As an open and complementary educational model, the Youth Night School in higher vocational colleges and universities connects students’ individualized learning needs and their entrepreneurial economic empowerment. The study categorizes students’ learning behaviors in youth night school into four dimensions of testing behaviors, forum interaction, content learning and resource searching, and 11 specific indicators of the observation system. And an improved deep neural network learning resource recommendation algorithm (UDN-CBR) is designed. It deeply mines the relationship between students’ learning behaviors and resource attributes by using a multilayer perceptron (MLP), and reads the resource text through a convolutional neural network (CNN). Ultimately, it fuses information from multiple sources to generate personalized learning resource recommendations for students. The study collected valid data and entrepreneurial profitability questionnaires from 354 participating night school students, whose average weekly study hours were above 2h in the middle and late semester, and whose average final test scores (87.67±10.99) were significantly higher than usual (77.95±13.58). Stepwise regression analysis and structural equation modeling validation pointed out that content learning was the most powerful factor in improving profitability with a path coefficient = 0.623, while forum interaction guided business stability with a standardized estimate = 0.588, and resource searching behaviors were the most prominent driver of self-growth performance with a path coefficient of 0.634.
