Meta Path-Based Collective Classification in Heterogeneous Information Networks

05/20/2013 ∙ by Xiangnan Kong, et al. ∙ 0

Collective classification has been intensively studied due to its impact in many important applications, such as web mining, bioinformatics and citation analysis. Collective classification approaches exploit the dependencies of a group of linked objects whose class labels are correlated and need to be predicted simultaneously. In this paper, we focus on studying the collective classification problem in heterogeneous networks, which involves multiple types of data objects interconnected by multiple types of links. Intuitively, two objects are correlated if they are linked by many paths in the network. However, most existing approaches measure the dependencies among objects through directly links or indirect links without considering the different semantic meanings behind different paths. In this paper, we study the collective classification problem taht is defined among the same type of objects in heterogenous networks. Moreover, by considering different linkage paths in the network, one can capture the subtlety of different types of dependencies among objects. We introduce the concept of meta-path based dependencies among objects, where a meta path is a path consisting a certain sequence of linke types. We show that the quality of collective classification results strongly depends upon the meta paths used. To accommodate the large network size, a novel solution, called HCC (meta-path based Heterogenous Collective Classification), is developed to effectively assign labels to a group of instances that are interconnected through different meta-paths. The proposed HCC model can capture different types of dependencies among objects with respect to different meta paths. Empirical studies on real-world networks demonstrate that effectiveness of the proposed meta path-based collective classification approach.

READ FULL TEXT VIEW PDF
POST COMMENT

Comments

There are no comments yet.

Authors

page 1

page 2

page 3

page 4

This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction

Collective classification methods that exploit the linkage information in networks to improve classification accuracies have been studied intensively in the last decade. Different from conventional supervised classification approaches that assume data are independent and identically distributed, collective classification methods aim at exploiting the label autocorrelation among a group of inter-connected instances and predict their class labels collectively, instead of independently. In many network data [19, 5], the instances are inter-related with complex dependencies. For example, in bibliographic networks, if two papers both cite (or are cited by) some other papers (i.e., bibliographic coupling or co-citation relationship) or one paper cites the other (i.e., citation relationship), they are more likely to share similar research topics than those papers without such relations. These dependencies among the related instances should be considered explicitly during classification process. Motivated by these challenges, collective classification problem has received considerable attention in the literature [13, 19, 10].

Figure 1: A Heterogeneous Information Network

Most approaches in collective classification focus on exploiting the dependencies among different interconnected objects, e.g., social networks with friendship links, webpage networks with hyper-links. With the recent advance in data collection techniques, many real-world applications are facing large scale heterogeneous information networks [16] with multiple types of objects inter-connected through multiple types links. These networks are multi-mode and multi-relational networks, which involves large amount of information. For example, a bibliographic network in Figure 1 involves five types of nodes (papers, author, affiliations, conference and proceedings) and five types of links. This heterogeneous information network is more complex and contain more linkage information than its homogenous sub-network, i.e., a paper network with only citation links.

In this paper, we focus on studying the problem of collective classification on one type of nodes within a heterogenous information network, e.g.

, classifying the paper nodes collectively in Figure 

1. Formally, the collective classification problem in heterogeneous information networks corresponds to predicting the labels of a group of related instances simultaneously. Collective classification is particularly challenging in heterogenous information networks. The reason is that, in the homogenous networks, conventional collective classification methods can classify a group of related instances simultaneously by considering the dependencies among instances inter-connected through one type of links. But in heterogeneous network, each instance can have multiple types of links, and the dependencies among related instances are more complex.

Notation Meta Path Semantics of the Dependency
1 PP Paper Paper Citation
2 PPP Paper Paper Paper Co-citation
3 PPP Paper Paper Paper Bibliographic coupling
4 PVP Paper Proceeding Paper Papers in the same proceeding
5 PVCVP Paper Proceeding Conference
Proceeding Paper Papers in the same conference
6 PAP Paper Author Paper Papers sharing authors
7 PAFAP Paper Author Institute
Author Paper Papers from the same institute
Table 1: Semantics of Meta Paths among Paper Nodes

If we consider collective classification and heterogeneous information networks as a whole, the major research challenges can be summarized as follows:

Multi-Mode and Multi-Relational Data: One fundamental problem in classifying heterogeneous information networks is the complex network structure that involves multiple types of nodes and multiple types of links. For example, in Figure 1, one paper node can be linked directly with different types of objects, such as authors, conference proceedings and other papers, through different types of links, such as citation, authoredBy, etc. Different types of links have totally different semantic meanings. Trivial application of conventional methods by ignoring the link types and node types can not fully exploit the structural information within a heterogeneous information network.

