# explainable_recommendation_through_attentive_multiview_learning__bcdf8e60.pdf

The Thirty-Third AAAI Conference on Artificial Intelligence (AAAI-19)

Explainable Recommendation through Attentive Multi-View Learning

Jingyue Gao,1,2 Xiting Wang,2* Yasha Wang,1 Xing Xie2

1Peking University, {gaojingyue1997, wangyasha}@pku.edu.cn 2Microsoft Research Asia, {xitwan, xingx}@microsoft.com

Recommender systems have been playing an increasingly important role in our daily life due to the explosive growth of information. Accuracy and explainability are two core aspects when we evaluate a recommendation model and have become one of the fundamental trade-offs in machine learning. In this paper, we propose to alleviate the trade-off between accuracy and explainability by developing an explainable deep model that combines the advantages of deep learning-based models and existing explainable methods. The basic idea is to build an initial network based on an explainable deep hierarchy (e.g., Microsoft Concept Graph) and improve the model accuracy by optimizing key variables in the hierarchy (e.g., node importance and relevance). To ensure accurate rating prediction, we propose an attentive multi-view learning framework. The framework enables us to handle sparse and noisy data by co-regularizing among different feature levels and combining predictions attentively. To mine readable explanations from the hierarchy, we formulate personalized explanation generation as a constrained tree node selection problem and propose a dynamic programming algorithm to solve it. Experimental results show that our model outperforms state-of-the-art methods in terms of both accuracy and explainability.

Introduction

Personalized recommendation has become one of the most effective techniques to help users sift through massive amounts of web content and overcome information overload. Recently, the recommendation community has reached a consensus that accuracy can only be used to partially evaluate a system. Explainability of the model, which is the ability to provide explanations for why an item is recommended, is considered equally important (Wang et al. 2018a). Appropriate explanations may persuade users to try the item, increase users trust, and help users make better decisions (Zhang et al. 2014a). Determining whether to optimize towards accuracy or explainability poses a fundamental dilemma for practitioners. Currently, their choices are limited to two types of models:

Deep but unexplainable. Deep learning-based recommendation models (Wang et al. 2017; Zheng, Noroozi,

*Xiting Wang is the corresponding author Copyright 2019, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved.

Bread Fruit

Shrimp Gelato

High Fiber Food

Pear Arugula

Figure 1: Multi-level user interest extracted using our method. The hierarchies correspond to a 26-year-old female Yelp user (left) and a 30-year-old male Yelp user (right). The features users care most about are highlighted in orange.

and Yu 2017) achieve state-of-the-art accuracy due to their capacity to model complex high-level features. However, the high-dimensional high-level features they learn are beyond the comprehension of ordinary users. Explainable but shallow. Although many explainable recommendation methods have been proposed, the explainable components in the models are usually shallow. Typical explainable components include one or two layers of attention networks (Chen et al. 2017), matrix factorization (Zhang et al. 2014a), and topic modeling (Mc Auley and Leskovec 2013). They have achieved considerable success in improving explainability. However, the lack of an effective mechanism to model high-level explicit features limits their accuracy and/or explainability. These methods can only select explanations from a pre-defined level of candidates. They cannot identify which level of features best represents a user s true interest. For example, they are not able to tell whether a user only likes shrimps (lower level) or is interested in seafood (higher level). In this paper, we aim to mitigate the trade-off between accuracy and explainability by developing an explainable deep model for recommendation. The model achieves stateof-the-art accuracy and is highly explainable. Moreover, it enables us to accurately portray hierarchical user interest. As shown in Figure 1, the model can automatically infer multilevel user profiles and identify which level of features best captures a user s true interest, e.g., whether s/he is interested in lower-level features such as shrimp or higher-level features such as seafood.

To design an explainable deep model as such, we are faced with two major technical challenges. The first challenge is to accurately model multi-level explicit features from noisy and sparse data. It is very difficult to model the relationships between high-level and low-level features since they have overlapping semantic meanings. For example, learning whether seafood is important to the user is challenging because the user may only mention shrimp or meat in the review. The second challenge is to generate explanations that are easy for common users to understand from the multilevel structure. To address these challenges, we develop a Deep Explicit Attentive Multi-View Learning Model (DEAML). The basic idea is to build an initial network based on an explainable deep structure (e.g., knowledge graph) and improve accuracy by optimizing key variables in the explainable structure (e.g., node importance and relevance). To improve model accuracy, we propose an attentive multi-view learning framework for rating prediction. In this framework, different levels of features are considered as different views. Adjacent views are connected by using the attention mechanism. Results from different views are co-regularized and attentively combined to make the final prediction. This framework helps improve accuracy since it is robust to noise and enables us to fully leverage the hierarchical information in the explainable structure. Second, we formulate personalized explanation generation as a constrained tree node selection problem. To solve this problem, we propose a dynamic programming algorithm, which finds the optimal features for explanation in a bottom-up manner. We conduct two experiments and a user study to evaluate our method. Numerical experiments show that DEAML outperforms state-of-art deep learning-based recommendation models in terms of accuracy. An evaluation with 20 human subjects has shown that the explanations generated by us are considered significantly more useful than state-of-theart aspect-based explainable recommendation method.

