Neural-Brane: Neural Bayesian Personalized Ranking for Attributed Network Embedding

Network embedding methodologies, which learn a distributed vector representation for each vertex in a network, have attracted considerable interest in recent years. Existing works have demonstrated that vertex representation learned through an embedding method provides superior performance in many real-world applications, such as node classification, link prediction, and community detection. However, most of the existing methods for network embedding only utilize topological information of a vertex, ignoring a rich set of nodal attributes (such as, user profiles of an online social network, or textual contents of a citation network), which is abundant in all real-life networks. A joint network embedding that takes into account both attributional and relational information entails a complete network information and could further enrich the learned vector representations. In this work, we present Neural-Brane, a novel Neural Bayesian Personalized Ranking based Attributed Network Embedding. For a given network, Neural-Brane extracts latent feature representation of its vertices using a designed neural network model that unifies network topological information and nodal attributes; Besides, it utilizes Bayesian personalized ranking objective, which exploits the proximity ordering between a similar node-pair and a dissimilar node-pair. We evaluate the quality of vertex embedding produced by Neural-Brane by solving the node classification and clustering tasks on four real-world datasets. Experimental results demonstrate the superiority of our proposed method over the state-of-the-art existing methods.


INTRODUCTION
The past few years have witnessed a surge in research on embedding the vertices of a network into a low-dimensional, dense vector space.The embedded vector representation of the vertices in such a vector space enables effortless invocation of off-the-shelf machine learning algorithms, thereby facilitating several downstream network mining tasks, including node classification [19], link prediction [8], community detection [21], job recommendation [5], and entity disambiguation [24].Most existing network embedding methods, including DeepWalk [14], LINE [17], Node2Vec [8], and SDNE [20], utilize the topological information of a network with the rationale that nodes with similar topological roles should be distributed closely in the learned low-dimensional vector space.While this suffices for node embedding of a bare-bone network, it is inadequate for most of today's network datasets which include useful information beyond link connectivity.Specifically, for most of the social and communication networks, a rich set of nodal attributes is typically available, and more importantly, the similarity between a pair of nodes is dictated significantly by the similarity of their attribute values.Yet, the existing embedding models do not provide a principled approach for incorporating nodal attributes into network embedding and thus fail to achieve the performance boost that may be obtained through modeling attribute based nodal similarity.Intuitively, joint network embedding that consider both attributional and relational information could entail complementary information and further enrich the learned vector representations.
We provide a few examples from real-life networks to highlight the importance of vertex attributes for understanding the role of the vertices and to predict their interactions.For example, users on social websites contain biographical profiles like age, gender, and textual comments, which dictate who they befriend with, and what are their common interests.In a citation network, each scientific paper is associated with a title, an abstract, and a publication venue, which largely dictates its future citation patterns.In fact, nodal attributes are specifically important when the network topology fails to capture the similarity between a pair of nodes.For example, in academic domain, two researchers who write scientific papers related to "machine learning" and "information retrieval" are not considered to be similar by existing embedding methods (say, DeepWalk or LINE) unless they are co-authors or they share common collaborators.In such a scenario, node attributes of the researchers (e.g., research keywords) are crucial for compensating for the lack of topological similarity between the researchers.In summary, by jointly considering the attribute homophily and the network topology, more informative node representations can be expected.
Recently, a few works have been proposed which consider attributed network embedding [11,22,25]; however, the majority of these methods use a matrix factorization approach, which suffers from some crucial limitations.For example, earliest among these works is Text-Associated DeepWalk (TADW) [22], which incorporates the text features of nodes into DeepWalk by factorizing a matrix M constructed from the summation of a set of graph transition matrices.But, SVD based matrix factorization is both time and memory consuming, which restricts TADW to scale up to large datasets.Furthermore, obtaining an accurate matrix M for factorization is difficult and TADW instead factorizes an approximate matrix, which reduces its representation capacity.Huang et al. [11] proposed another matrix factorization (MF) based method, known as, Accelerated Attributed Network Embedding (AANE).It suffers from the same limitation as TADW.Another crucial limitation of the above methods is that they have a design matrix which they factorize, but such a matrix cannot deal with nodal attributes of rich types.In summary, the representation power of a matrix factorization based method is found to be poorer than a neural network based method, as we will show in the experiment section of this paper.
We found two most recent attributed network embedding methods, GraphSAGE and Graph2Gauss, which use deep neural network methods.To generate embedding of a node, GraphSAGE [9] aggregates embedding of its multi-hope neighbors using a convolution neural network model.GraphSAGE has a high time complexity, besides such ad-hoc aggregation may introduce noise which adversely affects its performance.Recently, Bojchevski et al. [1] proposed the Graph2Gauss (G2G), where they embed each node as a Gaussian distribution.G2G uses a neural network based deep encoder to process the nodal attributes and obtains an intermediate hidden representation, which is then used to generate the mean vector and the covariance matrix of the learned Gaussian distribution of a node.As a result, in G2G's learning, the interaction between the attribute information and the topology information of a node is poor.On the other hand, the learning pipeline of our proposed Neural-Brane enables effective information exchange between the attribute and topology of a node, making it much superior than G2G while learning embedding for attributed networks.It is worth noting that some recent works have proposed semi-supervised attributed network embedding considering the availability of node labels [12,13], but our focus in this work is unsupervised attributed network embedding, for which vertex labels are not available.
Our solution and contribution: In this paper, we present Neural-Brane, a novel method for attributed network embedding.For a vertex of the input network, Neural-Brane infuses its network topological information and nodal attributes by using a custom neural network model, which returns a single representation vector capturing both the aspects of that vertex.The loss function of Neural-Brane utilizes BPR [15] to capture attribute and topological similarities between a pair of nodes in their learned representation vectors.Specifically, the BPR objective elevates the ranking of a vertex-pair having similar attributes and topology by embedding the vertices in close proximity in the representation space, in comparison to other vertex-pairs which are not similar.We summarize the key contributions of this work as follows: (

