In this paper, we propose a new neural network architecture for item recommendation with structural information. Our model, structural neural recommender (SNR) is based on neural networks and operates on a hierarchy paradigm, aiming to explore the effectiveness of incorporating different structural information for recommendation. Many recent state-of-the-art neural network based recommendation models exploit the nonlinear transformations for modeling the complex user-item interaction patterns and user historical behaviors, ignoring the item-item structural relationship. This structural information, however, is uncomplicated to derive and useful for inferring item characteristics. To utilize this information, SNR simultaneously learns representation from user-item interactions and item-item relationships. Empirical studies on eight real-world datasets demonstrate the effectiveness of incorporating such structural information, by outperforming classic and recent baselines. We also conduct detail ablation studies and hyper-parameter analysis to provide further understanding towards the behaviors of our model. Following the model development, we conduct a field experiment to demonstrate that the effectiveness of algorithmic recommender systems can further increase by using different types of message framing when communicating recommendations to consumers. Our results suggest that recommendations framed with a relevance appeal (e.g. “Top 5 brands for you”) are more effective in general, yet recommendations that are framed with a popularity appeal (“Top 5 most popular brands”) are more effective for customers who were acquired via social media (versus non-social media) advertising or for those who have stronger (versus weaker) social orientation.