DOI: https://doi.org/10.1145/3178876.3186064
WWW '18: Proceedings of The Web Conference 2018, Lyon, France, April 2018
Graph-based learning methods explicitly consider the relations between two entities (i.e., vertices) for learning the prediction function. They have been widely used in semi-supervised learning, manifold ranking, and clustering, among other tasks. Enhancing the expressiveness of simple graphs, hypergraphs formulate an edge as a link to multiple vertices, so as to model the higher-order relations among entities. For example, hyperedges in a hypergraph can be used to encode the similarity among vertices.
To the best of our knowledge, all existing hypergraph structures represent the hyperedge as an unordered set of vertices, without considering the possible ordering relationship among vertices. In real-world data, ordering relations commonly exist, such as in graded categorical features (e.g., users’ ratings on movies) and numerical features (e.g., monthly income of customers). When constructing a hypergraph, ignoring such ordering relations among entities will lead to severe information loss, resulting in suboptimal performance of the subsequent learning algorithms.
In this work, we address the inherent limitation of existing hypergraphs by proposing a new data structure named Partial-Order Hypergraph, which specifically injects the partially ordering relations among vertices into a hyperedge. We develop regularization-based learning theories for partial-order hypergraphs, generalizing conventional hypergraph learning by incorporating logical rules that encode the partial-order relations. We apply our proposed method to two applications: university ranking from Web data and popularity prediction of online content. Extensive experiments demonstrate the superiority of our proposed partial-order hypergraphs, which consistently improve over conventional hypergraph methods.
CCS Concepts: • Information systems → Web mining; Web applications; • Computing methodologies → Spectral methods; Semi-supervised learning settings; • Applied computing → Education;
ACM Reference Format:
Fuli Feng, Xiangnan He, Yiqun Liu, Liqiang Nie, and Tat-Seng Chua. 2018. Learning on Partial-Order Hypergraphs. In WWW 2018: The 2018 Web Conference, April 23–27, 2018, Lyon, France. ACM, New York, NY, USA 10 Pages. https://doi.org/10.1145/3178876.3186064
Graphs naturally represent relational data and have been widely used to model the relationships between entities. Simple graphs intuitively connect two vertices (i.e., entities of interest) with an edge (i.e., the relationship to model), which can be either undirected or directed depending on whether the pairwise relationship between entities is symmetric. For example, given a set of entities with feature vectors, we can construct an undirected graph by forming the adjacency matrix with a similarity metric [45, 48]. The World Wide Web is a well-known instance of directed graphs, where vertices represent webpages, and edges represent hyperlinks. With such graph representations of entities and their relations, many graph-based learning methods have been developed to address various tasks, such as semi-supervised learning [11, 32, 37, 47], manifold ranking [20, 30, 31, 49], and clustering [4, 8, 40], personalized recommendation [17, 23], and so on.
Hypergraph is a generalizaton of simple graph, in which an edge (aka. hyperedge) can connect any number of vertices rather than just two. As such, it can model high-order relations among multiple entities that cannot be naturally represented by simple graphs. Figure 1 shows an illustrative example of using graph methods to tackle the university ranking task [9]. Each university has two features: the located city and the salary level of its graduates (Figure 1 a). A simple graph can be constructed by connecting a university with its two nearest neighbors (Figure 1 b); then performing a manifold ranking on the simple graph can obtain a ranked list of universities. Further, we can build a hypergraph by connecting universities with a same attribute (Figure 1 c), e.g., universities that are located in the same city, which is a high-order relation among universities missed by the simple graph.
In existing research, hyperedges in a hypergraph are typically formed by linking similar entities — either globally similar such as a cluster of entities that are close to each other [27, 36, 38], or locally similar such as sharing a same attribute [3, 35, 43]. However, we argue that many real-world applications need to deal with far more complex relations than similarities. One particular type is the ordering relationship among entities, which commonly exists in graded categorical features and numerical features. Taking the university ranking task in Figure 1 as an example. Two universities u 5 and u 6 are located in the same city, while u 5 has a salary level much higher than u 6 — an evidence that u 5 might be ranked higher than u 6. Unfortunately, the hypergraph constructed in Figure 1 c encodes the similarity information only, thus fails to capture the ordering information on salary. To address this limitation, an intuitive solution is to incorporate the ordering relations by adding directed edges between entities of a hyperedge, as shown in Figure 1 d. It is worth noting that not every two entities within a hyperedge have an ordering relation; for example, two entities may have the same graded categorical feature (see u 1 and u 2 in Figure 1 d) or the difference on the target numerical feature is not significant enough. As such, we term such generalized hypergraph with partial-order relations on vertices as a Partial-Order Hypergraph (POH), which is a new data structure that has never been explored before.