RELATED WORK
There is a large body of works on representation learning on graphs (a.k.a.network embedding).Well known among these methods are DeepWalk [14] and Node2Vec [8], both of which capture local topology around a node through sequences of vertices obtained by uniform or biased random walk, and then use the Skip-Gram language model for obtaining the representation of each vertex.LINE [17] computes the similarity of a node to other nodes as a probability distribution by computing first and second order proximities, and design a KL-divergence based objective function which minimizes the divergence between empirical distribution from data and actual distribution from the embedding vectors.GraRep [2] is a matrix factorization based approach that leverages both local and global structural information.Furthermore, a few neural network based approaches are proposed for network embedding, such as [3,4,20].Interested readers can refer to the survey articles in [7,10], which present a taxonomy of various network embedding methods in the existing literature.Most of the aforementioned works only investigate the topological structure for network embedding, which is in fact only a partial view of an attributed network.To bridge this gap, a few attributed network embedding based approaches [6,11,13,16,22,25] are proposed.The general philosophy of such works is to integrate nodal features, such as text information and user profile, into topology-oriented network embedding model to enhance the performance of downstream network mining tasks.For example, TADW [22] performs low-rank matrix factorization considering graph structure and text features.Furthermore, TriDNR [13] adopts a two-layer neural networks to jointly learn the network representations by leveraging inter-node, node-word, and label-word relationships.Different from the existing methods, our proposed unsupervised embedding method (Neural-Brane) utilizes a designed neural network architecture and a novel Bayesian personalized ranking based loss function to learn better network representations.

PROBLEM STATEMENT
Throughout this paper, scalars are denoted by lowercase alphabets (e.g., n).Vectors are represented by boldface lowercase letters (e.g., x).Bold uppercase letters (e.g., X) denote matrices, and the i th row of a matrix X is denoted as x i .The transpose of the vector x is denoted by x T .The dot product of two vectors is denoted by ⟨a, b⟩.∥X∥ F is the Frobenius norm of matrix X. Finally calligraphic uppercase letter (e.g., X) is used to denote a set and |X| is used to denote the cardinality of the set X.
Let G = (V, E, A) be an attributed network, where V is a set of n nodes, and E is a set of edges, and A is a n × m binary attribute matrix such that the row a i denotes a row attribute vector associated with node i in G.Each edge (i, j) ∈ E is associated with a weight w i j .The neighbors of node i is represented as N (i).m is the number of node attributes in A. We use A(i) to denote the non-zero attribute set of node i.
Fig. 1.Neural-Brane architecture.Given a node u, a u is its binary attribute vector and n u is its adjacency vector.Our training uses node-triplets (u, i, j), such that (u, i) ∈ E and (u, j) E.
The attributed network embedding problem is formally defined as follows: given an attributed network G = (V, E, A), we aim to obtain the representation of its vertices as a n × d matrix , where f i is the row vector representing the embedding of node i.The representation matrix F should preserve the node proximity from both network topological structure E and node attributes A. Eventually, F serves as feature representation for the vertices of G, as such, that they can be used for various downstream network mining tasks.

