PDF Publication Title:
Text from PDF Page: 001
Recommending Related YouTube Videos Shantanu Thakoor, Anchit Gupta, Parth Shah {thakoor, anchitg, parth95}@stanford.edu Abstract Building a recommendation system for YouTube represents a problem of large scale and huge importance. In this paper we explore a recommendation system, which unlike previous approaches more directly relies on YouTube’s inherent graph structure. Using graph node embeddings we train a neural network for link prediction and subsequently tackle the challenge of scalability by combin- ing this with a locality sensitive hashing based approach. We also show how our approach extends to provide recommendation given a set of user rated videos. 1 Introduction Youtube is the world’s largest online video content provider. It is responsible for serving over 5 bil- lion videos every single day to a total user base of size 1.3 billion. Building a good quality recom- mendation engine to serve this enormous group of people and help them sift through the billions of minutes of videos is of huge importance. In this project, we seek to tackle the problem of YouTube Video Recommendations. Specifically, given a set of videos and their ratings, we wish to suggest other videos similar to the well-rated videos and dissimilar to the badly-rated videos - this could be used, for example, to suggest a new video to a user based on videos they have already liked, or to suggest a suitable additional video for a play-list a user has created. Traditionally this problem has been solved by using an approach based on classical recommender systems where we are trying to recommend an item to a user based on their preferences. Tradi- tional recommender systems use techniques like item based collaborative filtering [ASA+10], which use notions of distances between items. Methods like this rely on having utility matrices available and cannot scale well to the huge sizes involved in YouTube. Also these distances are calculated using feature vectors of items, and do not use any graph information. Our approach relies on using graph information to give better recommendations and also a much more scalable system. We cast YouTube recommendation as a network problem, where given a set of nodes we wish to predict other nodes that are closely related to those in our set. The bulk of our approach relies on using node2vec [GL16], where nodes are embedded into a low-dimensional vector space that has certain nice properties of distances between neighbors and graph structure being preserved. After computing these vectors for our graph we proceed to train a neural network for link prediction on these vectors. To make our approach more scalable we subsequently do locality sensitive hashing of these vectors for scalable nearest neighbor queries. We also investigate combining the neural network based link predictor with the hashing based nearest neighbor search to further filter out the videos and obtain good quality recommendations. The rest of the paper is organized as follows. In Section 2, we give a brief overview of past work tackling similar problems. In Section 3, we describe the dataset we use, and conduct preliminary analysis of the dataset with experiments relevant to our task. In Section 4, we begin addressing our main problem, by applying the node2vec technique to our dataset. We also present analysis on its Stanford University CS224W Final ProjectPDF Image | Recommending Related YouTube Videos
PDF Search Title:
Recommending Related YouTube VideosOriginal File Name Searched:
cs224w-66-final.pdfDIY PDF Search: Google It | Yahoo | Bing
Cruise Ship Reviews | Luxury Resort | Jet | Yacht | and Travel Tech More Info
Cruising Review Topics and Articles More Info
Software based on Filemaker for the travel industry More Info
The Burgenstock Resort: Reviews on CruisingReview website... More Info
Resort Reviews: World Class resorts... More Info
The Riffelalp Resort: Reviews on CruisingReview website... More Info
| CONTACT TEL: 608-238-6001 Email: greg@cruisingreview.com | RSS | AMP |