A Graph Neural Network Recommendation System for Joint Comment Text Representation

Aiming at the current insufficiency of search engines alone in adequately addressing information overload across various domains, a graph neural network recommendation system incorporating review text representations was designed. This work includes designing an attention mechanism-based review feature extraction module, which inputs text segments within a sliding window into an attention network to compute the attention weights of words in context. Simultaneously, the traditional random negative sampling method was improved by replacing the traversal step unfavorable for parallel computation in conventional random negative sampling algorithms with matrix computations, thereby enhancing the algorithm's computational efficiency on parallel platforms. In addition, a hybrid negative sampling strategy was designed for this model, which employs the improved random negative sampling algorithm in the early training phase to rapidly iterate model parameters and adopts the positive and pruned negative sampling algorithm in the later training phase to sample high-quality negative samples, providing reliable support for further improving model accuracy. During the experimental phase, the effectiveness of the model and the hybrid negative sampling strategy was validated on three real-world open-source recommendation datasets.