Smart objects recommendation based on pre-training with attention and the thing–thing​ relationship in social Internet of things

Abstract

In Internet of things (IoT) and Social Internet of things (SIoT), how to select or recommend suitable smart objects from an ocean of smart objects has become an increasingly critical issue. In this paper, we propose a novel neural network model called BLA (BERT and Bi-LSTM with attention) for smart objects scoring tasks to make recommendations in social Internet of things. The model uses a BERT network to obtain the sentence vectors for a smart object related text, and then uses Bi-LSTM with two types of attention mechanisms to extract representations of the smart object vectors. The devised attention mechanism contains a self-attention (SA) layer and a global-attention (GA) layer. The SA layer is able to estimate the importance of sentences or fields, which in a certain sense can substitute for manually defined features at the sentence and field level. The GA layer can measure the relationships between sentences (or fields) and user requirements, which further helps the model obtain more effective smart object vectors. The thing–thing relationship of Internet of things is introduced into the model to inprove the recommendation effect. Experimental results on the datasets demonstrate that our model outperforms other baseline methods.