NEURAL-BRANE: ATTRIBUTED NETWORK EMBEDDING FRAMEWORK
In this section, we discuss the proposed neural Bayesian personalized ranking model for attributed network embedding.The model uses a neural network architecture with embedding layer, hidden layer, output layer, and BPR layer from bottom to top, as illustrated in Figure 1.Specifically, the embedding layer learns a unified vector representation of a node from the vector representation of its nodal attributes and neighbors; the hidden layer applies nonlinear dimensionality reduction over the embedding vectors of the nodes, the output layer and the BPR layer enable model inference through back-propagation.

Embedding Layer
The embedding layer has two embedding matrices P, and P ′ ; each row of P is a d 1 dimensional vector representation of an attribute, and each row of P ′ is a d 2 dimensional vector representation of a vertex (both d 1 and d 2 are user-defined parameter).These matrices are updated iteratively during the learning process.For a given vertex u, embedding layer produces u's latent representation vector f u by learning from embedding vectors of u's attributes and neighbors, i.e., corresponding rows of P and P ′ , respectively; thus the neighbors and attributes of u are jointly involved in the construction of u's latent representation vector (f u ), which enables Neural-Brane to bring the latent representation vectors of nodes with similar attributes and neighborhood in close proximity in the latent space.We illustrate the vector construction process using a toy attributed graph in Figure 2. Given the vertex b from the toy graph, the embedding layer first takes its attribute and adjacency vectors (from P and P ′ ) as input and then generates its corresponding attributional and nodal embedding matrices (P (at t r ) b and P ′(nbr ) b ) by using the CONCAT-LOOKUP(•) function.After that, attributional and neighborhood embedding vectors are obtained from P (at t r ) b and P ′(nbr ) b by using the max-pooling operation respectively.Finally, the learned attributional and neighborhood embedding vectors are concatenated together to obtain the final embedding representation of the vertex b.Below we provide more details of the operations in embedding layer.

Encoding attributional information.
Given a node u ∈ V and the attribute matrix A, a u ∈ IR 1×m is A's row corresponding to u's binary attribute vector.We apply a row-wise concatenation based embedding lookup layer to transform a u into a latent matrix, P (at t r ) u , as shown below: where P ∈ IR m×d 1 is the attribute embedding matrix in which each row is a d 1 (user defined parameter) sized vector representation of an attribute.Lookup is performed by CONCAT-LOOKUP(•) function which first performs a row projection on P by selecting the rows corresponding to the attribute-set A(u) and then stacks the selected vectors row-wise into the matrix Then we apply a max-pooling operation on the generated P (at t r ) u matrix in order to transform it into a single vector.Specifically, max-pooling operation retains the most informative signal by extracting the largest value in each dimension (i.e., column) of the matrix P (at t r ) u to obtain v at t r u .
where v at t r u ∈ IR 1×d 1 is the latent vector representation of node u based on its attributional signals, and MP(•) denotes the max-pooling operation.
4.1.2Encoding network topology.Given a node u, we describe its neighborhood by using a binary adjacency vector, denoted as n u ∈ IR 1×n , in which u's neighbors are set to 1, and the rest of entries are set as 0. Similar to the operations we use for encoding the attributional information, we apply a row-wise concatenation based lookup layer to transform n u into a latent matrix P ′(nbr ) u and then apply max-pooling operation on the obtained latent matrix.Thus, where P ′ ∈ IR n×d 2 is the neighborhood embedding matrix for lookup (similar to matrix P), and P ′(nbr ) u ∈ IR | N(u) |×d 2 is the obtained latent matrix generated from the CONCAT-LOOKUP(•) function.Moreover, v nbr u ∈ IR 1×d 2 obtained from the MP(•) operation is the latent vector representation of node u based on its neighborhood topology.

