PDF Publication Title:
Text from PDF Page: 007
In our network, nodes having an edge is a proxy for the underlying videos being related. Hence, we have obtained a classifier that, given a pair of video, can very accurately predict whether they are related. In theory, this classifier could be used to find all related videos of a given node. How- ever, doing so would involve iterating through all the other nodes in the graph. This is not feasible for large datasets such as ours. Instead, we plan to use a fast, approximate method to generate a list of nodes which are likely candidates for being related videos, and then use this neural network to further filter this list. In Section 5, we describe a Locality Sensitive Hashing based nearest- neighbor approach, to generate this list of candidates, and further extend this classifier which only predicts whether a pair of nodes is related, to a system which recommends a new video given a set of videos and their ratings. 5 Locality Sensitive Hashing for Approximate Related Video Prediction In this section, we first provide motivation for using LSH as a scalable, approximate solution to our problem. Then, we conduct experiments showing that we can indeed find a hash function satisfy- ing our desired property of locality preservation. Finally, we evaluate this approach on the task of giving video recommendations based on a set of videos and their ratings. 5.1 LSH as a Scalable Recommendation Algorithm As detailed in [ASA+10], most recommender systems use a user-item utility matrix, forming a rich model of user preferences and item similarity. These approaches do not scale well - in partic- ular, for the YouTube platform where there are millions of users and videos, it would be infeasible to store this information. However, we can assume that we have a limited amount of information about a particular user’s preferences. Hence, given a set of videos a user has watched, and their ratings, we wish to provide the user with recommendations of new videos. To solve this problem, we first recall the original work on which node2vec was based - the Word2Vec model introduced in [MCCD13]. This model embeds words into a low-dimensional vector space, for use in downstream NLP tasks. The node2vec model is quite analogous to Word2Vec in how it is trained and used. Later works in NLP considered finding an efficient and expressive representation of sentences. In particular, [ZKFM10] suggests finding a sentence vector by taking a linear combination of the vector representations of the words it contains. We extend this approach, to representing a set of videos and their ratings, by taking a linear combi- nation of the vector representations of the videos, using weights based on their ratings. ′ i wivi v = i wi We use this weighted vector v′ and find videos in its neighborhood. The intuition is that we will be finding videos close to the well-rated videos, and in a sense far away from the badly-rated videos. Hence, we can use efficient nearest-neighbor algorithms to find the best videos, and thus make our recommendation. However, we face a further problem of scalability: nearest neighbor searches in a high-dimensional dataset are expensive. Traditional recommender systems can deal with this by taking advantage of certain properties of the utility matrix such as sparsity; since we do not have a full user-item util- ity matrix, we instead propose an efficient approximate solution. As detailed in [WSSJ14], can use Locality Sensitive Hashing, to find a single number that represents each node. We then con- duct a nearest-neighbor search on this one-dimensional space. Since our hash function is locality- preserving, a nearest-neighbor search in this space would be a good approximate solution to a nearest-neighbor search in the original space. Such a one-dimensional search may also be imple- mented using a distributed hash table similar to the work by [SMK+01] - hence, this approach would truly be scalable for a dataset the size of YouTube. 7PDF 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 |