DEVELOPMENT OF E-COMMERCE RECOMMENDER SYSTEM BASED ON LINK PREDICTION WITH MULTILAYER PERCEPTRON

• Main Author: Sekar Ayuningtyas, Annisa
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/51040
• Institution: Institut Teknologi Bandung

Recommender system is a tool to overcome information overload problem in e-commerce. One of the best practice tips in e-commerce recommender system is to provide recommendation based on the interaction of other users. This can be solved with the link prediction problem, which is a problem to detect the existence or weight of an invisible link between two nodes. The recommendation system based on link prediction with multilayer perceptron can solve recommendation system problems, namely cold start and sparsity data, besides that it also has better evaluation results in the research of Hou and Holder (2017) and in the research of He et al (2017). In this final project, an e-commerce recommendation system based on link prediction with multilayer perceptron was built. The recommendation system accepts input in the form of user identity and produces top-N item recommendations. The main part of the recommendation system is the multilayer perceptron model which generates the interaction weight between the customer and the goods. Then the weights are sorted from the largest and the top N products are taken to become the top-N product recommendations. The experiment in this final project was conducted to determine the parameters that most influence the evaluation results of the recommendation system with the evaluation metrics of RMSE, NDCG, and training duration. The parameters used are the number of hidden layers, learning rate, number of epochs, embedding size, and hidden layer size. The most influencing parameters are the embedding size and hidden layer size, which have a significant evaluation value of 0.02. Meanwhile, other parameters are only around 0.0001-0.0005.