Exploring Multifaced User Modelling in Textual Data Streams
Doctoral thesis
Permanent lenke
https://hdl.handle.net/11250/2830279Utgivelsesdato
2021Metadata
Vis full innførselSamlinger
Sammendrag
User modelling technologies play an important role in the success of many online applications such as recommender systems. However, it is far from enough to solve the cold-start issue and data sparsity problem commonly existing in the real-world dataset purely relying on user-item interactions. To this end, the objective of this doctoral thesis is to develop effective user modelling approaches to build high-quality user profiles for better mining users’ intrinsic and potential interests while alleviating cold-start and data sparsity issues raised from traditional collaborative filtering methods. Specifically, we focus on analyzing and exploiting user/item related attributes and auxiliary information knowledge from online data streams to obtain users’ needs or preferences.
To leverage attributes of users/items, such as time, location, news title and article content, we first proposed a neural time series forecasting model (NTSF) to draw users’ interest patterns over time on Twitter which takes emerging topics, users’ intrinsic interests, users’ recent behaviors and cyclic patterns of users into consideration. To jointly capture sequential patterns in streams of clicks and various item semantic features, we further devise a Deep Joint Neural Network (DeepJoNN) which consists of two parts of deep neural networks (CNN and RNN) coupled together in a hierarchical way. Considering the uncertainty of user behaviors in textual data streams, we propose a dynamic attention-integrated neural network to integrate spatial-temporal, semantic, inter- and intra-session features in a unified framework for modelling complex dynamic user interests.
We also study the auxiliary information, especially knowledge bases or knowledge graph (KG), in the user of improving user profiles for effective recommendations. Specifically, we firstly investigate the recent research progress about recommending on graphs. To explore the influence of semantic features inferenced from KGs on user modelling and multiple relations in KGs in revealing user intents, we then propose a novel Relational Knowledge-aware Heterogeneous Graph Attention Network, ReKaH_GAT, which fuses item sequential information within sessions and path connectivity with relations in KGs to understand user intents and improve the interpretability of recommender systems.
Through extensive evaluation, we show that our proposed user-modelling approaches perform better than traditional methods in user behavior prediction and recommendation tasks.