In this work, we formalize the concept of POHs and further develop regularization-based graph learning theories on them. We express the partial-order relations with logical rules, which can be used to encode prior domain knowledge. In the previous example of university ranking, one example of domain knowledge can be that for two universities ui and uj in the same city, ui tends to be ranked higher than uj if the salary level of ui is higher than that of uj . The corresponding logical rule can be written as:
The key contributions of the paper are summarized as follows.
The remainder of this paper is organized as follows. Section 2 introduces some preliminary knowledge. In Section 3, we present our proposed methods, followed by reviewing related work in Section 4. In Section 5 and 6, we employ POH learning to address the two tasks of university ranking and popularity prediction, respectively. We conclude the work in Section 7.
We introduce some notations first. We use bold capital letters (e.g., X) and bold lowercase letters (e.g., x) to denote matrices and vectors, respectively. Scalars and hyperparameters are respectively represented as normal lowercase letters (e.g., x) and Greek letters (e.g., λ). If not otherwise specified, all vectors are in a column form, and Xij denotes the entry at the i-th row and the j-th column of X.
As aforementioned, the vertices and hyperedges of a hypergraph represent the entities of interest and their relations, respectively. Given N entities with their features $\mathbf {X}=[\mathbf {x_1},\mathbf {x_2},\cdots ,\mathbf {x_N}]^{T}\in \mathbb {R}^{N \times D}$ , we can construct a hypergraph with N vertices and M hyperedges, for which the structure can be represented as an incidence matrix $\mathbf {H} \in \mathbb {R}^{N \times M}$ . Similar to the incidence matrix of a simple graph, H is a binary matrix, where Hij = 1 if the i-th vertex is connected by the j-th hyperedge, otherwise Hij = 0. There are two ways to define a hyperedge in existing work: attribute-based [3, 22, 35, 43] and neighbor-based [27, 36, 38]. An attribute-based hyperedge connects vertices with same value on one or multiple target attributes (i.e., features). A neighbor-based hyperedge connects vertices nearby, based on these vertices with similarity larger than a threshold or simply using the k-nearest neighbors.
Moreover, we use the diagonal matrix $\mathbf {E} \in \mathbb {R}^{M \times M}$ to denote the degrees of hyperedges, i.e., $E_{jj} = \sum _{i = 1}^{N} H_{ij}$ denotes the number of vertices connected by the j-th hyperedge. Analogous to simple graph that an edge typically has a weight to model the strength of the relation, a hyperedge in hypergraphs also has a weight to denote the density of the vertices it connected. Such weights are represented as a diagonal matrix $\mathbf {W} \in \mathbb {R}^{M \times M}$ . To estimate the hyperedge weight, many methods have been proposed, among which the most popular one is to use the average pairwise similarity between vertices connected by the hyperedge:
Graph-based learning has been applied to various machine learning tasks such as manifold ranking, semi-supervised learning, and clustering [1, 26, 44]. The general problem setting is to learn a prediction function $\mathbf {\hat{y}} = f(\mathbf {x})$ , which maps an entity from the feature space to the target label space. It is usually achieved by minimizing an objective function abstracted as:
= G + L, \eqno{3} where $\mathcal {L}$ is a task-specific loss function that measures the error between prediction $\mathbf {\hat{y}}$ and ground-truth y, $\mathcal {G}$ is a graph regularization term that smooths the prediction over the graph, and λ is a hyperparameter to balance the two terms. The regularization term typically implements the smoothness assumption that similar vertices tend to have similar predictions. For hypergraphs, a widely used $\mathcal {G}$ is the hypergraph Laplacian term, defined as:
Distinct from the typical problem setting of hypergraph learning, we further associate the problem with a set of logic rules, which can be used to encode the partial-order relations between entities:
• Irreflexivity: not pr (xi , xi ). • Asymmetry: if pr (xi , xj ) then not pr (xj , xi ). • Transitivity: pr (xi , xj ) and pr (xj , xk ) implies pr (xi , xk ).
