{"title": "Learning Disentangled Representations for Recommendation", "book": "Advances in Neural Information Processing Systems", "page_first": 5711, "page_last": 5722, "abstract": "User behavior data in recommender systems are driven by the complex interactions of many latent factors behind the users\u2019 decision making processes. The factors are highly entangled, and may range from high-level ones that govern user intentions, to low-level ones that characterize a user\u2019s preference when executing an intention. Learning representations that uncover and disentangle these latent factors can bring enhanced robustness, interpretability, and controllability. However, learning such disentangled representations from user behavior is challenging, and remains largely neglected by the existing literature. In this paper, we present the MACRo-mIcro Disentangled Variational Auto-Encoder (MacridVAE) for learning disentangled representations from user behavior. Our approach achieves macro disentanglement by inferring the high-level concepts associated with user intentions (e.g., to buy a shirt or a cellphone), while capturing the preference of a user regarding the different concepts separately. A micro-disentanglement regularizer, stemming from an information-theoretic interpretation of VAEs, then forces each dimension of the representations to independently reflect an isolated low-level factor (e.g., the size or the color of a shirt). Empirical results show that our approach can achieve substantial improvement over the state-of-the-art baselines. We further demonstrate that the learned representations are interpretable and controllable, which can potentially lead to a new paradigm for recommendation where users are given fine-grained control over targeted aspects of the recommendation lists.", "full_text": "Learning Disentangled Representations for\n\nRecommendation\n\nJianxin Ma1,2\u2217, Chang Zhou1\u2217, Peng Cui2, Hongxia Yang1, Wenwu Zhu2\n\n1Alibaba Group, 2Tsinghua University\n\nmajx13fromthu@gmail.com, ericzhou.zc@alibaba-inc.com,\n\ncuip@tsinghua.edu.cn, yang.yhx@alibaba-inc.com, wwzhu@tsinghua.edu.cn\n\nAbstract\n\nUser behavior data in recommender systems are driven by the complex interactions\nof many latent factors behind the users\u2019 decision making processes. The factors are\nhighly entangled, and may range from high-level ones that govern user intentions,\nto low-level ones that characterize a user\u2019s preference when executing an intention.\nLearning representations that uncover and disentangle these latent factors can bring\nenhanced robustness, interpretability, and controllability. However, learning such\ndisentangled representations from user behavior is challenging, and remains largely\nneglected by the existing literature. In this paper, we present the MACRo-mIcro\nDisentangled Variational Auto-Encoder (MacridVAE) for learning disentangled\nrepresentations from user behavior. Our approach achieves macro disentanglement\nby inferring the high-level concepts associated with user intentions (e.g., to buy\na shirt or a cellphone), while capturing the preference of a user regarding the\ndifferent concepts separately. A micro-disentanglement regularizer, stemming\nfrom an information-theoretic interpretation of VAEs, then forces each dimension\nof the representations to independently re\ufb02ect an isolated low-level factor (e.g.,\nthe size or the color of a shirt). Empirical results show that our approach can\nachieve substantial improvement over the state-of-the-art baselines. We further\ndemonstrate that the learned representations are interpretable and controllable,\nwhich can potentially lead to a new paradigm for recommendation where users are\ngiven \ufb01ne-grained control over targeted aspects of the recommendation lists.\n\n1\n\nIntroduction\n\nLearning representations that re\ufb02ect users\u2019 preference, based chie\ufb02y on user behavior, has been a\ncentral theme of research on recommender systems. Despite their notable success, the existing user\nbehavior-based representation learning methods, such as the recent deep approaches [49, 32, 31, 52,\n11, 18], generally neglect the complex interaction among the latent factors behind the users\u2019 decision\nmaking processes. In particular, the latent factors can be highly entangled, and range from macro\nones that govern the intention of a user during a session, to micro ones that describe at a granular level\na user\u2019s preference when implementing a speci\ufb01c intention. The existing methods fail to disentangle\nthe latent factors, and the learned representations are consequently prone to mistakenly preserve the\nconfounding of the factors, leading to non-robustness and low interpretability.\nDisentangled representation learning, which aims to learn factorized representations that uncover\nand disentangle the latent explanatory factors hidden in the observed data [3], has recently gained\nmuch attention. Not only can disentangled representations be more robust, i.e., less sensitive to the\nmisleading correlations presented in the limited training data, the enhanced interpretability also \ufb01nds\n\n\u2217Equal contribution. Work done at Alibaba.\n\n33rd Conference on Neural Information Processing Systems (NeurIPS 2019), Vancouver, Canada.\n\n\fFigure 1: Our framework. Macro disentanglement is achieved by learning a set of prototypes, based\non which the user intention related with each item is inferred, and then capturing the preference of a\nuser about the different intentions separately. Micro disentanglement is achieved by magnifying the\nKL divergence, from which a term that penalizes total correlation can be separated, with a factor of \u03b2.