A Hybrid Model Based on LFM and BiGRU Toward Research Paper Recommendation

Abstract

To improve the accuracy of user implicit rating prediction, we combine the traditional latent factor model (LFM) and bidirectional gated recurrent unit neural network (BiGRU) model to propose a hybrid model that deeply mines the latent semantics in the unstructured content of the text and generates a more accurate rating matrix. First, we utilize the user's historical behavior (favorites records) to build a user rating matrix and decompose the matrix to obtain the latent factor vectors of users and literature. We also apply the BERT model for word embedding of the research papers to obtain the sequence of word vectors. Then, we apply the BiGRU with the user attention mechanism to mine the research paper textual content and to generate the new literature latent feature vectors that are used to replace the original literature latent factor vectors decomposed from the rating matrix. Finally, a new rating matrix is generated to obtain users' ratings of noninteractive research papers and to generate the recommendation list according to the user latent factor vector. We design experiments on the real datasets and verify that the research paper recommendation model is superior to traditional recommendation models in terms of precision, recall, F1-value, coverage, popularity and diversity.INDEX TERMS recommender systems, deep learning, LFM, BiGRU, user attention I. INTRODUCTION

Scientific paper recommendation systems: a literature review of recent publications

“…Zhao et al [ 123 ] present a personalised approach focusing on diversity of results which consists of three parts. First LFM extracts latent factor vectors of papers and users from the users’ interactions history with papers.…”

“…[109] Graph-based [110] Hybrid [111] Hybrid [117] Hybrid [118] N e t w o r k × [123] Hybrid also utilise historic user-interaction data or descriptions of paper features (see, e.g. Li et al [57] who describe their approach as network-based while using a graph structure, textual components and user profiles) which would render them as either CF or CBF also.…”

“…We differentiate between profiles derived from user interactions and ones derived from papers. Most user profiles are constructed from users' actual interactions: unspecified historical interaction [30,37,56,57,64,118], the mean of the representation of interacted with papers [19], time decayed interaction behaviour [62], liked papers [69,123], bookmarked papers [84,119], read papers [111,113], rated papers [3,4,110], clicked on papers [24,26,49], categories of clicked papers [1], features of clicked papers [104], tweets [74][75][76], social interactions [65] and explicitly defined topics of interest tags [119].…”

“…Among the most common methods was their application on papers: in an unspecified representation [30,119], Word2Vec [19,37,44,45,55,104,113], Word2Vec of LDA top words [24,107], Doc2vec [21,28,48,62,63,107], Doc2Vec of word pairs [109], BERT [123] and SBERT [5,19]. Most times these approaches do not mention which part of the paper to use as input but some specifically mention the following parts: titles [37], titles and abstracts [28,45], titles, abstracts and bodies [48], keywords and paper [119].…”

“…We found a multitude of machine learning methods being used, from multi armed bandits [1], LSTM [24,37,113], multi-layer perceptrons [62,96,104], (bi-)GRU [37,69,71,123], matrix factorisation [4,62,69,110,111], gradient ascent or descent [41,57,63,116], some form of simple neural network [30,37,56], some form of graph neural network [19,49,104], autoencoder [4], neural collaborative filtering [62], learning methods [30,123] to DTW [48]. Three approaches ranked the papers to recommend [56,57,118] with, e.g.…”