TESTING AND DEVELOPING E-COMMERCE RECOMMENDATION SYSTEMS USING RNN BP AND BPTT ALGORITHM BASED ON USER SEARCH AND PURCHASE ORDER

• Main Author: Limassa, Alvin
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/49890
• Institution: Institut Teknologi Bandung

E-commerce is a digital platform that provides hundreds of millions of products to buy online. The large number of products makes it difficult for users to decide which product they want to buy. These problems can be solved using a recommendation system. The recommendation system can display relevant products to users so that it is easier for users to find and buy certain products. The recommendation system that is commonly used today is matrix factorization which does not pay attention to purchase order as a factor that determines recommendations. In this final project, deep learning-based RNN-BP and RNN-BPTT algorithms have been tested to make recommendations based on user search and purchase data sequences. The first test is done by comparing the recall value@20 and mrr@20 generated by each algorithm. The algorithm with the best accuracy value is optimized by changing the batch-size, drop-out, learning-rate values. The second test is carried out to see the effect of clustering on the accuracy value of the recommendations generated by the system. Testing of the RNN-BP and RNN-BPTT algorithms shows that the RNN-BP algorithm produces an accuracy value that is more optimal than the RNN-BPTT for the three datasets. The most optimal value is generated by the Yoochoose Gmb dataset with recall@20, which is 58.27% and mrr@20, which is 23.77%. The optimal parameter value of the RNN-BP algorithm was obtained using a mini-batch value of 100, a dropout value of 0.75, a learning rate of 0.001. The optimal parameter value of the RNN-BPTT algorithm was obtained using a mini-batch value of 50, a dropout value of 0.25, and a learning rate of 0.001. The clustering treatment of the dataset shows an insignificant decrease in the accuracy value in the experiment with the RNN-BP algorithm. However, the clustering treatment in the RNN-BPTT experiment showed an increase in the accuracy value. Keywords: E-Commerce, rnn, clusteri