LSH.9 Locality-sensitive hashing: how it works

[@ayushsinghal28]

how do we determine a good value of 'L' i.e. the number of tables. Is there some logic to get it?

[@mariusm5187]

Finally understand LSH

[@carles5601]

Great explanation. Thanks!

[@AnkitSharmaKumar]

Great explanation... Thanks for sharing!

[@gaaligadu148]

I don't understand how we found near duplicates inside a bucket.You can't compare using all the D-dimensions because obviously it will be different because they are near duplicates.Don't we use subsections of the D-dimensions like you said it before to check near duplicates ?

[@dani0qiu0china]

very good complexity analysis !

[@MrBertmsk]

it's little unclear how to eliminate duplicates? Each "table" (bucket?) contains different hash ids for the data. Should I do comparasions within one bucket or against all buckets? How to combine resulting hash then?

[@renzocoppola4664]

if you compare against all buckets then you would be comparing aganst all points

[@federicomagliani1]

Have you been understood? I repeated the hash process and I only concatenate the results.

[@gauravmenghani4]

You do comparisons within the buckets to remove false positives. You repeat the process with new random hyperplanes to consider points which were false negatives in the previous iteration.

[@renzocoppola4664]

I suppose you could take the adventage that the neraby buckets have 1 bit difference.

[@alihusen111]

would you pleas tell me what do you mean when you said we do the same comparison to eliminate d ??

[@harshgoyal5694]

What do u mean by D dimensional document??? Thanks in advance :)

[@RobertoMartin1]

D is the size of the dictionary. you represent each document with words from the dictionary, so a D dimensional document will have at most d unique words. Usually, each document will contain less than D unique words, but they're still represented with D dimensions, just that some of the dimensions have zero as values.

[@rahulat85]

8:39

[@gaaligadu148]

Hi harsh, each document can be represented numerically by D-dimensions. For ex: it can be image whose D-dimension vector would be all of it's pixel values