CS136, Lecture 32

1. Hashing functions
   1. Selecting hashing functions
      1. Digit selection
      2. Division
      3. Mid-Square:
      4. Folding
      5. String-valued keys
   2. What to do if you obtain hash clashes?
      1. Open Addressing:
         1. Linear rehashing
         2. Quadratic rehashing
         3. Double Hashing
      2. External chaining.
   3. Analysis

Hashing functions

This implementation assumes that the data has a key which is of a restricted type (some enumerated type in Pascal, integers in Java), which is not always the case.

Note also that the size requirements for this implementation could be prohibitive.

Ex. If the array held 2000 student records indexed by social security number it would be declared as ARRAY[0..999,999,999]

What if most of entries are empty? If we use a smaller array then all elements will still fit.

Suppose we have a lot of data elements of type EltType and a set of locations in which we could store data elements.

Consider a function H: EltType -> Location with the properties

1. H(elt) can be computed quickly

2. If elt₁ <> elt₂ then H( elt₁) <> H( elt₂). (H is one-to-one function)

This is called a perfect hashing function. Unfortunately, they are difficult to find unless you know all possible entries to the table in advance. This is not often the case.

Instead we use something that behaves well, but not necessarily perfectly:
i.e., keep property 1, but may give up 2.

The goal is to scatter elements through the array randomly so that they won't bump into each other.

Define a function H: Keys -> Addresses, and call H(element.key) the home address of element.

Of course now we can't list elements easily in any kind of order, but hopefully we can find them in time O(1).

Note that each entry in the table will need to include the actual key, since several different keys will likely get mapped to the same subscript.

Hash functions and hashtables are important enough that Hashtable is provided by Java as a utility class (in java.utility.* - see section 12.7 in Arnold & Gosling) and the method

        public int hashCode()

Most important property is that for all o and o', if o.equals(o') then must have o.hashCode() = o'.hashCode(). Otherwise hashTable algorithms won't work!

There are two problems to look at:

1. What are good hashing functions?

2. What do we do when two different elements get sent to same home address?

Selecting hashing functions

Here are some sample Hashing functions.

Presume for the moment that the keys are numbers.

a. Digit selection

Unfortunately it is easy to get a biased sample. We can carefully analyze keys to see which will work best. We must watch out for patterns - they should generate all possible table positions. (For example the first digits of SS#'s reflect the region in which they were assigned and hence usually would work poorly as a hashing function.)

b. Division

This is very efficient and often gives good results if the TableSize is chosen properly.

If it is chosen poorly then you can get very poor results. If TableSize = 2⁸ = 256 and the keys are integer ASCII equivalent of two letter pairs, i.e. Key(xy) = 2⁸ * (int)x + (int)y, then all pairs starting ending with the same letter get mapped to same address. Similar problems arise with any power of 2.

The best bet seems to be to let the TableSize be a prime number.

In practice if no divisor of the TableSize is less than 20, the hash function seems to be OK. (Text uses 997 in the sample program)

c. Mid-Square:

Example: Let the keys range between 1 and 32000 and let the TableSize be 2048 = 2¹¹.

Square the Key and remove the middle 11 bits. (Grabbing certain bits of a word is easy to do using shift operators in assembly language or can be done using the div and mod operators using powers of two.)

In general r bits gives a table of size 2^r.

d. Folding

This is often used if the key is too big. E.g., If the keys are Social security numbers, the 9 digits will generally not fit into an integer. Break it up into three pieces - the 1st digit, the next 4, and then the last 4. Then add them together.

Now you can do arithmetic on them.

This technique is often used in conjunction with other methods (e.g. division)

e. String-valued keys

If you use longints to hold the numbers, then you can get 4 letters into one number in this way. If they are all alphabetic (no special characters), then you can subtract (int)'a' from each ASCII code in order to reduce the size of the keys.

Here is a very simple-minded hash code for strings: Add together the ordinal equivalent of all letters and take the remainder mod tableSize.

Problem: Words with same letters get mapped to same places:

miles, slime, smile

This would be much improved if you took the letters in pairs before division.

Nevertheless, for simplicity we adopt this simple-minded (and thus relatively useless) hash function for the following discussion.

Here is a function which adds up ord of letters and then mod tableSize:

hash = 0;
for (int charNo = 0; charNo < word.length(); charNo++) 
    hash = hash + (int)(word.charAt(CharNo));
hash = hash % tableSize;  (* gives 0 <= hash < tableSize *)

Code is only a little more complex to multiply each succeeding character by 2*8.