Related Work

Many methods have been proposed to improve recommendation accuracy, including content-based (Kompan and Bielikov a 2010), collaborative filtering-based (Das et al. 2007) and hybrid methods (De Francisci Morales, Gionis, and Lucchese 2012). As an integration and extension of previous methods, deep learning-based models are proposed to further improve accuracy. For example, CDL (Wang, Wang, and Yeung 2015) jointly performs deep representation leaning and collaborative filtering by employing a hierarchical Bayesian model. He et al. propose a Neural Collaborative Filtering framework to learn nonlinear interactions between users and items (He et al. 2017). In Deep Co NN and NARRE (Zheng, Noroozi, and Yu 2017; Chen et al. 2018), convolutional neural networks are leveraged to extract features from textual reviews. Although these methods achieve significant improvement in accuracy, the high-level features extracted are beyond the understanding of common users. Except for unstructured textual data, researchers have also designed models that leverage structural

information. Famous examples are taxonomy-based methods that incorporate taxonomies of items into latent factor models (Kanagal et al. 2012; Zhang et al. 2014b). These methods can effectively mitigate data sparsity. However, their accuracy and explainability are limited due to the lack of a mechanism to model multi-level item features. Recently, researchers have discovered that providing explanations may improve persuasiveness, effectiveness, efficiency and user trust (Zhang et al. 2014a). Thus, many methods have been developed to improve the explainability of recommendation models (Ren et al. 2017; Peake and Wang 2018; Wang et al. 2018b). Pioneer works (Mc Auley and Leskovec 2013; Zhang et al. 2014a) focus on improving the explainability of collaborative filtering models. These works usually rely on shallow explainable components such as matrix factorization (Zhang et al. 2014a) and generative models (Diao et al. 2014; Wu and Ester 2015) for explanation generation. More recently, researchers discovered the attention mechanism s capability in improving the explainability of deep learning-based methods. By using the attention mechanism, the models can automatically learn the importance of explicit features and at the same time refine user and/or item embeddings (Chen et al. 2017; 2018). Researchers have also built explainable models based on interpretable structures such as graphs (He et al. 2015) and trees (Wang et al. 2018a). The aforementioned explainable recommendation methods have achieved considerable success in improving explainability. However, to the best of our knowledge, none of the existing explainable recommendation methods can model multi-level explicit (explainable) features. As a result, the accuracy and/or explainability of these methods are limited. Compared with these methods, our DEAML is an explainable deep model that achieves state-of-the-art accuracy and meanwhile is highly explainable. Moreover, our model can automatically learn multi-level user profile and infer which level of features best captures a user s interest.

Problem Definition We aim to build an explainable deep recommendation model by incorporating an explicit (explainable) feature hierarchy. Input: The input of our model includes a user set U, an item set V , and an explicit feature hierarchy Υ.

Definition 1. An explicit feature hierarchy Υ is a tree where each node Fl is an explicit feature or aspect (e.g., meat) of the items (e.g., restaurants). Each edge is a tuple (Fl1, Fl2), which represents that the child Fl1 (e.g., beef) is a sub-concept of the parent Fl2 (e.g., meat). We denote the set of nodes in Υ as F, where F = {F1, ..., FL} and L represents the total number of nodes in Υ.

To build Υ, we leverage the Microsoft Concept Graph1 (Wu et al. 2012; Wang et al. 2015), which is a widely used knowledge graph with over 5 million concepts and 85 million Is A relations (e.g., cat Is A animal). We first map n-grams in the reviews to concepts (or instances) in the concept graph. Only the frequently used concepts that are highly

1https://concept.research.microsoft.com/

Hierarchical Propagation of

User-Feature Interest

EFM++ at View 1

EFM++ at View H

EFM++ at View 2

User Embedding

Item Embedding

Co-Regularization

Co-Regularization

Attention Layer

Rating Prediction

Hierarchical Propagation of

Item-Feature Quality

Attentive Multi-View Learning

Attention Layer

Attention Layer

Figure 2: Architecture of our proposed DEAML model for rating prediction.

correlated with the ratings are kept and regarded as the explicit features. We then recursively search the explicit features in the concept graph and utilize the Is A relationships to build Υ. On average, the depth of Υ is 5. Next, we define the hierarchical level of each feature.

