PDF Publication Title:
Text from PDF Page: 002
efficiency and our ability to predict edges based on it. In Section 5, we further develop our main approach, by presenting our idea of using Locality Sensitive Hashing to provide a more scalable method, and commenting on its effectiveness. Finally, in Section 6, we provide a brief summary of our work, insights we gain from it, and possible future extensions. 2 Related Work Extensive work has been done on recommendation systems based on per user item utility func- tions. These can be broadly classified into three categories 1) Content based 2) User collaborative 3) Item collaborative . For a more detailed description, we refer the reader to [ASA+10]. Using neighborhood information for collaborative filtering recommender systems has previously been studied in [Dom07]. The authors in [SP14] propose a scalable peer to peer recommendation system based on distributed hash tables and locality sensitive hashing of user-item preference ma- trices. Based on ideas inspired from the above, [DLL+10] build a YouTube recommendation system by computing candidate videos based on relatedness scores between videos and then a ranking algo- rithm. More recently [CAS16] propose a algorithm having both candidate generation and ranking done using deep neural networks. These methods differ from our approach as they use complete user watch and click data unlike our more restricted dataset and also they do not incorporate graph information directly into their models. The graph based approach we adopt can also be seen in the light of link prediction. These methods mainly rely on graph heuristics based on graph properties like common neighbors, path lengths and neighborhood sets.[LNK03] provides good background on most of the classical approaches for the same. Recently node2vec [GL16], a node embedding approach based on random walks, has shown much better accuracy than the above methods for link prediction and we use the same in our work. In Section 4, we provide a more detailed explanation of the node2vec method. [RY17]looks at mak- ing a Reddit recommender system based on text data gathered from user comments and incorpo- rate the graph structure of the social network into their system using node2vec. For a more detailed treatment of the above papers, we refer the reader to our project proposal docu- ment. 3 Analysis of Dataset 3.1 Dataset Description We use the YouTube dataset found at [CDL08]. The dataset contains video IDs, along with some metadata including uploader, length, ratings, category, age, and a list of up to 20 IDs of related videos. The dataset was created by starting a breadth-first search crawl starting from videos in YouTube’s set of Most Viewed, Top Rated, Recommended, and other playlists. The dataset con- tains information from more than 60 such crawls, most of which discovered videos on the order of 106 to 107. Due to scalability issues, in this analysis, we restrict our attention to one particular crawl over YouTube videos. 3.2 Working with Maximum Strongly Connected Component We analyze the natural graph representation of this dataset, with nodes representing videos and an edge from a to b meaning that b was in the list of related videos for b. Our graph contains over 1.76M nodes, and 4.4M edges. In Table 1 we examine the out-degrees of nodes in the entire graph, and in the largest strongly connected component. We can see that in the entire graph a large majority of the nodes have out degree 0 and are sink nodes. For our approach, sink nodes are essentially nodes for which we have no information. Hence, node2vec (which finds related nodes using random walks starting at each node) will not be able to produce vectors that can easily be used in downstream tasks. Since the vast majority of nodes in the original graph have little to no information associated with them, we restrict our atten- 2PDF 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 |