Towards Robust Recommendation via Decision Boundary-aware Graph Contrastive Learning

Formally, let $\mathcal{U}=\{u_{1},u_{2},\dots,u_{M}\}$ and $\mathcal{I}=\{i_{1},i_{2},\dots,i_{N}\}$ denote the set of users and items, respectively, where $M$ and $N$ represent the number of users and items, respectively. Considering recommendation scenario with implicit feedback, a binary matrix $\mathbf{R}\in\mathbb{R}^{M\times N}$ are typically used to record user-item interactions (e.g., clicks or purchases), where $r_{u,i}=1$ indicates that user $u$ has interacted with item $i$, otherwise $r_{u,i}=0$. Following most GNN-based recommendation works (He et al., 2020, 2023; Huang et al., 2021), we formulate the interaction behaviors between users and items as a standard bipartite graph $\mathcal{G}=\{\mathcal{V},\mathbf{A}\}$, where $\mathcal{V}=\mathcal{U}\cup\mathcal{I}$ involves all graph nodes, and the adjacent matrix $\mathbf{A}$ is defined as follows:

Following the common practice (He et al., 2020; Cai et al., 2023), we encode the user $u$ and item $i$ as d-dimensional latent vectors $\mathbf{e}_{u}\in\mathbb{R}^{d}$ and $\mathbf{e}_{i}\in\mathbb{R}^{d}$, respectively. Besides, $\mathbf{E}=\{\mathbf{e}_{u}\mid u\in\mathcal{U}\}\cup\{\mathbf{e}_{i}\mid i\in\mathcal{I}\}$ is defined as the overall learnable embedding matrix for all nodes.

Similar to other GCL-based works (Yu et al., 2022; Wu et al., 2021; Wei et al., 2022), this paper adopts the LightGCN (He et al., 2020) as model backbone. Specifically, the comprehensive graph representations $\mathbf{z}_{u}$ and $\mathbf{z}_{i}$ for user $u$ and item $i$ in LightGCN are calculated by

where $\mathcal{N}_{u}$ and $\mathcal{N}_{i}$ indicate the neighboring nodes of user $u$ and item $i$, respectively. $\mathbf{h}^{(l)}_{u}$ and $\mathbf{h}^{(l)}_{i}$ means the $l$-th layer graph representation for user $u$ and item $i$, respectively. Here, $\mathbf{h}^{(0)}_{u}$ and $\mathbf{h}^{(0)}_{i}$ are initialized with the learnable embedding $\mathbf{e}_{u}$ and $\mathbf{e}_{i}$, respectively. The predicted score $\hat{r}_{u,i}$ for the $(u,i)$ pair is computed as the inner product of their graph representations, i.e., $\hat{r}_{u,i}=\Braket{\mathbf{z}_{u},\mathbf{z}_{i}}$. Finally, the BPR (Rendle et al., 2012) loss is adopted as the optimization objective:

where $\sigma(x)=1/(1+e^{-x})$, $\mathcal{I}_{u}^{+}$ and $\mathcal{I}_{u}^{-}$ represent the positive item and unobserved item set for user $u$, respectively.

In real-world scenarios, interaction behaviors between users and items are actually highly sparse, which can lead to severe overfitting and bias problems (Wu et al., 2021; jing2023contrastive). Graph contrastive learning (GCL), as a novel learning paradigm, helps mitigate the above problems (Yu et al., 2022; Cai et al., 2023). In specific, GCL firstly generates diverse graph views for each user and item (e.g., node dropout and feature masking). Then the different views of the same user (item) are treated as the positive pairs, while the different views of the different instances are treated as the negative pairs. Finally, contrastive learning loss is used to optimize the model parameters with paired users and items, where InfoNCE (Oord et al., 2018) is the most commonly adopted loss. Formally, the contrastive learning loss for the user side can be defined as follows:

where $\mathbf{x}_{u}$ and $\mathbf{y}_{u}$ denote the two different augmented views of user $u$, $sim(\cdot,\cdot)$ and $\tau$ represents the cosine similarity function and temperature hyper-parameter, respectively. Similarly, the contrastive learning loss of the item side is formulated as follows:

where $\mathbf{x}_{i}$ and $\mathbf{y}_{i}$ denote the two different views of item $i$.

