Amina Amara PhD thesis defense

Title: A deep learning-based approach for anchor user modeling across online social networks under a big data environment.
Supervisors: Dr. Mohamed BEN AOUICHA & Dr. Mohamed Ali HADJ TAEIB
Defense Date: 26 may 2022
An accurate and comprehensive user modeling has been proven fundamental for many personalized social media-based services including link prediction and recommendation systems. A major challenge lies in that data available in a single online social network are usually very limited and sparse. Depending on the purpose for which an online social platform is designed, each platform offers only a partial view of a user from a particular perspective. Therefore, users are usually involved in several online social networks simultaneously to enjoy specific services provided by different networks, thus bringing about the interconnection of online social networks via bridge users called anchor users. Accordingly, a comprehensive anchor user modeling requires taking into account the user’s social data scattered through several social networking platforms. However, online social networks are completely isolated. In this respect, the present research work aims to build a holistic and comprehensive anchor user model from his data distributed over multiple social information sources and thus serve for enhancing the quality of personalized social media-based services. For this reason, we first study different aspects related to the target social networking platforms selected in this dissertation including covered functionalities, data heterogeneity, and data accessibility. The study reveals also that the major challenge towards anchor user modeling is linking user’s identities across online social networks. Moreover, recent trend to solve the user modeling problem is to leverage network embedding techniques and build embedding vectors corresponding to user nodes in the social network. Thereby, this thesis firstly presents a detailed survey of more recent network embedding based approaches. Then, the problem of anchor user identification across social platforms is addressed by providing a feature-based approach exploiting various similarity measures. Indeed, the proposed algorithm leverages the profile information (i.e. the user’s name) with the network properties (i.e. friendship network) to compute the matching degree between the two user’s identities pertaining to two different social networks and then the obtained result is compared to a predefined threshold value to decide whether the two profiles pertain to the same real person or not. Several similarity measures are tested including Jaccard Index, Hub Promoted Index, Hub Depressed Index, and Sørenson Index. Experimental results conducted on two real-world datasets, collected from Facebook and Twitter social networks, showed that the Jaccard Index outperforms all other measures with a threshold value δ = 0.03. In addition, as social networks stand as typical sources of big data, a BIGUI (BIG User Identification) architecture for anchor users storing and processing is designed. The provided architecture is built on the basis of a set of adequate technologies in conformity with the social big graph data pipeline to cope with online social network-associated big data challenges. Thereafter, the next step consists on building an anchor user model encoding diverse network structures of the given user, i.e. intra-network and cross-network features. In this regard, the OVRAU model (Overall low-dimensional Vector Representation for Anchor Users) is implemented to learn an overall low-dimensional representation for modeling anchor users, from a multiplex heterogeneous social network, by investigating the intra-network as well as cross-network structural information. Unlike existing works, the proposed model considers the multi-network scenarios to capture diverse network structures of anchor users. For this reason, OVRAU model includes two types of embeddings for each anchor user: one high-dimensional base embedding and a low-dimensional social edge embedding for each social network. In particular, the model learns a function that generates social edge embeddings by sampling and aggregating structural features from an anchor user's neighborhood inside different social networks, through one of three candidate functions of aggregation, namely mean aggregator, max-pooling aggregator, and LSTM aggregator, with self-attention mechanism. Link prediction is used as a downstream task to evaluate the effectiveness of the learned embeddings. Experiments have been conducted on real-world social networks dataset and the experimental results demonstrate that the proposed model with all three variants can significantly outperform the existing approaches of network representation learning when it is applied on link prediction task and also achieves better performance over all compared baselines.