Heterogeneous Dependencies: Another problem is that objects in heterogeneous information networks can be linked indirectly through different types of relational paths. Each types of relational path corresponds to different types of indirect relationships between objects. For example, in Figure 1, paper nodes can be linked with each other indirectly through multiple indirect relationships, such as, 1) the “paper-author-paper” relation indicates relationships of two papers sharing same authors; 2) the “paper-author-institute-author-paper” relation denotes relationship between papers that are published from the same institute. Heterogenous information networks can encode various complex relationships among different objects. Thus, ignoring or treating all relations equally will loss information dependence information in a heterogeneous information network. Exploring such heterogeneous structure information has been shown useful in many other data mining tasks, such as ranking [9, 8], clustering [17, 18] and classification tasks [6].

In this paper, we study the problem of collective classification in heterogeneous information networks and propose a novel solution, called Hcc (meta-path based Heterogenous Collective Classification), to effectively assign class labels to one type of objects in the network. Different from conventional collective classification methods, the proposed Hcc model can exploit a large number of different types of dependencies among objects simultaneously. We define meta path-based dependencies to capture different types of relationships among objects. By explicitly exploiting these dependencies, our Hcc method can effectively exploit the complex relationships among objects. Empirical studies on real-world tasks demonstrate that the proposed approach can significantly boost the collective classification performances in heterogeneous information networks.

The rest of the paper is organized as follows. We start by a brief review on related work of collective classification and heterogeneous information networks. Then we introduce the preliminary concepts, give the problem definitions in Section 3 and present the Hcc algorithm in Section 4. Then Section 5 reports the experiment results. In Section 6, we conclude the paper.

2 Related Work

Our work is related to both collective classification techniques on relational data and heterogeneous information networks. We briefly discuss both of them.

Collective classification of relational data has been investigated by many researchers. The task is to predict the classes for a group of related instances simultaneously, rather than predicting a class label for each instance independently. In relational datasets, the class label of one instance can be related to the class labels (sometimes attributes) of the other related instances. Conventional collective classification approaches focus on exploiting the correlations among the class labels of related instances to improve the classification performances. Roughly speaking, existing collective classification approaches can be categorized into two types based upon the different approximate inference strategies: (1) Local methods: The first type of approaches employ a local classifier to iteratively classify each unlabeled instance using both attributes of the instances and relational features derived from the related instances. This type of approaches involves an iterative process to update the labels and the relational features of the related instances, e.g. iterative convergence based approaches [13, 10] and Gibbs sampling approaches [12]. Many local classifiers have been used for local methods, e.g.logistic regression [10]

, Naive Bayes

[13], relational dependency network [14], etc. (2) Global methods: The second type of approaches optimizes global objective functions on the entire relational dataset, which also uses both attributes and relational features for inference [19]. For a detailed review of collective classification please refer to [15].

Heterogeneous information networks are special kinds of information networks which involve multiple types of nodes or multiple types of links. In a heterogeneous information network, different types of nodes and edges have different semantic meanings. The complex and semantically enriched network possesses great potential for knowledge discovery. In the data mining domain, heterogeneous information networks are ubiquitous in many applications, and have attracted much attention in the last few years [18, 17, 6]. Sun et al. [18, 16] studied the clustering problem and top-k similarity problem in heterogeneous information networks. Ming et al. studied a specialized classification problem on heterogeneous networks, where different types of nodes share a same set of label concepts [6]. However, these approaches are not directly applicable in collective classification problems, since focus on convention classification tasks without exploiting the meta path-based dependencies among objects.

3 Problem Definition

In this section, we first introduce several related concepts and notations. Then, we will formally define the collective classification problem in heterogeneous information networks..

Definition 1. Heterogeneous Information Network: A heterogeneous information network [18, 16] is a special kind of information network, which is represented as a directed graph . is the set of nodes, including types of objects . is the set of links between the nodes in , which involves multiple types of links.

Example 1. ACM conference network: A heterogeneous information network graph is provided in Figure 1. This network involves five types of objects, i.e., papers (P), authors (A), institutes (F), proceedings (V) and conferences (C), and five types of links, i.e., citation, authoredBy, affiliation, publishedIn and collectedIn.

