TransRev: Modeling Reviews as Translations from Users to Items
- 1 Citations
- 3k Downloads
Abstract
The text of a review expresses the sentiment a customer has towards a particular product. This is exploited in sentiment analysis where machine learning models are used to predict the review score from the text of the review. Furthermore, the products costumers have purchased in the past are indicative of the products they will purchase in the future. This is what recommender systems exploit by learning models from purchase information to predict the items a customer might be interested in. The underlying structure of this problem setting is a bipartite graph, wherein customer nodes are connected to product nodes via ‘review’ links. This is reminiscent of knowledge bases, with ‘review’ links replacing relation types. We propose TransRev, an approach to the product recommendation problem that integrates ideas from recommender systems, sentiment analysis, and multi-relational learning into a joint learning objective.
TransRev learns vector representations for users, items, and reviews. The embedding of a review is learned such that (a) it performs well as input feature of a regression model for sentiment prediction; and (b) it always translates the reviewer embedding to the embedding of the reviewed item. This is reminiscent of TransE [5], a popular embedding method for link prediction in knowledge bases. This allows TransRev to approximate a review embedding at test time as the difference of the embedding of each item and the user embedding. The approximated review embedding is then used with the regression model to predict the review score for each item. TransRev outperforms state of the art recommender systems on a large number of benchmark data sets. Moreover, it is able to retrieve, for each user and item, the review text from the training set whose embedding is most similar to the approximated review embedding.
Keywords
Recommender systems Knowledge graphs Sentiment analysis1 Introduction
In addition, the availability of product reviews allows users to make more informed purchasing choices and companies to analyze costumer sentiment towards their products. The latter was coined sentiment analysis and is concerned with machine learning approaches that map written text to scores. Nevertheless, even the best sentiment analysis methods cannot help in determining which new products a costumer might be interested in. The obvious reason is that costumer reviews are not available for products they have not purchased yet.
With this paper we propose TransRev, a method that combines a personalized recommendation learning objective with a sentiment analysis objective into a joint learning objective. TransRev learns vector representations for users, items, and reviews jointly. The crucial advantage of TransRev is that the review embedding is learned such that it corresponds to a translation that moves the embedding of the reviewing user to the embedding of the item the review is about. This allows TransRev to approximate a review embedding at test time as the difference of the item and user embedding despite the absence of a review from the user for that item. The approximated review embedding is then used in the sentiment analysis model to predict the review score. Moreover, the approximated review embedding can be used to retrieve reviews in the training set deemed most similar by a distance measure in the embedding space. These retrieved reviews could be used for several purposes. For instance, such reviews could be provided to users as a starting point for a review, lowering the barrier to writing reviews.
2 TransRev: Modeling Reviews as Translations in Vector Space
We address the problem of learning prediction models for the product recommendation problem. A small example of the input data typical to such a machine learning system is depicted in Fig. 1. This reminds of knowledge bases, with ‘reviews’ replacing relation types. Two nodes in a knowledge base may be joined by a number of links, each representing one relation type from a small vocabulary. Here, if two nodes are connected they are linked by one single edge type, in which case it is represented by a number of words from a (very) large vocabulary.
There are a set of users \(\mathbf {U}\), a set of items \(\mathbf {I}\), and a set of reviews \(\mathbf {R}\). Each \(\mathtt {rev}_{(\mathtt {u,i})} \in \mathbf {R}\) represents a review written by user \(\mathtt {u}\) for item \(\mathtt {i}\). Hence, \(\mathtt {rev}_{(\mathtt {u,i})} = [\mathtt {t}_1, \cdots , \mathtt {t}_{n}]\), that is, each review is a sequence of n tokens. In the following we refer to \((\mathtt {u}, \mathtt {rev}_{(\mathtt {u,i})}, \mathtt {i})\) as a triple. Each such triple is associated with the review score \(\mathtt {r}_{(\mathtt {u,i})}\) given by the user \(\mathtt {u}\) to item \(\mathtt {i}\).
All parameters of the parts of the objective are jointly learned with stochastic gradient descent. More details regarding the parameter learning are contained in the experimental section.