Integration component.
Once we obtain the vector representation of node u from both its attributional information and topological structure as developed in Equations 1, 2, 3 and 4, we further integrate both latent vectors into a unified vector representation by vector concatenation, as shown below: where f u ∈ IR 1×d (d 1 + d 2 = d), and "||" denotes the vector concatenation operation.

Hidden Layer
Given the obtained embedding vector f u ∈ IR 1×d for node u in the attributed network G, the hidden layer aims to transform its embedding vector into another representation h u , in which signals from attributes and neighborhood of a vertex interact with each other.Formally, given f u , the hidden layer produces h u ∈ IR 1×h by the following formula: Here we use rectified linear function ReLU (x), defined as max(0, x), as the activation function for achieving better convergence speed.Parameters W ∈ IR h×d and b ∈ IR h×1 are weights and bias for the hidden layer, respectively; h is a user-defined parameter denoting the number of neurons in the hidden layer.It is worth mentioning that in the hidden layer, all the nodes share the same set of parameters {W, b}, which enables information sharing across different vertices (see the box denoted as "Hidden Layer" in Figure 1).

Output and BPR Layers
Given a node pair u and i, we use their corresponding representations h u and h i from hidden layer (Equation 6) as input for the output layer.The task of this layer is to measure the similarity score between a pair of vertices by taking the dot product of their representation vectors.Since this computation uses the vector representation of the vertices from the hidden layer, it encodes both attribute similarity and neighborhood similarity jointly.The similarity score between vertices u and i, defined as s ui , is calculated as ⟨h u , h i ⟩.
BPR layer implements the Bayesian personalized ranking objective.For the embedding task, the ranking objective is that the neighboring nodes in the graph should have more similar vector representations in the embedding space than non-neighboring nodes.For example, the similarity score between two neighboring vertices u and i, should be larger than the similarity score between two non-neighboring nodes u and j.As shown in Figure 1, given the vertex triplet (u, i, j), we model the probability of preserving ranking order s ui > s u j using the sigmoid function σ As we observe from Equation 7, the larger the difference between s ui and s u j , the more likely the ranking order s ui > s u j is preserved.By assuming that all the triplet based ranking orders generated from the graph G to be independent, the probability of all the ranking orders being preserved is defined as follows: where D represents training triplet sets generated from G and i > u j is a shorthand notation denoting s ui > s u j ; the notation is motivated from the concept that i is larger than j considering the partial order relation > u .
The goal of our attributed network embedding is to maximize the expression in Equation 8.For the computational convenience, we minimize the sum of negative-likelihood loss function, which is shown as below: where Θ = {P, P ′ , W, b} are model parameters used in all different layers, and λ • ||Θ|| 2  F is a regularization term to prevent model overfitting.

Model inference and optimization.
We employ the back propagation algorithm by utilizing mini-batch gradient descent to optimize the parameters Θ = {P, P ′ , W, b} in our model.The main process of mini-batch gradient descent is to first sample a batch of triplets from G. Specifically, given an arbitrary node u, we sample one of its neighbors i, i.e., i ∈ N (u), with the probability proportional to the edge weight w i j .On the other hand, we sample its non-neighboring node j, i.e., j N (u), with the probability proportional to the node degree in the graph.Then for each mini-batch training triplets, by using the chain rule, we compute the derivative and update the corresponding parameters Θ by walking along the descending gradient direction.In particular, by back-propagating from Bayesian personalized ranking layer to hidden layer, we update the gradients w.r.t.weight matrix W and bias vector b accordingly.Then in the embedding layer, we update the gradients of the corresponding embedding vectors (i.e., rows) in {P, P ′ } associated with all the neighboring nodes and attributes involved in each mini-batch training triplets respectively.Mathematically, where α is the learning rate.In addition, we initialize all model parameters Θ by using a Gaussian distribution with 0 mean and 0.01 standard deviation.The pseudo-code of the proposed Neural-Brane framework is summarized in Algorithm 1.