Symbol Definition
the set of nodes, involving types of nodes
the set of edges or links
the given attribute values for each node in target type
the set of variables for labels of the nodes in , and
and the sets for training nodes and testing nodes, and
the given label for node , and
the set of meta paths
the index set of all related instances to through meta path
Table 2: Important Notations.

Different from conventional networks, heterogeneous information networks involve different types of objects (e.g., papers and conference) that are connected with each other through multiple types of links. Each type of links represents an unique binary relation from node type to node type , where holds iff object and are related by relation . denotes the inverted relation of , which holds naturally for . Let denote the domain of relation , denotes its range. . For example, in Figure 1, the link type “authorBy” can be written as a relation between paper nodes and author nodes. holds iff author is one of the authors for paper . For convenience, we can write this link type as “” or “”.

In heterogenous information networks, objects are also inter-connected through indirect links, i.e., paths. For example, in Figure 1, paper 1 and paper 4 are linked through a sequence of edges: “”. In order to categorize these paths, we extend the definition of link types to “path types”, which are named as meta path, similar to [16, 9].

Definition 2. Meta Path: A meta path represents a sequence of relations with constrains that . The meta path can also be written as , i.e., corresponds to a composite relation between node type and . and . The length of is , i.e., the number of relations in .

Different meta paths usually represent different semantic relationships among linked objects. In Table 1, we show some examples of meta paths with their corresponding semantics. Most conventional relationships studied in network data can naturally be captured by different meta paths. For example, the paper co-citation relation [3] can naturally be represented by meta path “”, and the co-citation frequencies can be written as the number of path instances for the meta path. Here a path instance of , denoted as , is an unique sequence of nodes and links in the network that follows the meta path constrains. For convenience, we use the node type sequence to represent a meta path, i.e., . For example, we use to represent the meta path “”. Note that for meta paths involving citation links, we explicitly add arrows to represent the link directions, e.g., the paper co-citation path can be written as .

Collective Classification in Heterogeneous Information Networks
In this paper, we focus on studying the collective classification problem on one type of objects, instead of on all types of nodes in heterogeneous information networks. This problem setting exists in a wide variety of applications. The reasons are as follows: in heterogenous information networks, the label space of different types of nodes are quite different, where we can not assume all types of node share the same set of label concepts. For example, in medical networks, the label concepts for patient classification tasks are only defined on patient nodes, instead of doctor nodes or medicine nodes. In a specific classification task, we usually only care about the classification results on one type of node. Without loss of generality, we assume the node type is the target objects we need to classify. Suppose we have nodes in . On each node

, we have a vector of attributes

in the -dimensional input space, and . Let be the possible class labels. On each node , we also have a label variable indicating the class label assigned to node , .

Assume further that we are given a set of known values for nodes in a training set , and denotes the index set for training data. , where is the observed labels assigned to node . Then the task of collective classification in heterogeneous information networks is to infer the values of for the remaining nodes in the testing set ().

Figure 2: Meta path-based dependencies in collective classification for heterogeneous information networks. with double circles denotes the current label variable to be predicted. Each rectangle represents a group of instances following the same meta path. denotes the attribute values of the instance.

As reviewed in Section 2

, the inference problem in classification tasks is to estimate

given a labeled training set. Conventional classification approaches usually require i.i.d. assumptions, the inference for each instance is performed independently:

Homogeneous Link-based Dependency
In collective classification problems, the labels of related instances are not independent, but are closely related with each other. Conventional approaches focus on exploiting label dependencies corresponding to one types of homogeneous links to improve the classification performances, e.g., citation links in paper classification tasks, co-author links in expert classification tasks. These methods can model . Here denotes the vector containing all variable (), and denotes the index set of related instances to the -th instance through meta path . Hence, by considering the single type of dependencies, we will have

Meta Path-based Dependency
In heterogeneous information networks, there are complex dependencies not only among instances directly linked through links, but also among instances indirectly linked through different meta paths. In order to solve the collective classification problem more effectively, in this paper, we explicitly consider different types of meta-path based dependencies in heterogeneous information networks. Meta path-based dependences refer to the dependencies among instances that are inter-connected through a meta path.

To the best of our knowledge, meta path-based dependencies have not been studied in collective classification research before. Given a set of meta paths , the meta path-based dependency models are shown in Figure 2, i.e., . denotes the index set of related instances to the -th instance through meta path .