2.1 On the Choice of TransE as Modeling Assumption
The choice of TransE as underlying modeling assumption to this recommendation problem is not arbitrary. Given the user and item embeddings, and without further constraints, it allows to distinctively compute the approximate review embedding via Eq. (6). Another popular knowledge graph embedding method is DistMult [16]. In applying such modeling assumption to this problem one would obtain the approximate review embedding by solving the following optimization problem: \(\hat{\mathbf {h}}_{\mathtt {rev}_{\mathtt {(u,i)}}} = \max _{\mathbf {h}} (\mathbf {e}_{\mathtt {i}} \circ \mathbf {e}_{\mathtt {u}}) \mathbf {h}\), where \(\circ \) is the element-wise multiplication. The solution to that problem would be any vector with infinite norm. Therefore, one should impose constraints in the norm of the embeddings to obtain a non-trivial solution. However, previous work [11] shows that such constraint harms performance. Similarly, most of the knowledge graph embedding methods would require to impose constraints in the norm of the embeddings. The translation modeling assumption of TransE facilitates the approximation of the review embedding without additional constraints, while its performance is on par with, if not better, than most of all other translation-based knowledge graph embedding methods [11].
3 Related Work
There are three lines of research related to our work: knowledge graph completion, recommender systems and sentiment analysis.
The first research theme related to TransRev is knowledge graph completion. In the last years, many embedding-based methods have been proposed to infer missing relations in knowledge graphs based on a function that computes a likelihood score based on the embeddings of entities and relation types. Due to its simplicity and good performance, there is a large body of work on translation-based scoring functions [5, 13]. [14] propose an approach to large-scale sequential sales prediction that embeds items into a transition space where user embeddings are modeled as translation vectors operating on item sequences. The associated optimization problem is formulated as a sequential Bayesian ranking problem [28]. To the best of our knowledge, [14] is the first work in leveraging ideas from knowledge graph completion methods for recommender system. Whereas TransRev addresses the problem of rating prediction by incorporating review text, [14] addresses the different problem of sequential recommendation. Therefore the experimental comparison to that work is not possible. In TransRev the review embedding translates the user embedding to the product embedding. In [14], the user embedding translates a product embedding to the embedding of the next purchased product. Moreover, TransRev gets rid of the margin-based loss (and consequently of the negative sampling) due to the joint optimization of Eqs. (2) and (4), whereas [14] is formalized as a ranking problem in a similar way to [5]. Subsequently, there has been additional work on translation-based models in recommender systems [25, 33]. However, these works cannot incorporate users’ feedback other than ratings into the learning, which has been shown to boost performance [21].
There is an extensive body of work on recommender systems [1, 6, 29]. Singular Value Decomposition (SVD) [17] computes the review score prediction as the dot product between the item embeddings and the user embeddings plus some learnable bias terms. Due to its simplicity and performance on numerous data sets—including winning solution to the Netflix prize—it is still one of the most used methods for product recommendations. Most of the previous research that explored the utility of review text for rating prediction can be classified into two categories.
Semi-supervised Approaches. HFT [21] was one of the first methods combining a supervised learning objective to predict ratings with an unsupervised learning objective (e.g. latent Dirichlet allocation) for text content to regularize the parameters of the supervised model. The idea of combining two learning objectives has been explored in several additional approaches [3, 9, 19]. The methods differ in the unsupervised objectives, some of which are tailored to a specific domain. For example, JMARS [9] outperforms HFT on a movie recommendation data set but it is outperformed by HFT on data sets similar to those used in our work [36].
Supervised Approaches. Methods that fall into this category such as [31, 32] learn latent representations of users and items from the text content so as to perform well at rating prediction. The learning of the latent representations is done via a deep architecture. The approaches differences lie mainly in the neural architectures they employ.
There is one crucial difference between the aforementioned methods and TransRev. TransRev predicts the review score based on an approximation of the review embedding computed at test time. Moreover, since TransRev is able to approximate a review embedding, we can use this embedding to retrieve reviews in the training set deemed most similar by a distance metric in the embedding space.
Similar to sentiment analysis methods, TransRev trains a regression model that predicts the review rating from the review text. Contrary to the typical setting in which sentiment analysis methods operate, however, review text is not available at prediction time in the recommender system setting. Consequently, the application of sentiment analysis to recommender systems is not directly possible. In the simplest case, a sentiment analysis method is a linear regressor applied to a text embedding (Eq. (3)).
4 Experimental Setup
We conduct several experiments to empirically compare TransRev to state of the art methods for product recommendation. Moreover, we provide some qualitative results on retrieving training reviews most similar to the approximated reviews at test time.
