Traditional recommendation system has low recommendation accuracy and severe cold start problem. A hybrid recommendation method combining improved sparse-edge noise reduction automatic coding and matrix decomposition model of neighboring project influences is proposed

considering the review text and scores

through an improved sparse-edge noise reduction automatic coding model from commodity review texts. Product features are extracted

combining user reviews and ratings

taking into account the impact of neighboring users on target users

integrating the influence of neighboring projects into matrix decomposition algorithms

and compared with the traditional collaborative deep learning model（CDL） and hybrid SDAE model

it is increased highest by 8. 370 %.