High-order and Motif-based Graph Embedding in Service Recommendation
Nguyen, Thi Thuy Mo
MetadataShow full metadata
With the rapid increase of the number of diverse services and APIs on the Internet and the Web, recommender systems have become an indispensable tool for service discovery and selection. Existing probabilistic matrix factorization (PMF) recommender systems can effectively exploit the latent features of the invocations with the same weight but not all features are equally significant and predictive, and the useless features may bring noises to the model. Currently, Deep Neural Networks (DNN) based collaborative filtering is a popular method for service recommendation. However, many such approaches treat each mashup-API invocation as separate instances and overlook the intrinsic relationships among data. We have finished three major pieces of work as follows. First, we propose the Attentional PMF Model (AMF), which leverages a neural attentional network to learn the significance of latent features. We then inject the attentional scores and the mashup-API context similarity, and API co-invocation into the matrix factorization structure for training. Particularly, the we first apply an attentional mechanism with the neural network to the PMF-based framework. Later on, we then use the Doc2Vec technique to explore the latent features from document context of the mashups and APIs’ description and do some statistical distribution fitting to draw out the APIs co-invocation pattern. Such auxiliary information significantly regularize the learning model for better prediction. Second, inspired by the discovery that the autoencoder architecture can force the hidden representation to capture information on the structure of networked data, we propose a novel framework called Data Augmented High-order Graph Autoencoder (DHGA) that learns the latent high-order connectivity signals in the mashup-API invocation graph for service recommendation. Specifically, each mashup-API pair is augmented with their higher-order neighbourhood data, and such data is input into two sets of autoencoders, one set for the mashups and the other for the APIs. All the autoencoders in one set shared parameters, so increasing the number of autoencoders does not increase the model size. Finally, we propose a Motif-based Graph Convolution Layers with self-attention using attention motif-based neighbour mechanism to capture the high-order structure, and a Motif-based Graph Attention Collaborative Filtering model with MLP to generate the recommendation prediction. Specifically, we analysis some most common used sub-graphs or motifs in the mashup-API bipartite and apply the motif-based attentional mechanism for each motif type and put them in a convolution neural layers. Then, a dot product is used to calculate the invocation prediction value between a pair of mashup and API. In summary, we have explored three recommendation models for service composition including AMF, DHGA, and MGAT. All of them successfully explore the latent features of mashups and APIs in different sides of the data to support for the Collaborative Filtering Neural Network recommendation model. While the AMF focus on the context and co-invocation auxiliary information in a shallow framework, DHGA and MGAT pay attention on the intrinsic relation of the data structure through using high-order and motif-based connectivity with deep learning neural network layers. All of the proposed models obtain superior performance on service recommendation accuracy compared with existing defined baselines in the thesis.