| In recent years,the online video industry has been benefited with by the rapid development of mobile network.In accordance with the report from CNNIC,the mobile netizen in China has reached 817 million,accounting for 98.6% population of the whole netizen and the market value of online video industry has been over 201 billion,which is 39.1% greater than the number in the last year.The increasingly more customers have aroused much more demanding recommendation systems while they also contributed to tremendous profit.Related recommendation is an important application of video recommendation system and popular among online video sites.Users do not need to get back when they clicked into a video and most related videos will be listed aside automatically.The related recommendation module has much enhanced user experience and has come to be the one of the most significant product of video sites.The target of related recommendation is to calculate the proximity of two videos.Related recommendation is usually divided into two parts.The first step is transforming videos into embedding vectors and making the embedding vectors retain video information to the utmost extent is the key of related recommendation.The next step is proximity search,which means searching most similar videos for a specific video as the recommendation result according to similarity measures like cosine similarity.There are two important types of information for online videos.The first one is the video playing history of users which recorded the interaction between users and videos.It will indicate the similarity between two videos if they appear consecutively in a playing sequence.The another one is the attributes of videos and one of the most valuable attributes is tag.The relation between tags and videos is similar to the connection between keywords and papers.Apparently,if two videos share similar tags,they will be proximal to each other.As for now,researches have been focused on the exploit of playing history or video attributes and cannot jointly use both of the two types of information.Consequently,in this paper,we have proposed a video related recommendation method based on graph embedding.The key insight is a video embedding model integrated both playing history and video tags.After getting the video embedding,cosine similarity is adopted to implement video recommendation.The model is divided into two sub-models to integrate tag information.The first model is the tag embedding model based on Deep Walk,which models tag feature properly.The second one is the video embedding model which incorporates the out embedding of the first one.The video embedding model is able to handle video tag inputs while receive video sequences.Finally,the model’s performance is enhanced and cold-start can be alleviated because those two kinds of information compensates each other. |
