From Intrinsic to Counterfactual: On the Explainability of Contextualized Recommender Systems

Recommender systems point the online users to a certain list of items with potential interests, thereby increasing the cross-sales and customer engagement. Many efforts have been devoted to this direction with the booming growth of e-commerce and online streaming services. Recent years have seen significant improvement in performance with the emergence of more sophisticated models. However, only showing the recommended ranking list of items can hardly gain the trust and satisfaction of users. The reason is that many modern recommendation models are difficult to support the end-users in the decision-making processes by providing useful and meaningful explanations (Tintarev and Masthoff, 2015; Gedikli et al., 2014; Zhang and Chen, 2020). Traditional explainable models such as collaborative filtering (Herlocker et al., 2000; Sarwar et al., 2001), factorization-based model (Zhang et al., 2014; Wang et al., 2018b; Pan et al., 2020a), and tree-based models (Wang et al., 2018a) can leverage the explanations from user/item features (Polleti and Cozman, 2019), opinions (McAuley and Leskovec, 2013), and summarized topics (Wu and Ester, 2015). However, they usually have limited performance and lack the ability to extend to complicated scenarios. Recently, various types of neural network based explainable recommendation models have been proposed. A2CF (Chen et al., 2020) is an attribute-aware model that uses the residual network to model user-item reviews with joint sentiment analysis; SAERS (Hou et al., 2019) leveraged the power of gradient-weighted class activation mapping (Selvaraju et al., 2020) and used the backpropagated ranking loss to generate item image’s saliency map as visual explanations; AMCF (Pan et al., 2020b) maps the uninterpretable features into the interpretable aspect features by minimizing both ranking loss and interpretation loss, etc. Various kinds of explanation models are emerging every year, e.g., set operation based model (Balog et al., 2019), knowledge graph based explanation model (Ai et al., 2018), disentangled embedding explanation (Ma et al., 2019), and many others.

These existing approaches generate explanations in different ways but lack the systemic analysis regarding explainable recommendation models from multiple perspectives. Based on the classification of explainable models  (Zhang and Chen, 2020), they can be either information source dependent (displaying explainable content that is human-readable as the justification of the ranking model’s decision) or model architecture-dependent (deliberately designing the machine learning model with the ability to explain). In our work, we have considered both directions and studied three types of models that are coherently connected as well as complying with this classification taxonomy.

Attention models are one type of input processing technique that allows the neural network to focus on certain parts of a complex input by assigning various modular learned weights. In this way, it allows the neural networks to approximate the visual attention mechanism humans use and the output can be calculated as an attention-weighted sum of the input. It has gained a lot of successes in multiple domains especially natural language processing and computer vision. In recent years, various works (Chen et al., 2018; Wu et al., 2019; Li et al., 2020, 2021; Gao et al., 2019) have been proposed to use attention in recommender systems. Intuitively, different users may interact with the same item with attention to different aspects, e.g., words in the text, saliency areas in images, etc. NPA (Wu et al., 2019) and NRPA (Liu et al., 2019) use two-level attention models of both word-attention and document-attention to personalize news recommendation; NETE (Li et al., 2020) learned template-controlled sentences from data to describe word features; VECF (Chen et al., 2019) utilized both attention-weighted visual features and GRU-based weak supervised learning to increase the model explanation on personalized ranking. In this work, we adopt a similar strategy where the attention weights are learned with ID embeddings and feature embeddings. However, for the sake of explanation, the attentions are applied to features instead of their embeddings.

Despite the successes of recommendation models, numerous works have reported that they can be vulnerable when exposed to adversarial perturbations or even random perturbations (He et al., 2018; Deldjoo et al., 2021). Thus, two types of adversarial recommendation models have been proposed: adversarial learning (Goodfellow et al., 2015; Madry et al., 2018) based RS and GAN-based RS. Among the first type: APR (He et al., 2018) proposed to add adversarial perturbations on user/item embeddings to improve BPR(Rendle et al., 2009) performance; AMR (Tang et al., 2020) leverages the adversarial training by adding adversarial perturbations to target image features. For the second type: IRGAN (Wang et al., 2017), for the first time, formulated the IR tasks as a minimax game where the generator learns the discrete relevance distribution of entities and the discriminator is the ranking model that decides whether the user-item pair is relevant or not; CFGAN (Chae et al., 2018) relaxed the discrete constraints of IRGAN by introducing a generator with the ability to perform continuous embedding space synthesizing; PURE (Zhou et al., 2021) further enriched the CFGAN by performing a positive-unlabeled sampling strategy and reached state-of-the-art performance. Nevertheless, none of them have the ability to provide explanations for the decisions made by the ranking model. Our work bridges the gap by introducing an explanation to the ranking model via adversarial training.