In what follows, we first present how to construct and represent a POH. We then elaborate our proposed regularized learning on POHs. Lastly, we analyze its time complexity.
To jointly represent the partial-order relations and the higher-order relations among entities, we first construct a normal hypergraph, and then use directed edges to connect vertices that have any partial-order relation. Note that it is possible that there are multiple edges between two vertices, since multiple partial-order relations are considered. Concerning the efficiency of downtream graph-based learning applications, we constrain that directed edges only exist on vertices connected by at least one hyperedge. Such a constraint guarantees that the number of directed edges constructed from a partial-order relation is no larger than the number of nonzero entries in the hypergraph Laplacian matrix. As such, a learning algorithm that enumerates all directed edges will not increase the time complexity of calculating the hypergraph Laplacian term.
As described in Section 2.1, after constructing a hypergraph, we use several matrices to represent it, such as the incidence matrix H and hypergraph Laplacian matrix L. In addition to these matrices, we further introduce a partial incidence matrix $\mathbf {H^{r}} \in \mathbb {R}^{N \times N}$ to represent the directed edges of a partial-order relation r. As shown in Figure 2, given a partial-order relation r, we first construct a binary relation matrix $\mathbf {P^{r}} \in \mathbb {R}^{N \times N}$ , where $P^{r}_{ij} = 1$ if pr (xi , xj ) is true. Based on the incidence matrix H of the hypergraph, we further build a co-occurrence matrix $\mathbf {C} \in \mathbb {R}^{N \times N}$ , where each element Cij denotes the number of hyperedges connecting vertex i and j. Then the partial incidence matrix H r can be derived,
After constructing a POH, we have several matrices to represent it, including the general ones describing a conventional hypergraph (e.g., the incidence matrix H and edge weight matrix W), and the specific partial incidence matrices {Hr |r = 1, 2, ⋅⋅⋅, R} to model partial-order relations. We now consider how to extend the conventional hypergraph learning methods for POHs.
The key problem is the encoding of the partial-order relations and the corresponding logical rules into the learning framework of Equation (3). Generally speaking, there are two solutions — either injecting the rules into the predictive model (i.e., the definition of f(x)), or using the rules to regularize the learning (i.e., augmenting the objective function Γ). The first solution may need different ways to encode the rules for different predictive models, such as logistic regression, factorization machines [15], and neural networks [14]. As such, we opt for the second solution, aiming to provide a generic solution for POH learning. Specifically, we append an additional regularization term $\mathcal {P}$ that encodes partial-order rules to the objective function:
= r=1R ar| Hr | {i,j | Hrij 0 } 1 - qr(yi, yj),
\eqno{9} where $\mathbf {\hat{y_i}} = f(\mathbf {x_i})$ is the prediction of xi , |Hr | denotes the number of nonzero entries in Hr , and ar is the hyperparameter to control the importance of the logical rule of the r-th partial-order relation. The core idea of this regularization term is to enforce the predictions of vertices that have a partial-order relation to be consistent with the corresponding rule. To be more specific, small values of $\mathcal {P}_0$ can be achieved if $q^{r}(\mathbf {\hat{y_i}}, \mathbf {\hat{y_j}})$ is 1, meaning that pr (xi , xj ) is true (evidenced by $H^r_{ij}\ne 0$ ) and the rule pr (xi , xj ) → qr (f(xi ), f(xj )) is satisfied. However, this definition treats all vertex pairs of a partial-order relation equally, without considering their relative strengths. This assumption may decrease modelling fidelity for practical applications. To address this problem, we revise the regularizer as:
3.2.1 Optimization. By minimizing the objective function of Equation (8), we can achieve the prediction function that is smooth over the hypergraph and satisfies the logical rules on partial-order relations. Note that the regularization terms $\mathcal {L}$ and $\mathcal {P}$ operate on the predicted label space only and do not introduce extra model parameters. As such, the only model parameters to learn come from the predictive model f(x). Given that f is a differentiable function (e.g., logistic regression and neural networks), we can optimize the objective function with standard gradient-based methods, such as the stochastic (or batch) gradient descent [12]. Moreover, one can also directly learn f(x) without specifying an explicit form of the predictive model. This will make the prediction function comply with the regularization to the maximum degree. We will empirically study how would this affect the prediction performance for downstream applications in Section 6.