Definition 2. The hierarchical level hl of Fl is recursively defined based on its children: hl = max Fc children(FL) hc+ 1. If Fl is a leaf, its hierarchical level is set to 1. Let H denote the largest hierarchical level of internal nodes in Υ.

Besides Υ, two other types of input are the user set U and the item set V . Each user in U is represented as a tuple (i, xi), where i is the user id and xi RL is the observed user-feature interest vector. Here the l-th element in xi measures how much user i cares about feature Fl. Similarly, each item in V is represented as (j, yj) and the observed item-feature quality vector yj RL measures how well j performs on these features. We adopt the method proposed by (Zhang et al. 2014a) to calculate xi and yj. The basic idea is to compute xi and yj based on the mentioned times and the sentiments of the features in the reviews. Output: Given a user i and an item j, the output of our model includes the predicted rating rij and a personalized feature-level explanation E, where E = {Fl1, ..., Fl T } is a subset of F. Similar to (Zhang et al. 2014a), the final explanations presented to users are in the template as follows.

You might be interested in [features in E], on which this item performs well.

Approach Overview Given an explicit feature hierarchy Υ, we build a Deep Explicit Attentive Multi-View Learning Model (DEAML) that is able to 1) accurately predict ratings and 2) produce useful explanations. Our method consists of two steps. Step 1: Prediction with Attentive Multi-View Learning. In this step, we build the DEAML model by minimizing the rating prediction error. The architecture of the model is shown in Figure 2. For each user i, we first build a userfeature interest hierarchy (Figure 2A) by using Υ and xi.

While xi measures user interest to some extent, it is inaccurate due to data sparsity, noises in the data, and also overlapping semantics of the features. To improve accuracy, we propose a hierarchical propagation method that propagates user interest along the hierarchy based on user attention. Similarly, for each item j, we build an item-feature quality hierarchy (Figure 2C) by propagating yj along edges of Υ. To effectively predict ratings by using the two hierarchies, we propose an attentive multi-view learning framework. As shown in Figure 2B, features at different hierarchical levels are regarded as different views. Predictions from multiple views are co-regularized to enforce the agreement among them. Considering a user s interest is not equally distributed over different hierarchical levels, we calculate the final prediction by attentively integrating results from multiple views. Hierarchical propagation and attentive multi-view learning are jointly optimized, which allows us to simultaneously infer user profiles and predict ratings in an end-to-end manner. Step 2: Personalized explanation generation. In this step, we generate the explanations by using the DEAML model. To this end, we first define a utility function for candidate features, which jointly considers user-feature interest, item-feature quality and attention scores of different views. Then we formulate explanation generation as a constrained tree node selection problem and propose a dynamic programming algorithm to solve it efficiently.

Prediction with Attentive Multi-View Learning

In this section, we introduce how to predict the rating given xi and yj. We first describe the hierarchical propagation method used to build an accurate user-feature interest hierarchy. Then we propose the attentive multi-view learning framework that leverages complementary information in the feature hierarchy to improve rating prediction accuracy.

Hierarchical Propagation

A straightforward way to build the user-feature interest hierarchy is to map elements in xi to Υ. Specifically, for each

Attention Layer

1 ~i x c ~i x il x

User Embedding

Feature Embedding

Children Parent

Figure 3: Hierarchical propagation of user interest.

feature Fl in Υ, we can assume that user i s interest in Fl is equal to xil, which is the lth element in xi. Since xil is calculated by counting the number of times Fl appears in the reviews written by user i, this method captures user interest to some extent. However, due to data sparsity, noises in the data, and overlapping semantics of features in Υ, the constructed user-feature interest hierarchy can be inaccurate. For instance, a user interested in meat may comment frequently on beef and pork but rarely mentions the word meat. To solve this issue, we infer users true interest xi by propagating xi along Υ. For each parent Fl, we calculate xil by considering personalized user attention (Figure 3):

xil = αllxil + P

Fc children(Fl) αlc xic,

αlc = exp(α lc) P

Fc children(Fl) {Fl} exp(α lc ), (1)

α lc = h 1 Re LU(Wlel + Wcec + Wuui + b1) + b2,

where αlc is the attention score of feature Fc. el, ec Rd1 are pre-trained embeddings of Fl and Fc. We follow the method proposed in (Choi et al. 2017) to train feature embeddings that capture both semantic and hierarchical information. As there are H levels of features except FL in Υ, the user embedding at level h is uh i Rk and the global user embedding ui Rk H is the concatenation of uh i at all levels (shown in Figure 2). The hidden layer size of the attention network is d2. Wl Rd2 d1, Wc Rd2 d1, Wu Rd2 k H, h1 Rd2, b1 Rd2, b2 R and ui Rk H are model parameters to be learned. Propagation of yj is similar to that of xi. After that we obtain a more accurate item-feature quality vector yj.

