Using Siamese Neural Networks to Generate Image-Based Travel Recommendations

Abstract

Planning a trip is hard and time-consuming, but thanks to recommenders, it doesn’t have to be. Today, recommenders are used in practically every industry, making the user experience simple and stress-free. However, most recommenders have access to large volumes of data, something that is not always available when it comes to recommending travel destinations, leading to what is known as the cold start problem. For example, when recommending hotels or restaurants, recommenders can make matches based on user reviews. Recommending a destination presents a different kind of problem. There are little to no user reviews for an overall city, beach, or destination. This is why a different approach must be taken: using images to recommend travel destinations. Similar images of various types of destinations can be analyzed by a neural network in order to determine their similarity. But even then, a large amount of images would be needed in order to train an effective neural network capable of predicting accurate recommendations. So in order to minimize the amount of images needed, a Siamese Neural Network will be implemented. The model will be implemented using the TensorFlow framework with a Triplet Loss function. The dataset being used to train the network is a small custom-built dataset of city and beach images gathered from Wikipedia and Wikimedia Commons. The demonstrated Siamese model was able to accurately identify similar images of similar places.