Using DeepWalk to Link Prediction in Subgraph

This research presents an innovative application of the DeepWalk algorithm in conjunction with k-Nearest Neighbors (k-NN) for link prediction tasks, using the Amazon dataset as a case study. The DeepWalk algorithm, designed to learn latent representations of vertices within a network, leverages randomized path traversing to reveal localized structures within networks. This is achieved by employing random paths as sequences to train a Skip-Gram Language Model. This approach allows us to extract insights from a part of the graph and extend it to the entire network, enhancing computational efficiency. In this study, the dataset was preprocessed to handle duplicates and missing values. A graph representation of the dataset was constructed based on product similarities, with DeepWalk used to generate graph embeddings. A k-NN model was trained on these embeddings to predict product links. The performance of the approach was assessed using various evaluation metrics, with particularly high AUC and Recall values indicating strong predictive accuracy. Two feature selection techniques, LASSO regression and Recursive Feature Elimination (RFE), were employed to improve the model's performance. The results highlighted ‘Price’, ‘Offers’, ‘Stock Availability’, ‘Category’, and ‘MRP’ as the most impactful features.