For each meta path, one instance can be connected with multiple related instances in the network. For example, in Figure 3, Paper 1 is correlated with Paper 2, 3 and 4 through meta path , i.e., . Hence, by considering meta path-based dependencies, we will have

4 Meta Path-based Collective Classification

For classifying target nodes in a heterogeneous information network, the most naïve approach is to approximate with the assumptions that all instances are independent from each other. However, this approach can be detrimental to their performance for many reasons. This is particularly troublesome when nodes in heterogeneous networks have very complex dependencies with each other through different meta paths.

In this section, we propose a simple and effective algorithm for meta path-based collective classification in heterogeneous information networks. We aim to develop a model to estimate the probabilities

. We first introduce how the extract the set of meta paths from a heterogeneous information network, then propose our collective classification algorithm, called Hcc (Heterogeneous Collective Classification).

We first consider how to extract all meta paths in a heterogeneous information network of bounded length . When is small, we can easily generate all possible meta paths as follows: We can organize all the type-correct relations into a prefix tree, called dependence tree. In Figure 4, we show an example of dependence tree in ACM conference networks. The target nodes for classification are the paper nodes, and each paper node in the dependence tree corresponds to an unique meta path, indicating one type of dependencies among paper instances. However, in general the number of meta paths grows exponentially with the maximum path length . As it has been showed in [16], long meta paths may not be quite useful in capturing the linkage structure of heterogeneous information networks. In this paper, we only exploit the instance dependences with short meta paths ().

Figure 3: Path instances corresponding to the meta path .

In many really world network data, exhaustively extracting all meta paths may result in large amount of redundant meta paths, e.g., . Including redundant meta paths in a collective classification model can result in overfitting risks, because of additional noisy features. Many of the redundant meta paths are constructed by combining two or more meta paths, e.g., meta path can be constructed by two paths. In order to reduce the model’s overfitting risk, we extract all meta paths that cannot be decomposed into shorter meta paths (with at least one non-trivial meta paths). Here non-trivial meta paths refer to the paths with lengths greater than 1. For example, in ACM conference network, meta paths like can be decomposed into and , thus will be excluded from our meta path set. In Figure 5, we the meta path set extract process as the “Initialization” step of our proposed method. By breadth-first search on the dependence tree, our model first select shortest meta paths from the network. Then longer meta paths are incrementally selected into path set until we reach a meta path that can be decomposed into shorter meta paths in .

After the meta path set is extracted from the heterogeneous information network, we then show how to use these meta paths to perform collective classification effectively. Conventional collective classification based on iterative inference process, e.g. ICA (Iterative Classification Algorithm) [13, 10], provide a simple yet very effective method for collective classification in homogeneous networks. Inspired by the success of these iterative inference methods, in this paper, we propose a similar framework for meta path-based collective classification method. This approach is called Hcc (Heterogeneous Collective Classification), summarized in Figure 5.

Figure 4: An example of dependence tree for meta path-based dependencies. Each paper node corresponds to a unique type of path-based dependencies in the network.
Input:
: a heterogeneous information network, : maximum meta path length.
: attribute vectors for all target instances, : labels for the training instances.
: the index set for training instances, : the index set for testing instances.
A: a base learner for local model, : maximum # of iteration.
Initialize:
 - Construct the meta path set by searching the dependence tree on :
  Breadth first search on dependence tree by adding short meta paths into first:
   1. If the length of meta path in current tree node is greater than , exit the BFS;
   2. If the current meta path in current tree node cannot be reconstructed by the paths in ,
    Add into ; Otherwise, prune the current node from BFS.
Training:
 - Learn the local model :
  1. Construct an extended training set by converting each instance to as follows:
    
  2. Let be the local model trained on .
Bootstrap:
 - Estimate the labels, for
  1. Produce an estimated value for as follows:
     = using attributes only.
Iterative Inference:
 - Repeat until convergence or #iteration
  1. Construct the extended testing instance by converting each instance to () as follows:
    
  2. Update the estimated value for on each testing instance () as follows:
     = ().
Output:
: The labels of test instances ().
Figure 5: The Hcc algorithm

The general idea is as follows: we model the joint probability based on the following assumption: if instance and are not connected via any meta path in , the variable is conditional independent from given the labels of all ’s related instances, i.e., . Hence the local conditional probability each instance’s label can be modeled by a base learner with extended relational features built upon the predicted ’s (). And the joint probability can be modeled based on these local conditional probabilities by treating the instances as being independent.

