Adversarial auto-encoder for rating prediction with ratings and reviews

Abstract

Recommender systems have been widely used in our life in recent years to facilitate our life. And it is very important and meaningful to improve recommendation performance. Generally, recommendation methods use users’ historical ratings on items to predict ratings on their unrated items to make recommendations. However, with the increase of the number of users and items, the degree of data sparsity increases, and the quality of recommendations decreases sharply. In order to solve the sparsity problem, other auxiliary information is combined to mine users’ preferences for higher recommendation quality. Similar to rating data, review data also contain rich information about users’ preferences on items. This paper proposes a novel recommendation model, which harnesses an adversarial learning among auto-encoders to improve recommendation quality by minimizing the gap of the rating and review relation between a user and an item. The empirical studies on real-world datasets show that the proposed method improves the recommendation performance.

Keywords

Review recommendation adversarial learning auto-encoder

1. Introduction

Recommender systems provide users with their preferred items to solve the information overload problem in the era of the explosive growth of Web information. They have been playing an increasingly important role in Web applications in a variety of areas. Collaborative Filtering (CF) is a widely used recommendation approach [1]. The matrix factorization (MF) method [11] predicts ratings by the inner product of the user and item latent feature vectors learned from the historical user-item rating information. CF algorithms usually use users’ historical ratings on items to predict ratings on their unrated items to make recommendations. However, the rating data are sparse because a user may only rate several items. In order to solve the sparsity problem, we need other data sources to mine users’ preferences for higher recommendation quality. As user-generated reviews on items contain valuable information including users’ preferences on items and item features, many methods have been proposed to combine users’ historical reviews with real ratings on items to generate high-quality recommendations [5].

The main idea of these methods is to generate recommendations by transforming reviews, users, and items into a unified latent feature space. The latent feature space can be obtained from user-item rating information and user-item review information. Obviously, in the rating prediction task for review-based recommendations, on the one hand, we can draw a user-item rating relation from latent feature representations of a user and an item; on the other hand, we can also extract the user-item review relation from the review contents of the user on the item. The crucial problem is how to fuse the two relations to make recommendations. As both the rating relation and review relation between a user and an item can show the same preference of the user on the item, we should minimize their gap and make them indistinguishable.

The mechanism of generative adversarial nets (GANs) [6] trains a discriminator to distinguish real data from the generated data, and a generator to generate high-quality data to fool the discriminator. When the discriminator fails to discriminate between real data and fake data, we can say that the fake data are the same as the real data. So it is intuitive to employ such kind of min-max game to fuse the user-item rating relation and review relation by making them indistinguishable. In this paper, we train a discriminator to discriminate the two relations. As a result, the feature extractors represented by auto-encoders are optimized to minimize the difference of the two relations in an adversarial manner. Finally, we can use the learned user and item representations to predict ratings for review-based recommendations.

The remainder of this paper is organized as follows: Section 2 gives a brief overview of related work; Section 3 details the proposed model; Section 4 introduces the experimental setup and reports the results. Finally, Section 5 concludes the paper.

2. Related work

2.1. Rating and review-based recommendation methods

2.1.1. Traditional methods

In the traditional collaborative filtering algorithms, rating information is usually used to calculate the similarity of users or items. On the one hand, under the assumption that a user tends to select items similar to what the user has bought in the past, the item-based collaborative filtering (ICF) recommends items by using the similarities between different items. On the other hand, under the assumption that similar users have similar tastes, the user-based collaborative filtering (UCF) recommends items by using the similarities between different users [1]. The matrix factorization (MF) method [11] predicted ratings by the inner product of the user and item latent feature vectors learned from the historical user-item rating matrix. Many subsequent works added auxiliary information including users’ reviews on items to improve the recommendation performance under the MF framework. HFT [19] assumed that topic factors of an item share topics of all review documents of the item. Specially, the topic probability distribution of all reviews of the item can be measured by topic factors of the item. TopicMF [2] further combined user and item topic factors with reviews simultaneously.

2.1.2. Neural network-based methods

The neural networks for traditional recommendations can be divided into two categories [23]. One is how to use multi-layer non linear transformation functions to calculate the matching score of user and item embedding. For example, NeuMF [8] concatenated the user and item vectors and then feeded them into a multi-layer neural network to obtain the score values. But because of the huge differences in semantics between users and items, we cannot match them in the initial space. To solve this problem, the other methods are proposed to use neural networks for collaborative filtering based on representation learning. These methods focus on how to map users and items into a unified space by using neural networks and then match them. For example, ReDa [24] simultaneously learned the latent representations of users and items using auto-encoders only from historical user-item rating information. In ReDa, the encoder part transformed the user or item rating vector into a low-dimensional latent feature space, and the decoder part reconstructed user or item information.

