T2L2: A Tiny Three Linear Layers Model for Service Mashup Creation

Mashup creation is a classic problem in service computing and can be solved using service recommendation approaches. There are many service recommendation studies and have achieved remarkable results. However, there is a growing tendency for these studies to use multiplex data and more complicated models to improve the performance of recommendations, especially after the emergence of deep learning. This trend has led to a heavy reliance on computational resources and an increased cost of data acquisition, which limits the practical use of these methods, but the performance gains are still very limited. In this paper, we improve recommendation performance by rethinking the characteristics of the data in the mashup creation scenario, i.e. representation heterogeneity between services and mashup, rather than the use of multiplex data and more complicated models. To achieve this, we propose a Tiny Three Linear Layers (T2L2) model. T2L2 is a tiny model with three linear layers requiring only requires functional descriptions of services and mashups as input. The first two linear layers are used to align the representation space of services and mashups. The last linear layer is used to calculate the matching scores of services and mashups. Extensive experiments conducted on a real-world dataset from ProgrammableWeb show that T2L2 outperforms existing state-of-the-art methods in commonly-used evaluation metrics with a significant reduction in model complexity and required data.