Attentive Multi-View Learning After we obtain accurate estimates of user-feature interest and item-feature quality, we predicate how much user i likes item j based on xi and yj. To accurately predict the rating, we design an attentive multi-view learning framework. In this section, we first introduce how to predict the ratings based on the information in a single view. Then, we describe the co-regularization loss that enforces agreement among different views. Finally, we illustrate how to combine different views in a unified model and learn parameters from different views jointly. Loss in each view. In our framework, each view consists of features at the same hierarchical level. We first consider rating prediction in a single view. Suppose there are Lh features at level h. By only considering users interest on

these features, we obtain a local user-feature interest vector xh i RLh, which is a slice of xi. Similarly, we can obtain yh j RLh, which is a slice of yj. At level h, we employ an extended version of EFM (Zhang et al. 2014a) to predict the rating based on xh i and yh j . Compared with collaborative filtering, EFM enriches the user and item representations by adding an additional set of latent factors learned from the explicit features. Specifically, the user embedding for view h consists of two parts: uh i = ph i ch i Rk. Here ph i Rk1 are latent factors learned from explicit features (explicit factors), ch i Rk2 are implicit factors, and is the concatenation operator. Similarly, the item embedding at level h is vh j = qh j dh j , where qh j Rk1 denote the explicit factors and dh j Rk2

are the implicit factors. The explicit factors ph i and qh j are used to fit xh i with Zhph i and fit yh j with Zhqh j , where Zh RLh k1 is the projection matrix of features at level h. Traditional EFM directly predicts the rating of user i on item j as uh i vh j . We further extend it by incorporating global average rating bias µ, user bias oi and item bias oj:

rh ij = uh i vh j + oi + oj + µ. (2)

Let Ωdenote the set of training instances and rij denote the ground-truth rating of user i on item j, the loss in view h is:

Lh = λa P i,j Ω( rh ij rij)2 +

i xh i Zhph i 2 + λy P

j yh j Zhqh j 2, (3)

where λa, λx and λy are weights of corresponding items. Co-regularization loss. We then consider how the models in multiple views can complement each other. Based on the consensus principle which aims to maximize the agreement on multiple views (Xu, Tao, and Xu 2013), a common paradigm of multi-viewing learning is co-regularization. Actually, information distributed in different views describes the inherent characteristics of user-feature interest and itemfeature quality from various aspects. We can regularize the learning of multiple views by enforcing agreement among their predictions. This way, complementary information can be leveraged to benefit the learning of each single view. By enforcing the agreement between two adjacent views, the co-regularization loss is:

h=1 ( rh ij rh+1 ij )2. (4)

Joint learning. Predictions from multiple views should be combined for the final prediction. One method is to average the predictions of all views. However, this assumes that each view contributes equally to the rating. In reality, different views reveal user-feature interest and item-feature quality at different levels of abstractness. The levels of interest also vary among users and items. Some users may care more about general features (e.g. food and service) while others may be more attracted by a specific feature (e.g. beef). This is the same for items. The views are not equally useful and should be assigned different weights. The weights are supposed to be determined in a user-item-specific way to better consider users interest and items quality. Thus we use

the attention mechanism to combine multiple views. Let Ih denote the H-dimensional one-hot vector for view h. The attention network for view combination is:

w h = h 2 Re LU(W1ui + W2vj + W3Ih + b3) + b4,

wh = exp(w h) PH h =1 exp(w h ). (5)

where ui, vj Rk H are the global user embedding and item embedding used in hierarchical propagation. W1 Rd3 k H, W2 Rd3 k H, W3 Rd3 H, h2 Rd3, b3 Rd3, b4 R are model parameters and d3 is the hidden layer size of this attention network. The final prediction is a weighted sum of predictions in each view, where wh is the weight of view h:

h=1 wh rh ij. (6)

Let Θ 2 denote the L2 norm of all parameters in our model. We optimize the following objective function:

i,j Ω ( rij rij)2 +

h=1 Lh + λv Lv + λr Θ 2, (7)

where λv, λr are weights of corresponding items. We use the Adam optimizer (Kingma and Ba 2014) because it can automatically adjust the learning rate during the training phase.