In this subsection, we analyze the time complexity of POH learning by comparing with conventional hypergraphs. As discussed in [21] and Section 2.2, the computational complexity of conventional hypergraph learning methods are O(J), where J denotes the number of nonzero entries in the sparse hypergraph Laplacian matrix L. In contrast, the additional computational cost of our POH learning comes from the construction of the partial incidence matrices {Hr |r = 1, 2, ⋅⋅⋅, R} and the partial-order regularization term $\mathcal {P}$ . To compute Hr , we need to obtain the co-occurrence matrix C first, and then evaluate the element-wise product C⊙Pr . For C, we can achieve it by traversing all nonzero entries on the incidence matrix H, for which the complexity is O(J). Then for each nonzero element Cij in C, we check whether p r (x i , x j ) is true or not to obtain C⊙Pr . As such, the complexity of constructing a Hr is O(J), and the overall complexity of constructing all R partial incidence matrices is O(RJ). Similarly, the computation of $\mathcal {P}$ can be done in O(RJ) time. In a real-world application, the number of partial-order relations R is typically a small number, since we need to account for the most prominent numerical or graded categorical features only. As such, the overall time complexity of our POH learning is essentially O(J), which is the same as that of conventional hypergraph learning.
Our work is directly related to the recent work on hypergraphs, which can be seperated into two main categories: hypergraph construction and hypergraph utilization. Since proposed in [48], researchers have paid lots of attention on how to construct hypergraphs. For instances, Wang et al. leveraged a sparse representation of the entity feature space to generate hyperedges and learn the relationship among hyperedges and the connected vertices[38]. Instead of simply learning a sparse representation, Liu et al. employed an elastic net to regulate the representation learing [24]. Besides, Feng et al. jointly learned hypergraph representations of multiple hypergraphs by further encouraging the consitancy among different hypergraph representations [10]. However, none of the aforementioned work is able to incorporate the partial-order relations among entities that exist in graded categorical features and numerical features during the construction of hypergraphs. They thus lead to severe information loss and limit the expressiveness of the constructed hypergraphs.
Besides, hypergraph and hypergraph-based learning have been widely applied on many machine learning tasks, including clustering, embedding, ranking, semi-supervised classification and regression [3, 22, 27, 35, 36, 43]. For instance, the authors of [22] constructed a hypergraph to represent the correlations among news readers, news articles, topics and name entities and then ranked the news articles on the hypergraph to make recommendation for readers. In [3], the authors utilized a hypergraph to represent the relations among candidate sentences and made a graph-based extractive document summarization. Yoshida and Yuichi used a hypergraph to estimate the betweenness centrality and importance of vertices [42]. Huang et al. constrcuted a hypergraph of images and predicted the attributes of the images with hypergraph-based label propagation [19]. Tran et al.built a hypergraph where vertices and hyperedges respectively represent features and training samples to represent the sparse pattern of the training data [35]. Hmimida and Kanawati depicted the relation among social users, resources, and the tags of resources assigned by the users. They then employed hypergraph-based ranking to recommend candidate tags for resources [18]. These articles indicated the popularity and usability of hypergraphs and learning on hypergraphs. However they only utilized conventional hypergraphs instead of simultaneously improving the representation ability of the conventional hypergraphs and the corresponding learning methods.
