Learning-based recommendation system

Abstract

Site selection and land planning were critical activities for property development in various countries and regions. Decision-makers needed to consider several factors, such as demographics, transportation, and accessibility, when identifying and evaluating potential locations for commercial, retail, and residential purposes. Although efforts had been made to analyse the requirements of some specific Point-of-Interest (POI) categories, such as stores, a systematic approach to support POI requirements modelling was still lacking. The traditional multi-criteria decision-making framework for store site selection oversimplified the local characteristics that were complex, high-dimensional, and unstructured. This approach did not adequately capture the nuances and specificities of different subzones, making it difficult to achieve optimal site selection and land planning outcomes. Recent advancements in machine learning had provided new opportunities for addressing this challenge. However, there was currently no single machine learning model that could perform both site selection and land planning tasks simultaneously. Instead, these tasks typically required different parameters and modelling techniques, and the results of one task may inform the other. To this end, the model transformed the site selection and land planning problem into a recommendation problem. Using Google's POI dataset, the model was trained to predict ratings for various POIs in Singapore based on its subzone features and the site features. The site embedding vectors and subzone embedding vectors were generated using a modified Deep Semantic Similarity Model. These vectors were then used to compute similarities between subzones and sites, which were used to make recommendations. Therefore, the project recommended subzones to build for a specific site or recommended sites for a specific subzone. The model recommendation results were presented to the end user on a web page. The experimental results demonstrated that the low Mean Square Error score strongly indicated good performance, thereby reinforcing its validity and usefulness for site selection and land planning tasks.