Recently, deep neural networks have also been successfully applied to predicting ratings for review-based recommendations. These methods can be roughly divided into two categories. One is how to use user-item review information to generate latent feature representations of users and items to improve recommendation quality. For example, ConvMF [9] employed a convolutional neural network to learn item features from item review documents for rating prediction. TransNet [3] learned the latent representations of users and items by convolutional neural networks from their review documents respectively, and then the learned latent representations are concatenated to predict ratings. TARMF [16] co-learned matrix factorization and an attention-based GRU network to represent user and item features from ratings and user review documents. Following an adversarial training strategy, MT [17] employed a neural architecture of sequence-to-sequence learning to obtain user and item representations from user review documents and item review documents. DAML [15] utilized local and mutual attention of the convolutional neural network to jointly learn the user and item representations from user review documents and item review documents. The other is how to generate reviews by combining user-item rating information to enhance the performance of latent feature representations of users and items, and then achieve better rating prediction performance. For example, NRT [13] employed gated recurrent neural networks to transform user and item latent representations into a concise abstractive review, and simultaneously predict precise ratings.

2.2. Generative adversarial networks for recommendations

GANs are widely studied during recent several years for generative models. The basic idea is to play a min-max game between two adversarial methods: the generator is trained to generate high-quality data to fool the discriminator, while the discriminator is trained to discriminate real data and generated data [6]. [18] proposed an adversarial auto-encoder (AAE) in which the encoder learned the latent representation following an arbitrary prior distribution by the adversarial learning, while the decoder mapped the imposed prior to the reconstructed data distribution by minimizing the reconstruction loss.

Researchers have tried to exploit GANs to improve the performance of recommender systems by using the discriminator to distinguish true instances from fake instances based on the data space or the latent space. Based on the data space, [22] used the GAN mechanism to generate an item that a user may prefer, and obtained impressive results; [7] enhanced the pairwise ranking method by performing adversarial training on the user and item latent representations. Based on the latent space, [21] used the adversarial learning method to seek a common subspace of different models for retrieval; [12] used the discriminator to discriminate the same user’s different latent representations generated by two generators. [4] followed the direction of real-valued vector-wise GAN in order to fully exploit the advantage of adversarial training for higher recommendation accuracy in CF. Different from their recommendation methods by using GAN, our model plays a min-max game between the user-item rating relation and review relation, instead of between different latent representations of a user or an item.

3. Model

As shown in Fig. 1, our model (aae-RS for short) composes of three auto-encoders, namely $G_{r e}$ , $G_{r a U}$ , and $G_{r a I}$ , and one discriminative network D. In the process of learning these neural networks, $G_{r e}$ is used to extract the latent features from a review of user u on item j which forms a review relation between the user and the item. $G_{r a U}$ and $G_{r a I}$ are used to extract latent features of user u and item j respectively; latent features of user u and item j are combined to form a rating relation between them, and are also used to predicate the rating of user u on item j. The D is used to discriminate the two relations. In our model, the real and fake side are the rating and review relation respectively.

Fig. 1.

The framework of our model aae-RS.

3.1. Representations of reviews

Let Y be a triplet set in which each triplet $(u, r e_{u, j}, j)$ means user u gave a review $r e_{u, j}$ on item j. Specifically, for the review word set $r e_{u, j} = {t_{1}, t_{2}, \dots, t_{T}}$ , where T is the size of the dictionary. We obtain a binary vector $s_{r e} (u, j) = {s_{i}}_{i = 1}^{T}$ where $s_{i} = 1$ if $t_{i} \in r e_{u, j}$ , and $s_{i} = 0$ otherwise.

