Abstract: In view of inaccurate rating prediction of traditional recommendation model， a model based on side-information enhanced collaborative filtering auto-encoder was proposed， and the design principle， loss function and structure of the model were discussed in detail. The model was implemented by a single hidden layer auto-encoder network with ratings and side-information as input/output data. The model was directly trained by ratings and side-information. The proposed framework not only reduces the scale and complexity of the model， but also amends the user’s interest with side-information. At the same time， the representation ability of the trained model is increased by reasonable division and expansion of training datasets. The comparisons on real datasets show that the proposed model with side-information can effectively ameliorate the user’s interest predicted by the model without side-information， and improve the accuracy of rating prediction， which has a certain practicality.

Key words: auto-encoder, collaborative filtering, side-information, recommender system, rating prediction