  For each meta path :
    1. Get related instances
    2.
  Return relational feature
Figure 6: Constructing meta path-based relational features (PathRelFeature).

In collective classification, each instance may be linked with different number of instances through one meta path. In order to build a fixed number of relational features for each instance, we employs aggregation functions to combine the predictions on the labels of related instances. Many aggregation functions can be used here, such as COUNT and MODE aggregators [10]. In this paper, we use the weighted label fraction of the related instances as the relational feature for each meta path. We calculate the average fraction of each label appearing in the related instances. Each related instance in re-weighted by the number of path instances between from the current node, e.g., for meta path , the papers that share more authors in their author lists are more likely to share similar topics than those only share one author. In detail, given an aggregation function, we can get one set of relational features from the labels of related instances for each meta path, as shown in Figure 6.

Inspired by the success of ICA framework [10, 11, 12] in collective classification, we designed a similar inference procedure for our Hcc method as shown in Figure 5. (1) For inference steps, the labels of all the unlabeled instances are unknown. We first bootstrap an initial set of label estimation for each instance using content attributes of each node. In our current implementation, we simply set the relational features of unlabeled instances with zero vectors. Other strategies for bootstrap can also be used in this framework. (2) Iterative Inference: we iteratively update the relational features based on the latest predictions and then these new features are used to update the prediction of local models on each instance. The iterative process terminates when convergence criteria are met. In our current implementation, we update the variable in the -th iteration ( say ) using the predicted values in the -th iteration () only.

Data Sets
Characteristics ACM-A ACM-B DBLP SLAP
# Feature 1,903 376 1,618 3,000
# Instance 12,499 10,828 4,236 3714
# Node Type 5 5 2 10
# Link Type 5 5 2 11
# Class 11 11 4 306
Table 3: Summary of experimental datasets.

5 Experiments

5.1 Data Collection

In order to validate the collective classification performances, we tested our algorithm on four real-world heterogeneous information networks (Summarized in Table 3).

(a) ACM Conference Datasets
(b) DBLP Dataset
(c) SLAP Dataset
Figure 7: Schema of datasets
  • ACM Conference Dataset: Our first dataset studied in this paper was extracted from ACM digital library111http://dl.acm.org/ in June 2011. ACM digital library provides detailed bibliographic information on ACM conference proceedings, including paper abstracts, citation, author information etc. We extract two ACM sub-networks containing conference proceedings before the year 2011.

    • The first subset, i.e., ACM Conference-A, involves 14 conferences in computer science: SIGKDD, SIGMOD, SIGIR, SIGCOMM, CIKM, SODA, STOC, SOSP, SPAA, MobiCOMM, VLDB, WWW, ICML and COLT. The network schema is summarized in Figure 7(a), which involves five types of nodes and five types of relations/links. This network includes 196 conference proceedings (e.g., KDD’10, KDD’09, etc.), 12.5K papers, 17K authors and 1.8K authors’ affiliations. On each paper node, we extract bag-of-words representation of the paper title and abstract to use as content attributes. The stop-words and rare words that appear in less than 100 papers are removed from the vocabulary. Each paper node in the network is assigned with a class label, indicating the ACM index term of the paper including 11 categories. The task in this dataset is to classify the paper nodes based on both local attributes and the network information.

    • The second subset, i.e., ACM Conference-B, involves another 12 conferences in computer science: ACM Multimedia, OSDI, GECCO, POPL, PODS, PODC, ICCAD, ICSE, ICS, ISCA, ISSAC and PLDI. The network includes 196 corresponding conference proceedings, 10.8K papers, 16.8K authors and 1.8K authors’ affiliations. After removing stop-words in the paper title and abstracts, we get 0.4K terms that appears in at least 1% of the papers. The same setups with ACM Conference-A dataset are also used here to build the second heterogeneous network.

  • DBLP Dataset: The third dataset, i.e., DBLP four areas222http://www.cs.illinois.edu/homes/mingji1/DBLP_four_area.zip [7], is a bi-type information network extracted from DBLP333http://www.informatik.uni-trier.de/~ley/db/, which involves 20 computer science conferences and authors. The relationships involve conference-author links and co-author links. On the author nodes, a bag-of-words representation of all the paper titles published by the author is used as attributes of the node. Each author node in the network is assigned with a class label, indicating research area of the author. The task in this dataset is to classify the author nodes based on both local attributes and the network information. For detailed description of the DBLP dataset, please refer to [7].

  • SLAP Dataset: The last dataset is a bioinformatic dataset SLAP [2], which is a heterogeneous network composed by over nodes and edges. As shown in Figure 7(c), the SLAP dataset contains integrated data related to chemical compounds, genes, diseases, side effects, pathways etc. The task we studied is gene family prediction, where we treat genes as the instances, and gene family as the labels. In SLAP dataset, each gene can belong to one of the gene family. The task of gene family prediction is that, we are given a set of training gene instances, and for each unlabeled gene instance, we want to predict which gene family the gene belongs to. In details, we extracted gene ontology terms (GO terms) and used them as the features of each gene instance.

Method Type of Classification Dependencies Exploited Publication
Svm Multi-Class Classification All independent [1]
Ica Collective Classification Citation links [10]
Cp Combined Relations Combine multiple relations [4]
Cf Collective Fusion Ensemble learning on multiple relations [4]
Hcc Multiple paths Meta-path based dependencies This paper
Table 4: Summary of compared methods.

5.2 Compared Methods

In order to validate the effectiveness of our collective classification approach, we test with following methods:

  • Heterogeneous Collective Classification (Hcc): We first test our proposed method, Hcc, for collective classification in heterogeneous information networks. The proposed approach can exploit dependencies based on multiple meta paths for collective classification.

  • Homogeneous Collective Classification (Ica): This method is our implementation of the ICA (Iterative Classification Algorithm) [10] by only using homogeneous network information for collective classification. In the homogeneous information networks, only paper-paper links in ACM datasets and author-author links in DBLP dataset are used.

  • Combined Path Relations (Cp): We compare with a baseline method for multi-relational collective classification [4]: We first convert the heterogeneous information networks into multiple relational networks with one type of nodes and multiple types of links. Each link type corresponds to a meta path in the Hcc method. Then, the Cp method combines multiple link types into a homogeneous network by ignoring the link types. We then train one Ica model to perform collective classification on the combined network.

  • Collective Ensemble Classification (Cf): We compare with another baseline method for multi-relational collective classification. This method is our implementation of the collective ensemble classification [4], which trains one collective classification model on each link types. We use the same setting of the Cp method to extract multi-relational networks. Then we use Ica as the base models for collective classification. In the iterative inference process, each model vote for the class label of each instance, and prediction aggregation was performed in each iteration. Thus this process is also called collective fusion, where each base model can affect each other in the collective inference step.

  • Ceiling of Hcc (Hcc-ceiling): One claim of this paper is that Hcc can effectively infer the labels of linked unlabeled instances using iterative inference process. To evaluate this claim, we include a model which use the ground-truth labels of the related instances during the inference. This method illustrate a ceiling performance of Hcc can possibly achieve by knowing the true label of related instances.

  • Hcc with all meta-paths (Hcc-all): Another claim of this paper is that selected meta path in Hcc can effectively capture the dependencies in heterogeneous information networks and avoiding overfitting. To evaluate this claim, we include a model which uses all possible meta paths with a maximum path length of 5. This method illustrates the performance of Hcc if we exhaustively involves all possible path-based dependencies without selection.

We use LibSVM with linear kernel as the base classifier for all the compared methods. The maximum number of iteration all methods are set as 10. All experiments are conducted on machines with Intel Xeon™Quad-Core CPUs of 2.26 GHz and 24 GB RAM.

(a) ACM Conference-A
(b) ACM Conference-B
(c) DBLP
(d) SLAP
Figure 8: Collective classification results.

5.3 Performances of Collective Classification

In our first experiment, we evaluate the effectiveness of the proposed Hcc method on collective classification. 10 times -fold cross validations are performed on each heterogeneous information network to evaluate the collective classification performances. We report the detailed results in Figure 8

. It shows the performances of the six methods on three datasets with box plots, including the smallest/largest results, lower quartile, median and upper quartile. In DBLP dataset, note that

Hcc-all is equivalent to Hcc method due to the fact that the schema graph of DBLP is relatively simple. In SLAP dataset, the schema is much more complex than all the other datasets, and in this case, Hcc-all is too computationally expensive. And we didn’t show Cf in SLAP dataset, because the performance is not as good as other baselines.

The first observation we have in Figure 8 is as follows: almost all the collective classification methods that explicitly exploit the label dependencies from various aspects, can achieve better classification accuracies than the baseline Svm, which classify each instance independently. These results can support the importance of collective classification by exploiting the different types of dependencies in network data. For example, Ica outperformances Svm by exploiting autocorrelation among instances while considering only one type of links, i.e., citation links in ACM datasets and co-author links in the DBLP dataset. Cf and Cp methods can also improve the classification performances by exploiting multiple types of dependencies. Similar results have also been reported in collective classification literatures.