Taking the obtained binary vector $s_{r e} (u, j)$ as input, the network $G_{r e}$ consists of several non-linear hidden layers as follows: $\begin{array}{l} (1) & \begin{aligned} h_{u, j}^{(1)} = σ (W_{r e}^{(1)} s_{r e} (u, j) + b_{r e}^{(1)}), \\ h_{u, j}^{(l)} = σ (W_{r e}^{(l)} h_{u, j}^{(l - 1)} + b_{r e}^{(l)}), l = 2, \dots, K . \end{aligned} \end{array}$ Here, σ is a sigmoid function, and K represents the number of hidden layers. $h_{u, j}^{(l)}$ , $W_{r e}^{(l)}$ , and $b_{r e}^{(l)}$ represent the hidden vector, weight matrix and bias vector in the l-th layer respectively. In the output layer, we can get the reconstructed output ${\hat{s}}_{r e} (u, j)$ of the input vector $s_{r e} (u, j)$ . The reconstruction loss is $\begin{array}{rcl} L_{recon 1} & = & \frac{1}{2 \times | Y |} \sum_{Y} ‖ (s_{r e} (u, j) - {\hat{s}}_{r e} (u, j)) \\ (2) & ⊙ s_{r e} (u, j) ‖_{2}^{2} . \end{array}$ As the sparsity of the input results in that the number of zero elements in $s_{r e} (u, j)$ is much larger than that of non-zero elements, the Hadamard product ⊙ in Eq. (2) makes the auto-encoder tend to reconstruct the non-zero elements [20].

In the above method, we actually employ a deep auto-encoder to construct the review representations. The encoder part composes of non-linear hidden layers from the first hidden layer to the $K / 2$ -th hidden layer to transform the word set of a review into a low-dimensional latent feature space in which each element is a latent feature of a review. It is worth mentioning that we employ the $h_{u, j} = h_{u, j}^{(K / 2)}$ as the representations of review $r e_{u, j}$ . The size of the latent feature space is the number of nodes in the $K / 2$ -th hidden layer. Moreover, the reconstruction process of the decoder part from the $K / 2$ -th hidden layer to the output layer makes the representation preserve all the word information.

Obviously, we don’t need to recommend an item to a user if the user has written a review on the item. So a separate review can not be used to generate recommendation. In addition, we need the representations of users and items. In the next section, we will use the rating information of users on items to represent users and items.

3.2. Representations of users and items

The rating matrix is denoted $R \in R^{M \times N}$ , where M and N are respectively the numbers of users and items. If we take items as the features of each user, or users as the features of each item, we can obtain a user rating vector $x^{U} (u) = {r_{u j}}_{j = 1}^{N}$ and an item rating vector $x^{I} (j) = {r_{u j}}_{u = 1}^{M}$ where $r_{u j}$ is an element of R which denotes the user u’s rating on item j [24]. So the auto-encoder is still used to learn the latent representations of users and items respectively.

Taking the vector $x^{U} (u)$ and $x^{I} (j)$ as inputs, the network $G_{r a U}$ and $G_{r a I}$ be shown in Eq. (3) and (4) respectively. $\begin{array}{l} (3) & \begin{aligned} f_{u}^{(1)} = σ (W_{r a U}^{(1)} x^{U} (u) + b_{U}^{(1)}), \\ f_{u}^{(l)} = σ (W_{r a U}^{(l)} f_{u}^{(l - 1)} + b_{U}^{(l)}), l = 2, \dots, K^{'}, \end{aligned} \\ (4) & \begin{aligned} g_{j}^{(1)} = σ (W_{r a I}^{(1)} x^{I} (j) + b_{I}^{(1)}), \\ g_{j}^{(l)} = σ (W_{r a I}^{(l)} g_{j}^{(l - 1)} + b_{I}^{(l)}), l = 2, \dots, K^{″}, \end{aligned} \end{array}$ where $K^{'}$ and $K^{″}$ represent the number of layers and each node is activated by the sigmoid function σ. In the output layer, we get the reconstructed output ${\hat{x}}^{U} (u)$ and ${\hat{x}}^{I} (j)$ with the sigmoid function. We still use the Hadamard product ⊙ to reconstruct the non-zero elements [24], and then the corresponding reconstruction loss functions about users and items are shown in Eq. (5) and (6) respectively. $\begin{array}{l} L_{recon 2} = \frac{1}{2 \times | U |} \\ (5) & \times \sum_{u \in U} {‖ (x^{U} (u) - {\hat{x}}^{U} (u)) ⊙ x^{U} (u) ‖}_{2}^{2}, \\ L_{recon 3} = \frac{1}{2 \times | I |} \\ (6) & \times \sum_{j \in I} {‖ (x^{I} (j) - {\hat{x}}^{I} (j)) ⊙ x^{I} (j) ‖}_{2}^{2} . \end{array}$ Similar to the representation of a review, we employ the $f_{u} = f_{u}^{(K^{'} / 2)}$ as the representations of user u, and $g_{j} = g_{j}^{(K^{″} / 2)}$ as the representations of item j

