Mining%20di%20Dati%20Web

1
Mining di Dati Web

• Web Search Engines Query Log Mining
• A.A 2006/2007

2
Whats Recorded in a WSE Query Log?

• Each component of a WSE records information about
  its operations.
• We are mainly concerned with frontend logs.
• They record each query submitted to the WSE.

3
Data Recorded

• Among other information WSEs record
• The query topic.
• The first result wanted.
• The number of results wanted.
• Some examples
• q(fabrizio silvestri)f(1)n(10)
• q(information retrieval)f(5)n(15)
• Some other information
• The language.
• Results folded? (Y/N).
• Etc.

Commonly referred to as the query
4
What Can We Look For?

• The most popular queries.
• How queries are distributed.
• How queries are related.
• How subsequent queries are related.
• How topics are distributed.
• How topics change throughout the 24 hours.
• Can we exploit this information?

5
Lets Start Looking at Some Interesting Items

• What are the most popular queries?

6
Most Popular Topics
7
Most Popular Terms
8
What Are Users Doing?

• Not typing many words!
• Average query was 2.6 words long (in 2001), up
  from 2.4 words in 1997.
• Moving toward e-commerce
• Less sex (down from 17 to 9), more business (up
  from 13 to 25).
• Spink A., et al. From e-Sex to e-Commerce Web
  Search Changes, Computer, March 2002.

9
Why Are Queries so Short?

• Users minimize effort.
• Users dont realize more information is better.
• Users learn that too many words belongs to fewer
  results. (Since implicit AND)
• Query Boxes are Small.
• Belkin, N.J., et al. Rutgers TREC 2001
  Interactive Track Experience, in Voorhees
  Harmon, The Tenth Text Retrieval Conference.

10
Different Kind of Queries
11
Distribution of Query Types
12
Hourly Analysis of a Query Log

• Steven M. Beitzel, Eric C. Jensen, Abdur
  Chowdhury, David Grossman, Ophir Frieder, "Hourly
  Analysis of a Very Large Topically Categorized
  Web Query Log", Proceedings of the 2004 ACM
  Conference on Research and Development in
  Information Retrieval (ACM-SIGIR), Sheffield, UK,
  July 2004.

13
Frequency Time Distribution
14
Query Repetition
15
Query Categories
16
Categories over Time
17
Analysis of Three Query Logs

• Tiziano Fagni, Salvatore Orlando, Raffaele
  Perego, Fabrizio Silvestri. Boosting the
  Performance of Web Search Engines Caching and
  Prefetching Query Results by Exploiting
  Historical Usage Data. ACM Transactions on
  Information Systems. 24(1). January 2006.

18
Temporal Locality
?0.66
19
Query Submission Distance
20
Page Requested
21
Subsequent Page Requests
22
Query Caching
Index
WSE
Francesca, 1
Francesca, 1
Results
Francesca
23
Caching Who Care?!?

• Successful caching of query results can
• Lower the number/cost of query executions.
• Shorten the engines response time.
• Increase the engines throughput.

24
Caching How-To?

• Caching can exploit locality of reference in the
  query streams search engines are faced with.
• Query popularity follows a power-law and vary
  widely, from the extremely popular to the very
  rare.

25
Caching What to Measure?

• Hit Ratio
• Let N be the number of requests to the WSE
• Let H be the number of hits - i.e. the number of
  queries that can be answered by the cache.
• The Hit Ratio HR is defined as HR H/N. Usually
  is expressed in percentage.
• E.g. HR 30 means that the thirty percent of
  the queries are satisfied using the cache.
• Alternatively we could define the Miss Ratio MR
  1 - HR M/N. Where M is the number of miss -
  i.e. the number of queries that cannot be
  answered by the query.

26
What About the Throughput?

• The throughput is defined as the number of
  queries answered per-second.
• Caching, in general, rises the throughput.
• The lower the hit-ratio the lower the throughput.
• The lower the cache response-time the higher the
  throughput.

27
Caching Complexity

• The caching response time depends on the
  replacement policy complexity.
• The complexity usually depends on the cache size
  K.
• There exists policies that are
• O(1) - i.e. constant. They dont depend on the
  size of the cache.
• O(log K).
• O(N).

28
Is There Only Caching?

• No!!!!
• Theres also PREFETCHING!
• Whats Prefetching
• Anticipating users query by exploiting query
  stream properties
• Uhuuuu! Sounds like kind of Usage Mining!
• For instance lets have a look at the probability
  of subsequent page requests.
• Prefetching factor p is the number of pages
  prefetched.