\n\ndirect application in recommendation-related tasks, such as transparent advertising [33], customer-\nrelationship management, and explainable recommendation [51, 17]. Moreover, the controllability\nexhibited by many disentangled representations [19, 14, 10, 8, 9, 25] can potentially bring a new\nparadigm for recommendation, by giving users explicit control over the recommendation results and\nproviding a more interactive experience. However, the existing efforts on disentangled representation\nlearning are mainly from the \ufb01eld of computer vision [28, 15, 20, 30, 53, 14, 10, 39, 19].\nLearning disentangled representations based on user behavior data, a kind of discrete relational\ndata that is fundamentally different from the well-researched image data, is challenging and largely\nunexplored. Speci\ufb01cally, it poses two challenges. First, the co-existence of macro and micro factors\nrequires us to to separate the two levels when performing disentanglement, in a way that preserves\nthe hierarchical relationships between an intention and the preference about the intention. Second,\nthe observed user behavior data, e.g., user-item interactions, are discrete and sparse in nature, while\nthe learned representations are continuous. This implies that the majority of the points in the high-\ndimensional representation space will not be associated with any behavior, which is especially\nproblematic when one attempts to investigate the interpretability of an isolated dimension by varying\nthe value of the dimension while keeping the other dimensions \ufb01xed.\nIn this paper, we propose the MACRo-mIcro Disentangled Variational Auto-Encoder (MacridVAE)\nfor learning disentangled representations based on user behavior. Our approach explicitly models\nthe separation of macro and micro factors, and performs disentanglement at each level. Macro\ndisentanglement is achieved by identifying the high-level concepts associated with user intentions,\nand separately learning the preference of a user regarding the different concepts. A regularizer\nfor micro disentanglement, derived by interpreting VAEs [27, 44] from an information-theoretic\nperspective, is then strengthened so as to force each individual dimension to re\ufb02ect an independent\nmicro factor. A beam-search strategy, which handles the con\ufb02ict between sparse discrete observations\nand dense continuous representations by \ufb01nding a smooth trajectory, is then proposed for investi-\ngating the interpretability of each isolated dimension. Empirical results show that our approach can\nachieve substantial improvement over the state-of-the-art baselines. And the learned disentangled\nrepresentations are demonstrated to be interpretable and controllable.\n\n2 Method\n\nIn this section, we present our approach for learning disentangled representations from user behaivor.\n\n2.1 Notations and Problem Formulation\nA user behavior dataset D consists of the interactions between N users and M items. The interaction\nbetween the uth user and the ith item is denoted by xu,i \u2208 {0, 1}, where xu,i = 1 indicates that user\nu explicitly adopts item i, whereas xu,i = 0 means there is no recorded interaction between the two.\nFor convenience, we use xu = {xu,i : xu,i = 1} to represent the items adopted by user u. The goal\nis to learn user representations {zu}N\nu=1 that achieves both macro and micro disentanglement. We\nuse \u03b8 to denote the set that contains all the trainable parameters of our model.\n\n2\n\nEncoder621477325172643535461\u00b5(3)\u00b5(2)(1)(2)(3)z(1)\u00b5(1)z(3)N(\u00b5(1),(1))N(\u00b5(3),(3))z(2)N(\u00b5(2),(2))N(0,20I)N(0,20I)N(0,20I)DKLDKLDKLReconstruction LossEncoderEncoderDecoderDecoderDecoderPrototypeyy Positive ItemNegative ItemNot Encoded\fMacro disentanglement Users may have very diverse interests, and interact with items that belong\nto many high-level concepts, e.g., product categories. We aim to achieve macro disentanglement,\nu ] \u2208 Rd(cid:48)\nu ; . . . ; z(K)\nby learning a factorized representation of user u, namely zu = [z(1)\n, where\nd(cid:48) = Kd, assuming that there are K high-level concepts. The kth component z(k)\nu \u2208 Rd is for\ncapturing the user\u2019s preference regarding the kth concept. Additionally, we infer a set of one-hot\nvectors C = {ci}M\ni=1 for the items, where ci = [ci,1; ci,2; . . . ; ci,K]. If item i belongs to concept k,\nthen ci,k = 1 and ci,k(cid:48) = 0 for any k(cid:48)\n\nu=1 and C unsupervisedly.\n\nu ; z(2)\n\n(cid:54)= k. We jointly infer {zu}N\n\nMicro disentanglement High-level concepts correspond to the intentions of a user, e.g., to buy\nclothes or a cellphone. We are also interested in disentangling a user\u2019s preference at a more granular\nlevel regarding the various aspects of an item. For example, we would like the different dimensions\nof z(k)\n\nu to individually capture the user\u2019s preferred sizes, colors, etc., if concept k is clothing.\n\n2.2 Model\n\nWe start by proposing a generative model that encourages macro disentanglement. For a user u, our\ngenerative model assumes that the observed data are generated from the following distribution:\n\n(cid:21)\n\n,\n\n(1)\n\n(2)\n\n(cid:20)(cid:90)\n\n(cid:89)\n\np\u03b8(xu) = Ep\u03b8 (C)\n\np\u03b8 (xu | zu, C) p\u03b8(zu) dzu\n\np\u03b8 (xu | zu, C) =\n\np\u03b8(xu,i | zu, C).\n\nxu,i\u2208xu\n\nThe meanings of xu, zu, C are described in the previous subsection. We have assumed p\u03b8(zu) =\np\u03b8(zu | C) in the \ufb01rst equation, i.e., zu and C are generated by two independent sources. Note that\nZu =(cid:80)M\nci = [ci,1; ci,2; . . . ; ci,K] is one-hot, since we assume that item i belongs to exactly one concept. And\np\u03b8(xu,i | zu, C) = Z\u22121\n\u03b8 (z(k)\nu ) is a categorical distribution over the M items, where\nu \u00b7\n\u03b8 : Rd \u2192 R+ is a shallow neural network that estimates\nhow much a user with a given preference is interested in item i. We use sampeld softmax [23] to\nestimate Zu based on a few sampled items when M is very large.