3.3. Fusion

Given user u, item j, and a review of the user on the item $r e (u, j)$ , we have the corresponding representation $f_{u}$ , $g_{j}$ and $h_{u, j}$ respectively. Auto-encoders have been used to obtain the distribution of latent features of their inputs. User u, item j, and review $r e (u, j)$ can share the same latent feature space. So we can obtain two relation representations between user u and item j. One is the review relation vector $h_{u, j}$ , the other is the rating relation vector calculated by $f_{u} ⊙ g_{j}$ . Our goal is to detect whether the rating and review relation have the same distribution. The mechanism of adversarial learning provides a solution to solve the problem. In the adversarial mechanism, we introduce a discriminative network D to discriminate rating and review relation. So we have the adversarial loss as $\begin{array}{rcl} L_{adv} & = & \frac{1}{| Y |} \sum_{r_{u j} \neq 0} (log (D (o (f_{u} ⊙ g_{j})) \\ (7) & + log (1 - D (h_{u, j}))) . \end{array}$ In Eq. (7), the adversarial mechanism requires that D maximizes the $L_{adv}$ to make D discriminate the two relations, and make $G_{r e}$ , $G_{r a U}$ , and $G_{r a I}$ to fool network D. The iteratively learning process will end until D can not discriminate the two relations. Unlike the adversarial auto-encoder [18] which generates true instances from a fixed prior distribution, our model uses the inferred latent representations as a true side to regularize another side. Then, on the one hand, $h_{u, j}$ could be drawn from the encoder conditioned on $s_{r e} (u, j)$ and taken as the fake side of D. On the other hand, the value of each element in $h_{u, j}$ that can represent a review $r e (u, j)$ should be sampled with a greater probability, which follows a Gaussian distribution. So we don’t directly use $f_{u} ⊙ g_{j}$ as the real side of D, but use a transformation [10] as $\begin{matrix} (8) & o (f_{u} ⊙ g_{j}) = f_{u} ⊙ g_{j} + λ 1 ⊙ ϵ, \end{matrix}$ where $ϵ$ is drawn from a Gaussian distribution with a mean of $0$ and a standard deviation of $1$ , namely, $ϵ \sim N (0, 1)$ . Obviously, we use Eq. (8) to obtain the real side following a Gaussian distribution $N (f_{u} ⊙ g_{j}, λ 1)$ and also make $L_{adv}$ be differentiable. The adversarial learning ensures that both real and fake side of D have the same probability distribution, which results in that $h_{u, j}$ follows a Gaussian distribution as $\begin{matrix} (9) & h_{u, j} \sim N (f_{u} ⊙ g_{j}, 1) . \end{matrix}$ In Eq. (9), the distribution is a member of the parametric family of Gaussian distribution, therefore all $f_{u}$ and $g_{j}$ can be also inferred by maximizing likelihood over the $h_{u, j}$ which is equivalent to minimizing the sum-of-squared-errors objective function $\begin{matrix} (10) & L_{recon 4} = \frac{1}{2} \sum_{Y} ‖ h_{u, j} - f_{u} ⊙ g_{j} ‖_{2}^{2} . \end{matrix}$ Minimizing equation (10) intuitively means that $s_{r e} (u, j)$ can be positively correlated with the corresponding value $f_{u} ⊙ g_{j}$ in the latent feature space. Such optimization minimizes the gap among the different latent representations of the same pair of user and item.

3.4. Rating prediction

The obtained latent feature representations of user u and item j are fed into $K^{‴}$ fully connected layers with $K^{‴}$ weight matrices and bias vectors as $\begin{array}{rcl} e^{(K^{‴})} & = & σ (W_{P}^{(K^{‴})} (\dots σ (W_{P}^{(1)} (f_{u} ⊙ g_{j}) + b_{P}^{(1)}) \dots) \\ (11) & + b_{P}^{(K^{‴})}) . \end{array}$

At last, the predicted rating of user u on item j is obtained via a regression layer with weight matrix $W_{P}$ and bias vector $b_{P}$ by $\begin{array}{rcl} (12) & {\hat{r}}_{u j} = W_{P} e^{(K^{‴})} + b_{P} . \end{array}$

As our task is rating prediction, the square loss between the predicted ratings and the real ratings is used for training. The loss function is $\begin{matrix} (13) & L_{pre} = \frac{1}{| Y |} \sum_{r_{u j} \neq 0} {(r_{u j} - {\hat{r}}_{u j})}^{2} . \end{matrix}$

