Due to the spread of online web services about videos such as YouTube and Tencent, the number of videos posted to them is increasing. We can now watch a large amount of videos with ease, however, explosive videos available frequently overwhelmed users, leading them to make poor decisions. Recommendation systems are designed to predict the future preferences of users' based on their previous interactions with the items. For example, a user's previous viewing information can be used to make recommendations on his/her future viewing video. The vast amount of information produced by the users can be used by different methods for recommendations, with either neighborhood based methods, or machine-learning based methods or matrix-factorization based methods. These all use low-rank approximation of input data, however, may suffer from data sparsity or noise in data. Recently, word embedding methods in deep learning are successfully used to learn linguistic regularities and semantic information from large text datasets. It has made a big innovation in natural language processing and text mining fields. They can learn low-dimensional vector space representation of input elements with effect, which lays the foundation of the recommendation research in online video services. Actually, user's viewing history is one of the important factors reflecting the user's attention and also it has good scalability, high accuracy and flexibility comparing to rating data and reviews written by users. In this work, we aim to recommend next videos by adopting word embedding techniques proposed in Word2Vec framework via analyzing user's playback behavior. Unlike the previous works that use Word2Vec for recommendation, a non-textual feature namely the past viewing videos of the users, is used to make recommendations. By representing each user (according to her/his history viewing behavior) with a high dimensional vector space, we can filter the target user potential interest in candidate video, thus a personalized recommendation strategy based on the user's playback behavior sequence for online video service websites is proposed. This strategy maps the video into feature vectors and extracts the semantic features of the video via user's video playing behavior. Then, the user's interest distribution matrix is modeled by clustering the feature vectors of the user's video playing history. A recommendation list is generated in combination with user interest preferences and user viewing history aging. Offline experiments were conducted in a large-scale video service system. Compared with random algorithms, item-based collaborative filtering recommendations, and user-based collaborative filtering recommendations, it improves the average accuracy of Top-N recommendation for users watching videos 22.3%, 30.7% and 934% respectively. The relative increase of 52.8%, 41% and 1065% respectively in the recall rate indicator is achieved. Moreover, compared with the matrix decomposition algorithm SVD++, the model based on the bidirectional LSTM+Attention and the deep interest network DIN, the proposed method achieves different levels of improvement in both the Top-N recommendation accuracy rate and the recall rate. The recommendation strategy not only obtains better performance as a whole, but also attempts to solve the problems of data requirements, data sparsity and data noise faced by traditional recommendation algorithms. © 2020, Science Press. All right reserved.