\n\n(cid:80)K\n(cid:80)K\nk=1 ci,k \u00b7 g(i)\nu ) and g(i)\n\u03b8 (z(k)\nk=1 ci,k \u00b7 g(i)\n\ni=1\n\nMacro disentanglement We assume above that the user representation zu is suf\ufb01cient for predict-\ning how the user will interact with the items. And we further assume that using the kth component\nz(k)\nu alone is already suf\ufb01cient if the prediction is about an item from concept k. This design explicitly\nencourages z(k)\nu to capture preference regarding only the kth concept, as long as the inferred concept\nassignment matrix C is meaningful. We will describe later the implementation details of p\u03b8(C),\np\u03b8(zu) and g(i)\nu ). Nevertheless, we note that p\u03b8(C) requires careful design to prevent mode\ncollapse, i.e., the degenerate case where almost all items are assigned to a single concept.\n\n\u03b8 (z(k)\n\noptimize \u03b8 by maximizing a lower bound of(cid:80)\nln p\u03b8(xu) \u2265 Ep\u03b8 (C)\n\n(cid:2)Eq\u03b8 (zu|xu,C)[ln p\u03b8(xu | zu, C)] \u2212 DKL(q\u03b8(zu | xu, C)(cid:107)p\u03b8(zu))(cid:3) .\n\nVariational inference We follow the variational auto-encoder (VAE) paradigm [27, 44], and\nu ln p\u03b8(xu), where ln p\u03b8(xu) is bounded as follows:\n(3)\nSee the supplementary material for the derivation of the lower bound. Here we have introduced\na variational distribution q\u03b8(zu | xu, C), whose implementation also encourages macro disentan-\nglement and will be presented later. The two expectations, i.e., Ep\u03b8 (C)[\u00b7] and Eq\u03b8 (zu|xu,C)[\u00b7], are\nintractable, and are therefore estimated using the Gumbel-Softmax trick [22, 41] and the Gaussian\nre-parameterization trick [27], respectively. Once the training procedure is \ufb01nished, we use the mode\nof p\u03b8(C) as C, and the mode of q\u03b8(zu | xu, C) as the representation of user u.\n(cid:81)d\nMicro disentanglement A natural strategy to encourage micro disentanglement is to force statisti-\ncal independence between the dimensions, i.e., to force q\u03b8(z(k)\nu,j | C), so that\nFortunately, the Kullback\u2013Leibler (KL) divergence term in the lower bound above does provide a\nway to encourage independence. Speci\ufb01cally, the KL term of our model can be rewritten as:\nEpdata(xu) [DKL(q\u03b8(zu | xu, C)(cid:107)p\u03b8(zu))] = Iq(xu; zu) + DKL(q\u03b8(zu | C)(cid:107)p\u03b8(zu)).\n\neach dimension describes an isolated factor. Here q\u03b8(zu | C) =(cid:82) q\u03b8(zu | xu, C)pdata(xu) dxu.\n\nu | C) \u2248\n\nj=1 q\u03b8(z(k)\n\n(4)\n\n3\n\n\findependence between the dimensions, if we choose a prior that satis\ufb01es p\u03b8(zu) =(cid:81)d(cid:48)\n\nSee the supplementary material for the proof. Similar decomposition of the KL term has been\nnoted for the original VAEs previously [1, 25, 9]. Penalizing the latter KL term would encourage\nj=1 p\u03b8(zu,j).\nOn the other hand, the former term Iq(xu; zu) is the mutual information between xu and zu under\nq\u03b8(zu | xu, C)\u00b7pdata(xu). Penalizing Iq(xu; zu) is equivalent to applying the information bottleneck\nprinciple [47, 2], which encourages zu to ignore as much noise in the input as it can and to focus\non merely the essential information. We therefore follow \u03b2-VAE [19], and strengthen these two\nregularization terms by a factor of \u03b2 (cid:29) 1, which brings us to the following training objective:\n\n(cid:2)Eq\u03b8 (zu|xu,C)[ln p\u03b8(xu | zu, C)] \u2212 \u03b2 \u00b7 DKL(q\u03b8(zu | xu, C)(cid:107)p\u03b8(zu))(cid:3) .\n\nEp\u03b8 (C)\n\n(5)\n\nImplementation\n\ni=1 \u2208 RM\u00d7d used by the decoder, M context representations {ti}M\n\n2.3\nIn this section, we describe the implementation of p\u03b8(C), p\u03b8(xu,i | zu, C) (the decoder), p\u03b8(zu)\n(the prior), q\u03b8(zu | xu, C) (the encoder), and propose an ef\ufb01cient strategy to combat mode collapse.\nk=1 \u2208 RK\u00d7d, M item\nThe parameters \u03b8 of our implementation include: K concept prototypes {mk}K\ni=1 \u2208 RM\u00d7d\nrepresentations {hi}M\nused by the encoder, and the parameters of a neural network fnn : Rd \u2192 R2d. We optimize \u03b8 to\nmaximize the training objective (see Equation 5) using Adam [26].\n(cid:81)M\nPrototype-based concept assignment A straightforward approach would be to assume p\u03b8(C) =\ni=1 p(ci) and parameterize each categorical distribution p(ci) with its own set of K \u2212 1 parameters.\nThis approach, however, would result in over-parameterization and low sample ef\ufb01ciency. We instead\npropose a prototype-based implementation. To be speci\ufb01c, we introduce K concept prototypes\ni=1 from the decoder. We then assume ci is a\n{mk}K\none-hot vector drawn from the following categorical distribution p\u03b8(ci):\n\nk=1 and reuse the item representations {hi}M\n\n\u03c4 , 1\n\nsi,k = COSINE(hi, mk)/\u03c4,\n\nci \u223c CATEGORICAL (SOFTMAX([si,1; si,2; . . . ; si,K])) ,\n\n(6)\nwhere COSINE(a, b) = a(cid:62)b/((cid:107)a(cid:107)2 (cid:107)b(cid:107)2) is the cosine similarity, and \u03c4 is a hyper-parameter that\n\u03c4 ]. We set \u03c4 = 0.1 to obtain a more skewed distribution.\nscales the similarity from [\u22121, 1] to [\u2212 1\nPreventing mode collapse We use cosine similarity, instead of the inner product similarity adopted\nby most existing deep learning methods [32, 31, 18]. This choice is crucial for preventing mode\ncollapse. In fact, with inner product, the majority of the items are highly likely to be assigned\nto a single concept mk(cid:48) that has an extremely large norm, i.e., (cid:107)mk(cid:48)(cid:107)2 \u2192 \u221e, even when the\nitems {hi}M\ni=1 correctly form K clusters in the high-dimensional Euclidean space. And we observe\nempirically that this phenomenon does occur frequently with inner product (see Figure 2e). In\ncontrast, cosine similarity avoids this degenerate case due to the normalization. Moreover, cosine\nsimilarity is related with the Euclidean distance on the unit hypersphere, and the Euclidean distance\nis a proper metric that is more suitable for inferring the cluster structure, compared to inner product.