5:
Calculate h u , h i , h j based on the Equation 6.
For the time complexity analysis, given the sampled training triplet set D, the total costs of calculating and updating gradients of L w.r.t.corresponding embedding vectors involved in {P, P ′ } are O(d).Similarly, the total costs of computing and updating gradients of L w.r.t.parameters {W, b} in the hidden layer are O(hd + h).Therefore, the total computational complexity of the proposed methodology for Neural-Brane is |D | * O(d) + O(hd + h) .As time complexity of the Neural-Brane is linear to the embedding size and hidden layer dimension, it is extremely fast.For example, it takes only 10 minutes to learn embedding for our largest dataset Arnetminer (see Table 1).

EXPERIMENTS AND RESULTS
In this section, we first introduce the datasets and baseline comparisons used in this work.Then we thoroughly evaluate our proposed Neural-Brane through two downstream data mining tasks (node classification and clustering) on four real-world networks, for which node attributes are available.Finally, we analyze the quantitative experimental results, investigate parameter sensitivity, convergence behavior, and the effect of pooling strategy of Neural-Brane.

Experimental Setup
Datasets.We perform experiments on four real-world datasets, whose statistics are shown in Table 1.The largest among these networks has around 5.5K vertices, and 18K edges.Note that, publicly available networks exist, which are larger than the networks that we use in this work, but those larger networks are neither attributed nor they have class label for the vertices, so we cannot use those in our experiment.Nevertheless, our largest dataset Arnetminer, has more nodes, edges and attributes than datasets used by recent attribute embedding papers [22,25].More description of the datasets is given below.
CiteSeer1 is a citation network, in which nodes refer to papers and links refer to citation relationship among papers.Selected keywords from the paper are used as nodal attributes.Additionally, the papers are classified into 6 categories according to its research domain, namely Artificial Intelligence (AI), Database (DB), Information Retrieval (IR), Machine Learning (ML), Human Computer Interaction (HCI), and Multi-Agent Analysis.Caltech36 and Reed98 [18] are two university Facebook networks.Specifically, each node represents a user from the corresponding university and edge represents user friendship.The attributes of each node is represented by a 64-dimensional one-hot vector based on gender, major, second major/minor, dorm/house, and year.We use student/faculty status of a node as the class label.

Baseline Comparison.
To validate the benefit of our proposed Neural-Brane, we compare it against 10 different methods.Among all the competing methods, DeepWalk, LINE, and Node2Vec are topology-oriented network embedding approaches.NNMF, DeepWalk + NNMF, GraphSAGE, PTE-KL, TADW, AANE and G2G are state-of-the-arts for combining both network structure and nodal attributes for network representation learning.Note that PTE-KL is a semisupervised embedding approach, and we hold the label information out for a fair comparison.
AANE [11]: Accelerated Attributed Network Embedding learns low-dimensional representation of nodes from network linkage and content information through a joint matrix factorization.(10) G2G [1]: Graph2Gauss learns node representation such that each node vector is a Gaussian distribution.

Parameter Setting and Implementation Details.
There are a few user-defined hyperparameters in our proposed embedding model.We fix the embedding dimension d = 150 (same for all baseline methods) with d 1 = d 2 = 75.For the number of neurons in hidden layer h, we set it to be 150.For the regularization coefficient λ in the embedding model (see Equation 9), we set it as 0.00005.In addition to that, we fix the learning rate α = 0.5 (see Equation 10) and batch size to be 100 during the model learning and optimization.For baseline methods such as GraphSAGE, PTE-KL, AANE, G2G and others, we select learning rate α from the set {0.01, 0.05, 0.1, 0.5}3 using grid search.Similarly for PTE-KL, TADW and other baseline methods regularization coefficient λ is selected from the set {0.01, 0.001, 0.0001}.For random walk based baselines (DeepWalk and Node2Vec), we select the best walk length from the set {20, 40, 60, 80}.For the rest of hyper-parameters, we use default parameter values as suggested by their original papers.