Personalized Explanation Generation The goal of personalized explanation generation is to select T features from Υ that are most useful in helping user i decide whether s/he will try item j. Selecting features distributed in multiple hierarchical levels is challenging because 1) users interest is not evenly distributed over different hierarchical levels and 2) we need to avoid selecting features with overlapping semantics. In this section, we first introduce our utility function, which estimates the usefulness of each feature. Then, we formulate the feature selection as a constrained tree node selection problem and propose a dynamic programming method to solve it efficiently. Utility function. Given user i and item j, three factors are considered when judging whether a feature Fl is useful:

Whether user i is interested in Fl. How well item j performs on Fl. The weight of the view that Fl belongs to. In view h, the explicit factors ph i and qh j are used to fit the user-interest vector xh i with Zhph i and fit yh j with Zhqh j . Let Fl be the ˆl-th feature at level h. We define the utility score of Fl in the recommendation instance (i, j) as:

Ψ(Fl) = Zhph i ˆl Zhqh j

where wh is the user-item-specific attention weight for view h, which is calculated by using Equation (5). The utility score for the root feature is set to 0 since it does not have a concrete meaning. This utility function jointly incorporates three factors mentioned above. Constrained tree node selection. Next, we select T features to be included in the explanation. The goal is to maximize

the total utility score of the selected features. We further require that features (e.g. beef) cannot be selected simultaneously with their ancestors in Υ (e.g. meat) since they are semantically overlapping. Mathematically, we formulate personalized explanation generation as a constrained tree node selection problem:

argmax φ1,...,φL

PL l=1 (φlΨ(Fl)) ,

s.t. PL l=1 φl = T, φl {0, 1}, (9) φl1φl2Anc(l1, l2) = 0, hl1 < hl2, where φl = 1 (φl = 0) means feature Fl is (not) selected and Anc(l1, l2) returns 1 if Fl2 is an ancestor of Fl1. We solve the above problem by using dynamic programming. Let G(l, t) denote the maximum utility we can obtain by selecting t nodes in the subtree rooted at Fl. Let us use IDls to represent the id of the s-th child feature of Fl, employ J(l, s, t) to denote the maximum utility we can obtain by selecting t nodes from the first s children of Fl, and use S to denote the total number of children of Fl. We have the following transition equations:

G(l, t) = max (Ψ(Fl), J(l, S, t)) t = 1 J(l, S, t) t > 1 (10)

J(l, s, t) = max 0 t t (J(l, s 1, t ) + G(IDls, t t )) .

Here, we need to carefully consider whether t = 1 because if we select more than one node (i.e., t >1) from the subtree rooted at Fl, Fl itself can not be selected. Otherwise, the constraint will be violated. Supposing that the root of Υ is FL, we can obtain the final solution with the maximum utility G(L,T) by using Equation (10) iteratively. Time complexity. If the maximum number of children of a feature is M, the time complexity of our dynamic programming algorithm is O(LMT 2). It is more efficient than the brute force algorithm whose time complexity is O(LT ).

Experiments We conduct two experiments and a user study to evaluate the effectiveness of our method. First, we demonstrate that our method performs better than the state-of-the-art methods in terms of accuracy. Then, we conduct parameter sensitivity analysis to validate the robustness of our model. Finally, we evaluate the explainability of DEAML with 20 human subjects in a user study.

Table 1: Statistics of three public datasets.

Dataset #Users #Items #Reviews

Toys and Games 19,412 11,924 167,597 Digital Music 5,541 3,568 64,706 Yelp 8,744 14,082 212,922

Experimental Settings Datasets. We use three datasets from different domains for evaluation. Table 1 summarizes the statistics of the datasets.

Toys and Games is the part of the Amazon dataset2 that

2http://jmcauley.ucsd.edu/data/amazon

Table 2: RMSE comparison with baselines on three datasets. Best results are highlighted in bold.

G1 G2 G3 Ours NMF PMF SVD++ CKE HFT EFM Deep Co NN NARRE DEAML-V DEAML

Toys and Games 1.1489 1.1832 0.9071 0.9923 0.9958 0.9534 0.9199 0.9084 0.9062 0.9040 Digital Music 1.1520 1.2619 0.9211 0.9849 1.0910 0.9696 0.9212 0.9209 0.9190 0.9118 Yelp 1.2678 1.2413 1.1561 1.2279 1.2738 1.2019 1.1503 1.1348 1.1343 1.1333

focuses on Toys and Games. We choose the 5-core version where all users and items have at least 5 reviews. Digital Music is also from the Amazon 5-core dataset. It focuses on the domain of digital music. Yelp consists of restaurant reviews from Yelp Challenge 20183. Since the raw data is very large and sparse, we follow (Zhang et al. 2014a) to preprocess the data. Specifically, we select restaurants located in the Phoenix city and ensure that all users and items have at least 10 ratings. Baselines. We select eight competitive methods for comparison. These methods can be divided into three groups according to the type of data they use.