Following the previous work [9], we formulated university ranking as an unsupervised ranking (i.e., re-ranking) problem. Given N universities with a feature matrix $\mathbf {X} \in \mathbb {R}^{N \times M}$ and a historical ranking result $\mathbf {y} \in \mathbb {R}^{N}$ , the target is to learn a new ranking $\mathbf {f} \in \mathbb {R}^N$ . To solve the problem, we manually selected several partial-order relations and constructed a partial-order hypergraph (POH) to represent the given universities. Upon the constructed POH, we learned f by minimizing a ranking instantiation of the POH learning objective function. Specifically, we set the loss term in Equation (8) as $\mathcal {L} = \Vert \mathbf {y} - \mathbf {f}\Vert ^2_F$ , which encourages the learned ranking to be consistent and smooth with the historical one. Besides, we set the soft logic functions {sr |r = 1, 2, ⋅⋅⋅, R} as sr (fi , fj ) = fi − fj , i.e., university i is encouraged to be ranked ahead of university j if the two universities have the selected partial-order relations. By specifying the above application-specific terms, we derived the objective function for the task,
5.1.1 Dataset. To investigate the effectiveness of the proposed method, we employed the dataset of Chinese university ranking collected by [9]. This dataset contains 743 Chinese universities with data collected between January 1st, 2015 and May 1st, 2016. For each university, Web data from five channels were collected, including official, mass media, academic, employment, and general user channels. The official channel contains the primary information of a university, such as student quality, official activities, and development plans. In the mass media channel, they collected news reports mentioning the given university from mass media. The academic and employment channels contain university's academic status and graduate students employment statistics, respectively. The general user channel contains public impressions, attitudes, and sentiment polarities of universities shared in social media posts. To describe the universities, we also employed the rich set of features extracted by the authors of [9]. Among the 743 universities, we selected 438 top-tier ones in China1 since they have more academic and research activities. Towards the historical ranking result y, we merged the ranking results in 2015 from four well-known ranking systems of Chinese universities, namely, CUAA, WSL, WH, and iPIN2. The ranking ground-truth is generated in the same way, but based on the rankings of the four systems in the year 2016. As such, the task can be understood as using the past year's ranking and this year's features to predict the ranking of universities in this year. We normalized the constructed historical ranking result and the ground-truth into range [0, 1] by scaling them with 1/N.
5.1.2 Evaluation. We performed 5-fold cross-validation, employing three metrics to evaluate the ranking performance: mean absolute error (MAE) [41], Kendall's tau (Tau) [28], and Spearman's rank3 (Rho) [34]. The three metrics have been widely used to evaluate pointwise, pairwise, and listwise ranking methods [25]. Note that better performance is evidenced by smaller MAE, larger Tau and Rho scores. Moreover, we carried out the student's t-test and reported the p-values where necessary.
5.1.3 Methods. We compared with following baselines4:
We evaluate several POH methods that incorporate different partial-order relations on the same hypergraph structure of Hypergraph:
Methods | MAE | Tau | Rho |
Simple Graph | 0.074 ± 9e-3 | 0.870 ± 2e-2 | 0.970 ± 8e-3 |
Hypergraph | 0.067 ± 7e-3 | 0.876 ± 9e-3 | 0.974 ± 5e-3 |
GMR | 0.065 ± 7e-3 | 0.871 ± 3e-2 | 0.970 ± 1e-2 |
POH-Salary | 0.054 ± 1e-2* | 0.892 ± 1e-2* | 0.979 ± 5e-3* |
POH-NCEE | 0.055 ± 1e-2* | 0.893 ± 9e-3* | 0.978 ± 5e-3* |
POH-All | $\mathbf {0.053\pm \text{{1e-2}}}^{*}$ | $\mathbf {0.898\pm \text{{1e-2}}}^{**}$ | $\mathbf {0.980\pm \text{{6e-3}}}^{**}$ |
* and ** denote that the corresponding performance is significantly better (p-value < 0.05) than all baselines and all other methods, respectively. |
We employed grid search to select the optimal hyperparameters for POH methods based on the results of Tau. The optimal hyperparameter setting and implementation of the compared methods can be publicly accessed5. For POH-Salary and POH-NCEE, we tuned one implicit (k) and two explicit hyperparameters (λ and β). To validate the strength of the proposed POH over traditional hypergraph, we set k and λ as the optimal ones of the baseline Hypergraph, and then searched β in the range of [1e-4, 1e1]. For POH-All, we tuned one more hyperparameter a 1, which controls the importance of logical rules and is in the range of [0, 1].
Note that we have intentionally fixed k and λ to the optimal ones of Hypergraph, which also simplifies the tuning process. Further tuning k and λ based on the performance of POH methods can lead to even better performance (see Figure 4). Figure 3 shows the performance of POH-All w.r.t. β and a 1. This was accomplished by varying one parameter and fixing the other to the optimal value. As can be seen, our method is rather insensitive to hyperparameters around their optimal settings.