Efficient way using Horner's rule:

hash = 0;
for (in CharNo = word.length()-1;charNo >= 0; charNo--)
    hash = (256*hash + (int)(word.charAt(CharNo))) % tableSize;

Notice we mod by tableSize each time we update hash to prevent overflows.

Efficient way of calculating uses only word.length() multiplications, while normal way would involve O(word.length()²) multiplications.

What to do if you obtain hash clashes?

A hash clash occurs when two different keys have the same home location.

There are two main options for getting out of trouble:

1. Rehash to try to find an open slot (open addressing). This must be done in such a way that one can find the element again quickly!

2. External chaining.

    Object [] table = new Object[TableSize]; 

1. Open Addressing:

There will be three types of entries in the table:

• they are empty, represented by null;

• deleted entries - marked by inserting a special "reserved" object", and

• normal objects which contain a reference to an object.

Here are some variations:

a. Linear rehashing

This is about as simple a function as possible. Successive rehashing will eventually try every slot in the table for an empty slot.

Ex. Table to hold strings, TableSize = 26

H(key) = ASCII code of first letter of string - ASCII code of first letter of 'a'.

Strings to be input are GA, D, A, G, A2, A1, A3, A4, Z, ZA, E

Look at how get inserted with linear rehashing:

0   1   2   3   4   5   6   7   8   9   10      ...     26
A   A2  A1  A3  D   A4  ZA  GA  G   E   ..      ...     Z

In example happened with A, A2, A1, A3, A4.

Secondary clustering occurs when a key has a home address which is occupied by an element which originally got hashed to a different home address, but in rehashing got moved to the address which is the same as the home address of the new element.

In example, happened with E

What happens when delete A2 & then search for A1?

Must mark deletions (not just make them empty) and then try to fill when possible.
Can be quite complex.

See code in text for how this is handled.

Minor variants of linear rehashing would include adding any number k (which is relatively prime to TableSize) to i rather than 1.

If the number k is divisible by any factor of TableSize (i.e., k is not relatively prime to TableSize), then not all entries of the table will be explored when rehashing. For instance, if TableSize = 100 and k = 50, the Probe function will only explore two slots no matter how many times it is applied to a starting location.

Often the use of k=1 works as well as any other choice.

b. Quadratic rehashing

This variant helps with secondary clustering but not primary clustering. (Why?) It can also result in instances where in rehashing you don't try all possible slots in table.

For example, suppose the TableSize is 5, with slots 0 to 4. If the home address of an element is 1, then successive rehashings result in 2, 0, 0, 2, 1, 2, 0, 0, ... The slots 3 and 4 will never be examined to see if they have room. This is clearly a disadvantage.

c. Double Hashing

E.g. compute delta(Key) = Key % (TableSize -2) + 1, and add delta for successive tries.

If the TableSize is chosen well, this should alleviate both primary and secondary clustering.

For example, suppose the TableSize is 5, and H(n) = n % 5. We calculate delta as above. Thus, while H(1) = H(6) = H(11) = 1, the jump sizes differ since delta(1) = 2, delta(6) = 1, and delta(11) = 3.

2. External chaining.

The easiest way to do this is to let each slot be the head of a linked list of elements.

Draw picture with strings to be input as GA, D, A, G, A2, A1, A3, A4, Z, ZA, E

The simplest way to represent this is to allocate a table as an array of pointers, with each non-nil entry a pointer to the linked list of elements which got mapped to that slot.

We can organize these lists as ordered, singly or doubly linked, circular, etc.

We can even let them be binary search trees if we want.

Of course with good hashing functions, the size of lists should be short enough so that there is no need for fancy representations (though one may wish to hold them as ordered lists).

See figure 14.2 in text on efficiency

There are some advantages to this over open addressing:

1. Deleting not a big deal

2. The number of elements stored in the table can be larger than the table size.

3. It avoids all problems of secondary clustering (though primary clustering can still be a problem)

Analysis

The load factor of a table is defined as a = F(number of elts in table,size of table)

a = 1 means the table is full, a = 0 means it is empty.

Larger values of a lead to more collisions.

(Note that with external chaining, it is possible to have a > 1).

The following table summarizes the performance of our collision resolution techniques in searching for an element. The value in each slot represents the average number of compares necessary for the search. The first column represents the number of compares if the search is ultimately unsuccessful, while the second represents the case when the item is found:

Double hashing is similar, but not so bad, whereas external increases not very rapidly at all (linearly).

In particular, if a = .9, we get

The space requirements (in words of memory) are roughly the same for both techniques:

• TableSize + n (Objectsize) --- open addressing

• TableSize + n (Objectsize + 1) --- external chaining

General rule of thumb: small elts, small load factor, use open addressing.

If large elts then external chaining gives good performance at a small cost in space.