3.5. Model inference

By fixing $G_{r a U}$ , $G_{r a U}$ , and $G_{r e}$ , training D aims to distinguish two representations drawn from users’ review actions and rating actions, and then we have $\begin{matrix} (14) & L_{D} = λ_{1} max_{D} L_{adv} . \end{matrix}$

We train $G_{r e}$ by fixing $G_{r a U}$ , $G_{r a U}$ , and D. As the outputs of the $G_{r e}$ is a part of Eq. (2) and (10) and also a part of adversarial training in Eq. (7), the optimization with respect to $G_{r e}$ should be: $\begin{matrix} (15) & L_{G_{r e}} = min_{G_{r e}} λ_{1} L_{adv} + λ_{3} L_{recon 1} + λ_{4} L_{recon 4} . \end{matrix}$

By fixing $G_{r e}$ and D, we train $G_{r a U}$ and $G_{r a U}$ by $\begin{array}{l} L_{G_{r a U}} = min_{G_{r a U}} λ_{1} L_{adv} + λ_{3} L_{recon 2} \\ (16) & + λ_{4} L_{recon 4} + λ_{5} L_{pre}, \\ L_{G_{r a I}} = min_{G_{r a I}} λ_{1} L_{adv} + λ_{3} L_{recon 3} \\ (17) & + λ_{4} L_{recon 4} + λ_{5} L_{pre} . \end{array}$

Furthermore, in order to avoid the overfitting problem, we use the $L 2$ regularization method to regularize all neural variables weighted by $λ_{2}$ .

4. Experiments

In this section, we conduct experiments on two real-world data sets to answer the following questions which can certify the effectiveness of our model.

Q1: How does our proposed model aae-RS perform compared to other recommendation methods?

Q2: How do the key components impose influence on the performance of our model aae-RS?

Q3: Is our model aae-RS sensitive to the key hyper-parameters?

4.1. Experimental settings

4.1.1. Data set

We adopt Yelp dataset and Amazon video game dataset to perform a fair evaluation of our models. Yelp is the largest comment website in the USA. Users can rate merchants, submit comments, and exchange shopping experiences on the Yelp website. The Yelp dataset used in this paper includes $4, 852$ reviews. And Amazon is the largest ecommerce company in the USA. It is a multi-category online mall in which users can shop and write reviews. The Amazon dataset used in this paper includes $17, 286$ video game reviews. Ratings in these data sets scale from 1 to 5 stars. We remove stop words and select the about $7, 000$ and $9, 000$ most frequent words as the vocabulary on the two datasets. Then, we remove users and items that occur fewer than 10 times. The sparsity proportion are $98.82 %$ and $97.73 %$ which refer to the ratio of the number of zero elements to the number of all elements on the Amazon video game and Yelp dataset respectively. The statistics of the two datasets are presented in Table 1.

Table 1
Statistics of the datasets

Dataset Yelp Amazon Video Game

Numbers of Users 894 847

Numbers of Items 461 899

Numbers of Words 6,969 8,966

Numbers of Reviews 4,852 17,286

Sparsity 98.82% 97.73%

Dataset	Yelp	Amazon Video Game
Numbers of Users	894	847
Numbers of Items	461	899
Numbers of Words	6,969	8,966
Numbers of Reviews	4,852	17,286
Sparsity	98.82%	97.73%

4.1.2. Evaluation metrics

We randomly select $40 %$ , $60 %$ and $80 %$ of the observations as the training set and use the remaining as the test set. We build the model on the training sets and measure the prediction accuracy on the test sets. The predictive accuracy is reported as the root-means square error (RMSE) calculated by $\begin{matrix} (18) & RMSE = \sqrt{\frac{1}{| τ |} \sum_{(u, r e_{u, j}, j) \in τ} {(r_{u j} - {\hat{r}}_{u j})}^{2}}, \end{matrix}$ where τ is the test set.

4.1.3. Implementation details

In our experiments, the $G_{r e}$ , $G_{r a U}$ , $G_{r a I}$ are three fully connected networks with one hidden layer consisting of 20 nodes. And the discriminative network is a three-layer full connected network with one hidden layer consisting of 10 nodes. We choose $sigmoid$ function as the activation function for all networks. We train these neural networks by using the mini-batch gradient descent method. The batch size is 2048. The learning rate of the discriminator is 0.5 and the learning rate of others is 1. The hyper parameter $λ_{1}$ is 0.0001, $λ_{2}$ is 0.0001, $λ_{3}$ is 30, $λ_{4}$ is 20, and $λ_{5}$ is 0.001. Impacts of hyper parameters can be found in Section 4.4.

Fig. 2.

Comparisons of aae-RS with different methods.