Counterfactual explanation is a specific class of explanation model that provides the reasoning between decision making and input modifications. In recommendations, counterfactual explanations provide perturbations to user input features and these modifications can help the ranking model’s transition to the desirable decision boundary, e.g., flipping the original relevance preference over pair of items. Multiple efforts have been made to categorize and evaluate the counterfactual explanation models (Verma et al., 2020a; Wachter et al., 2017) and they have found successful applications in multiple domains, including natural language processing, computer vision (Goyal et al., 2019), and recommender systems (Tan et al., 2021; Wang et al., 2021; Tran et al., 2021; Ghazimatin et al., 2020). This is an emerging domain in recent years, and our work extends the existing work with multi-pass counterfactual augmentations.

In this paper, we study the contextual recommendation problem where $\mathcal{U}$, $\mathcal{I}$, and $\mathcal{F}$ are the set of users, items, and features, respectively. Given a user $u\in\mathcal{U}$, a list of relevant items and their corresponding interactions are recorded. For each user-item pair $(u,i)$, there exists a group of raw features that can be used as the source of explanation to describe the interaction between the user $u$ and the item $i$. A good source of features is the review text from the user for the specific items. Compared with many existing works that only utilized the isolated features from either the user (e.g., biographical information) or the item (e.g., item title description or item image), the text review features have the explicit interactions between user and item, thus, it is more informative in terms of building an explainable ranking model. We further assume $\Omega$ to be the index set of observed user-item-feature interactions $(u,i,\bm{f}_{u,i})\in\Omega$ where we denote $\bm{f}_{u,i}$ as the observed review features within one specific application domain (e.g., electronics, fashion). Here, the corresponding user id, item id, as well as their review features should satisfy $u\in\mathcal{U},i\in\mathcal{I}$ and $\bm{f}_{u,i}\subset\mathcal{F}$.

Then, our recommendation problem is formulated as follows:

It should be noted that under this problems setting, the features being utilized for users are not isolated from these ones from items, this is different from the traditional contextualized recommender systems (Guo et al., [n.d.]; He and McAuley, [n.d.]; Rendle, 2010). Similar to (Zhang et al., 2014; Tan et al., 2021), we decompose the review features $\bm{f}_{u,i}\in\mathbb{R}^{p}$ (where $p<P$) into user feature vector $\bm{f}_{u}\in\mathbb{R}^{p}$ and item feature vector $\bm{f}_{i}\in\mathbb{R}^{p}$ w.r.t. the $k$-th review feature $w_{k}$ as shown in Equation (1). Intuitively, the user tends to comment on the features they care more about, thus, the user feature is merely a frequency-based vector. However, for items, their qualities can be accurately reflected in the review sentiments from multiple users. Therefore, we formulate the item features by considering both review feature frequency and average sentiment.

where $T$ is the maximum review scores a user can give to an item, $\bm{f}_{u,i}^{k}$ is the frequency that user $u$ mentioned review feature $k$ toward item $i$, and $s_{i}^{k}$ is the average sentiment on review feature $k$ toward item $i$. After transformation, all entries of these feature vectors are mapped into the range of $[1,T]$ and the the indexed set becomes $\Omega=\{(u,i,\bm{f}_{u},\bm{f}_{i})\}$.

In this subsection, we introduce the neural attention recommendation (NAR) in detail. NAR is a white box recommendation model with intrinsically explainable designs and it has two major components: a user (item) network to learn user (item) representations, and a final prediction network to predict the relevance scores. The overview of the NRA model is shown in Figure 1.

In this section, we introduce the contextualized adversarial ranking model (CAR) in detail. CAR is a gray box recommendation model with a partial transparency requirement on the feature gradients and it allows the ranking model to be retrained with adversarial augmentations. Specifically, CAR has the same user (item) embedding network as NAR. In addition, CAR also has two distinct components: an adversarial augmentation network, and the final prediction network. The overview of the NRA model is shown in Figure 2.

In this section, we introduce the counterfactual neural recommendation model (CNR) in detail. CNR is a blackbox recommendation model with model-agnostic designs to generate explanations for rankings. Similar to the CAR model, CNR has exactly the same user (item) embedding network and final prediction network. Nevertheless, its augmentation network is designed by following the counterfactual thinking (Tan et al., 2021; Verma et al., 2020b): ‘‘User X purchased item A over item B because feature variables have values ($f_{1},f_{2},...$) associated with X. If user X has a few feature variables ($f_{1}^{\prime},f_{2}^{\prime},...$) changed and all other variables remained the same, user X would purchase item B over item A instead’’.

