The cold start problem is one of the most challenging issues in recommender systems. In Walmart’s learning recommender system, a large number of learning items are frequently created and updated. Accurately delivering these items, which often lack user interaction, to Walmart associates is crucial. In this scenario, optimizing the algorithm to address the cold start problem becomes especially important. In this article, we propose a pioneering two-stage, content-aware deep learning recommendation architecture specifically optimized for this challenge. In the candidate generation stage, we develop a novel semantic similarity algorithm to identify latent connections between learning items. This algorithm efficiently selects low interaction items related to Walmart associates’ job responsibilities, preparing the inputs for the algorithm in the subsequent ranking stage. In the ranking stage, we make two key enhancements to the Deep Factorization Machine (DeepFM) ranking algorithm to improve prediction accuracy, particularly in cold start scenarios. First, we train a neural language model to generate embedding vectors and input to both the wide and deep components of the ranking algorithm. Second, the semantic similarity score of candidate items and the user’s browsing history are incorporated into DeepFM, further enhancing the personalized ranking algorithm’s prediction accuracy. This deep learning recommendation system has been successfully launched in the Walmart Academy App.