Table 2

Performance of aae-RS compared with its two variants

Dataset	Percent of Training	40	60	80
Yelp	aae-RS without D	0.9953	0.9741	0.9233
	aae-RS without $ϵ$	0.9740	0.9681	0.9069
	aae-RS	0.9706	0.9636	0.8980
Amazon	aae-RS without D	1.0098	0.9811	0.9684
	aae-RS without $ϵ$	1.0035	0.9706	0.9641
	aae-RS	0.9974	0.9628	0.9615

4.2. Performance comparison (Q1)

We compare our aae-RS with ReDa, mDA-CF, ConvMF- and HFT on these data sets.

ReDa [24] simultaneously learns latent representations of users and items using auto encoders only from historical user-item rating information, and values of their dot product are the predicted ratings.

mDA-CF [14] learns latent representations of users and items using auto encoders only from review information, and values of their dot product are the predicted ratings.

ConvMF- [9] learns item latent features from item review documents which are close to ones from the matrix factorization method. ConvMF- refers to the ConvMF model uses the full connected network instead of a convolutional neural network.

HFT [19] co-learns topic features from review documents and matrix factorization techniques for rating prediction.

All comparison results on RMSE are summarized in Fig. 2. In this figure, the performance of each method generally becomes better when the sampling training data increases in the two datasets. On Yelp, aae-RS and ReDa, both of which use the encoder-decoder to reconstruct rating information, achieve higher performance. Specially, our model aac-RS achieves the best results by additional review information. Furthermore, we can observe that the effects of mDA-CF which only reconstructs review information and ConvMF- which does not reconstruct review information are not stable in different proportion of training sets. Therefore, we can conclude that it is necessary for aae-RS to reconstruct both rating and review information. On Amazon, we can see that ReDa and mDA-CF, which only reconstruct either the rating information or review information, are not as effective as ConvMF- and HFT which only consider the items’ review information. Our method achieves the best performance by comprehensively considering users’ ratings, items’ ratings, users’ reviews and items’ reviews.

4.3. Effect of adversarial learning and noise (Q2)

To deeply demonstrate the contributions of adversarial learning and noises for aae-RS, we compare the performance of aae-RS with its two variants, namely aae-RS without D and aae-RS without $ϵ$ . For aae-RS without D and $ϵ$ , we remove the discriminator D and noises $ϵ$ respectively. As shown in Table 2, the aae-RS achieves best performance on the two datasets. On Yelp with $40 %$ , $60 %$ and $80 %$ as the training set, the relative improvement over aae-RS without D is $2.48 %$ , $1.08 %$ and $2.74 %$ respectively, and the relative improvement over aae-RS without $ϵ$ is $0.35 %$ , $0.46 %$ and $0.98 %$ respectively. On Amazon with $40 %$ , $60 %$ and $80 %$ as the training set, aae-RS outperforms aae-RS without D by $1.23 %$ , $1.87 %$ and $0.71 %$ , respectively. And compared with aae-RS without $ϵ$ , aae-RS obtains $0.61 %$ , $0.80 %$ and $0.27 %$ relative improvement respectively. These results confirm the usefulness of two components in aae-RSs.

Fig. 3.

Training loss trend.

On Yelp, we can also see that the recommendation performance improves significantly by adversarial learning and noise for the $80 %$ training set, which shows that we need more data to train our model to achieve the best performance. On the Amazon dataset, we can see that the improvement of aae-RS is not significant for the $80 %$ training set. One of the possible reasons is that the data set is large enough to ensure that Eq. (9) is naturally established. However, in practical applications, we cannot clearly determine how large the data set is. In addition, user privacy may limit the size of data. In this case, we may use a small training set to train aae-RS for recommendations.

We further explore the trend of training loss to show the effect of adversarial learning in aae-RS. Let other loss denote the sum of three reconstruction losses and prediction loss. We sample values of the adversarial loss and the other loss from epoch 1 to 500 on the two datasets as shown in Fig. 3. If the value of adversarial loss would have a remarkable increasing or decreasing trend, the discriminator can distinguish two relations well and can not fuse them to achieve effective recommendations. Fortunately, such results do not happen in Fig. 3. In Fig. 3, we can see that the other loss decreases almost monotonously and converges smoothly during training, while the adversarial loss decreases almost around the initial 10 epochs and then stabilizes. So we can conclude that results in Fig. 3 are in accordance with the expectation of the fusion strategy in our aae-RS framework.