4.1 Data Sets
We evaluate the various methods on data sets from the Amazon Product Data^{1}, which has been extensively used in previous works [21, 22, 23]. The data set consists of reviews and product metadata from Amazon from May 1996 to July 2014. We focus on the 5-core versions (which contain at least 5 reviews for each user and item) of those data sets. There are 24 product categories from which we have randomly picked 18. As all previously mentioned works, we treat each of these resulting 18 data sets independently in our experiments. Ratings in all benchmark data sets are integer values between 1 and 5. As in previous work, we randomly sample 80% of the reviews as training, 10% as validation, and 10% as test data. We remove reviews from the validation and test splits if they involve either a product or a user that is not part of the training data.
4.2 Review Text Preprocessing
Performance comparison (MSE) on 18 datasets. The asterisk \(*\) indicates the macro MSE across all the Amazon data sets.
HFT | SVD | DeepCoNN | T-Nets | MPCN | TransRev | |
---|---|---|---|---|---|---|
Amazon Instant Video | 0.888 | 0.904 | 0.943 | 1.007 | 0.997 | 0.884 |
Automotive | 0.862 | 0.857 | 0.853 | 0.946 | 0.861 | 0.855 |
Baby | 1.104 | 1.108 | 1.154 | 1.338 | 1.304 | 1.100 |
Cds and Vinyl | 0.854 | 0.863 | 0.888 | 1.010 | 1.005 | 0.854 |
Grocery and Gourmet Food | 0.961 | 0.964 | 0.973 | 1.129 | 1.125 | 0.957 |
Health and personal care | 1.014 | 1.016 | 1.081 | 1.249 | 1.238 | 1.011 |
Kindle Store | 0.593 | 0.607 | 0.648 | 0.797 | 0.775 | 0.599 |
Musical Instruments | 0.692 | 0.694 | 0.723 | 1.100 | 0.923 | 0.690 |
Office Products | 0.727 | 0.727 | 0.738 | 0.840 | 0.779 | 0.724 |
Patio, Lawn and Garden | 0.956 | 0.950 | 1.070 | 1.123 | 1.011 | 0.941 |
Pet Supplies | 1.194 | 1.198 | 1.281 | 1.346 | 1.328 | 1.191 |
Tools and Home Improvement | 0.884 | 0.884 | 0.946 | 1.122 | 1.096 | 0.879 |
Toys and Games | 0.784 | 0.788 | 0.851 | 0.974 | 0.973 | 0.784 |
Beauty | 1.165 | 1.168 | 1.184 | 1.404 | 1.387 | 1.158 |
Digital Music | 0.793 | 0.797 | 0.835 | 1.004 | 0.970 | 0.782 |
Video Games | 1.086 | 1.093 | 1.133 | 1.276 | 1.257 | 1.082 |
Sports and Outdoors | 0.824 | 0.828 | 0.882 | 0.994 | 0.980 | 0.823 |
Cell Phones and Accessories | 1.285 | 1.290 | 1.365 | 1.431 | 1.413 | 1.279 |
\(0.926^*\) | \(0.930^*\) | \(0.969^*\) | \(1.116^*\) | \(1.079^*\) | 0.921\(^{*}\) |
4.3 Baselines
We compare to the following methods: a SVD matrix factorization; HFT, which has not often been benchmarked in previous works; and DeepCoNN [38], which learns user and item representations from reviews via convolutional neural networks. We also include MPCN [34] (which stands for multi-pointer co-attention networks) in the comparison, however, as indicated in previous work [8] MPCN is a non-reproducible work^{2}. Therefore, we simply copy numbers from [34], since they used the same data sets as the ones used in this work. Additionally, we also include performance for TransNets (T-Nets) [7], whose numbers are also copied from [34]. T-Nets is similar to TransRev in that it also infers review latent representations from user and item representations. Different to TransRev, it does not have any underlying graph-based modeling assumption among users, items and reviews.