One should note that many such explanations exist, our objective is to find a counterfactual perturbation that alters features as little as possible but returns the closest world where the ranking decision has been flipped. Based on the above principle, we can find the counterfactual perturbation $\bm{\delta}_{cf}$ and an optimum solution $\theta$ for the ranking model by solving the following minimax problem:

where $\bm{\delta}_{cf}$ is the counterfactual perturbation of user features by fixing the model parameter $\theta$ and minimizing this objective. $d(\cdot,\cdot)$ is the distance function that measures the difference between original user features and perturbed user features. Meanwhile, the ranking loss should be maximized since the ranking decision is designed to be flipped after the feature perturbations are applied. In practice, we should initialize $\theta$ by training the ranking model without any perturbations, and then we can iteratively update $\bm{\delta}_{cf}$ and $\theta$ until a perturbation has been found that is sufficiently small to flip the ranking predictions. For example, under the pairwise learning setting, $R_{u,i}>R_{u,j}$ will become $R_{u,i}<R_{u,j}$ if the counterfactual perturbation is applied on user features $\bm{f}_{u}+\bm{\delta}_{cf}$.

The distance term $\mathrm{dist}(\cdot,\cdot)$ is critical in counterfactual recommendations, especially when we aim to generate human interpretable explanations. Namely, only a small number of features should be changed and the rest remain untouched. In this way, these counterfactual perturbations are easier for humans to interpret. Thus, various distance functions can be defined to support our goal. One straightforward definition is L2 distance $\mathrm{dist}=||\bm{\delta}_{cf}||_{2}^{2}$, which can be used to guarantee the user features are minimally perturbed. For the sake of introducing sparsity to $\bm{\delta}_{cf}$ so that humans can understand such explanation, we also introduce the L1 norm $\mathrm{dist}=||\bm{\delta}_{cf}||_{2}^{2}+||\bm{\delta}_{cf}||_{1}$. This is also called elastic net regularization, which combines ridge and lasso regularization.

For the recommendation evaluation, the comparison results are summarized in Tables 4-8. We observe that regardless of the data set size or the application domain, the adversarial perturbation and augmentation based method CAR always outperforms all other baselines. Meanwhile, we also observe that the counterfactual-based models, such as CER and CNR, also have competitive results. Specifically, CNR consistently beats CER in every aspect. This is because CNR is the minimax version of CER and it has multiple rounds of optimizations to find the best counterfactual perturbations and keeps improving the ranking decision boundaries constantly. As a direct comparison, CAR also generates perturbations and retrains the ranking model on top. Nevertheless, the objective of CAR is to directly optimize the ranking performance. With more near-boundary perturbation examples generated, the embedded feature space will be further covered for the scenarios where the data set is extremely sparse. Similar phenomena have been observed in multiple recent works that use adversarial augmentation (He et al., 2018; Zhou et al., 2021; Chae et al., 2018). For the attention-based model NAR, due to its high model complexity, it does not perform well for small-scaled data sets because of model overfitting. But its performance gets significantly improved on large-scale data sets and becomes very competitive on them. The VBPR and NCF models in general are quite stable in terms of their performance, but due to the lack of a module for explanation, their prediction results are hardly interpretable.

In terms of the user-orientated evaluation w.r.t. the explanations, we mainly focus on horizontally comparing all three proposed models. As shown in Figure 3, we observe that neural attention model NAR and counterfactual model CNR perform relatively well, but the adversarial perturbation model CAR has low scores across all five data sets. The reason is that NAR and CNR are more suitable for model explanations: NAR has an explicitly designed architecture for learning the attention weights of the input features; CNR uses the counterfactual perturbation to alter the ranking decision, therefore, these perturbations are also explainable in terms of making a ranking decision. As a sharp contrast, CAR learns its adversarial perturbations to maximize the ranking loss and they have weak correlations with explanation because these perturbations do not guarantee to change the model ranking.

In this section, we compare and connect all three explanation models, NAR, CAR, and CNR, by conducting various visualizations. These visualization show that quantitative evaluations and qualitative visualization are well aligned. Since the users are inclined to give reviews to items with positive or negative sentiment, then, the word-sentiment pairs of each user are one reliable proxy of ground truth (GT) for explanation evaluations.