The first group (G1) only relies on the observed rating matrix for rating prediction. This group consists of three methods: NMF (Lee and Seung 2001), PMF (Mnih and Salakhutdinov 2008) and SVD++ (Koren 2008). The second group (G2) contains a knowledge-based method, CKE (Zhang et al. 2016). Here, we regard the feature hierarchy Υ as the structural knowledge in CKE. The third group (G3) consists of four methods that leverage the textual reviews for rating prediction: HFT (Mc Auley and Leskovec 2013), EFM (Zhang et al. 2014a), Deep Co NN (Zheng, Noroozi, and Yu 2017) and NARRE (Chen et al. 2018). Among them, Deep Co NN and NARRE are deep learning-based. EFM is the state-ofthe-art method for mining feature-level explanations from a single layer of features (no hierarchical structure). Evaluation metric. We adopt the widely-used Root Mean Square Error (RMSE) to evaluate the accuracy of all algorithms on rating prediction. A lower RMSE indicates a better performance. This is calculated as:

i,j ( rij rij)2, (11)

where N is the total number of testing instances. Details. We randomly split the dataset into training (70%), validation (15%) and test (15%) sets. We tune the hyperparameters of each algorithm on the validation set and evaluate their performance on the test set. The number of latent factors k for algorithms is searched in [8,16,32,64,128]. After parameter tuning, we set k = 8 for NMF, PMF and HFT, and k = 16 for SVD++. We set k = 32 for EFM, CKE, Deep Co NN, NARRE and DEAML. We reuse most hyperparameters of Deep Co NN and NARRE reported by their authors. For simplicity, the number of explicit and implicit factors are set equally in EFM and DEAML. When constructing Υ in DEAML, we select the top-2000 frequently mentioned concepts in reviews, sort them by their correlations with the

3https://www.yelp.com/dataset/challenge

ratings and keep only top-500 of them. We set d1, d2, d3, λv, and λa to 20, 10, 10, 10.0, and 3.0, respectively.

Study on Model Accuracy

Overall performance. The accuracy of our approach and the baselines on three datasets are shown in Table 2. After analyzing the results, we made the following conclusions. First, methods that incorporate review information (G3) generally perform better than others. Compared with those only relying on a rating matrix (G1) and the method that incorporates the feature hierarchy (G2), methods in G3 respectively achieve 8.9% and 3.0% improvement in accuracy on average. We ascribe this to the fact that reviews provide more information about users interest and items quality. Second, deep recommendation models (Deep Co NN, NARRE and ours) perform significantly better than those with shallow models (10.7% improvement in accuracy on average). This is because deep models can model high-level features more effectively and better capture non-linear interactions. This again demonstrates the necessity of developing a deep version of the explainable models. Third, our DEAML model achieves the best performance on all three datasets. Our approach performs significantly better than EFM, achieving 5.7%, 5.2% and 6.0% improvements on three datasets. This demonstrates the effectiveness of our attentive multi-view learning framework. While EFM predicts ratings from a single view, our design of attentive multi-view learning allows different views to complement and benefit each other. Compared with CKE that simply learns item embeddings from knowledge bases, our approach fully leverages the feature hierarchy and improves accuracy by 8.0%. The infusion of explicit feature hierarchy provides our approach with additional knowledge that Deep Co NN and NARRE cannot leverage. Given a set of explicit features, DEAML can differentiate low-level features from high-level ones and learn the associations between the features, while Deep Co NN and NARRE treat all features as if they are at the same level. On average, our method is 1.0% more accurate than Deep Co NN and NARRE. Effectiveness of the attention mechanism. In DEAML, we employ the attention mechanism to combine predictions from different views. To validate the effectiveness of the attention mechanism, we design DEAML-V, which is a variant that assigns same weights to all views. From Table 2 we can observe that DEAML constantly performs better than DEAML-V. It verifies our consideration that different views contain features at different levels of abstractness and should be assigned weights according to users interest and items quality. The attention mechanism enables DEAML to combine multiple views more adaptively and perform better.

8 16 32 64 128 Number of latent factors

DEAML SVD++ Deep Co NN NARRE

Figure 4: RMSE on Digital Music with different numbers of latent factors (compared with three competitive baselines).

0 10 20 30 40 50

0 1 2 3 4 5

Figure 5: RMSE on Digital Music with different weights λv of co-regularization (left) and weights λa of errors in each view (right).

Parameter Sensitivity Analysis

In this section, we study the sensitivity of several important parameters in our approach. We only report results on Digital Music here due to the space limitation. Results on other datasets are similar.