\n\nu ; z(2)\n\n(cid:80)K\nu ; . . . ; z(K)\nk=1 ci,k \u00b7 g(i)\n\nu ) = exp(COSINE(z(k)\nu }N\n\nDecoder The decoder predicts which item out of the M ones is mostly likely to be clicked by a user,\nwhen given the user\u2019s representation zu = [z(1)\nu ] and the one-hot concept assignments\n\u03b8 (z(k)\nu ) is a categorical distribution over\ni=1. We assume that p\u03b8(xu,i | zu, C) \u221d\n{ci}M\n\u03b8 (z(k)\nthe M items, and de\ufb01ne g(i)\nu , hi)/\u03c4 ). This design implies that {hi}M\ni=1\nwill be micro-disentangled if {z(k)\nu=1 is micro-disentangled, as the two\u2019s dimensions are aligned.\nPrior & Encoder The prior p\u03b8(zu) needs to be factorized in order to achieve micro disentan-\n0I). The encoder q\u03b8(zu | xu, C) is for comput-\nglement. We therefore set p\u03b8(zu) to N (0, \u03c32\ning the representation of a user when given the user\u2019s behavior data xu. The encoder main-\ntains an additional set of context representations {ti}M\ni=1, rather than reusing the item represen-\ntations {hi}M\ni=1 from the decoder, which is a common practice in the literature [32]. We assume\n| xu, C) as a multivariate\nnormal distribution with a diagonal covariance matrix N (\u00b5(k)\nu )]2), where the mean and\n\nq\u03b8(zu | xu, C) = (cid:81)K\n\n| xu, C), and represent each q\u03b8(z(k)\nu , [diag(\u03c3(k)\n\nk=1 q\u03b8(z(k)\n\nu\n\nu\n\n4\n\n\f(cid:18)\n\n1\n2\n\n\u2212\n\nb(k)\nu\n\n(cid:19)\n\n. (7)\n\na(k)\nu\n(cid:107)a(k)\nu (cid:107)2\n\nthe standard deviation are parameterized by a neural network fnn : Rd \u2192 R2d:\n(a(k)\n\n, \u03c3(k)\n\nu ) = fnn\n\nu , b(k)\n\nu =\n\nu \u2190 \u03c30 \u00b7 exp\n\n\uf8f6\uf8f8 , \u00b5(k)\n\n\uf8eb\uf8ed(cid:80)\n(cid:113)(cid:80)\n\ni:xu,i=+1 ci,k \u00b7 ti\ni:xu,i=+1 c2\ni,k\n\nThe neural network fnn(\u00b7) captures nonlinearity, and is shared across the K components. We\nnormalize the mean, so as to be consistent with the use of cosine similarity which projects the\nrepresentations onto a unit hypersphere. Note that \u03c30 should be set to a small value, e.g., around 0.1,\nsince the learned representations are now normalized.\n\n2.4 User-Controllable Recommendation\n\nThe controllability enabled by the disentangled representations can bring a new paradigm for recom-\nmendation. It allows a user to interactively search for items that are similar to an initial item except for\nsome controlled aspects, or to explicitly adjust the disentangeld representation of his/her preference,\nlearned by the system from his/her past behaviors, to actually match the current preference. Here we\nformalize the task of user-controllable recommendation, and illustrate a possible solution.\nTask de\ufb01nition Let h\u2217 \u2208 Rd be the representation to be altered, which can be initialized as either\nan item representation or a component of a user representation. The task is to gradually alter its jth\ndimension h\u2217,j, while retrieving items whose representations are similar to the altered representation.\nThis task is nontrivial, since usually no item will have exactly the same representation as the altered\none, especially when we want the transition to be smooth, monotonic, and thus human-understandable.\n\nSolution Here we illustrate our approach to this task. We \ufb01rst probe the suitable range (a, b) for\nh\u2217,j. Let us assume that prototype k\u2217 is the prototype closest to h\u2217. The range (a, b) is decided such\nthat: prototype k\u2217 remains the prototype closest to h\u2217 if and only if h\u2217,j \u2208 (a, b). We can decide each\nendpoint of the range using binary search. We then divide the range (a, b) into B subranges, a = a0 <\na1 < a2 . . . < aB = b. We ensure that the subranges contain roughly the same number of items from\ndecide the B items by maximizing(cid:80)\nconcept k\u2217 when dividing (a, b) . Finally, we aim to retrieve B items {it}B\nt=1 \u2208 {1, 2, . . . , M}B\nthat belong to concept k\u2217, each from one of the B subranges, i.e., hit,j \u2208 (at\u22121, at]. We thus\nt(cid:48) ,\u2212j )\nwhere hi,\u2212j = [hi,1; hi,2; . . . ; hi,j\u22121; hi,j+1; . . . ; hi,d] \u2208 Rd\u22121 and \u03b3 is a hyper-parameter. We\napproximately solve this maximization problem sequentially using beam search [36].\nIntuitively, selecting items from the B subranges ensures that the items change monotonously in\nterms of the jth dimension. On the other hand, the \ufb01rst term in the maximization problem forces the\nretrieved items to be similar with the initial item in terms of the dimensions other than j, while the\nsecond term encourages any two retrieved items to be similar in terms of the dimensions other than j.\n\nCOSINE(hit,\u2212j ,h\u2217,\u2212j )\n\n1\u2264t<t(cid:48)\u2264B e\n\nCOSINE(hit,\u2212j ,hi\n\n1\u2264t\u2264B e\n\n(cid:80)\n\n+\u03b3\u00b7\n\n\u03c4\n\n\u03c4\n\n,\n\n3 Empirical Results\n\n3.1 Experimental Setup\n\nDatasets We conduct our experiments on \ufb01ve real-world datasets. Speci\ufb01cally, we use the large-\nscale Net\ufb02ix Prize dataset [4], and three MovieLens datasets of different scales (i.e., ML-100k,\nML-1M, and ML-20M) [16]. We follow MultVAE [32], and binarize these four datasets by keeping\nratings of four or higher while only keeping users who have watched at least \ufb01ve movies. We\nadditionally collect a dataset, named AliShop-7C 2, from Alibaba\u2019s e-commerce platform Taobao.\nAliShop-7C contains user-item interactions associated with items from seven categories, as well as\nitem attributes such as titles and images. Every user in this dataset clicks items from at least two\ncategories. The category labels are used for evaluation only, and not for training.\n\nBaselines We compare our approach with MultDAE [32] and \u03b2-MultVAE [32], the two state-of-\nthe-art methods for collaborative \ufb01ltering. In particular, \u03b2-MultVAE is similar to \u03b2-VAE [19], and\nhas a hyper-parameter \u03b2 that controls the strength of disentanglement. However, \u03b2-MultVAE does\nnot learn disentangled representations, because it requires \u03b2 (cid:28) 1 to perform well.