Deep Graph-Based Novelty Analysis in Social Media Data

<p dir="ltr">Analysing novelty in social media data has attracted significant attention due to its wide-ranging applications, including market research, understanding customer preferences, personalised recommendations, and improving customer service. For example, in live-streaming e-commerce events, novel content tends to capture audience attention, making the detection of such novelty essential. From the perspective of influencers, they can learn how to create more novel content by analysing detected novelties and continuously evaluate the novelty level of their current creations. From the platform's perspective, they can gain insights into users' preferences for specific content and understand the trend-driven behaviours within the user community. From the users' perspective, in addition to enjoying user experience improvements based on novelty analysis, they can also interact on social platforms, generating a significant amount of new data. Obviously, it is crucial to discover these novelties based on social media data, recommend them, and effectively plan their utilisation. However, existing detection, recommendation, and planning algorithms fail to consider the multi-party interactions on social platforms and the unique characteristics of social media data. Therefore, in this thesis, three approaches are proposed to handle novelty analysis in social media data: detecting novelties over online video streams, making high-quality influence-aware group recommendations over online streams, and making session planning over heterogeneous social platforms. </p><p dir="ltr">We propose a general framework, the Long-term Information REconstruction-based Model (LIREM) for online novelty detection. Specifically, we design a novel outlier detection method to filter noisy features, thereby enhancing the model’s learning capability. Based on the cleaned features, we construct an LSTM-Decoder model to predict the reconstruction error, which serves as the novelty score for each video segment. To accommodate continuously evolving data streams, we introduce an incremental model updating strategy that dynamically adapts the model without requiring full retraining. Additionally, we develop a bounding-based technique to improve detection efficiency by reducing unnecessary computations. An adaptive optimisation strategy is further proposed to dynamically select optimal bounds for filtering, ensuring efficient and effective identification of novelty candidates. </p><p dir="ltr">Next, we propose a comprehensive framework for Influence-aware Group Recommendation (IGR) tailored for high-speed social streams. GroupGCN is proposed to mitigate data sparsity and effectively capture media dynamics, enabling superior data representation. Building on this, TGGCN-RA enhances the accuracy of group interest predictions for future time points, making it well-suited for dynamic environments. In addition, we propose the DYIC model, which captures key aspects of group behaviour, including group activeness, similarity, and their willingness to propagate items. This is achieved through a novel dynamic item-aware graph, DI2PROG, which simulates information propagation more effectively. Finally, we present the GES algorithm for edge sampling in GREG. GES preserves the distribution of the sampled dataset and tracks interest drifts within groups, enabling efficient and effective training of the TGGCN-RA model. </p><p dir="ltr">Furthermore, we propose the Motivation-Aware Session Planning (MASP) framework for session planning across heterogeneous social platforms. MASP introduces HeterBERT, which addresses attribute-level heterogeneity by managing attribute uncertainty and capturing correlations in heterogeneous items. To improve accuracy, we develop a motivation-aware user preference prediction method that leverages the motivations behind user activities. Additionally, we design a multi-constraint session generation algorithm with optimisation strategies to ensure efficient and effective planning under various constraints. Experimental results confirm MASP's strong performance in complex social environments.</p>