Node Classification.
For fair comparison between network embedding methods, we purposely choose a linear classifier to control the impact of complicated learning approaches on the classification performance.Specifically, we treat the node representations learned by different approaches as features, and train a logistic regression classifier for multi-class / binary classification.In each dataset, p% ∈ {30%, 50%, 70%} of nodes are randomly selected as training set and the rest as test set.We use the widely used metric Macro-F1 [23] for classification assessment.Each method is executed 10 times and the average value is reported.For Neural-Brane, we also report standard Table 2 shows results for node classification, where each column is an embedding method and rows represent different train splits (p).As we observe from Table 2, performance of the last four (PTE-KL, TADW, AANE, G2G) baseline methods are highly competitive among each others.But, our proposed Neural-Brane consistently outperforms all these and other baseline methods under all training ratios.Moreover, the overall performance improvement that our Neural-Brane delivers over the second best method is significant.For example, in Citeseer dataset, when training ratio p ranges from 30% to 70%, Neural-Brane outperforms the G2G by 8.8%, 8.6%, 8.4% in terms of Macro-F1, respectively.Furthermore, the improvement over G2G is statistically significant (paired t-test with p-value ≪ 0.01).The relatively good performance of our proposed Neural-Brane across various training ratios is due to the fact that our proposed neural Bayesian personalized ranking framework is able to generate high-quality latent features by capturing crucial ordering information between nodes and incorporating nodal attributes and network topology into network embedding.Furthermore, BPR is shown to be better suited than other loss functions, such as point-wise square loss in TADW and K-L divergence based objective in LINE and PTE-KL, for placing similar nodes in the embedding space for the downstream node classification task.
Among the competing methods, topology-oriented network embedding approaches such as LINE and DeepWalk perform fairly poor on all datasets.This is mainly because the network structure is rather sparse and only contains limited information.On the other hand, TADW is much better than DeepWalk due to the fact that textual contents contain richer signals compared to the network structure.When concatenating the embedding vectors from DeepWalk and NNMF, the classification performance is relatively improved compared to a single DeepWalk.However, the naive combination between DeepWalk and NNMF is far from optimal, compared to our proposed Neural-Brane.Note that, GraphSAGE for Arnetminer dataset is not able to complete after 2 days on contemporary server having 64 cores with 2.3 GHz and 132 GB memory.

Visualization and Node
Clustering.The primary goal of graph embedding approaches is to put similar nodes closer in their corresponding latent space, hence a desirable embedding method should generate clusters of similar nodes in the embedding space.Visualization for large number of classes in two dimensional space is impractical.Instead, in Figure 3, we plot 2D representation of learned vector representations for Caltech36 and Reed98 datasets.Note that both of these datasets contain only 2 classes and hence provide interpretable visualization.Specifically, we plot embedding representations of Neural-Brane along with two best competing methods, namely TADW and AANE.These figures clearly demonstrate that Neural-Brane provides better discrimination of classes through clustering in the latent space compared to both TADW and AANE.
For the other two larger datasets (CiteSeer and Arnetminer), we use k-means clustering approach to the learned vector representations of nodes and utilize both Purity and Normalized Mutual Information (NMI) [23] to assess the quality of clustering results.Furthermore, we match the groundtruth number of clusters as input for running k-means, execute the clustering process 10 times to alleviate the sensitivity of centroid initialization, and report the average results.
The clustering results for both CiteSeer and Arnetminer datasets are depicted in Figure 4.As we can see, our proposed Neural-Brane consistently achieves the best clustering results in contrast to all competing baselines.For example, in Citeseer dataset, our proposed Neural-Brane achieves 0.3524 NMI.However, the best competing method PTE-KL only obtains 0.2653 NMI, indicating more than 32.8% gains.Similarly, for Arnetminer dataset, Neural-Brane obtains 34.5% improvements over the best competing approach DeepWalk in terms of NMI.The possible explanation for higher performance of Neural-Brane could be due to the fact that our proposed Bayesian ranking formulation directly optimizes the pairwise distance between similar and dissimilar nodes, thus making their corresponding vectors cluster-aware in the embedded space.