\n\n2The dataset and our code are at https://jianxinma.github.io/disentangle-recsys.html.\n\n5\n\n\fTable 1: Collaborative \ufb01ltering. All methods are constrained to have around 2M d parameters, where\nM is the number of items and d is the dimension of each item representation. We set d = 100.\n\nDataset\n\nMethod\nAliShop-7C MultDAE\n\nML-100k\n\nML-1M\n\nML-20M\n\nNet\ufb02ix\n\nOurs\n\nOurs\n\nOurs\n\nNDCG@100\n\nOurs\n\nOurs\n\n0.23923 (\u00b10.00380)\n\u03b2-MultVAE 0.23875 (\u00b10.00379)\n0.29148 (\u00b10.00380)\nMultDAE\n0.24487 (\u00b10.02738)\n\u03b2-MultVAE 0.27484 (\u00b10.02883)\n0.28895 (\u00b10.02739)\nMultDAE\n0.40453 (\u00b10.00799)\n\u03b2-MultVAE 0.40555 (\u00b10.00809)\n0.42740 (\u00b10.00789)\nMultDAE\n0.41900 (\u00b10.00209)\n\u03b2-MultVAE 0.41113 (\u00b10.00212)\n0.42496 (\u00b10.00212)\n0.37450 (\u00b10.00095)\nMultDAE\n\u03b2-MultVAE 0.36291 (\u00b10.00094)\n0.37987 (\u00b10.00096)\n\nMetrics\n\nRecall@20\n\n0.15242 (\u00b10.00305)\n0.15040 (\u00b10.00302)\n0.18616 (\u00b10.00317)\n0.23794 (\u00b10.03605)\n0.24838 (\u00b10.03294)\n0.30951 (\u00b10.03808)\n0.34382 (\u00b10.00961)\n0.33960 (\u00b10.00919)\n0.36046 (\u00b10.00947)\n0.39169 (\u00b10.00271)\n0.38263 (\u00b10.00273)\n0.39649 (\u00b10.00271)\n0.33982 (\u00b10.00123)\n0.32792 (\u00b10.00122)\n0.34587 (\u00b10.00124)\n\nRecall@50\n\n0.24892 (\u00b10.00391)\n0.24589 (\u00b10.00387)\n0.30256 (\u00b10.00397)\n0.32279 (\u00b10.04070)\n0.35270 (\u00b10.03927)\n0.41309 (\u00b10.04503)\n0.46781 (\u00b10.01032)\n0.45825 (\u00b10.01039)\n0.49039 (\u00b10.01029)\n0.53054 (\u00b10.00285)\n0.51975 (\u00b10.00289)\n0.52901 (\u00b10.00284)\n0.43247 (\u00b10.00126)\n0.41960 (\u00b10.00125)\n0.43478 (\u00b10.00125)\n\nHyper-parameters We constrain the number of learnable parameters to be around 2M d for each\nmethod so as to ensure fair comparison, which is equivalent to using d-dimensional representations\nfor the M items. Note that all the methods under investigation use two sets of item representations,\nand we do not constrain the dimension of user representations since they are not parameters. We set\nd = 100 unless otherwise speci\ufb01ed. We \ufb01x \u03c4 to 0.1. We tune the other hyper-parameters of both our\napproach\u2019s and our baselines\u2019 automatically using the TPE method [6] implemented by Hyepropt [5].\n\n3.2 Recommendation Performance\n\nWe evaluate the performance of our approach on the task of collaborative \ufb01ltering for implicit feedback\ndatasets [21], one of the most common settings for recommendation. We follow the experiment\nprotocol established by the previous work [32] strictly, and use the same preprocessing procedure as\nwell as evaluation metrics. The results on the \ufb01ve datasets are listed in Table 1.\nWe observe that our approach outperforms the baselines signi\ufb01cantly, especially on small, sparse\ndatasets. The improvement is likely due to two desirable properties of our approach. Firstly, macro\ndisentanglement not only allows us to accurately represent the diverse interests of a user using\nthe different components, but also alleviates data sparsity by allowing a rarely visited item to\nborrow information from other items of the same category, which is the motivation behind many\nhierarchical methods [50, 38]. Secondly, as we will show in Section 3.4, the dimensions of the\nrepresentations learned by our approach are highly disentangled, i.e., independent, thanks to the\nmicro disentanglement regularizer, which leads to more robust performance.\n\n3.3 Macro Disentanglement\n\nWe visualize the high-dimensional representations learned by our approach on AliShop-7C in order to\nqualitatively examine to which degree our approach can achieve macro disentanglement. Speci\ufb01cally,\nwe set K to seven, i.e., the number of ground-truth categories, when training our model. We visualize\nthe item representations and the user representations together using t-SNE [40], where we treat the\nK components of a user as K individual points and keep only the two components that have the\ni:xu,i>0 ci,k, where ci,k is\n\nhighest con\ufb01dence levels. The con\ufb01dence of component k is de\ufb01ned as(cid:80)\n\nthe value inferred by our model, rather than the ground-truth. The results are shown in Figure 2.\n\nInterpretability Figure 2c, which shows the clusters inferred based on the prototypes, is rather\nsimilar to Figure 2d that shows the ground-truth categories, despite the fact that our model is trained\n\n6\n\n\f(a) Items and users. Item i is colored according to arg maxk ci,k,\ni.e., the inferred category. Each component of a user is treated as an\nindividual point, and the kth component is colored according to k.\n\n(b) Users only, colored in the\nsame way as Figure 2a.\n\n(c) Items only, colored in the same\nway as Figure 2a.\n\n(d) Items only, colored according\nto their ground-truth categories.\n\n(e) Items, obtained by training a\nnew model that uses inner product\ninstead of cosine, colored accord-\ning to the value of arg maxk ci,k.\nFigure 2: The discovered clusters of items (see Figure 2c), learned unsupervisedly, align well with\nthe ground-truth categories (see Figure 2d, where the color order is chosen such that the connections\nbetween the ground-truth categories and the learned clusters are easy to verify). Figure 2e highlights\nthe importance of using cosine similarity, rather than inner product, to combat mode collapse.\n\n(a) Bag size.\n\n(b) Bag color.\n\n(c) Styles of phone cases.\n\n(d) Bag size. The same dimension\nas Figure 3a.\nFigure 3: Starting from an item representation, we gradually alter the value of a target dimension,\nand list the items that have representations similar to the altered representations (see Subsection 2.4).\n\n(e) Bag color. The same dimension\nas Figure 3b.\n\n(f) Chicken \u2192 beef \u2192 mutton \u2192\nseafood.