Deep Contextual of Document Using Deep LSTM Meet Matrix Factorization to Handle Sparse Data: Proposed Model

Abstract

Recommender system is important tool in big data era. It has responsible to make suggestion about product or service automatically for web application or mobile. In everyday utility, we cannot escape for information about food, travelling, social network, ticketing, news and etc. What the best choice for customer necessary is recommender system task to provide relevant information. Collaborative filtering is most useful recommender system technique in which considering user behaviour in the past to calculate recommendation. The first generation of collaborative filtering exploit statistical approach to calculate product recommendation. However, traditional collaborative filtering facing serious problem in scalability, accuracy and shortcoming in large data. Model based in the second generation of collaborative filtering to produce product recommendation where this model rely on matrix factorization to produce recommendation. Model based proven better performance over memory based. However, model-based performance degrades significantly when met with sparse data due the number of rating are very small. This problem popular called sparse data problem. Several methods proposed by researchers to handle sparse data problem. Mostly of them exploit text document to increase recommendation performance. However, majority of model fail to gain text document understanding. This study proceeds ongoing process with several stage. First, develop model to interpreted text document using LSTM aims to capture contextual understanding of document. Second, integrated LSTM with matrix factorization. This step aims to produce rating prediction considering text document of the product. The first step completely finished. According to experiment report, this model success to capture contextual of the document then transform into 2D space text document representation. For the further research, we are going to integrated with matrix factorization and evaluation result of rating prediction using RMSE metric evaluation.