Mining di Dati Web Search Engine s Query Log

Mining di Dati Web Search Engine’s Query Log Mining A. A 2006/2007

What’s 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.

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”

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?

Let’s Start Looking at Some Interesting Items What are the most popular queries?

Most Popular Terms

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.

Why Are Queries so Short? Users minimize effort. Users don’t 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.

Different Kind of Queries

Distribution of Query Types

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.

Frequency Time Distribution

Query Repetition

Query Categories

Categories over Time

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.

Temporal Locality =0. 66

Query Submission Distance

Page Requested

Subsequent Page Requests

Query Caching WSE Francesca, 1 ce Fran sca, 1 Results Francesca Index

Caching: Who Care? !? Successful caching of query results can: Lower the number/cost of query executions. Shorten the engine’s response time. