- : Ms Word, Ms Word Format
- : 70 Pages
- : ₦3,000 | $25 | ₵60 | Ksh 2720
- : 1-5 Chapters
- Click to DOWNLOAD Materials
LABELING ACTORS AND UNCOVERING CAUSAL ACCOUNTS OF THEIR STATES IN SOCIAL NETWORKS AND SOCIAL MEDIA
Abstract
The emergence of social networks and social media has resulted in exponential increase in the amount of data that link diverse types of richly structured digital objects e.g., individuals, articles, images, videos, music, etc. Such data are naturally represented as heterogeneous networks with multiple types of objects e.g., actors, video, postings, images, etc. and multiple types of links e.g., links that connect actors to items e.g., photos, videos, articles that denote relationships between individuals and items, and links between actors that denote social ties e.g., friendship, etc. Such data present a number of research challenges in machine learning, network analysis, and causal discovery. The data are incomplete, heterogeneous, highly sparse, and exhibit complex relationships and temporal structure. It is often useful to associate with each actor or individual, one or more labels that represent the memberships of the actor in specific groups e.g., political, social, or religious groups. Labels of actors in a social network can be used in a variety of ways: recommending specific items (e.g., music, movies), activities. Furthermore, labels or states of actors change over time due to the temporal nature of the real-world social networks and social media. Uncovering the causal accounts of the state dynamics of actors has an important impact in explaining why people change their preference, emotion, or interest in a specific activity and in predicting actors’ future behaviors. Against this background, this dissertation aims to develop (i) novel machine learning approaches for labeling actors in social networks, with particular emphasis on methods for dealing with a diversity of actors, objects, and relationships and with data sparsity;
(ii) uncovering the causal accounts of the social network and social media dynamics (with emphasis on causal accounts of sentiment change and related applications). The introduced algorithms are implemented and experimentally evaluated on a number of real-world data sets. The proposed methodologies for actor labeling and analysis of causal accounts have broad applicability in a variety of areas such as bioinformatics and neuroscience.
Chapter 1 | Labeling Actors and Uncovering Causal Accounts of Their States in Social Networks and Social Media: An Overview
1.1 Introduction
Social networks, e.g., Facebook[1] , and social media, e.g., Youtube[2] , have provided large amounts of network data that connect actors (individuals) with other actors, as well as diverse types of media items e.g., videos, images, postings, articles, etc. There exist several types of objects and relationships among objects in such data which are naturally represented as heterogeneous networks. It is often useful to associate with each actor, one or more labels that represent the memberships of the actor in specific groups e.g., political, social, or religious groups [Bui and Honavar (2014); Lo et al. (2016); Pfeiffer III et al. (2015)] . Labels of actors in a social network can be used in a variety of ways: recommending specific items (e.g., music, movies), activities, etc [Cheng et al. (2015); Lo et al. (2016)] . However, in many real-world social networks, the labels of many actors are unknown for a variety of reasons e.g., incompleteness of actor profiles.
Machine learning algorithms [Grover and Leskovec (2016); Kong et al. (2011); Lu and Getoor (2003); Macskassy and Provost (2007); Sen et al. (2008)] on classifying actors in a homogeneous network, i.e., single type of objects and links network, have been explored in several studies. However, such approaches are not appropriate in the cases of social networks and social media since such data are heterogeneous in both object types, e.g., actor, image, video, and link types, e.g., links that connect actor to actor and actor to video. Moreover, in social networks, links that connect actors denote different types of social ties [Berlingerio et al. (2013); Dickison et al. (2016)] , e.g., links that connect actors who are relatives or links that connect actors who are classmates. Social networks [Ellison et al. (2007)] and social media [Kietzmann et al. (2011); Kwak et al. (2010)] can be in the forms which includes:
(i) networks of actors where links that connect actors denote different social ties
[Dickison et al. (2016); Gollini and Murphy (2016); Tang and Liu (2009b)] , e.g., Google+[3] , Facebook; (ii) networks containing actors and media items, e.g., photos, videos, postings and links between actors and links that connects actors to only a small subset of items but the number of distinct items is quite large, e.g., in the millions [Bui and Honavar (2013)] , e.g., Youtube, Flickr[4] ; (iii) networks containing of multiple types of objects and links [Angelova et al. (2011); Bangcharoensap et al. (2016); Bui and Honavar (2014); Ji et al. (2011); Wan et al. (2015b)] , e.g., LinkedIn[5] , DBLP[6] ; and (iv) those that have the combined characteristics from the three above forms. Social media and social networks in form (ii) in which we do not consider links among media items (e.g., there is no relation among image and video) are considered as special cases of those in form (iii).
Existing approaches on labeling actors in heterogeneous networks include: approaches that explore the semantic meanings of social relations and social contexts [Bui and Honavar (2014); Tang and Liu (2009a,b); Wang and Sukthankar (2013); Zafarani et al. (2014)] ; approaches that aggregate the ranking of objects to label actors [Chen et al. (2013); Ji et al. (2011)] ; approaches that base on the meta-data path of different types of objects in heterogeneous networks [Hu et al. (2016); Kong et al. (2013, 2012); Liang et al. (2016); Sun et al. (2011, 2012)] ; and others that use transductive learning methods [Bangcharoensap et al. (2016); Ji et al. (2010); Wan et al. (2015b)] and random walk-based methods [Angelova et al. (2011); Bui and Honavar (2014); Wang et al. (2012)] . However, there is a substantial room for introducing the novel algorithms for efficiently labeling actors in variety forms of heterogeneous networks.
Learning to label actors regardless of the relationship meanings of social ties between actors in a social network will output an inaccurate predictive model. The relationship types in social network can be either explicit or implicit [Berlingerio et al. (2013); Dickison et al. (2016)] . Eldardiry and Neville [Eldardiry and Neville (2011)] proposed an ensemble of relational learners wherein each relational learner focuses on a single social dimension or view of the network. EdgeCluster [Tang and Liu (2009b)] and SCRN [Wang and Sukthankar (2013)] extract the implicit views in the network and uses the actors’ social dimensions, i.e., social relations (view affiliation) as features to label actors. Several authors have proposed latent space models to handle the multiple views in multi-view networks. Examples of such models [Salter-Townshend and McCormick (2013)] include hypergraph regularized generative models [Wang et al. (2016b)] , partially shared latent factor models [Liu et al. (2015)] . Latent space joint model [Gollini and Murphy (2016)] models multi-view network to learn the latent representation which can be used to learn to label nodes. However, the approaches that make use of relationship types of links among actors to the purpose of labeling actors in social networks and social media are limited.
Social networks containing actors and media items and links between actors and links that connect actors to only a small subset of items (e.g., in Flickr, YouTube) often lead to data sparsity which causes over-fitting and hence poor performance in predicting labels of actors. To our knowledge, there are few studies on handling the sparsity of items in real-world multiple types of objects and links networks for the labeling purpose such as the approach in [Bui and Honavar (2013)] . Multiple types of objects and links social networks, e.g., LinkedIn, DBLP, provide a rich source of relations among objects of several types but it is challenging to capture that informative characteristic of the data to build a predictive model in a robust way. Hence, there is a necessity to introduce novel algorithms [Angelova et al. (2011); Bangcharoensap et al. (2016); Bui and Honavar (2014); Ji et al. (2011)] to efficiently take into account the rich and complex structure of such data to learn an accurate model to label actors.
Real-world social media and social networks evolve over time, e.g., actors
in an online health forum can change their emotion (i.e., sentiment state) by communicating with others via postings [Bui et al. (2015, 2016); Portier et al. (2013); Qiu et al. (2011); West et al. (2014)] ; actors in a social network can change their friend list by making new friend or disconnecting with old friends. Hence, actors’ states (i.e., labels) change due to the dynamic interaction among them. It is compelling to dynamically label actors in such evolving data [Anderson et al. (2015); Bertsimas et al. (2003)] and discover the causal accounts of label dynamics of actors. Uncovering the causal accounts of the label dynamics of actors has an important impact in explaining why people change their preference, emotion, or interest in a specific activity and in predicting actors’ future behaviors [Bui et al. (2015, 2016)] . Existing machine learning approaches are limited in their applicability in uncovering causal accounts of label or state dynamics of actors.
Against this background, this dissertation first aims to propose a set of novel algorithms to learn a predictive model to efficiently label actors in social networks and social media, with particular emphasis on approaches for dealing with a diversity of actors, objects, and relationships and with data sparsity. Secondly, the dissertation introduces a novel framework to uncover the causal accounts of the sentiment label dynamics of actors in social media.
1.2 Dissertation Contributions
Chapter 1: Labeling Actors and Uncovering Causal Accounts of their States in Social Networks and Social Media: an Overview
We present a general introduction and motivation for the problem of labeling actors and uncovering the causal accounts of their state dynamics in social networks and social media, as well as an overview and a summary of contributions of this dissertation.
Chapter 2: On the Utility of Abstraction in Labeling Actors in Social Networks
This chapter presents our first set of results, showing an approach to explore the utility of abstraction in learning classifiers to label actor in a social network. Social networks are naturally represented as heterogeneous networks with multiple types of objects and multiple types of links. The number of distinct items represented in real-world social networks is quite large (in the millions) although only a small subset of them are linked to a specific actor, say, John Smith. This leads to data sparsity which in turn leads to over-fitting by overly complex models [Lu and Getoor (2003)] and hence poor predictive performance in labeling the unlabeled actors in a network. The situation is further complicated by the fact thatitems linked to actors in real-world networks are often described at varying levels of granularity or detail. This calls for an effective means of choosing an abstract yet sufficiently informative representation of network data to achieve accurate and reliable labeling. In this chapter, we propose an approach that induces hierarchical taxonomies over items and uses the resulting taxonomies as a basis for selecting abstract and hence parsimonious representations of network data. This offers a means of striking a compromise between using the finest granularity identity-level representation of item objects of a social network on the one hand and the coarsest granularity type-level representation (in which objects are simply encoded by their
types).
This chapter has been published in [Bui and Honavar (2013)] .
Chapter 3: Labeling Actors in Social Networks using a Heterogeneous Graph Kernel
In this chapter, we also consider the problem of labeling actors in social networks where the labels correspond to membership in specific interest groups, or other attributes of the actors. Actors in a social network are linked to not only other actors but also items (e.g., video and photo) which in turn can be linked to other items or actors. Given a social network in which only some of the actors are labeled, our goal is to predict the labels of the remaining actors. We introduce a variant of the random walk graph kernel to deal with the heterogeneous nature of the network (i.e., presence of a large number of node and link types). We show that the resulting heterogeneous graph kernel (HGK) can be used to build accurate classifiers for labeling actors in social networks. Specifically, we describe results of experiments on two real-world data sets that show HGK classifiers often significantly outperform or are competitive with the state-of-the-art methods for labeling actors in social networks.
This chapter has been published in [Bui and Honavar (2014)] .
Chapter 4: Labeling Actors in Multi-view Social Networks by Integrating Information From Within and Across Multiple Views
Real world social networks are characterized by agents who are linked to other agents or entities via links of multiple types. Each relationship type denotes one view of the network. In this chapter, we consider the problem of labeling actors in such multi-view networks based on the connections among them. Given a social network in which only some of the actors are labeled, as in Chapter 2 and Chapter 3 our goal is to predict the labels of the remaining actors. We introduce a new random walk kernel, namely Inter-Graph Random Walk Kernel (IRWK), to deal with multi-view social networks. IRWK not only uses random walks within the view but also exploits the random walks across other views. We use the resulting IRWKs to train classifiers for labeling actors in a multi-view social network. The results of our experiments on two real-world multi-view social networks show that: (i) IRWK classifiers outperform or are competitive with several state-of-the-art methods for labeling actors in a social network; (ii) IRWKs are robust with respect to different choices of user-specified parameters; and (iii) IRWK kernel computation converges very fast within a few iterations.
This chapter was accepted in 2016 IEEE BigData. Ngot P. Bui is the first author of the accepted work.
Chapter 5: Temporal Causality of Social Support in an Online Community for Cancer Survivors
Social media and social network change over time due to the dynamic interaction of actors. Likewise, the actors’ labels or states that denote their preference, emotion, and behavior also change over time. Uncovering the causal accounts of the label dynamics of actors has an important impact in explaining why people change their preference, emotion, or interest in a specific activity and in predicting actors’ future behaviors. This chapter focuses on the work of analyzing the causal accounts on dynamic sentiment labels of actors in a social media. We use a case study online health community forum data set for this work, the American Cancer Society’s Cancer Survivor Network.
This chapter has been published in [Bui et al. (2015)] .
Chapter 6: Incorporating Uncertainty of Actor States Into Temporal Causality Analysis
In this chapter we extend our work from Chapter 5 on temporal causality on social support in online health communities along several directions. First, we analyze the temporal causality relationship as a function of the discussion topic. Second, we investigate the robustness of the framework with respect to the choice of the classification threshold of the sentiment classifier and the choice of the specific sentiment classifier used. Third, we offer a modification of the basic framework for temporal causality analysis and hence uncertainty in the states of the probabilistic Kripke structure resulting from the use of an imperfect state transducer (e.g., an imperfect sentiment classifier in an analysis of temporal causality of sentiment dynamics). Our experiments show that the conclusions on temporal causality in Chapter 5 are robust with respect to the choice of the (i) classification threshold of the sentiment classifier; (ii) and the choice of the specific sentiment classifier used.
This chapter has been published in [Bui et al. (2016)] .
Chapter 7: Conclusions and Future Directions
We conclude with a summary of the dissertation, the contributions, and broader directions for future work.
LABELING ACTORS AND UNCOVERING CAUSAL ACCOUNTS OF THEIR STATES IN SOCIAL NETWORKS AND SOCIAL MEDIA