Effect of number of latent factors. We first compare our approach with three competitive baselines (SVD++, Deep Co NN and NARRE) at different numbers of latent factors. The results are shown in Figure 4. We can see that DEAML consistently achieves the best performance with varying numbers of latent factors. This demonstrates the robustness of our approach. We further observe that increasing the number of latent factors does not necessarily improve accuracy, because too many latent factors may lead to over fitting.

Effectiveness of co-regularization. We then study how RMSE changes with λv, which is the weight of the coregularization loss term. The left part of Figure 5 shows that our approach achieves the best performance when λv is 10.0. Smaller or larger λv will hurt performance. This demonstrates the importance of involving agreement (coregularization) among different views.

Effectiveness of enforcing accurate prediction in each view. The right part of Figure 5 shows how RMSE changes with increasing λa, which is the weight of rating prediction errors in each single view. The lowest RMSE is achieved when λa = 3.0. When λa is 0, the performance degrades since we fail to train a good predictor in each single view.

Study on Explainability In this section, we evaluate the usefulness of our explanations in helping real-world users make better decisions.With the help of a data annotation company, we recruit 20 participants who have written at least 15 Yelp reviews. The participants are independent and do not know each other. All reviews on the restaurants they mentioned are crawled and merged with other reviews in the Yelp dataset. Two baselines are used for comparison. The first baseline is PAV, which is the famous People Also Viewed explanation. The second baseline is EFM, which is the state-of-the-art method for generating feature-level explanations. We apply our method and baselines on the merged dataset and randomly sample 100 restaurants for testing. We ensure that none of the participants go to these restaurants. For each restaurant, we generate three explanations for each participant using PAV, EFM, and DEAML. The template of EFM is same as DEAML and PAV has it own format: People also viewed this item . T is set to 3 for both EFM and DEAML. The participants are required to give annotations from 1 to 5 on each explanation according to its usefulness in helping them decide whether they will go to the restaurants. We randomly shuffle the order of explanations so that they do not know which explanation comes from which method4.

Table 3: Average score on explanation usefulness. <30 and 30 refer to two age groups.

Male Female <30 30 Overall

PAV 1.35 1.51 1.65 1.11 1.41 EFM 3.18 3.13 3.03 3.32 3.16 DEAML 3.69 3.52 3.58 3.68 3.63

Table 3 shows the average score on the usefulness of explanations generated by the three methods. We observe that PAV performs the worst since it only suggests that one restaurant is popular and is not personalized. Our DEAML outperforms EFM in explanation usefulness (14.9% improvement). This suggests that DEAML can accurately identify a user s interest and a restaurant s quality on multi-level features while EFM suffers from the sparse and noisy data. Additionally, EFM often selects features that are semantically overlapped (e.g., food and cheese) since it ignores the hierarchical structure of features while DEAML considers such constraints to make the generated explanation more informative. We also study the annotations given by participants of different profiles such as gender and age (see Table 3). DEAML consistently performs the best for both male and female participants. It is the same for participants in the two age groups. This suggests that DEAML is robust to various user profiles and is personalized enough to generate useful explanations that help each user make better decisions.

Conclusions In this paper, we propose a Deep Explicit Attentive Multi View Learning Model (DEAML) for explainable rec-

4Explanations generated are at https://www.microsoft.com/enus/research/uploads/prod/2018/10/explanations.csv

ommendation, which combines the advantages of deep learning-based methods and existing explainable methods. The proposed attentive multi-view learning frameworks enables the model to accurately predict ratings and infer multilevel user profiles. To support personalized explanation generation from multi-level features, we formulate the problem as constrained tree node selection and solve it efficiently by using dynamic programming. Experimental results show that our model performs better than state-of-the-art methods in both accuracy and explainability.

Acknowledgments This work is supported by the National Natural Science Foundation of China (Grant No. 61772045).