4.4 Parameter Setting
We set the dimension k of the embedding space to 16 for all methods. We evaluated the robustness of TransRev to changes in Sect. 4.6. Alternatively, one could use off-the-shelf word embeddings (e.g. word2vec [24] or ELMO [26]), but this would require to assume the existence of a large collection of text for effectively learning good word representations in an unsupervised manner. However, such a corpus may not be available for some low-resource languages or domain-specific use cases. For HFT we used the original implementation of the authors^{3} and validated the trade-off term from the values [0.001, 0.01, 0.1, 1, 10, 50]. For TransRev we validated \(\lambda \) among the values [0.05, 0.1, 0.25, 0.5, 1] and the learning rate of the optimizer and regularization term (\(\mu \) in our model) from the values [0.001, 0.005, 0.01, 0.05, 0.1] and [0.00001, 0.00005, 0.0001, 0.0005, 0.001], respectively. TransRev ’s parameters were randomly initialized [12] and learned with vanilla stochastic gradient descent. A single learning iteration performs SGD with all review triples in the training data and their associated ratings. For TransRev we used a batch size of 64. We ran TransRev for a maximum of 500 epochs and validated every 10 epochs. For SVD we used the Python package Surprise^{4}, and chose the learning rate and regularization term from the same range of values. Parameters for HFT were learned with L-BFGS, which was run for 2,500 learning iterations and validated every 50 iterations. For DeepCoNN the original authors’ code is not available and we used a third-party implementation^{5}. We applied the default hyperparameters values for dropout and L2 regularization and used the same embedding dimension as for all other methods. All methods are validated according to the Mean Squared Error (MSE).
4.5 Results
The experimental results are listed in Table 1 where the best performance is in bold font. TransRev achieves the best performance on all data sets with the exception of the Kindle Store and Automotive categories. Surprisingly, HFT is more competitive than more recent approaches that also take advantage of review text. Most of these recent approaches do not include HFT in their baselines. TransRev is competitive with and often outperforms HFT on the benchmark data sets under consideration. To quantify that the rating predictions made by HFT and TransRev are significantly different we have computed the dependent t-test for paired samples and for all data sets where TransRev outperforms HFT. The p-value is always smaller than 0.01.
Sensitivity to latent dimension.
k | Baby | Digital Music | Office products | Tools& Home Improv. |
---|---|---|---|---|
4 | 1.100 | 0.782 | 0.724 | 0.880 |
8 | 1.100 | 0.782 | 0.723 | 0.878 |
16 | 1.100 | 0.782 | 0.724 | 0.879 |
32 | 1.102 | 0.785 | 0.722 | 0.888 |
64 | 1.099 | 0.787 | 0.726 | 0.888 |
4.6 Hyperparameters
We randomly selected the 4 data sets Baby, Digital Music, Office and Tools&Home Improvement from the Amazon data and evaluated different values of k for user, item and word embedding sizes. We increase k from 4 to 64 and always validate all hyperparameters, including the regularization term. Table 2 list the MSE scores. We only observe small differences in the corresponding model’s performances. This observation is in line with [21].
4.7 Visualization of the Word Embeddings
Review embeddings, which are learned from word embeddings, are learned to be good predictors of user ratings. As a consequence the learned word embeddings are correlated with the ratings. To visualize the correlation between words and ratings we proceed as follows. First, we assign a score to each word that is computed by taking the average rating of the reviews that contain the word. Second, we compute a 2-dimensional representation of the words by applying t-SNE [20] to the 16-dimensional word embeddings learned by TransRev. Figure 4 depicts these 2-dimensional word embedding vectors learned for the Amazon Beauty data set. The corresponding rating scores are indicated by the color.
The clusters we discovered in Fig. 4 are interpretable. They are meaningful with respect to the score, observing that the upper right cluster is mostly made up of words with negative connotations (e.g. horrible, useless\(\dots \)), the lower left one contains neutral words (e.g. with, products\(\dots \)) and the lower right one contains words with positive connotations (e.g. awesome, excellent\(\dots \)).
4.8 Suggesting Reviews to Users
Reviews retrieved from the Beauty (upper) and Digital Music (lower) data sets. In parenthesis the ratings associated to the reviews.
Actual test review | Closest training review in embedding space |
---|---|
skin improved (5) | makes your face feel refreshed (5) |
love it (5) | you’ll notice the difference (5) |
best soap ever (5) | I’ll never change it (5) |
it clumps (2) | gives me headaches (1) |
smells like bug repellent (3) | pantene give it up (2) |
fake fake fake do not buy (1) | seems to be harsh on my skin (2) |
saved my skin (5) | not good quality (2) |
another great release from saliva (5) | can’t say enough good things about this cd (5) |
a great collection (5) | definitive collection (5) |
sound nice (3) | not his best nor his worst (4) |
a complete massacre of an album (2) | some great songs but overall a disappointment (3) |
the very worst best of ever (1) | overall a pretty big disappointment (2) |
what a boring moment (1) | overrated but still alright (3) |
great cd (5) | a brilliant van halen debut album (5) |
We believe that more sophisticated sentence and paragraph representations might lead to better results in the review retrieval task. As discussed, a promising line of research has to do with learning representations for reviews that are aspect-specific (e.g. “ease of use” or “price”).
