Recommender systems are one of the most significant field of studies in recent years. As more and more people opt for self-guided trips, the demand for an accurate recommender has risen. Although there have been countless past studies in this field, some of them have drawbacks. For example, machine learning methods such as K nearest neighbors suffer from the repercussions of data sparcity and lack of data. To address problems discussed, scholars have come up with a GAN(generative adversarial networks)-based method. However, the generated image is far from acceptable. Therefore, I've brough up adjustments to the model architecture and the selection of loss functions to boost the performance and ultimately generate clear visulized recommendations.
This diagram illustrates an enhanced Generative Adversarial Network (GAN) designed for trajectory prediction with improved error modeling. The input data consists of matrices initialized with -1, where each positive integer within the matrix represents a sequence value along a specific travel path. For example, if a user visits the coordinate (20, 21), the corresponding cell in the matrix is filled with the sequence value corresponding to that location.
The condition data is a masked version of the input matrix, where all positive integers are replaced by 1. This masking ensures that the model uses these hints effectively to reconstruct the original travel sequence.
The model architecture consists of a Generator and a Discriminator. The generator predicts travel paths based on the condition data, while the discriminator assesses the authenticity of these predicted paths. The reconstruction loss is computed by comparing the predicted sequences with the actual travel paths, providing feedback on the generator’s accuracy. Simultaneously, the discriminator generates an authenticity score, representing how realistic the predicted paths appear.
The two losses—reconstruction error and authenticity loss—are weighted and summed to compute the final loss used for backpropagation. This weighted loss ensures that the generator improves in both reconstructing accurate sequences and creating realistic outputs, enhancing the overall predictive capability of the model. This improved GAN framework leverages both reconstruction and authenticity feedback, balancing them to create accurate and realistic predictions for human mobility patterns.