
BroadCF consists of two main stages: a preprocessing step to generate collaborative vectors and the application of a Broad Learning System (BLS) for prediction.
1. Collaborative Vector Preprocessing
Directly feeding sparse, high-dimensional rating vectors into a shallow network is ineffective. BroadCF creates a low-dimensional user-item rating collaborative vector ($X$):
- KNU (k-Nearest Users): For a user $u$, the system finds the $k$ most similar users and extracts their ratings for item $v$ to form vector $p_v^u$.
- LNI (l-Nearest Items): Similarly, it finds the $l$ most similar items to $v$ and extracts user $u$'s ratings for them to form vector $q_v^u$.
- Smart Filling: Instead of simple mean imputation, a similarity-based threshold strategy fills gaps for neighbors who haven't rated an item, preserving high-quality signals.
2. Broad Learning System (BLS)
Instead of stacking layers vertically, BLS expands the input horizontally:
- Mapped Features ($Z_i$): The input is projected into a feature space using random weights, refined via a sparse autoencoder to ensure compactness.
- Enhanced Features ($H_i$): These features are further expanded using additional random weights and nonlinear activations (ReLU).
- The Solver: Unlike DNNs that use iterative backpropagation, BLS solves for the output weights using ridge regression and the Moore-Penrose pseudo-inverse. This one-step calculation eliminates the need for lengthy training epochs.
Experimental Results
The authors evaluated BroadCF on seven real-world datasets, including MovieLens (ml-la) and six Amazon subsets (Digital Music, Grocery, Patio/Lawn, Automotive, Baby, and Instant Video). The Amazon-IV dataset, with over 429,000 users, served as a stress test for scalability.
- Accuracy: BroadCF consistently outperformed eight state-of-the-art baselines (including NeuMF, NGCF, and LightGCN) in RMSE and MAE across almost all datasets.
- Training Speed: On the Amazon-DM dataset, BroadCF was 4x faster than the fastest baseline (SHT) and roughly 30x faster than attention-based models like BPAM.
- Scalability and Memory: On the largest dataset (Amazon-IV), most DNN-based models failed due to out-of-memory (OOM) errors. BroadCF handled it easily, maintaining a constant, small number of trainable parameters (3,125 in experiments) regardless of dataset size, whereas models like DeepCF scale to over 137 million parameters.
Conclusion
BroadCF demonstrates that "broadening" a shallow network with random features can approximate complex functions as well as deep architectures but with significantly lower overhead. By keeping hidden weights random and only training the final output layer, the model retains nonlinear modeling capabilities while achieving massive speedups. Future work will focus on incorporating auxiliary data like text reviews or images and exploring cross-domain recommendation tasks.
Summary updated