Research On Attribute Embedding-based Deep Recommendation Algorithm

The rapid development of Internet technology and industry has not only changed the socio-economic and lifestyle of people,but also changed the objects and means of scientific research.With the increasing of Internet access and the dynamic,complexity and quantity of information resources,the recommender system,as an important means of filtering information,is the most effective solution to overcome information overload in recent years.Among them,the rating prediction is one of the most important tasks of the recommender system,which deduces the unknown rating according to the existing rating,thereby implementing recommendations.The recommender system makes new recommendations by learning the historical ratings of user-items.This is also known as collaborative filtering and is the most mainstream and basic method in the recommendation field today.However,the simple use of the collaborative filtering method for ratings often does not work well because the user 's ratings of items are often sparse.For example,the sparseness of Movielens1 M dataset is 95.35%,and the sparseness of Movielens10 M has reached 98.59%.Cold start due to sparsity is often an obstacle to providing high-quality recommendations.In order to make up for this problem,this thesis proposes a recommendation method combining attribute information and rating matrix to effectively merge attribute information into explicit feedback.The attribute information may be the attributes of the user,such as gender,age,occupation,etc.or the attributes of the item,such as the director,actor,country,and type of the movie,or the price,color,manufacturer,and production date of the product.This thesis explores an effective method of fusing user ratings and item attributes to improve the recommendation accuracy,and proposes a deep recommendation model based on attribute embedding that uses attention mechanisms and deep feedforward networks to model the representation of users,items,and attributes of item,and a generalized collaborative filtering based on deep non-linear interactions between users and attributes of items is realized.In order to effectively utilize the rich attribute information beyond the rating,this thesis uses an end-to-end network to model user-item ratings and all attribute information,and uses the attention mechanism to dynamically assign different weights to each attribute.In the end,all components complement and inspire each other,so that feature representations richer in semantic information can be learned.In order to verify the feasibility and effectiveness of the general recommendation model proposed in this thesis,extensive experiments have been conducted on the Hetrec2011-movielens-2k data-set,and the proposed method has made substantial progress compared with other classical methods in attribute representation learning and recommendation performance.