5.2.1 Method Comparison. Table 1 summarizes the performance comparison on university ranking, from which we have the following observations: (1) Hypergraph performs better than Simple Graph, which verifies that considering the higher-order relations among universities is effective for the ranking task. (2) All POH-based methods outperform baselines by a large margin (e.g., POH-All ourperforms GMR with an improvement of 18.46%, 3.10%, and 1.03% w.r.t. MAE, Tau, and Rho, respectively). It demonstrates the effectiveness of our proposed POH and regularized learning in integrating partial-order relations. (3) POH-All outperforms both POH-Salary and POH-NCEE. It further verifies the advantage of POH-based learning methods and reflects that jointly modelling multiple partial-order relations and rules is helpful. (4) The p-values of student's t-test between POH-based methods and all the other methods are smaller than 0.05, indicating the significance of the performance improvements.
As we constructed hypergraphs by connecting a vertex with its k-nearest vertices, larger k makes the POH-based methods consider more vertex pairs with the given partial-order relations (these pairs would be eliminated if the vertex pair is not connected by any hyperedge). It is thus interesting to see how does the setting of k impact the performance of POH learning. Figure 4 shows the performance of Tau of Hypergraph and our POH-based methods on different k. Note that other hyperparameters have been fairly tuned for each setting of k. As can be seen, all POH-based methods outperform the Hypergraph on all settings. It demonstrates that the proposed POH learning consistently outperforms the conventional hypergraph, regardless of the underlying hypergraph structure. Moreover, all POH-based methods achieve performances better than those reported in Table 1, which shows the performance of POH methods on the optimal k of Hypergraph only. It reveals the potential of POH-based methods on further improvements if a better hyperparameter tuning strategy is applied.
To understand the results better, we performed finer-grained error analysis. Given the result generated by a method, we generated an array of rank positions (integers from 1 to N) for universities, computing the absolute error on the rank position on each university.
Figure 5 depicts the distribution of the absolute rank errors as a boxplot. As can be seen, the rank error distribution of POH-based methods is more dense and centralizes at smaller medians than that of Simple Graph and Hypergraph. It provides sufficient evidence on the better ranking generated by the POH-based methods. Moreover, we find that Simple Graph and Hypergraph make errors larger than 5 on about 25% of the universities, which is rarely seen from POH-based methods. Meanwhile, the largest error made by Simple Graph and Hypergraph is almost two times that of POH-based methods. These results demonstrate that POH-based methods are more robust, thus being more applicable in real-world applications. Among the baselines, GMR achieves the smallest rank error, which is comparable with POH-Salary. This signifies the usefulness of modelling the relations among data from different channels, which could be a future direction to be explored by POH-based methods.
Besides the investigation on pointwise rank errors, we further performed an analysis on pairwise ranks. For each method, we counted the number of university pairs that are ranked correctly, and drew the percentage of correct pairs in Figure 6. As shown, POH-based methods manage to generate ranks with correct order on 1% more university pairs than Simple Graph, Hypergraph, and GMR, further demonstrating the accuracy and advantage of POH-based methods. Considering that there are more than 80,000 university pairs, an improvement of 1% (correctly ranking 800+ pairs) is a significant improvement.
Tau
|
Rho
|
|||||||||
CUAA | WSL | WH | iPIN | POH | CUAA | WSL | WH | iPIN | POH | |
CUAA | - | 0.763 | 0.778 | 0.431 | 0.765 | - | 0.903 | 0.903 | 0.573 | 0.886 |
WSL | 0.763 | - | 0.823 | 0.459 | 0.806 | 0.903 | - | 0.950 | 0.627 | 0.943 |
WH | 0.778 | 0.823 | - | 0.469 | 0.832 | 0.903 | 0.950 | - | 0.645 | 0.961 |
iPIN | 0.431 | 0.459 | 0.469 | - | 0.569 | 0.573 | 0.627 | 0.645 | - | 0.750 |
POH | 0.765 | 0.806 | 0.832 | 0.569 | - | 0.886 | 0.943 | 0.961 | 0.750 | - |
Finally, we studied whether the ranking generated by our method POH-All is consistent with the four popular Chinese university ranking systems of year 2016 (cf. Section 5.1.1). We evaluated the pairwise correlation (Tau) and listwise correlation (Rho) between the ranked lists of the five ranking systems (POH-All and the four existing systems). The results are shown in Table 2. We can see that our method achieves rather high correlations with existing ranking systems, and it correlates most with WH. This shows that our result is relatively consistent with these manually devised ranking systems, implying that our POH-based re-ranking method shall be acceptable by the general readers.
