House price prediction with Convolutional Neural Network (CNN)

In this paper, we examine the applicability of Convolutional Neural Networks (CNNs) for predicting the cost of houses with an inclusion of visual and non-visual elements. Traditional end-to-end patterns of supervised machine learning usually incorporate a finite set of numerical or categorical features, ignoring spatial and aesthetic information. CNNs, which can parse complex patterns in unstructured data, represent a way to improve predictive accuracy at the price of including property images with numerical and geographical coordinates. This variety of inputs is integrated within the proposed model to reflect the visual and contextual aspects of real estate valuation. The training and testing set used property images, georeferenced information, and typical numerical features. Incorporation of architectural features, including design, quality and neighbourhood aesthetics, showed a higher variance than the baseline machine learning model. Analysis of the outcomes obtained in experiments showed that CNN-based models performed better than the benchmark models, such as regression, gradient boosting, and random forest regression, in identifying the subtle interconnection between attractiveness and property values. The results further support the ability of CNNs to use multimodal information to solve diverse prediction problems within real estate. This research emphasizes the potential value of disseminating multiple data sources and merging relatively sophisticated neural structures into real estate applications, indicating the future viability of such paradigms in broader property market technologies. Subsequent studies can look at topics such as increasing the capacity of model structures and expanding the dataset to incorporate time dependency and other regions.