29
Prefetching PROS and CONS

• Prefetching enhance hit-ratio.
• Prefetching reduce the query load on the query
  server.
• The cost for computing p pages of results is
  approx the same of computing only one page
• Prefetching is very likely to load pages that
  will never be requested in future.

30
Adaptive Prefetching
31
Theoretical Bounds
32
Some Classical Caching Policies

• LRU
• Last Recently Used.
• Evict from Cache the query results that have been
  accessed farthest in the past.
• SLRU
• Two segments
• Probationary
• Protected.
• Lines in each segment are ordered from the most
  to the least recently accessed. Data from misses
  is added to the cache at the most recently
  accessed end of the probationary segment. Hits
  are removed from wherever they currently reside
  and added to the most recently accessed end of
  the protected segment. Lines in the protected
  segment have thus been accessed at least twice.
  The protected segment is finite, so migration of
  a line from the probationary segment to the
  protected segment may force the migration of the
  LRU line in the protected segment to the most
  recently used (MRU) end of the probationary
  segment, giving this line another chance to be
  accessed before being replaced.

33
Problems

• Classical Replacement Policies do not care about
  stream characteristics.
• They are not designed using usage mining
  investigation techniques.
• They offer godd performance, though!
• Uhmmm. Are you sure?!? Stay tuned!

34
Caching May be Attacked from two Directions

• Architecture of the caching system
• Two-level caching
• Three-level caching
• SDC
• Replacement policy
• PDC
• SDC
• Both
• SDC

35
Two-level Caching

• Cache of Query Results
• Cache of Inverted Lists
• Both

36
Throughput
37
Three-level Caching

• Long, X. and Suel, T. 2005. Three-level caching
  for efficient query processing in large Web
  search engines. In Proceedings of the 14th
  international Conference on World Wide Web
  (Chiba, Japan, May 10 - 14, 2005). WWW '05. ACM
  Press, New York, NY, 257-266.

38
Probability Driven Caching

• Lempel, R. and Moran, S. 2003. Predictive caching
  and prefetching of query results in search
  engines. In Proceedings of the 12th international
  Conference on World Wide Web (Budapest, Hungary,
  May 20 - 24, 2003). WWW '03. ACM Press, New York,
  NY, 19-28.
• Tanks to Ronny for his original slides!

39
Static-Dynamic Caching

• Tiziano Fagni, Salvatore Orlando, Raffaele
  Perego, Fabrizio Silvestri. Boosting the
  Performance of Web Search Engines Caching and
  Prefetching Query Results by Exploiting
  Historical Usage Data. ACM Transactions on
  Information Systems. 24(1). January 2006.
• Idea
• Divide the cache in two sets
• Static Set
• Dynamic Set.
• Fill the Static Set using the most frequently
  submitted query in the past.
• The Static Set is read-only good in
  multithreaded architectures.

40
Inside SDC

• Static-Dynamic Caching.
• The cache is divided into two sets
• Static Set contains the results of the queries
  most frequently submitted so far.
• Dynamic Set is implemented using a classical
  caching replacement policy like, for instance,
  LRU, SLRU, PDC.
• The Static Set size is given by fstaticN. Where
  0lt fstatic lt 1 is the fraction of the total
  entries (N) of the cache devoted to the Static
  Set.
• Adaptive Prefetching is adopted.

41
Benefits in Real-World Caches
SDC Cache
SDC Cache Thread
SDC Cache Thread
SDC Cache Thread
SDC Cache Thread
Static Set
Dynamic Set
Mutex
WSE
42
SDC Performance

• Linux PC 2GHz Pentium Xeon - 1GB RAM
• Single process.
• fstatic 0.5. No prefetching.

43
SDC Hit-Ratio
44
SDC Hit-Ratio
45
SDC Hit-Ratio
46
SDC Hit-Ratio
47
SDC Hit-Ratio
48
SDC Hit-Ratio
49
SDC Hit-Ratio
50
SDC Hit-Ratio
51
Why Static Set Helps?
52
Concurrent Caching
53
Freshness of the Training Data

• How frequently should we perform mining again on
  the usage data?
• Does performance of Usage-Mining-based caching
  degrades gracefully as time goes by?
• Do time-of-day patterns exist in query stream.

54
Daily Patterns