5 Conclusion
TransRev is a novel approach for product recommendation combining ideas from knowledge graph embedding methods, recommender systems and sentiment analysis. TransRev achieves state of the art performance on the data sets under consideration while having fewer (hyper)parameters than more recent works.
Most importantly, one main characteristic of TransRev is its ability to approximate the review representation during inference. This approximated representation can be used to retrieve reviews in the training set that are similar with respect to the overall sentiment towards the product. Such reviews can be dispatched to users as a starting point for a review, and thus lowering the barrier to writing new reviews. Given the known influence of product reviews in the purchasing choices of the users [2, 30], we think that recommender systems will benefit from such mechanism.
Footnotes
Notes
Acknowledgements
The research leading to these results has received funding from the European Union’s Horizon 2020 innovation action programme under grant agreement No 786741 – SMOOTH project. This publication reflects only the author’s views and the European Community is not liable for any use that may be made of the information contained herein.
References
- 1.Allen, R.B.: User models: theory, method, and practice. Int. J. Man Mach. Stud. 32(5), 511–543 (1990)CrossRefGoogle Scholar
- 2.
- 3.Bao, Y., Fang, H., Zhang, J.: TopicMF: simultaneously exploiting ratings and reviews for recommendation. In: AAAI, pp. 2–8 (2014)Google Scholar
- 4.Bermingham, A., Smeaton, A.F.: Classifying sentiment in microblogs: is brevity an advantage? In: CIKM, pp. 1833–1836 (2010)Google Scholar
- 5.Bordes, A., Usunier, N., García-Durán, A., Weston, J., Yakhnenko, O.: Translating embeddings for modeling multi-relational data. In: NIPS, pp. 2787–2795 (2013)Google Scholar
- 6.Breese, J.S., Heckerman, D., Kadie, C.M.: Empirical analysis of predictive algorithms for collaborative filtering. In: UAI, pp. 43–52 (1998)Google Scholar
- 7.Catherine, R., Cohen, W.W.: TransNets: learning to transform for recommendation. In: RecSys, pp. 288–296 (2017)Google Scholar
- 8.Dacrema, M.F., Cremonesi, P., Jannach, D.: Are we really making much progress? A worrying analysis of recent neural recommendation approaches. In: RecSys (2019)Google Scholar
- 9.Diao, Q., Qiu, M., Wu, C., Smola, A.J., Jiang, J., Wang, C.: Jointly modeling aspects, ratings and sentiments for movie recommendation (JMARS). In: KDD, pp. 193–202 (2014)Google Scholar
- 10.García-Durán, A., Bordes, A., Usunier, N.: Composing relationships with translations. In: EMNLP, pp. 286–290. The Association for Computational Linguistics (2015)Google Scholar
- 11.García-Durán, A., Bordes, A., Usunier, N., Grandvalet, Y.: Combining two and three-way embedding models for link prediction in knowledge bases. J. Artif. Intell. Res. 55, 715–742 (2016)MathSciNetCrossRefGoogle Scholar
- 12.Glorot, X., Bengio, Y.: Understanding the difficulty of training deep feedforward neural networks. In: AISTATS. JMLR Proceedings, vol. 9, pp. 249–256 (2010)Google Scholar
- 13.Guu, K., Miller, J., Liang, P.: Traversing knowledge graphs in vector space. In: EMNLP, pp. 318–327. The Association for Computational Linguistics (2015)Google Scholar
- 14.He, R., Kang, W., McAuley, J.: Translation-based recommendation. In: RecSys, pp. 161–169 (2017)Google Scholar
- 15.
