SEQUENTIAL PATTERN-BASED ENHANCED GRAPH EMBEDDING WITH SIDE INFORMATION FOR RECOMMENDATION SYSTEM

• Main Author: Ayu Chandra Kemala, Shinta
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/49943
• Institution: Institut Teknologi Bandung

This final project focuses on utilizing Sequential Pattern Mining (SPM) as a preprocess method for Enhanced Graph Embedding with Side Information or EGES input (Wang et al., 2018). The purpose of this project is to build a graph embedding that has a high transaction certainty by replacing the input that was originally in the form of an item page access history data into an item purchase history data of e-commerce. EGES accepts input in the form of page access history data and side information for each item. Although its performance is better than other graph embedding methods, EGES requires large data to perform well. The use of purchase history data which is less in number will affect the resulting embedding. By using SPM, transaction data is converted into a sequential pattern that has a relatively high occurrence frequency, so the input data is more certain. Embedding results from EGES are processed using logistic regression, then their performance is compared to the results of embedding from EGES without SPM to determine the effect of using sequential patterns. Based on the test results, the most suitable methods for EGES input preprocesses are SPADE (Zaki, 2001) and SPAM (Ayres et al., 2002) with AUC values reaching 0.5757, 5-6% higher than EGES without SPM, and AUPR values reaching 0.7867, 40% higher than EGES without SPM. Based on analysis and experiments, SPM is an appropriate method for preprocessing EGES input, and sequential pattern forms are suitable for use as EGES input. In addition to the SPM method, things that affect the quality of the prediction of the model from EGES are the amount of data, the balance of the number of classes, and the selection of EGES hyperparameters.