<p>The growth of IoT has enabled the world to have easy access to various devices and systems. One of the most important systems is the Indoor Positioning System (IPS) which has various uses such as navigation, tracking, and entertainment. It is applicable to almost all aspects of daily life. With the introduction of Bluetooth Low Energy (BLE), IPS using BLE proliferated rapidly because of its low cost while still maintaining good accuracy. Thus, nowadays research efforts focus on improving the BLE-based IPS performance. The latest advances in deep learning concepts include the Graph Neural Network (GNN) models, which fit the IPS problem as the indoor layout, beacon, and receiver can be mapped as graphs. We proposed a novel GNN edge regression for IPS which applies hybrid fingerprinting and multilateration methods to achieve the final coordinate output. The model takes the simplest input, that is the Received Signal Strength Indicator (RSSI) readings of the BLE devices. It predicts the distance between measured point and reference point and enables the multilateration to calculate the position. The proposed model achieves state-of-the-art performance with average Euclidean positioning error of 94.70 cm, representing a 39% mean absolute error (MAE) improvement over Convolutional Neural Network (CNN) method, a 47% MAE improvement compared to MultiLayer Perceptron (MLP) baseline method, and a 50% MAE improvement compared to the traditional fingerprinting weighted sum method. It is achievable due to the accuracy of the distance on the edge prediction which is the most crucial input to the multilateration algorithm.</p>