- 16.Kadlec, R., Bajgar, O., Kleindienst, J.: Knowledge base completion: baselines strike back. arXiv preprint arXiv:1705.10744 (2017)
- 17.Koren, Y., Bell, R.M., Volinsky, C.: Matrix factorization techniques for recommender systems. IEEE Comput. 42(8), 30–37 (2009)CrossRefGoogle Scholar
- 18.Lawlor, A., Muhammad, K., Rafter, R., Smyth, B.: Opinionated explanations for recommendation systems. In: Bramer, M., Petridis, M. (eds.) Research and Development in Intelligent Systems XXXII, pp. 331–344. Springer, Cham (2015). https://doi.org/10.1007/978-3-319-25032-8_25CrossRefGoogle Scholar
- 19.Ling, G., Lyu, M.R., King, I.: Ratings meet reviews, a combined approach to recommend. In: RecSys, pp. 105–112 (2014)Google Scholar
- 20.Maaten, L., Hinton, G.: Visualizing data using t-SNE. J. Mach. Learn. Res. 9, 2579–2605 (2008)zbMATHGoogle Scholar
- 21.McAuley, J.J., Leskovec, J.: Hidden factors and hidden topics: understanding rating dimensions with review text. In: RecSys, pp. 165–172 (2013)Google Scholar
- 22.McAuley, J.J., Pandey, R., Leskovec, J.: Inferring networks of substitutable and complementary products. In: KDD, pp. 785–794 (2015)Google Scholar
- 23.McAuley, J.J., Targett, C., Shi, Q., van den Hengel, A.: Image-based recommendations on styles and substitutes. In: SIGIR, pp. 43–52 (2015)Google Scholar
- 24.Mikolov, T., Sutskever, I., Chen, K., Corrado, G.S., Dean, J.: Distributed representations of words and phrases and their compositionality. In: Advances in Neural Information Processing Systems, pp. 3111–3119 (2013)Google Scholar
- 25.Palumbo, E., Rizzo, G., Troncy, R., Baralis, E., Osella, M., Ferro, E.: An empirical comparison of knowledge graph embeddings for item recommendation. In: DL4KGS@ ESWC, pp. 14–20 (2018)Google Scholar
- 26.Peters, M.E., et al.: Deep contextualized word representations. arXiv preprint arXiv:1802.05365 (2018)
- 27.Qureshi, M.A., Greene, D.: Lit@EVE: explainable recommendation based on wikipedia concept vectors. In: Altun, Y., et al. (eds.) ECML PKDD 2017. LNCS (LNAI), vol. 10536, pp. 409–413. Springer, Cham (2017). https://doi.org/10.1007/978-3-319-71273-4_41CrossRefGoogle Scholar
- 28.Rendle, S., Freudenthaler, C., Schmidt-Thieme, L.: Factorizing personalized Markov chains for next-basket recommendation. In: WWW, pp. 811–820 (2010)Google Scholar
- 29.Rennie, J.D.M., Srebro, N.: Fast maximum margin matrix factorization for collaborative prediction. In: ICML. ACM International Conference Proceeding Series, vol. 119, pp. 713–719 (2005)Google Scholar
- 30.
- 31.Seo, S., Huang, J., Yang, H., Liu, Y.: Interpretable convolutional neural networks with dual local and global attention for review rating prediction. In: RecSys, pp. 297–305 (2017)Google Scholar
- 32.Seo, S., Huang, J., Yang, H., Liu, Y.: Representation learning of users and items for review rating prediction using attention-based convolutional neural network. In: 3rd International Workshop on Machine Learning Methods for Recommender Systems (MLRec) (2017)Google Scholar
- 33.Tay, Y., Anh Tuan, L., Hui, S.C.: Latent relational metric learning via memory-based attention for collaborative ranking. In: Proceedings of the 2018 World Wide Web Conference on World Wide Web, pp. 729–739. International World Wide Web Conferences Steering Committee (2018)Google Scholar
- 34.Tay, Y., Tuan, L.A., Hui, S.C.: Multi-pointer co-attention networks for recommendation. In: KDD (2018)Google Scholar
- 35.
- 36.Wu, C., Beutel, A., Ahmed, A., Smola, A.J.: Explaining reviews and ratings with PACO: poisson additive co-clustering. In: WWW (Companion Volume), pp. 127–128 (2016)Google Scholar
- 37.Zhang, Y., Lai, G., Zhang, M., Zhang, Y., Liu, Y., Ma, S.: Explicit factor models for explainable recommendation based on phrase-level sentiment analysis. In: SIGIR, pp. 83–92 (2014)Google Scholar
- 38.Zheng, L., Noroozi, V., Yu, P.S.: Joint deep modeling of users and items using reviews for recommendation. In: WSDM, pp. 425–434 (2017)Google Scholar