Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
CE 221 Data Structures and Algorithms Chapter 5: Hashing Text: Read Weiss, §5.1-5.5 Izmir University of Economics 1 Introduction • Search Tree ADT was disccussed. • Now, Hash Table ADT • Hashing is a technique used for performing insertions and deletions in constant average time. • Thus, findMin, findMax and printAll are not supported. Izmir University of Economics 2 Tree Structures find, insert, delete worst case average case BST N log N AVL log N log N Izmir University of Economics 3 Goal • Develop a structure that will allow users to insert / delete / find records in constant average time – – – – structure will be a table (relatively small) table completely contained in memory implemented by an array capitalizes on ability to access any element of the array in constant time Izmir University of Economics 4 Hash Function • Determines position of a key in the array. • Assume table (array) size is N (TableSize). • Hash function hash(x) maps any key x to an int between 0 and N−1 • For example, assume that N=15, that key x is a non-negative integer between 0 and MAX_INT, and hash function hash(x) = x % 15. Izmir University of Economics 5 Choosing the Hash Functions (1) • If keys are integers key % TableSize ex:all keys=10*i,TableSize=10 So, It’s a good idea to make TableSize a prime • If keys are strings, then ASCII codes of chars keySize1 key[i] % TableSize, when TableSize=10,007 and i 0 keys are at most 8 chars in length (127*8=1,016) • The base 26+1=27 representation of the first 3 letters of the key (key[0] +27* key[1] + 729* key[2] ) % TableSize (263=17,576 but only 2,851 combinations in English) Izmir University of Economics 6 Choosing the Hash Functions (2) • Another hash function involving all characters keySize 1 key[keySize i 1] * 37 i i 0 • compute this by Horner’s Rule Izmir University of Economics 7 Hash Function Let hash(x) = x % 15. Then, if x = 25 129 35 2501 f(x) = 10 9 5 11 47 2 36 6 Storing the keys in the array is straightforward: 0 _ 1 _ 2 47 3 _ 4 _ 5 35 6 36 7 _ 8 9 _ 129 10 11 25 2501 12 _ 13 _ 14 _ • Thus, delete and find can be done in O(1), and also insert, except… Izmir University of Economics 8 Hash Function What happens when you try to insert: x = 65 ? x = 65 hash(x) = 5 0 _ 1 _ 2 47 3 _ 4 _ 5 6 35 36 65(?) 7 _ 8 9 _ 129 10 11 25 2501 12 _ 13 _ 14 _ • If, when an element is inserted, it hashes to the same value as an already inserted element, this is called a collision. Izmir University of Economics 9 Handling Collisions • Separate Chaining • Open Addressing – Linear Probing – Quadratic Probing – Double Hashing Izmir University of Economics 10 Handling Collisions Separate Chaining Izmir University of Economics 11 Separate Chaining • Keep a list of elements that hash to the same value • New elements can be inserted at the front of the list • ex: x=i2 and hash(x)=x%10 Izmir University of Economics 12 Performance of Separate Chaining • Load factor of a hash table, λ • λ = N/M (# of elements in the table/TableSize) • So the average length of list is λ • Search Time = Time to evaluate hash function + the time to traverse the list • Unsuccessful search= λ nodes are examined • Successful search=1 + ½* (N-1)/M (the node searched + half the expected # of other nodes) =1+1/2 *λ • Observation: Table size is not important but load factor is. • For separate chaining make λ 1 Izmir University of Economics 13 Separate Chaining: Disadvantages • Parts of the array might never be used. • As chains get longer, search time increases to O(N) in the worst case. • Constructing new chain nodes is relatively expensive (still constant time, but the constant is high). • Is there a way to use the “unused” space in the array instead of using chains to make more space? Izmir University of Economics 14 Handling Collisions Open Addressing Izmir University of Economics 15 Handling Collisions Linear Probing An alternative to resolving collisions with linked lists is to try alternative cells until an empty cell is found. Alternative Cells are h0(x), h1(x), h2(x),... where hi(x)=(hash(x)+f(i)) % TableSize with f(0)=0, f is the collision resolution strategy. Because all the data go inside the table For Linear probing, λ < 0.5 Izmir University of Economics 16 Linear Probing In linear probing f(i)=i (Popular Choice) Let key x be stored in element hash(x)=t of the array 0 1 2 47 3 4 5 6 35 36 65(?) 7 8 9 129 10 11 25 2501 12 13 14 What do you do in case of a collision? • If the hash table is not full, attempt to store key in the next array element (in this case (t+1)%N, (t+2)%N, (t+3)%N …) until you find an empty slot. Izmir University of Economics 17 Linear Probing Where do you store 65 ? 