Adversarial training (AT) stands out as one of the most promising approaches for bolstering adversarial robustness (Moosavi-Dezfooli et al., 2016; Madry et al., 2017; Goodfellow et al., 2014). The goal of AT is to increase model robustness by generating adversarial examples through well-designed perturbations, which purposefully induce the neural network to error. Formally, the optimal perturbation for data sample $(x,y)$ is found by maximizing the loss function $\mathcal{L}(\cdot):\delta^{*}=\arg\max\mathcal{L}(x+\delta,y;\bm{\theta})$ where $\delta$ represents an adversarial perturbation of $\ell_{p}$ norm smaller than $\epsilon$. Then, the model is trained on a mixture of both original clean examples and generated adversarial examples to enhance the robustness ability.

Discussion. Adversarial robustness uncovers the root cause of the model’s adversarial vulnerability, that is, the non-smooth feature space near data samples (Jiang et al., 2020). In other words, small input perturbations likely result in large changes in the potential semantics, subsequently affecting the model output, which is the basis challenge that adversarial defense algorithms strive to resolve. Actually, this particularly fits well with graph contrastive learning, which aims to maximize the consistency of the given instance under different augmentation views. More importantly, adversarial robustness provides the maximum boundary of feature perturbations that the model can tolerate (cf. Sec 3.2), which effectively restrains the exploration space for contrastive augmentation and guides the generation of optimal view-generator.

The overall framework of RGCL is presented in Figure 2. In specific, we calculate the maximum feature perturbations to guide the subsequent generation of both contrastive examples and adversarial examples. For contrastive examples, we firstly generate two random-augmented views $\mathbf{Z}^{\prime}$ and $\mathbf{Z}^{\prime\prime}$ using random perturbations. Besides, the third view $\mathbf{Z}^{ac}$, which we refer to as adversarial-contrastive view, is generated through maximizing relation-aware contrastive function. On the foundation of these contrastive samples, we employ multi-view contrastive learning to prompt high-quality representations. Furthermore, to safeguard the model robustness against potential compromises arising from the uniformity optimization of graph contrastive learning, we generate adversarial examples using maximum perturbation to strenuously enlarge the distances between data points and the decision boundary. Finally, the model is updated by employing a joint optimization objective with augmented contrastive and adversarial data.

To build our contrastive samples, we first derive perturbations that the original samples can maximally tolerate to maintain user preferences. Ideally, the perturbations should satisfy two conditions: (1) the perturbations should be as large as possible, such that the obtained contrastive samples are hard enough (hardness requirement). (2) The augmented samples after incorporating the perturbations should be still aligned with the user’s original preferences (rationality requirement).

Different from traditional adversarial learning problems based on classification settings, recommender system is basically a ranking problem, and the perturbations should be learned to maintain user preference rankings. To this end, we propose to learn the maximum perturbations that can maintain item pair-wise rankings. Furthermore, given that different orders of graph representations possess different levels of expressive capacity, that is, higher-layer representations aggregate richer structure information and reflect more complex connectivity patterns. Consequently, we tailor the maximum perturbation for each high-order graph representation independently. In specific, for each user $u$ and a positive-negative item pair $(i^{+},i^{-})$, suppose their original representations are $\mathbf{{z}}_{u}=\sum_{l=0}^{L}\mathbf{h}_{u}^{(l)}$, $\mathbf{z}_{i^{+}}^{\top}=\sum_{l=0}^{L}\mathbf{h}_{i^{+}}^{(l)}$, and $\mathbf{z}_{i^{-}}^{\top}=\sum_{l=0}^{L}\mathbf{h}_{i^{-}}^{(l)}$, respectively. We define the pair-wise ranking function as $g(u,i^{+},i^{-})=\Braket{\mathbf{\tilde{z}}_{u}^{(k)},\mathbf{{z}}_{i^{+}}}-\Braket{\mathbf{\tilde{z}}_{u}^{(k)},\mathbf{{z}}_{i^{-}}}$, where $\mathbf{\tilde{z}}_{u}^{(k)}=\sum_{l=0,l\neq k}^{L}\mathbf{h}_{u}^{(l)}+(\mathbf{h}_{u}^{(k)}+\bm{\Delta})$ is the user embedding after incorporating perturbation $\bm{\Delta}\in\mathbb{R}^{d}$ to $k$-th layer graph representation $\mathbf{h}_{u}^{(k)}$, and $<\cdot,\cdot>$ means inner product. Then, the learning objective of perturbation $\bm{\Delta}$ is designed as follows:

where $\|\cdot\|_{p}$ means the vector’s p-norm. Here, pair-wise ranking function $g(\cdot)$ is linearized around the $k$-th representation $\mathbf{h}_{u}^{(k)}$, thus the maximum perturbation $\bm{\Delta}_{u}^{(k)}$ is exactly corresponding to the orthogonal projection of $\mathbf{h}_{u}^{(k)}$ onto the model decision hyperplane.