Analysis of Parameter Sensitivity and Algorithm Convergence
We conduct experiments to demonstrate how the embedding dimension affects the node classification task using our proposed Neural-Brane.Specifically, we vary the number of embedding dimension parameter d as {50, 100, 150, 200, 250, 300} and set the training ratio p = 70%.We report the Macro-F1 results on all four datasets, which is shown in Figure 5a.As we observe, as the embedding dimension d increases, the classification performance in terms of Macro-F1 first increases and then tends to stabilize.The possible explanation could be that when the embedding dimension is too small, the embedding representation capability is not sufficient.However, when the embedding dimension becomes sufficiently large, it captures all necessary information from the data, leading to the stable classification performance.Furthermore, we investigate the convergence trend of Neural-Brane.As shown in Figure 5b, Neural-Brane converges approximately within 10 epochs and achieves promising convergence results in terms of the objective function value on all four datasets.We finally investigate the effect of the pooling strategy in the embedding layer for the task of node classification.For the comparison, we consider taking a sum rather than the max pooling and hold the rest of neural architecture and hyper-parameter settings constant.We report the Macro-F1 results on all four datasets with training ratio p = 70%, which is shown in Figure 6.As we observe, max pooling consistently performs better than alternative sum pooling strategy for the task of node classification across all datasets.The possible explanation is due to the fact that the max-pooling operation returns the strongest signal for each embedding dimension, which alleviates noisy signals.On the other hand, the sum pooling operation considers accumulated signals from each input embedding dimension, which leads to inaccurate information aggregation.

Fig. 2 .
Fig. 2. The figure shows the mechanism of the embedding layer for the vertex b of a toy attributed graph.The graph contains 5 vertices and 6 edges, where each vertex is associated with a collection of nodal attributes.For example, vertex b is connected to vertices {a, c, d } and associated with attributes {x 2 , x 6 }, respectively.The cardinality of the attribute set {x 1 , • • • , x 7 } is 7.

( 1 )
DeepWalk [14]: It utilize Skip-Gram based language model to analyze the truncated uniform random walks on the graph.(2) LINE [17]: It embeds the network into a latent space by leveraging both first-order and second-order proximity of each node.(3) Node2Vec [8]: Similar to DeepWalk, Node2Vec designs a biased random walk procedure for network embedding.(4) Non-Negative Matrix Factorization (NNMF): The model captures both node attributes and network structure to learn topic distributions of each node.(5) DW+NNMF: It simply concatenates the vector representations learned by DeepWalk and NNMF.(6) GraphSAGE [9]: GraphSAGE presents an inductive representation learning framework that leverages node feature information (e.g., text attributes) to efficiently generate node embeddings in the network.(7) PTE-KL [16]: Predictive Text Embedding framework aims to capture the relations of paperpaper and paper-attribute under matrix factorization framework.The objective is based on KL-divergence between empirical similarity distribution and embedding similarity distribution.(8) TADW [22]: Text-associated DeepWalk combines the text features of each node with its topology information and uses the MF version of DeepWalk.

Fig. 3 .
Fig. 3.The visualization comparison among various embedding methodologies for Caltech36 and Reed98 datasets

9 Fig. 5 .
Fig. 5. Analysis of the embedding dimension and convergence at embedding layer Max-Pooling at embedding layer

Fig. 6 .
Fig. 6.The pooling strategy comparison for the task of node classification 1)We propose Neural-Brane, a custom neural network based model for learning node embedding vectors by integrating local topology structure and nodal attributes.The source code (with datasets) of the Neural-Brane is available at: https://git.

Table 1 .
Statistics of Four Real-World DatasetsArnetminer 2 is a paper relation network consisting of scientific publications from 5 distinct research areas.Specifically, we select a list of representative conferences and journals from each of them. 1) Data Mining (KDD, SDM, ICDM, WSDM, PKDD); 2) Medical Informatics (JAMIA, J. of Biomedical Info., AI in Medicine, IEEE Tran. on Medical Imaging, IEEE Tran. on Information and Technology in Biomedicine); 3) Theory (STOC, FOCS, SODA); 4) Computer Vision and Visualization (CVPR, ICCV, VAST, TVCG, IEEE Visualization and Information Visualization) 5) Database (SIGMOD, VLDB, ICDE).Authors and keywords similarity between two papers are used for building edges.Keywords from paper title and abstract are used as attributes.

Table 2 .
Quantitative results of Macro-F1 between our proposed Neural-Brane and other baselines for the node classification task using logistic regression on various datasets (embedding dimension = 150).[ * GraphSAGE for Arnetminer is not able to complete after 2 days.]