Basic Estimate 2

Loading...

En provenance du cours de University of California San Diego

Advanced Algorithms and Complexity

204 notes

University of California San Diego

Advanced Algorithms and Complexity

204 notes

Cours 5 sur 6 dans Specialization Data Structures and Algorithms

You've learned the basic algorithms now and are ready to step into the area of more complex problems and algorithms to solve them. Advanced algorithms build upon basic ones and use new ideas. We will start with networks flows which are used in more typical applications such as optimal matchings, finding disjoint paths and flight scheduling as well as more surprising ones like image segmentation in computer vision. We then proceed to linear programming with applications in optimizing budget allocation, portfolio optimization, finding the cheapest diet satisfying all requirements and many others. Next we discuss inherently hard problems for which no exact good solutions are known (and not likely to be found) and how to solve them in practice. We finish with a soft introduction to streaming algorithms that are heavily used in Big Data processing. Such algorithms are usually designed to be able to process huge datasets without being able even to store a dataset.

À partir de la leçon

Streaming Algorithms (Optional)

In most previous lectures we were interested in designing algorithms with fast (e.g. small polynomial) runtime, and assumed that the algorithm has random access to its input, which is loaded into memory. In many modern applications in big data analysis, however, the input is so large that it cannot be stored in memory. Instead, the input is presented as a stream of updates, which the algorithm scans while maintaining a small summary of the stream seen so far. This is precisely the setting of the streaming model of computation, which we study in this lecture. The streaming model is well-suited for designing and reasoning about small space algorithms. It has received a lot of attention in the literature, and several powerful algorithmic primitives for computing basic stream statistics in this model have been designed, several of them impacting the practice of big data analysis. In this lecture we will see one such algorithm (CountSketch), a small space algorithm for finding the top k most frequent items in a data stream.

Introduction5:23

Heavy Hitters Problem7:33

Reduction 14:35

Reduction 26:30

Basic Estimate 18:04

Basic Estimate 27:14

Final Algorithm 15:27

Final Algorithm 212:57

Rencontrer les enseignants

Alexander S. Kulikov
Visiting Professor
Department of Computer Science and Engineering
Michael Levin
Lecturer
Computer Science
Daniel M Kane
Assistant Professor
Department of Computer Science and Engineering / Department of Mathematics
Neil Rhodes
Adjunct Faculty
Computer Science and Engineering

Well, now that we have been able to show that the expectation of

our randomized estimate for fi is exactly fi.

We would like to bound the variance and show that our randomized estimate for

fi actually is close to fi with high probability.

So, one way to show this is to use Chebyshev's inequality,

which is a quantity aversion of the law of large numbers.

Chebyshev's inequality says the following.

That if we have a random variable X with mean mu and variance sigma squared.

Then the probability that this random variable deviates from its mean

by more than a factor t times its standard deviation sigma is at most, 1/t squared.

So, our plan is as follows.

We would like to apply Chebyshev's inequality to our randomized estimator.

So, we want to let X, the random variable, equal the counter times the sine of i,

where i is a fixed date item.

And to do that, we need to bound the variance,

because we know that the mean is fi already.

So, let's move onto variance.

For that, we want to recall the expressions that we have for

our estimator.

We know that the counter times s(i) converting as fi

plus the contributions from the other data items j distinct from i,

f(j) times its sine times the sine of i.

We also know that the expectation of the counter times s(i) is exactly fi.

So, what we want to bound is the variance of our estimator,

which is nothing but the expectation of the sum over j

different from I that is contribution from other elements in the universe.

Of their frequency times their sine times their sine of i in try quantity squared.

Now it turns out that this is not too hard to compute which we will now do.

Before we apply the expectation, I want to design a somewhat simpler form that

will be easier to apply the expectation to.

So the counter times s(i)- f(i) the entire thing squared can do it in

estimation over J different from I.

J prime, different from I of the products of the frequencies, fJ and

fJ prime times the product of sine times the sine of i squared.

Now of course the sine of i squared is nothing but 1.

So, what we get is a double estimation over a pair of items j and

j prime that are distinct from i of the product of their frequencies and

the product of their sines.

So, now if we take the expectation of this quantity,

we'll see that the following happens.

If we look at a term in this double summation that corresponds to j

different from j prime.

Then the expectation of the product of the sines is exactly zero by the same

token in our mean analysis.

So, what we get is that all terms with j distinct from j

prime are zero in expectation.

And so the variance, is exactly the summation of all data items in

the universe other than i of their frequency squared.

Okay, so we have been able to construct a randomized estimate for

the frequency of any fixed item i.

And we have been able to balance variants, so we would like to now conclude

by Chebyshev's inequality our estimate is good.

Meaning that our estimate is close to its mean

the frequency of i with high probability.

But in order to do that,

we need to check that the square root of the summation over other data items j of

their frequency squared is actually small compared to the frequency of our item i.

Now, this is not generally true.

This depends on the actual stream, and

on the actual item that we're trying to estimate.

So, let's look at some examples,

because these examples will show when our estimate works well, and when it does not.

This will inform us in using our basic estimate that we just constructed as

a building block for the full algorithm in the rest of the lecture.

Okay, so this is our expression for the variance.

And again, think of item I as fixed, this is the item that we're trying to estimate.

Now suppose for a second that our stream

of data items has the following frequency histogram shown on the slide.

In this histogram,

item number 1 which is the most frequent item in the stream dominates the stream.

Well, if that is the case, and

if we're trying to estimate item one, then our estimate should work fairly well.

Indeed, our estimate works well exactly when the frequency of item i is larger

than roughly the square root of the sum of squares of other frequencies.

On the other hand, if this is our stream and instead of estimating item one.

We're trying to estimate our item six, our estimate is quite poor because

the error in estimating item six will include a contribution from item one.

And that item dominates the stream, so that is a problem.

So, to summarize, this very, simple basic estimate that we constructed works

really well if we're trying to estimate elements that dominate the stream.

And works quite poorly when we're trying to estimate elements that don't.

Especially in the presence of very large elements in the stream.

Okay, and so now our final algorithm.

Our full version of ApproximatePointQuery and the final CountSketch algorithm

will fix these deficiencies of our best basic estimate.

The way our algorithm works is as follows.

In order to find the top k elements approximately, what we will do is this.

We're first given our data string we'll hash the data

items in our universe into k buckets.

Well, this can be viewed as the process of constructing substreams.

Now we will run our basic estimate that we just designed

on every substream that is obtained via these hashing posts.

Now finally to boost the confidence of our estimates,

we'll repeat this procedure a logarithmic number of times.

And this will be our algorithm.

That the main intuition for why this works is exactly this.

This hashing process will allow us to insure and

when we're estimating a large item, then most of the substreams

that we estimated from will look like the one on the top of the slide.

We'll show that our current sketch algorithm will be estimating the large

items mostly from substreams that are dominated by these large items.

Now, if you want to estimate a small item,

you should also be able to figure out that this item has a small frequency.

The algorithm, as we will show,

actually estimates the small items mostly from streams that look like this,

that don't have any of the large items contaminating the estimate.

And in the next part of the lecture we will see the details of the algorithm and

some of the main ideas behind the analysis.

Explorer notre catalogue

Rejoignez-nous gratuitement et obtenez des recommendations, des mises à jour et des offres personnalisées.

Coursera propose un accès universel à la meilleure formation au monde, en partenariat avec des universités et des organisations du plus haut niveau, pour proposer des cours en ligne.

© 2018 Coursera Inc. Tous droits réservés.

Télécharger dans l'App Store

Disponible sur Google Play