Single-behaviour and multi-behaviour streaming recommender systems
In this information age, Recommender Systems (RSs) have played an increasingly important role in providing users with tailored suggestions that match their preferences. However, conventional offline RSs cannot well deal with the ubiquitous data streams of user-item interactions. This is because offline RSs are periodically trained with large-volume historical interaction data, and thus cannot well capture the latest preferences of users embedded in their recent interactions. To address this issue, Streaming Recommender Systems (SRSs) have emerged in recent years, which commonly train recommendation models with newly coming interaction data to capture the latest user preferences for streaming recommendations. For further improving the accuracies of streaming recommendations, this thesis proposes the following three approaches.
Firstly, training recommendation models with newly coming data only benefits overcoming the preference drift problem, but overlooks the long-term user preferences embedded in the historical data. Moreover, the common heterogeneity of users and items makes it more challenging to deliver accurate streaming recommendations, as different types of users (or items) have different preferences (or characteristics). To address these two issues, we propose a Variational and Reservoir-enhanced Sampling based Double-Wing Mixture-of-Experts framework, called VRS-DWMoE. Specifically, in VRS-DWMoE, we first devise variational and reservoir-enhanced sampling to wisely complement newly coming data with historical data for capturing long-term user preferences while addressing the issue of preference drift. Then, we propose a double-wing mixture-of-experts model to effectively learn the heterogeneous user preferences and item characteristics with two mixture of experts, respectively, where each individual expert model specialises in one type of users or items.
Secondly, the commonly existing underload (or overload) scenarios, where the data receiving speed is lower (or higher) than the data processing speed, should be well dealt with for accurate streaming recommendations. Therefore, we propose a Stratified and Time-aware Sampling based Adaptive Ensemble Learning framework, called STSAEL. Specifically, in STS-AEL, we first devise stratified and time-aware sampling to extract training data from both newly coming data and historical data. This practice not only benefits utilising the idle resources in underload scenarios more effectively, but also helps capture long-term user preferences while addressing the preference drift issue. After that, we propose adaptive ensemble learning to first leverage multiple individual recommendation models for concurrently learning from the prepared training data, and then fuse the results of these individual models with a sequential adaptive mechanism for accurate streaming recommendations.
Thirdly, all the existing SRSs have been devised for dealing with data streams of user-item interactions w.r.t. a single behaviour type (e.g., purchases) and commonly suffer from the data sparsity issue caused by the limited number of such singlebehaviour interactions. Therefore, we propose the first Multi-behaviour Streaming Recommender System, called MbSRS, to exploit more sufficient multi-behaviour interactions (e.g., purhcases, add-to-carts and views) for further improving the accuracies of streaming recommendations. In MbSRS, confronting data streams of multibehaviour interactions, we first propose the Multi-behaviour Learning Module (MbLM) to accurately learn the short-term user preferences and stable item characteristics. Then, we propose the Attentive Memory Network (AMN) to effectively maintain the long-term user preferences. After that, these learnt short-term user preferences and long-term user preferences are merged by the elaborately devised User Preference Merging Module (UPMM). Note that MbLM, AMN and UPMM all effectively leverage the multi-behaviour interactions to further improve the accuracies of streaming recommendations.
The superiorities and the effectiveness of all the above three approaches proposed in this thesis have been validated by both the theoretical analysis and extensive experiments that are conducted on real-world datasets.