\n\nwithout the ground-truth category labels. This demonstrates that our approach is able to discover and\ndisentangle the macro structures underlying the user behavior data in an interpretable way. Moreover,\nthe components of the user representations are near the correct cluster centers (see Figure 2a and\nFigure 2b), and are hence likely capturing the users\u2019 separate preferences for different categories.\n\nCosine vs. inner product To highlight the necessity of using cosine similarity instead of the more\ncommonly used inner product similarity, we additionally train a new model that uses inner product\nin place of cosine, and visualize the learned item representations in Figure 2e. With inner product,\nthe majority of the items are assigned to the same prototype (see Figure 2e). In comparison, all\nseven prototypes learned by the cosine-based model are assigned a signi\ufb01cant number of items (see\nFigure 2c). This \ufb01nding supports our claim that a proper metric space, such as the one implied by the\ncosine similarity, is important for preventing mode collapse.\n\n3.4 Micro Disentanglement\nIndependence We vary the hyper-parameters related with micro disentanglement (\u03b2 and \u03c30 for our\napproach, \u03b2 for \u03b2-MultVAE), and plot in Figure 4 the relationship between the level of independence\nachieved and the recommendation performance. Each method is evaluated with 2,000 randomly\n\n7\n\n\fFigure 4: Micro disentanglement vs. recommendation performance. (d, d(cid:48)) indicates d-dimensional\nitem representations and d(cid:48)-dimensional user representations. Note that d(cid:48) = Kd. We observe that\n(1) our approach outperforms the baselines in terms of both performance and micro disentanglement,\nand (2) macro disentanglement bene\ufb01ts micro disentanglement, as K = 7 is better than K = 1.\n\n(cid:80)\nsampled con\ufb01gurations on ML-100k. We quantify the level of independence achieved by a set of\nd-dimensional representations using 1 \u2212 2\n1\u2264i<j\u2264d |corri,j|, where corri,j is the correlation\nbetween dimension i and j. Figure 4 suggests that high performance is in general associated with a\nrelatively high level of independence. And our approach achieves a much higher level of independence\nthan \u03b2-MultVAE. In addition, the improvement brought by using K = 7 instead of K = 1 reveals\nthat macro disentanglement can possibly help improve micro disentanglement.\n\nd(d\u22121)\n\nInterpretability We train our model with K = 7, d = 10, \u03b2 = 50 and \u03c30 = 0.3, on AliShop-7C,\nand investigate the interpretability of the dimensions using the approach illustrated in Subsection 2.4.\nIn Figure 3, we list some representative dimensions that have human-understandable semantics.\nThese examples suggest that our approach has the potential to give users \ufb01ne-grained control over\ntargeted aspects of the recommendation lists. However, we note that not all dimensions are human-\nunderstandable. In addition, as pointed out by Locatello et al. [34], well-trained interpretable models\ncan only be reliably identi\ufb01ed with the help of external knowledge, e.g., item attributes. We thus\nencourage future efforts to focus more on (semi-)supervised methods [35].\n\n4 Related Work\n\nLearning representations from user behavior Learning from user behavior has been a central\ntask of recommender systems since the advent of collaborative \ufb01ltering [43, 42, 46, 12, 21]. Early\nattempts apply matrix factorization [29, 45], while the more recent deep learning methods [49, 32,\n31, 52, 11, 18] achieve massive improvement by learning highly informative representations. The\nentanglement of the latent factors behind user behavior, however, is mostly neglected by the black-box\nrepresentation learning process adopted by the majority of the existing methods. To the extent of our\nknowledge, we are the \ufb01rst to study disentangled representation learning on user behavior data.\n\nDisentangled representation learning Disentangled representation learning aims to identify and\ndisentangle the underlying explanatory factors [3]. \u03b2-VAE [19] demonstrates that disentanglement\ncan emerge once the KL divergence term in the VAE [27] objective is aggressively penalized. Later\napproaches separate the information bottleneck term [48, 47] and the total correlation term, and\nachieve a greater level of disentanglement [9, 25, 8]. Though a few existing approaches [14, 10, 7,\n13, 24] do notice that a dataset can contain samples from different concepts, i.e., follow a mixture\ndistribution, their settings are fundamentally different from ours. To be speci\ufb01c, these existing\napproaches assume that each instance is from a concept, while we assume that each instance interacts\nwith objects from different concepts. The majority of the existing efforts are from the \ufb01eld of computer\nvision [28, 15, 20, 30, 53, 14, 10, 39, 19]. Disentangled representation learning on relational data,\nsuch as graph-structured data, was not explored until recently [37]. This work focus on disentangling\nuser behavior, another kind of relational data commonly seen in recommender systems.\n\n5 Conclusions\n\nIn this paper, we studied the problem of learning disentangled representations from user behavior,\nand presented our approach that performs disentanglement at both a macro and a micro level. An\ninteresting direction for future research is to explore novel applications that can be enabled by the\ninterpretability and controllability brought by the disentangled representations.\n\n8\n\n0.140.160.180.200.220.240.26Performance (Recall@20)0.20.30.40.50.60.70.80.9Disentanglement (Uncorrelatedness)Representations = Items'0.140.160.180.200.220.240.26Performance (Recall@20)Representations = Users'MethodOurs(100,700)Ours(100,100)beta-MultVAE(100,700)beta-MultVAE(100,100)\fAcknowledgments\n\nThe authors from Tsinghua University are supported in part by National Program on Key Basic\nResearch Project (No. 2015CB352300), National Key Research and Development Project (No.