PDF Publication Title:
Text from PDF Page: 008
5.2 Analysis of the LSH Results We use the Locality Preserving Hash function provided by the scikit-learn LSHForest module 1. We wish to verify that this hash function is indeed locality-preserving. To this end, we conducted the following experiment: for a node of degree d, we calculated the 1.5d nearest neighbors, the extra 0.5d being a fudge factor to allow for the approximate nature of the LSH function. We then calculated the overlap between the node’s d true neighbors, and the 1.5d neighbors predicted by our algorithm. We calculated this both using the video category feature as part of the input to the LSH, and without. Finally, we plotted this function as amount of overlap as a function of the node degree in Figure 5.2. As we can see from the figure, the LSH is quite a reasonable approximate method for finding the neighbors of a node. For nodes of all degrees, it predicts around half of their true neighbors. An interesting fact is that adding category information actually worsens our results - but this is not surprising, considering that in Section 3 we found that video category is not a reliable feature for edge prediction. As a final experiment, given a node and its list of neighbors predicted by the LSH algorithm, we fed them to our classifier (as detailed at the end of Section 4). We obtained very good results from thisexperiment-onaverage, 82.1% oftheneighborspredictedbytheLSHalgorithm,wereclas- sified as having edges to the test node. In other words, the LSH algorithm predicts related videos with 82.1% accuracy. Hence, our LSH method is indeed an efficient yet accurate method for link prediction. 5.3 Experimental results based on a set of rated videos The approaches detailed in prior sections solved the problem of predicting related videos for a sin- gle given video. We now extend this to predicting related videos for a set of videos and their rat- ings. Since our dataset is almost 10 years old, most of the videos discovered in its crawls are no longer available on YouTube. We hence find the 500 top viewed videos that are still available by scraping the YouTube website, and use them to gain some insight into the data. Since our dataset contains edges between pairs of nodes, we had data regarding single-video rec- ommendations. Hence, we are able to present quantitative analysis in prior sections. However, we have no data regarding recommendations from a set of videos, and hence quantitative analysis is not possible. Hence, we present qualitative analysis to argue that our method gives reasonable re- sults. The results are summarized in Table 5.3. As a first test, we start with a seed set of 5 videos, all in the sports category and uploaded by the NBA channel. We take all the ratings (i.e, the weights in our linear combination) as equal. The 5 1scikit-learn.org/stable/modules/generated/sklearn.neighbors.LSHForest.html 8PDF 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 |