The result revealed Generative SSL models outperformed other models in datasets with relative serious data sparsity issue, suggesting that Generative models have the ability to capture the meaningful representations hidden in the graph-structured data with their complex model architectures.
The contrastive SSL models in the study performed relatively poorly, especially for LightGCL, whose best model is trained with epoch (training iteration) 0, literally untrained).
Nevertheless, SGL outperformed other models for yelp dataset, a dataset with characteristics of dense interaction. This finding suggested that the contrastive model can be also effective given sufficient level of interaction density for training datasets.
The scope of the study focused on collaborative filtering based recommendation algorithm. However, it still provided insights and foundations for further researches in other fields of recommendation system such as social recommendation, sequential recommendation and knowledge-enhanced recommendation.