\n2018AAA0102004), National Natural Science Foundation of China (No. 61772304, No. 61521002,\nNo. 61531006, No. U1611461), Beijing Academy of Arti\ufb01cial Intelligence (BAAI), and the Young\nElite Scientist Sponsorship Program by CAST. All opinions, \ufb01ndings, and conclusions in this paper\nare those of the authors and do not necessarily re\ufb02ect the views of the funding agencies.\n\nReferences\n[1] Alessandro Achille and Stefano Soatto. Information dropout: Learning optimal representations\nthrough noisy computation. IEEE transactions on pattern analysis and machine intelligence,\n40(12):2897\u20132905, 2018.\n\n[2] Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. Deep variational\n\ninformation bottleneck. In International Conference for Learning Representations, 2015.\n\n[3] Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and\nnew perspectives. IEEE transactions on pattern analysis and machine intelligence, 35(8):1798\u2013\n1828, 2013.\n\n[4] James Bennett and Stan Lanning. The net\ufb02ix prize. In Proceedings of KDD Cup and Workshop\n\n2007, 2007.\n\n[5] James Bergstra, Dan Yamins, and David D Cox. Hyperopt: A python library for optimizing the\nhyperparameters of machine learning algorithms. In Proceedings of the 12th Python in science\nconference, pages 13\u201320. Citeseer, 2013.\n\n[6] James S Bergstra, R\u00e9mi Bardenet, Yoshua Bengio, and Bal\u00e1zs K\u00e9gl. Algorithms for hyper-\nparameter optimization. In Advances in neural information processing systems, pages 2546\u2013\n2554, 2011.\n\n[7] Diane Bouchacourt, Ryota Tomioka, and Sebastian Nowozin. Multi-level variational autoen-\ncoder: Learning disentangled representations from grouped observations. In Thirty-Second\nAAAI Conference on Arti\ufb01cial Intelligence, 2018.\n\n[8] Christopher P Burgess, Irina Higgins, Arka Pal, Loic Matthey, Nick Watters, Guillaume Des-\njardins, and Alexander Lerchner. Understanding disentangling in beta-vae. arXiv preprint\narXiv:1804.03599, 2018.\n\n[9] Tian Qi Chen, Xuechen Li, Roger B Grosse, and David K Duvenaud. Isolating sources of\ndisentanglement in variational autoencoders. In Advances in Neural Information Processing\nSystems, pages 2610\u20132620, 2018.\n\n[10] Xi Chen, Yan Duan, Rein Houthooft, John Schulman, Ilya Sutskever, and Pieter Abbeel. Infogan:\nInterpretable representation learning by information maximizing generative adversarial nets. In\nAdvances in neural information processing systems, pages 2172\u20132180, 2016.\n\n[11] Paul Covington, Jay Adams, and Emre Sargin. Deep neural networks for youtube recommenda-\ntions. In Proceedings of the 10th ACM Conference on Recommender Systems, pages 191\u2013198.\nACM, 2016.\n\n[12] Mukund Deshpande and George Karypis. Item-based top-n recommendation algorithms. ACM\n\nTransactions on Information Systems (TOIS), 22(1):143\u2013177, 2004.\n\n[13] Nat Dilokthanakul, Pedro AM Mediano, Marta Garnelo, Matthew CH Lee, Hugh Salimbeni,\nKai Arulkumaran, and Murray Shanahan. Deep unsupervised clustering with gaussian mixture\nvariational autoencoders. arXiv preprint arXiv:1611.02648, 2016.\n\n[14] Emilien Dupont. Learning disentangled joint continuous and discrete representations.\n\nAdvances in Neural Information Processing Systems, pages 710\u2013720, 2018.\n\nIn\n\n9\n\n\f[15] SM Ali Eslami, Danilo Jimenez Rezende, Frederic Besse, Fabio Viola, Ari S Morcos, Marta\nGarnelo, Avraham Ruderman, Andrei A Rusu, Ivo Danihelka, Karol Gregor, et al. Neural scene\nrepresentation and rendering. Science, 360(6394):1204\u20131210, 2018.\n\n[16] F Maxwell Harper and Joseph A Konstan. The movielens datasets: History and context. Acm\n\ntransactions on interactive intelligent systems (tiis), 5(4):19, 2016.\n\n[17] Xiangnan He, Tao Chen, Min-Yen Kan, and Xiao Chen. Trirank: Review-aware explainable\nIn Proceedings of the 24th ACM International on\n\nrecommendation by modeling aspects.\nConference on Information and Knowledge Management, pages 1661\u20131670. ACM, 2015.\n\n[18] Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua. Neural\ncollaborative \ufb01ltering. In Proceedings of the 26th International Conference on World Wide Web,\npages 173\u2013182, 2017.\n\n[19] Irina Higgins, Loic Matthey, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick,\nShakir Mohamed, and Alexander Lerchner. beta-vae: Learning basic visual concepts with a\nconstrained variational framework. In International Conference on Learning Representations,\nvolume 3, 2017.\n\n[20] Jun-Ting Hsieh, Bingbin Liu, De-An Huang, Li F Fei-Fei, and Juan Carlos Niebles. Learning\nto decompose and disentangle representations for video prediction. In Advances in Neural\nInformation Processing Systems, pages 517\u2013526, 2018.\n\n[21] Yifan Hu, Yehuda Koren, and Chris Volinsky. Collaborative \ufb01ltering for implicit feedback\n\ndatasets. In ICDM, volume 8, pages 263\u2013272. Citeseer, 2008.\n\n[22] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax.\n\narXiv preprint arXiv:1611.01144, 2016.\n\n[23] S\u00e9bastien Jean, Kyunghyun Cho, Roland Memisevic, and Yoshua Bengio. On using very large\ntarget vocabulary for neural machine translation. In Proceedings of the 53rd Annual Meeting of\nthe Association for Computational Linguistics (ACL 2015), 2015.\n\n[24] Zhuxi Jiang, Yin Zheng, Huachun Tan, Bangsheng Tang, and Hanning Zhou. Variational deep\nembedding: an unsupervised and generative approach to clustering. In Proceedings of the 26th\nInternational Joint Conference on Arti\ufb01cial Intelligence, pages 1965\u20131972. AAAI Press, 2017.\n\n[25] Hyunjik Kim and Andriy Mnih. Disentangling by factorising. In International Conference on\n\nMachine Learning, pages 2654\u20132663, 2018.\n\n[26] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Interna-\n\ntional Conference for Learning Representations, 2014.\n\n[27] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint\n\narXiv:1312.6114, 2013.\n\n[28] Nikos Komodakis and Spyros Gidaris. Unsupervised representation learning by predicting\n\nimage rotations. In International Conference on Learning Representations (ICLR), 2018.