References Chen, J.; Zhang, H.; He, X.; Nie, L.; Liu, W.; and Chua, T.-S. 2017. Attentive collaborative filtering: Multimedia recommendation with item-and component-level attention. In SIGIR, 335 344. ACM. Chen, C.; Zhang, M.; Liu, Y.; and Ma, S. 2018. Neural attentional rating regression with review-level explanations. In WWW, 1583 1592. International World Wide Web Conferences Steering Committee. Choi, E.; Bahadori, M. T.; Song, L.; Stewart, W. F.; and Sun, J. 2017. Gram: graph-based attention model for healthcare representation learning. In SIGKDD, 787 795. ACM. Das, A. S.; Datar, M.; Garg, A.; and Rajaram, S. 2007. Google news personalization: scalable online collaborative filtering. In WWW, 271 280. ACM. De Francisci Morales, G.; Gionis, A.; and Lucchese, C. 2012. From chatter to headlines: harnessing the real-time web for personalized news recommendation. In WSDM, 153 162. ACM. Diao, Q.; Qiu, M.; Wu, C.-Y.; Smola, A. J.; Jiang, J.; and Wang, C. 2014. Jointly modeling aspects, ratings and sentiments for movie recommendation. In SIGKDD, 193 202. He, X.; Chen, T.; Kan, M.-Y.; and Chen, X. 2015. Trirank: Review-aware explainable recommendation by modeling aspects. In CIKM, 1661 1670. ACM. He, X.; Liao, L.; Zhang, H.; Nie, L.; Hu, X.; and Chua, T.- S. 2017. Neural collaborative filtering. In WWW, 173 182. International World Wide Web Conferences Steering Committee. Kanagal, B.; Ahmed, A.; Pandey, S.; Josifovski, V.; Yuan, J.; and Garcia-Pueyo, L. 2012. Supercharging recommender systems using taxonomies for learning user purchase behavior. Proceedings of the VLDB Endowment 5(10):956 967. Kingma, D. P., and Ba, J. 2014. Adam: A method for stochastic optimization. ar Xiv preprint ar Xiv:1412.6980. Kompan, M., and Bielikov a, M. 2010. Content-based news recommendation. In International conference on electronic commerce and web technologies, 61 72. Springer. Koren, Y. 2008. Factorization meets the neighborhood: a multifaceted collaborative filtering model. In SIGKDD, 426 434. ACM.

Lee, D. D., and Seung, H. S. 2001. Algorithms for nonnegative matrix factorization. In NIPS, 556 562. Mc Auley, J., and Leskovec, J. 2013. Hidden factors and hidden topics: understanding rating dimensions with review text. In Rec Sys, 165 172. ACM. Mnih, A., and Salakhutdinov, R. R. 2008. Probabilistic matrix factorization. In NIPS, 1257 1264. Peake, G., and Wang, J. 2018. Explanation mining: Post hoc interpretability of latent factor models for recommendation systems. In SIGKDD, 2060 2069. ACM. Ren, Z.; Liang, S.; Li, P.; Wang, S.; and de Rijke, M. 2017. Social collaborative viewpoint regression with explainable recommendations. In WSDM, 485 494. ACM. Wang, Z.; Wang, H.; Wen, J.-R.; and Xiao, Y. 2015. An inference approach to basic level of categorization. In CIKM, 653 662. ACM. Wang, X.; Yu, L.; Ren, K.; Tao, G.; Zhang, W.; Yu, Y.; and Wang, J. 2017. Dynamic attention deep model for article recommendation by learning human editors demonstration. In SIGKDD, 2051 2059. ACM. Wang, X.; He, X.; Feng, F.; Nie, L.; and Chua, T.-S. 2018a. Tem: Tree-enhanced embedding model for explainable recommendation. In WWW, 1543 1552. International World Wide Web Conferences Steering Committee. Wang, X.; Chen, Y.; Yang, J.; Wu, L.; Wu, Z.; and Xie, X. 2018b. A reinforcement learning framework for explainable recommendation. In ICDM. IEEE. Wang, H.; Wang, N.; and Yeung, D.-Y. 2015. Collaborative deep learning for recommender systems. In SIGKDD, 1235 1244. ACM. Wu, Y., and Ester, M. 2015. Flame: A probabilistic model combining aspect based opinion mining and collaborative filtering. In WSDM, 199 208. Wu, W.; Li, H.; Wang, H.; and Zhu, K. Q. 2012. Probase: A probabilistic taxonomy for text understanding. In SIGMOD, 481 492. ACM. Xu, C.; Tao, D.; and Xu, C. 2013. A survey on multi-view learning. ar Xiv preprint ar Xiv:1304.5634. Zhang, Y.; Lai, G.; Zhang, M.; Zhang, Y.; Liu, Y.; and Ma, S. 2014a. Explicit factor models for explainable recommendation based on phrase-level sentiment analysis. In SIGIR, 83 92. ACM. Zhang, Y.; Ahmed, A.; Josifovski, V.; and Smola, A. 2014b. Taxonomy discovery for personalized recommendation. In WSDM, 243 252. ACM. Zhang, F.; Yuan, N. J.; Lian, D.; Xie, X.; and Ma, W.-Y. 2016. Collaborative knowledge base embedding for recommender systems. In SIGKDD, 353 362. ACM. Zheng, L.; Noroozi, V.; and Yu, P. S. 2017. Joint deep modeling of users and items using reviews for recommendation. In WSDM, 425 434. ACM.