0 1 2 47 3 4 5 6 7 35 36 65 attempts 8 9 129 Izmir University of Economics 10 11 25 2501 12 13 14 18 Linear Probing Performance (1) • If the table is relatively empty, blocks of occupied cells start forming (primary clustering) • Expected # of probes • for insertions and unsuccessful searches is ½(1+1/(1-λ)2) • for successful searches is ½(1+1/(1-λ)) Izmir University of Economics 19 Linear Probing Performance (2) • Assumptions: If clustering is not a problem, large table, probes are independent of each other • Expected # of probes for unsuccessful searches (=expected # of probes until an empty cell is found) • Expected # of probes for successful searches i * (1 ) * i 1 i 0 1 (fraction of empty cells = 1 -λ) 1 (=expected # of probes when an element was inserted) (=expected # of probes for an unsuccessful search) 1 • Average cost of an insertion I ( ) 1 1 1 1 1 dx ln (earlier insertion are cheaper) 1 x 1 0 Izmir University of Economics 20 Linear Probing • Eliminates need for separate data structures (chains), and the cost of constructing nodes. • Leads to problem of clustering. Elements tend to cluster in dense intervals in the array. • Search efficiency problem remains. • Deletion becomes trickier…. Izmir University of Economics 21 Deletion Problem -- SOLUTION • Standard deletion cannot be performed in a probing hash table, because the cell might have caused a collison to go past it. “Lazy” deletion • Each cell is in one of 3 possible states: – active – empty – deleted • For Find or Delete – only stop search when EMPTY state detected (not DELETED) Izmir University of Economics 22 Deletion-Aware Algorithms • Insert – call Find, if cell = active – if Cell = deleted or empty already there insert, cell = active • Find – cell empty – cell deleted – cell active NOT found if key == key -> NOT FOUND else H = (H + 1) mod TS if key == key -> FOUND else H = (H + 1) mod TS • Delete – – – – call Find, cell active cell deleted cell empty DELETE; cell=deleted NOT found NOT found Izmir University of Economics 23 Handling Collisions Quadratic Probing Izmir University of Economics 24 Quadratic Probing In quadratic probing f(i)=i2 (Popular Choice) Let key x be stored in element hash(x)=t of the array 0 1 2 47 3 4 5 6 35 36 65(?) 7 8 9 129 10 11 25 2501 12 13 14 What do you do in case of a collision? • If the hash table is not full, attempt to store key in array elements (t+12)%N, (t+22)%N, (t+32)%N … until you find an empty slot. Izmir University of Economics 25 Quadratic Probing Where do you store 65 ? hash(65)=t=5 0 1 2 47 3 4 5 6 7 35 36 t t+1 attempts 8 9 129 t+4 Izmir University of Economics 10 11 25 2501 12 13 14 65 t+9 26 Quadratic Probing • Theorem: If quadratic probing is used, and the table size is prime, then a new element can always be inserted if the table is at least half empty. • Proof: Let TableSize be an odd prime > 3, -prove first TableSize / 2 alternative locations are all distinct. -Assume two of these locations are the same and i j . Then, h( x) i 2 h( x) j 2 (mod TableSize) i 2 j 2 (mod TableSize) i 2 j 2 0 (mod TableSize) (i j )(i j ) 0 (mod TableSize) • Since TableSize is prime and i and j are distinct (also less than floor(TableSize)), this is not possible. It follows that the first M/2 alternative are all distinct, and an insertion must succeed if the table is at least half full. Izmir University of Economics 27 Quadratic Probing • Tends to distribute keys better than linear probing, alleviates the problem of clustering (primary clustering. • There remains the problem of secondary clustering in which elements that hash to the same position will probe the same alternative cells • Runs the risk of an infinite loop on insertion, unless precautions are taken. • E.g., consider inserting the key 16 into a table of size 16, with positions 0, 1, 4 and 9 already occupied. • Therefore, table size should be prime. Izmir University of Economics 28 Handling Collisions Double Hashing Izmir University of Economics 29 Double Hashing • f(i)=i*hash2(x) is a popular choice • hash2(x)should never evaluate to zero • Now the increment is a function of the key – The slots visited by the hash function will vary even if the initial slot was the same – Avoids clustering • Theoretically interesting, but in practice slower than quadratic probing, because of the need to evaluate a second hash function. Izmir University of Economics 30 Double Hashing • Typical second hash function hash2(x)=R − ( x % R ) where R is a prime number, R < N Izmir University of Economics 31 Double Hashing Where do you store 99 ? hash(99)=t=9 Let hash2(x) = 11 − (x % 11), hash2(99)=d=11 Note: R=11, N=15 • Attempt to store key in array elements (t+d)%N, (t+2d)%N, (t+3d)%N … Array: 0 14 1 2 16 47 t+22 attempts 3 4 5 35 t+11 6 36 7 65 8 9 129 t 10 11 25 2501 12 99 t+33 13 14 29 Where would you store: 127? Izmir University of Economics 32 Double Hashing Let f2(x)= 11 − (x % 11) hash2(127)=d=5 Array: 0 14 1 16 2 47 t+10 attempts 3 4 5 35 6 36 7 65 t 8 9 129 10 11 25 2501 12 99 t+5 13 14 29 Infinite loop! Izmir University of Economics 33 Rehashing • If the table gets too full, the running times for the operations will start taking too long. • When the load factor exceeds a threshold, double the table size (smallest prime > 2 * old table size). • Rehash each record in the old table into the new table. • Expensive: O(N) work done in copying. • However, if the threshold is large (e.g., ½), then we need to rehash only once per O(N) insertions, so the cost is “amortized” constant-time. Izmir University of Economics 34 Factors affecting efficiency – Choice of hash function – Collision resolution strategy – Load Factor • Hashing offers excellent performance for insertion and retrieval of data. Izmir University of Economics 35 Comparison of Hash Table & BST Average Speed Find Min/Max Items in a range Sorted Input BST O(log2N) Yes Yes Very Bad HashTable O(1) No No No problems • Use HashTable if there is any suspicion of SORTED input & NO ordering information is required. Izmir University of Economics 36 Homework Assignments • 5.1, 5.2, 5.12, 5.14 • You are requested to study and solve the exercises. Note that these are for you to practice only. You are not to deliver the results to me. Izmir University of Economics 37