\n\n[29] Yehuda Koren, Robert Bell, Chris Volinsky, et al. Matrix factorization techniques for recom-\n\nmender systems. Computer, 42(8):30\u201337, 2009.\n\n[30] Adam Kosiorek, Hyunjik Kim, Yee Whye Teh, and Ingmar Posner. Sequential attend, infer,\nrepeat: Generative modelling of moving objects. In Advances in Neural Information Processing\nSystems, pages 8606\u20138616, 2018.\n\n[31] Xiaopeng Li and James She. Collaborative variational autoencoder for recommender systems.\nIn Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery\nand Data Mining, pages 305\u2013314. ACM, 2017.\n\n[32] Dawen Liang, Rahul G. Krishnan, Matthew D. Hoffman, and Tony Jebara. Variational autoen-\ncoders for collaborative \ufb01ltering. In Proceedings of the 2018 World Wide Web Conference,\nWWW \u201918, pages 689\u2013698, 2018.\n\n10\n\n\f[33] Bin Liu, Anmol Sheth, Udi Weinsberg, Jaideep Chandrashekar, and Ramesh Govindan. Adre-\nveal: Improving transparency into online targeted advertising. In Proceedings of the Twelfth\nACM Workshop on Hot Topics in Networks, 2013.\n\n[34] Francesco Locatello, Stefan Bauer, Mario Lucic, Sylvain Gelly, Bernhard Sch\u00f6lkopf, and\nOlivier Bachem. Challenging common assumptions in the unsupervised learning of disentangled\nrepresentations. In Proceedings of the 36th International Conference on Machine Learning\n(ICML 2019), 2019.\n\n[35] Francesco Locatello, Michael Tschannen, Stefan Bauer, Gunnar R\u00e4tsch, Bernhard Sch\u00f6lkopf,\nand Olivier Bachem. Disentangling factors of variation using few labels. arXiv preprint\narXiv:1905.01258, 2019.\n\n[36] B LOWERE. The harpy speech recognition system. PhD thesis, Carnegie Mellon University,\n\n1976.\n\n[37] Jianxin Ma, Peng Cui, Kun Kuang, Xin Wang, and Wenwu Zhu. Disentangled graph convolu-\ntional networks. In Proceedings of the 36th International Conference on Machine Learning\n(ICML 2019), 2019.\n\n[38] Jianxin Ma, Peng Cui, Xiao Wang, and Wenwu Zhu. Hierarchical taxonomy aware network\nembedding. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge\nDiscovery and Data Mining, 2018.\n\n[39] Liqian Ma, Qianru Sun, Stamatios Georgoulis, Luc Van Gool, Bernt Schiele, and Mario Fritz.\nDisentangled person image generation. In Proceedings of the IEEE Conference on Computer\nVision and Pattern Recognition, pages 99\u2013108, 2018.\n\n[40] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine\n\nlearning research, 9(Nov):2579\u20132605, 2008.\n\n[41] Chris J Maddison, Andriy Mnih, and Yee Whye Teh. The concrete distribution: A continuous\n\nrelaxation of discrete random variables. arXiv preprint arXiv:1611.00712, 2016.\n\n[42] Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt-Thieme. Bpr:\nIn Proceedings of the twenty-\ufb01fth\n\nBayesian personalized ranking from implicit feedback.\nconference on uncertainty in arti\ufb01cial intelligence, pages 452\u2013461. AUAI Press, 2009.\n\n[43] Paul Resnick, Neophytos Iacovou, Mitesh Suchak, Peter Bergstrom, and John Riedl. Grouplens:\nan open architecture for collaborative \ufb01ltering of netnews. In Proceedings of the 1994 ACM\nconference on Computer supported cooperative work, pages 175\u2013186. ACM, 1994.\n\n[44] Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation\nand approximate inference in deep generative models. In Proceedings of the 31st International\nConference on International Conference on Machine Learning-Volume 32, pages II\u20131278.\nJMLR. org, 2014.\n\n[45] Ruslan Salakhutdinov and Andriy Mnih. Probabilistic matrix factorization. In NIPS, volume 20,\n\npages 1\u20138, 2011.\n\n[46] Badrul Sarwar, George Karypis, Joseph Konstan, and John Riedl. Item-based collaborative\n\ufb01ltering recommendation algorithms. In Proceedings of the 10th international conference on\nWorld Wide Web, pages 285\u2013295. ACM, 2001.\n\n[47] Naftali Tishby, Fernando C Pereira, and William Bialek. The information bottleneck method.\n\narXiv preprint physics/0004057, 2000.\n\n[48] Naftali Tishby and Noga Zaslavsky. Deep learning and the information bottleneck principle. In\n\n2015 IEEE Information Theory Workshop (ITW), pages 1\u20135. IEEE, 2015.\n\n[49] Hao Wang, Naiyan Wang, and Dit-Yan Yeung. Collaborative deep learning for recommender\nsystems. In Proceedings of the 21th ACM SIGKDD International Conference on Knowledge\nDiscovery and Data Mining, pages 1235\u20131244. ACM, 2015.\n\n11\n\n\f[50] Yi Zhang and Jonathan Koren. Ef\ufb01cient bayesian hierarchical user modeling for recommen-\ndation system. In Proceedings of the 30th annual international ACM SIGIR conference on\nResearch and development in information retrieval, 2007.\n\n[51] Yongfeng Zhang, Guokun Lai, Min Zhang, Yi Zhang, Yiqun Liu, and Shaoping Ma. Explicit\nfactor models for explainable recommendation based on phrase-level sentiment analysis. In\nProceedings of the 37th international ACM SIGIR conference on Research & development in\ninformation retrieval, pages 83\u201392. ACM, 2014.\n\n[52] Chang Zhou, Jinze Bai, Junshuai Song, Xiaofei Liu, Zhengchao Zhao, Xiusi Chen, and Jun\nGao. Atrank: An attention-based user behavior modeling framework for recommendation. In\nThirty-Second AAAI Conference on Arti\ufb01cial Intelligence, 2018.\n\n[53] Jun-Yan Zhu, Zhoutong Zhang, Chengkai Zhang, Jiajun Wu, Antonio Torralba, Josh Tenen-\nbaum, and Bill Freeman. Visual object networks: Image generation with disentangled 3d\nrepresentations. In Advances in Neural Information Processing Systems, pages 118\u2013129, 2018.\n\n12\n\n\f", "award": [], "sourceid": 3075, "authors": [{"given_name": "Jianxin", "family_name": "Ma", "institution": "Alibaba Group"}, {"given_name": "Chang", "family_name": "Zhou", "institution": "Alibaba Group"}, {"given_name": "Peng", "family_name": "Cui", "institution": "Tsinghua University"}, {"given_name": "Hongxia", "family_name": "Yang", "institution": "Alibaba Group"}, {"given_name": "Wenwu", "family_name": "Zhu", "institution": "Tsinghua University"}]}