For the sake of simplicity and better interpretation, we denote that $f(\mathbf{h}_{u}^{(k)})=\partial{g(u,i^{+},i^{-})}/\partial{\mathbf{h}_{u}^{(k)}}$. The maximum perturbation $\mathbf{\Delta}_{u}^{(k)}$ is equivalent to solving for the directional vector from $\mathbf{h}_{u}^{(k)}$ to the decision boundary, which is formally given as follows:

where $\text{sign}(\cdot)$ is the sign function, and $\odot$ denotes element-wise product. The value of $q$ depends on the choice of perturbation norm $\ell_{p}$ ($1\leq p\leq\infty$), and satisfies that $\frac{1}{p}+\frac{1}{q}=1$ by following Holder’s Inequality’s constraint (Moosavi-Dezfooli et al., 2016). In our work, $p$ is set as $\infty$ and $q$ is set as 1, as we empirically found that perturbation constraints under the $\ell_{\infty}$ norm have better model performance.

Following that, since users often interact with multiple items in real-world recommendation scenarios, we extend the above method to all interactions of user $u$ for deriving the final optimal perturbation constraint, which can be rewritten as follows:

Note that we only focus on perturbing the high-order graph representations for users and items, while skipping the beginning features, i.e., $1\leq k\leq L$. This is because the original features contain the most abundant semantic information, and polluting these features could lead to a severe performance decrease. On the other hand, by perturbing higher-order representations, we subtly and implicitly disrupt the potential semantic and structural characteristics. Intuitively, it can efficaciously simulates the noise encountered in real-world scenarios, thereby further enhancing the model robustness. Similarly, we can obtain the graph perturbations of item nodes from a dual perspective.

As highlighted in Sec. 1, existing GCL-based recommenders struggle to achieve a harmonious balance between contrastive hardness and rationality, both of which are pivotal to acquire high-quality user (item) representations. To this end, in this subsection, we meticulously design the relation-aware adversarial-contrastive objective, which utilizes the global relationships among user-user and item-item to create more challenging positive and hard negative pairs under perturbation constraints. Finally, we optimize the representations through multi-view contrastive learning.

However, excessive pursuit of representation uniformity in GCL may lead to reduced distances between data points and the decision boundary, potentially compromising the model robustness. We attribute such dilemma is caused by the inherent deficiency that the GCL’s essence is unsupervised learning paradigm, which pushes all different instances apart while ignoring task-specific semantic relations (Wang and Liu, 2021). To tackle the above issue, we propose to use adversarial examples for achieving margin maximization. Specifically, we generate adversarial examples using the maximum adverasrial perturbation defined in Eq. (6), which can be formulated as follows:

We then utilize the generated adversarial examples to optimize the BPR objective (i.e., Eq. (1)), which is given as follows:

By explicitly creating adversarial examples around the model’s decision boundary, the model optimized with both original and adversarial data can more effectively boost the confidence of input data, thereby enhancing the model’s overall robustness.

The core motivation of this paper is to construct semantic preserving and hardness enhancing view-generator for contrastive learning. For the former, we capitalize on the decision boundary-aware constraint to help build rationality-aware views. For the latter, we carefully construct more challenging contrastive paired data because their hardness significantly affects the optimization process of model parameters.

To further explain, we give a proof that contrastive loss is essentially hardness-aware learning mechanism. Specifically, taking the side of users as an example, given a set of users $\mathcal{U}=\{u_{1},u_{2},\dots,u_{M}\}$, we denote the similarity of user $u_{i}$ under different augmented views (e.g., random-augmented view or adversarial-contrastive view) as $s_{i,i}$, and the similarity between user $u_{i}$ and $u_{j}$ as $s_{i,j}$. The probability of $u_{i}$ being identified as $u_{j}$ is formulated as:

Thus, the objective of contrastive learning is rewritten as follows:

Then, the expression of updating model parameters $\bm{\theta}$ is

where we give the derivation results for $\frac{\partial\varphi(u_{i})}{\partial{s_{i,i}}}$ and $\frac{\partial\varphi(u_{i})}{\partial{s_{i,j}}}$:

where we can observe that the gradients of the contrastive loss w.r.t. both positive and negative pairs are proportional to the corresponding exponential form of their similarity scores. This means that smaller positive pair similarity $s_{i,i}$ and larger negative pair similarity $s_{i,j}$ will have a greater impact on the model parameter optimization. Therefore, our proposed RGCL can learn the high-quality representations by constructing the challenging positive pairs and hard negative pairs, which fits to guide model optimization through hardness-aware contrastive learning.