Then we find that our meta path-based collective classification method (Hcc) consistently and significantly outperform other baseline methods. Hcc can utilize the meta path-based dependencies to exploit the heterogenous network structure more effectively. These results support our claim that in heterogeneous information networks, instances can be correlated with each other through various meta paths. Exploiting the complex dependencies among the related instances (i.e., various meta path-based dependencies) can effectively extract the heterogenous network structural information and thus boost the classification performance. Furthermore, Hcc is able to improve the classification performance more significantly in datasets with complex network schemas (ACM datasets) than that with simpler network structure (DBLP dataset).

We further observe that the Hcc models perform comparably with the Hcc-ceiling models which had access to the true label of related instances. This indicates that the Hcc model reach its full potential in approximated inference process. In addition, Hcc method with a small set of representative paths can achieve also comparable performances with Hcc-all models which includes all meta path combinations with path length . And the performances of Hcc are more stable than those Hcc-all in ACM datasets. In ACM Conference-B dataset, Hcc method with fewer meta-paths can even outperform Hcc-all method. These results support our second claim that the heterogeneous dependencies can be captured effectively by selecting a small set of representative meta-paths and thus our Hcc model can avoid overfitting than using all meta paths.

dataset method # path accuracy train time (se) test time (se)
ACM-Conf-A Svm 0 0.649 60.0 25.2
Ica 1 0.663 59.9 111.0
Hcc *6 *0.717 *88.3 *440.7
Hcc-all 50 0.722 332.0 1352.7
ACM-Conf-B Svm 0 0.557 9.5 15.2
Ica 1 0.581 16.7 206.6
Hcc *6 *0.658 *36.7 *325.6
Hcc-all 50 0.643 202.0 1130.2
Table 5: Results of running time. “# path” represents the number of meta paths explored by the method.

5.4 Running Time

In Table 5, we show the running time results of collective classification methods with different number of meta paths explored. In heterogeneous information networks, there are a large number of possible meta-paths. The more meta-paths are considered in the methods, the training and testing time will both be longer. For example, the Hcc-all method incorporates all possible meta paths with path length , i.e., 50 meta-paths in ACM datasets. The training time will be much slower with many meta-paths considered, because the additional computational costs for aggregating class labels through meta paths and the additional dimensions of features for the base learners. The testing times are also significantly affected by the number meta paths. When more paths are considered, the method needs to aggregate more labels from neighboring instances in each iteration during the inference. Moreover, when more dependencies are considered, the convergence rate of collective classification methods will also be slower. Based upon the above observation, we can see that our Hcc method only uses a small set of meta paths, and can achieve comparable or better performances than Hcc that uses all meta paths. These results support our motivation on using and selecting meta paths during the collective classification.

5.5 Influence of Meta Paths

In this subsection, we study the influence of different meta paths on the collective classification performance of our Hcc model. In heterogeneous information networks, different types of meta path correspond to different types of auto-correlations among instances, thus have different semantic meanings. In order to illustrate the influence of each path, we compare 6 different versions of Hcc model which exploit different paths separately.

We denote the Hcc model with only “paper-author-paper” path as “PAP”, and it can exploit the auto-correlation among papers which share authors. Similarly, “PVP” represents Hcc with “paper-proceeding-paper” path only. Here “PP*” denotes the Hcc with the set of paths that are composed by citation links: , , , , and . In this baseline, complex paths composed with citation links () are full exploited. The “iid” method represents the i.i.d. classification model using Svm.

(a) ACM Conference-A
(b) ACM Conference-B
Figure 9: Influences of the meta paths.

Figure 9 shows the collective classification performances using different meta paths. One can see that two paths are most relevant to the collective classification tasks in ACM dataset: 1) : papers in the same proceeding. It indicates that the topics of papers within the same conference proceeding (also published in the same year) are more likely to be similar from each other. 2) : papers published in the same conference (across different years). Since the topics papers in one conference can slightly changes year after year, but overall the paper topics within a same conference are relatively consistent. The most irrelevant path is , i.e., papers from the same institute. It’s reasonable that usually each research institute can have researchers from totally different research areas, such as researchers in the operating system area and those in bioengineer area. Moreover, we observe that the performance of PP* that involve different combination of citation links, such as co-citation relationships can achieve better performances than which only use the citation relationship. This support our intuition that meta path is very expressive and can represent indirect relationships that are very important for collective classification tasks.

