Locality Sensitive Hashing (LSH) and Min Hash

[Indyk-Motwani’98] Many distance related questions (nearest neighbor, closest x, ..) can be answered more efficiently by using locality sensitive hashing, where the main idea is that similar objects hash to the same bucket.   LSH function: Probability of collision higher for similar objects Hash data using several LSH functions At query time, get all objects…


Set Similarity and Min Hash

Given two sets S1, S2, find similarity(S1, S2) – based not hamming distance (not Euclidean). Jaccard Measure View sets at a bit-array. Indexes representing each possible element, and 1/0 representing presence/absence of the element in the set. Then Jaccard measure = What happens when: n element in each set from a possible universe u, s.t….


Information Retrieval & Search – Basic IR Models

Our focus in the database world has primarily been on retrieving information from a structured source, through a structured query. Relational DBs –> structured data, structured query XML DBs –> semi-structured data, structured query search –> unstructured data, unstructured query Actually, search –> data with hidden structure, query with possibly hidden structure Recent efforts on…


Information Theory (1) – The Science of Communication

IT is a beautiful sub-field of CS with applications across the gamut of scientific fields: coding theory and communications (under unreliable channels), cryptography, physics, biomedical engineering, computer graphics, machine learning, statistics, and even gambling and stock trading. It is truly a marvelous through process. Framework: "Bit" as the unit of communication. Arbitrary. Ok as long…


Random Sampling over Joins

Source: On Random Sampling over Joins. Surajit Chaudhuri, Rajeev Motwani, Vivek Narasayya, Sigmod 1999. What? Random sampling as a primitive relational operator: SAMPLE(R, f) where R is the relation and f the sample fraction. SAMPLE(Q, f) is a tougher problem, where Q is a relation produced by a query In particular, focus on sampling over…


Converting Between Random Sampling Methods

  Sampling f fraction out of n records: Sampling with replacement Sample is a multi-set of fn records. Any record could be samples multiple times. Sampling without replacement Each successive sample is uniformly at random from the remaining records Independent Coin flips: choose a record with probability f. The sample has s distinct records where…

Reservoir Sampling

A simple random sampling strategy to produce a sample without replacement from a stream of data – that is, in one pass: O(N) Want to sample s instances – uniformly at random without replacement – from a population size of n records, where n is not known. Figuring out n would require 2 passes. Reservoir…