Finding similar items–“mining of massive datasets” chp3

“Mining of Massive Datasets”

Jure Leskovec, Anand Rajaraman, Jeff Ullman

 


 

Summary of chapter 3 — Finding similar items

✦  Jaccard Similarity: The Jaccard similarity of sets is the ratio of the size of the intersection of the sets to the size of the union. This measure of similarity is suitable for many applications, including textual similarity of documents and similarity of buying habits of customers.

jaccard_sim

✦  Shingling: A k-shingle is any k characters that appear consecutively in a document. If we represent a document by its set of k-shingles, then the Jaccard similarity of the shingle sets measures the textual similarity of documents. Sometimes, it is useful to hash shingles to bit strings of shorter length, and use sets of hash values to represent documents.

✦  Minhashing: A minhash function on sets is based on a permutation of the universal set. Given any such permutation, the minhash value for a set is that element of the set that appears first in the permuted order.

min_hashing

✦  Minhash Signatures: We may represent sets by picking some list of per- mutations and computing for each set its minhash signature, which is the sequence of minhash values obtained by applying each permutation on the list to that set. Given two sets, the expected fraction of the permutations that will yield the same minhash value is exactly the Jaccard similarity of the sets.

minhash_signature

✦  Efficient Minhashing: Since it is not really possible to generate random permutations, it is normal to simulate a permutation by picking a random hash function and taking the minhash value for a set to be the least hash value of any of the set’s members.

✦  Locality-Sensitive Hashing for Signatures: This technique allows us to avoid computing the similarity of every pair of sets or their minhash signatures. If we are given signatures for the sets, we may divide them into bands, and only measure the similarity of a pair of sets if they are identical in at least one band. By choosing the size of bands appropriately, we can eliminate from consideration most of the pairs that do not meet our threshold of similarity.

✦  Distance Measures: A distance measure is a function on pairs of points in a space that satisfy certain axioms. The distance between two points is 0 if the points are the same, but greater than 0 if the points are different. The distance is symmetric; it does not matter in which order we consider the two points. A distance measure must satisfy the triangle inequality: the distance between two points is never more than the sum of the distances between those points and some third point.

 Euclidean Distance: The most common notion of distance is the Euclidean distance in an n-dimensional space. This distance, sometimes called the L2-norm, is the square root of the sum of the squares of the differences between the points in each dimension. Another distance suitable for Euclidean spaces, called Manhattan distance or the L1-norm is the sum of the magnitudes of the differences between the points in each dimension.

✦  Jaccard Distance: One minus the Jaccard similarity is a distance measure, called the Jaccard distance.

✦  Cosine Distance: The angle between vectors in a vector space is the cosine distance measure. We can compute the cosine of that angle by taking the dot product of the vectors and dividing by the lengths of the vectors.

cosine_sim

✦  Edit Distance: This distance measure applies to a space of strings, and is the number of insertions and/or deletions needed to convert one string into the other. The edit distance can also be computed as the sum of the lengths of the strings minus twice the length of the longest common subsequence of the strings.

edit_dis1

✦  Hamming Distance: This distance measure applies to a space of vectors. The Hamming distance between two vectors is the number of positions in which the vectors differ.

hamming_dis

✦  Generalized Locality-Sensitive Hashing: We may start with any collection of functions, such as the minhash functions, that can render a decision as to whether or not a pair of items should be candidates for similarity checking. The only constraint on these functions is that they provide a lower bound on the probability of saying “yes” if the distance (according to some distance measure) is below a given limit, and an upper bound on the probability of saying “yes” if the distance is above another given limit. We can then increase the probability of saying “yes” for nearby items and at the same time decrease the probability of saying “yes” for distant items to as great an extent as we wish, by applying an AND construction and an OR construction.

✦  Random Hyperplanes and LSH for Cosine Distance: We can get a set of basis functions to start a generalized LSH for the cosine distance measure by identifying each function with a list of randomly chosen vectors. We apply a function to a given vector v by taking the dot product of v with each vector on the list. The result is a sketch consisting of the signs (+1 or −1) of the dot products. The fraction of positions in which the sketches of two vectors agree, multiplied by 180, is an estimate of the angle between the two vectors.

✦  LSH For Euclidean Distance: A set of basis functions to start LSH for Euclidean distance can be obtained by choosing random lines and project- ing points onto those lines. Each line is broken into fixed-length intervals, and the function answers “yes” to a pair of points that fall into the same interval.

✦  High-Similarity Detection by String Comparison: An alternative approach to finding similar items, when the threshold of Jaccard similarity is close to 1, avoids using minhashing and LSH. Rather, the universal set is ordered, and sets are represented by strings, consisting their elements in order. The simplest way to avoid comparing all pairs of sets or their strings is to note that highly similar sets will have strings of approximately the same length. If we sort the strings, we can compare each string with only a small number of the immediately following strings.

✦  Character Indexes: If we represent sets by strings, and the similarity threshold is close to 1, we can index all strings by their first few characters. The prefix whose characters must be indexed is approximately the length of the string times the maximum Jaccard distance (1 minus the minimum Jaccard similarity).

✦  Position Indexes: We can index strings not only on the characters in their prefixes, but on the position of that character within the prefix. We reduce the number of pairs of strings that must be compared, because if two strings share a character that is not in the first position in both strings, then we know that either there are some preceding characters that are in the union but not the intersection, or there is an earlier symbol that appears in both strings.

✦  Suffix Indexes: We can also index strings based not only on the characters in their prefixes and the positions of those characters, but on the length of the character’s suffix – the number of positions that follow it in the string. This structure further reduces the number of pairs that must be compared, because a common symbol with different suffix lengths implies additional characters that must be in the union but not in the intersection.

 


Note: Pictures quoted from Google

1 条评论

  • manicure 2017年4月23日 回复

    Hey very interesting blog!

发表评论

电子邮件地址不会被公开。 必填项已用*标注