4.4. Effect of model parameters (Q3)

In this section, we investigate the influence of the hyper parameters $λ_{1}$ , $λ_{2}$ , $λ_{3}$ , $λ_{4}$ and $λ_{5}$ on the data set with $60 %$ as the training set. In these experiments, when tuning one parameter, we fix the values of the other four. Specifically, parameter $λ_{1}$ is tuned amongst {0.0001, 0.001, 0.01, 0.1, 1}, $λ_{2}$ amongst {0.00001, 0.0001, 0.001, 0.01, 0.1}, $λ_{3}$ amongst {0.03, 0.3, 3, 30, 50}, $λ_{4}$ amongst {0.02, 0.2, 2, 20, 50}, and $λ_{5}$ amongst {0.001, 0.01, 1, 10, 100}. We report all the results in Fig. 4 and Fig. 5. We can see that different parameter values bring different experimental results. Our aim is to select the best parameter values to achieve the best experimental results. Finally, we set $λ_{1} = 0.0001$ , $λ_{2} = 0.0001$ , $λ_{3} = 30$ , $λ_{4} = 20$ and $λ_{5} = 0.001$ as the default values for all datasets.

Fig. 4.

The study of parameter influence of aae-RS on yelp.

Fig. 5.

The study of parameter influence of aae-RS on Amazon.

5. Conclusions

In general, we use user-item rating information to make recommendations. In addition, auxiliary information can be used to improve the recommendation performance. In this paper, we propse the aae-RS model to fuse user-item rating and review information. In aae-RS, we simultaneously learn the latent factors of users and items from rating and review information. Specially, by adversarial learning, rating and review information can be better fused for recommendations when discriminators cannot distinguish them. Experiments on real-world data sets validate the superiority of our proposed framework. In the future, ass-RS will use more effective neural networks including Convolutional Neural Network (CNN) to represent user review information and make better recommendations.

Footnotes

Acknowledgement

This work was partly supported by the National Natural Science Foundation of China (61562009, 61420106005), and Program of Guizhou Provincial Science and Technology Department (No. [2019]2502). This paper is extended from our paper in the 2018 IEEE/WIC/ACM International Conference on Web Intelligence (WI 2018)

References

Adomavicius and

Tuzhilin , Toward the next generation of recommender systems: A survey of the state-of-the-art and possible extensions, IEEE Transactions on Knowledge and Data Engineering 17(6) (2005), 734–749. doi:10.1109/TKDE.2005.99.

Bao ,

Fang and

Zhang , TopicMF: Simultaneously exploiting ratings and reviews for recommendation, in: Proceedings of the 28th AAAI Conference on Artificial Intelligence, 2014, pp. 2–8.

Catherine and

W.W.

Cohen , TransNets: Learning to transform for recommendation, in: Proceedings of the 11th ACM Conference on Recommender Systems, 2017, pp. 288–296.

D.-K.

Chae ,

J.-S.

Kang ,

S.-W.

Kim and

J.-T.

Lee , CFGAN: A generic collaborative filtering framework based on generative adversarial networks, in: Proceedings of the 27th ACM International Conference on Information and Knowledge Management, 2018, pp. 137–146. doi:10.1145/3269206.3271743.

Chen ,

Chen and

Wang , Recommender systems based on user reviews: The state of the art, Journal of User Modeling and User-Adapted Interaction 25(2) (2015), 99–154. doi:10.1007/s11257-015-9155-5.

I.J.

Goodfellow ,

Pouget-Abadie ,

Mirza ,

Xu ,

Warde-Farley ,

Ozair ,

A.C.

Courville and

Bengio , Generative adversarial nets, in: Proceedings of the Annual Conference on Neural Information Processing Systems, 2014, pp. 2672–2680.

He ,

Du and

T.-S.

Chua , Adversarial personalized ranking for recommendation, in: Proceedings of the 41st International ACM SIGIR Conference on Research and Development in Information Retrieval, 2018, pp. 355–364.

He ,

Liao ,

Zhang ,

Nie ,

Hu and

T.-S.

Chua , Neural collaborative filtering, in: Proceedings of the 26th International Conference on World Wide Web, 2017, pp. 173–182. doi:10.1145/3038912.3052569.

D.H.

Kim ,

Park ,

Oh ,

Lee and

Yu , Convolutional matrix factorization for document context-aware recommendation, in: Proceedings of the 10th ACM Conference on Recommender Systems, 2016, pp. 233–240. doi:10.1145/2959100.2959165.

10.

D.P.

Kingma and

Welling , Auto-Encoding Variational Bayes, 2013, [Online], Available arXiv:1312.6114.