Predicting the popularity of online content is a hot research topic in social media mining and has varying problem statements [16, 33]. Following the recent work on micro-video popularity prediction [5], we formulated the task as a semi-supervised regression problem. Given N + U items with a feature matrix $\mathbf {X} \in \mathbb {R}^{(N + U) \times M}$ and the ground-truth popularity of the N items $\mathbf {y} \in \mathbb {R}^{N}$ , the objective is to learn a function $\hat{y_i} = f(\mathbf {x_i})$ that maps an item from the feature space to the popularity space. To solve the problem, we first constructed a POH with partial-order relations on some important numerical features (detailed later in experiments). We then derived an instantiation of the general framework Equation (8) for the semi-supervised regression task as follows:
6.1.1 Dataset. We employed the same dataset as [5] for experiments. It contains 9,719 micro-videos collected from Vine6, posted between July 1st and October 1st, 2015. Each micro-video has visual, audio, and textual contents, as well as the profile of the user who posted it. With these data, the authors of [5] extracted a rich set of popularity-oriented features, such as user activities, object distribution, aesthetic description, sentence embedding, and textual sentiment polarity, to represent a micro-video. To measure the popularity of a micro-video, they employed four popularity-related indicators, namely, the number of comments ($n\_comments$ ), the number of likes ($n\_likes$ ), the number of reposts ($n\_reposts$ ), and the number of loops ($n\_loops$ ); the four indicators were averagely fused ($(n\_comments + n\_likes + n\_reposts + n\_loops) / 4$ ) as the popularity ground-truth for a micro-video.
6.1.2 Evaluation. We performed 10-fold cross-validation and evaluated the performance in terms of three metrics. From the regression perspective, we followed the previous work [5] and employed normalized mean square error (nMSE). Meanwhile, we utilized two ranking-oriented metrics, Tau and Rho correlation coefficients. Besides, we carried out the student's t-test and reported the p-values where necessary.
6.1.3 Methods. We compare with following baselines:
We evaluated several POH methods on the same hypergraph structure of Hypergraph:
6.1.4 Hyperparameter Tuning. We employed the same procedure as described in Section 5.1.4 to tune the hyperparameters of POH methods. Optimal hyperparameter settings of each compared method will be released together with their codes. We investigated the sensitivity of our proposed POH-based methods by taking POH-All as an example. Figure 7 illustrates the performance of POH-All while varying one hyperparameter and fixing the others with optimal values. Again, the results demonstrate that our model is not sensitive to the parameters around their optimal settings.
Methods | nMSE | Tau | Rho |
Simple Graph | 0.999 ± 1e-3 | 0.137 ± 2e-2 | 0.200 ± 2e-2 |
Hypergraph | 1.000 ± 4e-5 | 0.165 ± 3e-2 | 0.240 ± 4e-2 |
TMALL7 | 0.979 ± 9e-3 | - | - |
POH-Follow | 1.000 ± 4e-4 | 0.393 ± 3e-2* | 0.562 ± 3e-2* |
POH-Loop | 0.997 ± 2e-3 | 0.376 ± 2e-2* | 0.540 ± 3e-2* |
POH-All | 0.989 ± 9e-3 | 0.419 ± 2e − 2 ** | 0.592 ± 3e − 2 ** |
GCN | 0.919 ± 6e-2 | 0.171 ± 2e-2 | 0.252 ± 3e-2 |
LR-HG | 0.846 ± 1e-1* | 0.117 ± 2e-2 | 0.182 ± 3e-2 |
LR-POH | 0.724 ± 2e − 1 ** | 0.350 ± 2e-2* | 0.496 ± 3e-2* |
* and ** denote that the corresponding performance is significantly better (p-value < 0.05) than all baselines and all other methods, respectively. |
We first investigated the effectiveness of the proposed methods. Table 3 shows the performance of all the compared methods. We have the following findings:
(1) Hypergraph outperforms Simple Graph w.r.t. Tau and Rho, although they achieve the same performance level on nMSE. It verifies that considering the higher-order relations among videos leads to popularity prediction with more accurate relative orders.
(2) POH-Follow and POH-Loop further surpass Hypergraph with an average improvement of 133.03% and 129.58% on the pairwise and listwise ranking metrics; meanwhile, slight improvement is obtained on the pointwise regression metric nMSE. This indicates that considering meaningful partial-order relations is particularly helpful for better predicting the relative order of the videos.