6 Conclusion

In this paper, we studied the collective classification problem in heterogeneous information networks. Different from conventional collective classification approaches in homogeneous networks which only involve one type of object and links, collective classification in heterogeneous information networks consider complex structure with multiple types of objects and links. We propose a novel solution to collective classification in heterogeneous information networks, called Hcc (Heterogeneous Collective Classification), which can effectively assign labels to a group of interconnected instances involving different meta path-based dependencies. The proposed Hcc model is able to capture the subtlety of different dependencies among instances with respect to different meta paths. Empirical studies on real-world heterogeneous information networks demonstrate that the proposed meta path-based collective classification approach can effectively boost classification performances in heterogeneous information networks.

References

  • [1] C.-C. Chang and C.-J. Lin.

    LIBSVM: a library for support vector machines

    , 2001.
  • [2] B. Chen, Y. Ding, and D. Wild. Assessing drug target association using semantic linked data. PLoS Comput Biology, 8(7), 2012.
  • [3] Y. Ding, E. Yan, A. Frazho, and J. Caverlee. PageRank for ranking authors in co-citation networks. Journal of the American Society for Information Science and Technology, 60(11):2229–2243, 2009.
  • [4] H. Eldardiry and J. Neville. Across-model collective ensemble classification. In AAAI, San Franciscol, CA, 2011.
  • [5] J. Gao, F. Liang, W. Fan, C. Wang, Y. Sun, and J. Han.

    Community outliers and their efficient detection in information networks.

    In KDD, pages 913–822, Washington, DC, 2010.
  • [6] M. Ji, J. Han, and M. Danilevsky. Ranking-based classification of heterogeneous information networks. In KDD, pages 1298–1306, San Diego, CA, 2011.
  • [7] M. Ji, Y. Sun, M. Danilevsky, J. Han, and J. Gao. Graph regularized transductive classification on heterogeneous information networks. In ECMLPKDD, pages 570–586, Barcelona, Spain, 2010.
  • [8] N. Lao and W. Cohen. Fast query execution for retrieval models based on path-constrained random walks. In KDD, pages 881–888, Washington, DC, 2010.
  • [9] N. Lao and W. Cohen. Relational retrieval using a combination of path-constrained random walks. Machine Learning, 81(2):53–67, 2010.
  • [10] Q. Lu and L. Getoor. Link-based classification. In ICML, pages 496–503, Washington, DC, 2003.
  • [11] L. K. McDowell, K. M. Gupta, and D. W. Aha. Cautious inference in collective classification. In

    Proceedings of the 22nd AAAI Conference on Artificial Intelligence

    , pages 596–601, Vancouver, Canada, 2007.
  • [12] L. K. McDowell, K. M. Gupta, and D. W. Aha. Cautious collective classification. Journal of Machine Learning Research, 10:2777–2836, 2009.
  • [13] J. Neville and D. Jensen. Iterative classification in relational data. In AAAI’10 Workshop on Learning Statistical Models from Relational Data, Austin, TX, 2000.
  • [14] J. Neville and D. Jensen. Collective classification with relational dependency networks. In KDD’03 Workshop on Multi-Relational Data Mining, pages 77–91, Washington, DC, 2003.
  • [15] P. Sen, G. Namata, M. Bilgic, L. Getoor, B. Gallagher, and T. Eliassi-Rad. Collective classification in network data. AI Magazine, 29(3):93–106, 2008.
  • [16] Y. Sun, J. Han, X. Yan, P. Yu, and T. Wu. PathSim: Meta path-based top-k similarity search in heterogeneous information networks. In VLDB, Seattle, WA, 2011.
  • [17] Y. Sun, J. Han, P. Zhao, Z. Yin, H. Cheng, and T. Wu. RankClus: integrating clustering with ranking for heterogeneous information network analysis. In EDBT, pages 565–576, Saint-Petersburg, Russia, 2009.
  • [18] Y. Sun, Y. S. Yu, and J. Han. Ranking-based clustering of heterogeneous information networks with star network schema. In KDD, pages 797–806, Paris, France, 2009.
  • [19] B. Taskar, P. Abbeel, and D. Koller. Discriminative probabilistic models for relational data. In UAI, pages 482–492, Edmonton, Alberta, 2002.