11.

Koren ,

R.M.

Bell and

Volinsky , Matrix factorization techniques for recommender systems, IEEE Computer 42(8) (2009), 30–37. doi:10.1109/MC.2009.263.

12.

Lee ,

Song and

I.-C.M.

Moon , Augmented variational autoencoders for collaborative filtering with auxiliary information, in: Proceedings of the 2017 ACM on Conference on Information and Knowledge Management, 2017, pp. 1139–1148. doi:10.1145/3132847.3132972.

13.

Li ,

Wang ,

Ren ,

Bing and

Lam , Neural rating regression with abstractive tips generation for recommendation, in: Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval, 2017, pp. 345–354. doi:10.1145/3077136.3080822.

14.

Li ,

Kawale and

Fu , Deep collaborative filtering via marginalized denoising auto-encoder, in: Proceedings of the 24th ACM International on Conference on Information and Knowledge Management, 2015, pp. 811–820. doi:10.1145/2806416.2806527.

15.

Liu ,

Li ,

Du ,

Chang and

Gao , DAML: Dual attention mutual learning between ratings and reviews for item recommendation, in: Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2019, pp. 344–352. doi:10.1145/3292500.3330906.

16.

Lu ,

Dong and

Smyth , Coevolutionary recommendation model: Mutual learning between ratings and reviews, in: Proceedings of the 2018 World Wide Web Conference on World Wide Web, 2018, pp. 773–782.

17.

Lu ,

Dong and

Smyth , Why I like it: Multi-task learning for recommendation and explanation, in: Proceedings of the 12th ACM Conference on Recommender Systems, 2018, pp. 4–12. doi:10.1145/3240323.3240365.

18.

Makhzani ,

Shlens ,

Jaitly ,

Goodfellow and

Frey , Adversarial Autoencoders, (2015), [Online], Available arXiv:1511.05644.

19.

J.J.

McAuley and

Leskovec , Hidden factors and hidden topics: Understanding rating dimensions with review text, in: Proceedings of the 7th ACM Conference on Recommender Systems, 2013, pp. 165–172. doi:10.1145/2507157.2507163.

20.

Tu ,

Zhang ,

Liu and

Sun , Transnet: Translation-based network representation learning for social relation extraction, in: Proceedings of the 26th International Joint Conference on Artificial Intelligence, 2017, pp. 2864–2870.

21.

Wang ,

Yang ,

Xu ,

Hanjalic and

Tao Shen , Adversarial cross-modal retrieval, in: Proceedings of the 2017 ACM on Multimedia Conference, 2017, pp. 154–162. doi:10.1145/3123266.3123326.

22.

Wang ,

Yu ,

Zhang ,

Gong ,

Xu ,

Wang ,

Zhang and

Zhang , IRGAN: A minimax game for unifying generative and discriminative information retrieval models, in: Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval, 2017, pp. 515–524. doi:10.1145/3077136.3080786.

23.

Xu ,

He and

Li , Deep learning for matching in search and recommendation, in: Proceedings of the 41st International ACM SIGIR Conference on Research and Development in Information Retrieval, 2018, pp. 1365–1368.

24.

Zhuang ,

Zhang ,

Qian ,

Shi ,

Xie and

He , Representation learning via dual-autoencoder for recommendation, Neural Networks 90 (2017), 83–89. doi:10.1016/j.neunet.2017.03.009.

Adversarial auto-encoder for rating prediction with ratings and reviews

Abstract

Keywords

1. Introduction

2. Related work

2.1. Rating and review-based recommendation methods

2.1.1. Traditional methods

2.1.2. Neural network-based methods

2.2. Generative adversarial networks for recommendations

3. Model

3.2. Representations of users and items

3.3. Fusion

3.4. Rating prediction

3.5. Model inference

4. Experiments

4.1. Experimental settings

4.1.1. Data set

Table 1 Statistics of the datasets Dataset Yelp Amazon Video Game Numbers of Users 894 847 Numbers of Items 461 899 Numbers of Words 6,969 8,966 Numbers of Reviews 4,852 17,286 Sparsity 98.82% 97.73%

4.1.3. Implementation details

4.3. Effect of adversarial learning and noise (Q2)

Footnotes

Acknowledgement

References

Table 1
Statistics of the datasets

Dataset Yelp Amazon Video Game

Numbers of Users 894 847

Numbers of Items 461 899

Numbers of Words 6,969 8,966

Numbers of Reviews 4,852 17,286

Sparsity 98.82% 97.73%