(3) POH-All outperforms POH-Follow and POH-Loop with a significant average improvement on Tau (+8.97%) and Rho (+7.44%) as well as a slight improvement on nMSE. It validates that jointly considering multiple partial-order relations is useful.
(4) Comparing Hypergraph with LR-HG, we can see that better nMSE can be achieved by using LR as the predictive model, but the two ranking metrics become worse. The same situation can be observed for POH-All and LR-POH. This provides evidence that using a sophisticated model can better fit the labels and help to minimize the regression loss, however, the ranking performance may not be necessarily improved. The same finding has been observed before in popularity prediction [16] and another orthogonal application of item recommendation [7]. In our case of graph-based learning, the regularizers (for smoothness and partial-order rules) carry strong signals for learning the relative orders between vertices. However, the regularization effects might be weakened when a specialized model is used to fit the label in the meantime. We leave more detailed exploration of this hypothesis as future work. (5) GCN outperforms POH-All w.r.t. nMSE, while Tau and Rho indicate that its ranking performance is worse. The lower nMSE of GCN can be credit to the strong representation power of the underlying neural network, which can fit the labels well. However, GCN may overfit the data and fail to predict the popularity ranking well without regularization on the relative orders of vertices.
(6) LR-POH achieves the best performance with significantly better nMSE than all the other compared methods as well as tremendously better Tau and Rho than all the baseline methods. This further demonstrates the effectiveness of our proposed POH learning.
We further studied whether the performance improvements of the proposed POH-based methods are consistent under different hypergraph settings. We compared the optimal performance of Hypergraph, POH-Follow, POH-Loop, and POH-All under different values of k, which controls the number of videos connected by a hyperedge. As illustrated in Figure 8, all POH-based methods outperform the Hypergraph under all the values of k by a large margin. It is worth noting that the optimal performance of POH methods are better than that shown in Table 3 (Table 3 shows the results of POH on the optimal setting of Hypergraph). This is consistent with the university ranking task, which implies the potential of further improving POH learning with a better hyperparameter tuning strategy.
In this paper, we proposed a novel partial-order hypergraph that improves conventional hypergraphs by encoding the partial-order relations among vertices. We then generalized existing graph-based learning methods to partial-order hypergraphs by integrating the second-order logic rules that encode the partial-order relations; moreover, the time complexity of learning remains unchanged. Experimental results on university ranking and video popularity prediction demonstrate the effectiveness of our proposed methods.
In future, we will explore the proposed POH to address more graph-based applications. Besides, we will further improve POH learning by replacing the linear prediction function from feature to label spaces with the advanced deep neural networks. Moreover, we plan to optimize the spatial complexity of POH with discrete hashing techniques like [46]. Furthermore, we would like to investigate the automatic extraction of partial-order relations and logical rules to construct POH.
⁎Xiangnan He is the corresponding author. This research is part of the NExT++ project, supported by the National Research Foundation, Prime Minister's Office, Singapore under its IRC@Singapore Funding Initiative.
1In China, universities were officially separated into three tiers by Ministry of Education (https://tinyurl.com/moe-univ-list/.).
2CUAA: http://www.cuaa.net/cur/. WSL: http://edu.sina.com.cn/gaokao/wushulian/. WH: http://www.nseac.com/html/168/. iPIN: https://www.wmzy.com/api/rank/schList/.
3Note that we omitted listwise ranking evaluation metrics like Precision@K [6], Recall@K [39], and NDCG@K [29] since they are sensitive to the selection of K
4Note that we omitted the comparison with TMALL and GCN mentioned in Section 6.1.3. Because [9] has shown that TMALL (similar to the JL baseline in their paper) is less effective than GMR; GCN is not fit for this unsupervised ranking task as it is designed for semi-supervised and supervised tasks [21]
5https://github.com/hennande/Partial_Order_Hypergraph
7This result is directly copied from the corresponding paper since we exactly followed their experimental settings.
This paper is published under the Creative Commons Attribution 4.0 International (CC-BY 4.0) license. Authors reserve their rights to disseminate the work on their personal and corporate Web sites with the appropriate attribution.
WWW '18, April 23-27, 2018, Lyon, France
© 2018; IW3C2 (International World Wide Web Conference Committee), published under Creative Commons CC-BY 4.0 License. ACM ISBN 978-1-4503-5639-8/18/04.
DOI: https://doi.org/10.1145/3178876.3186064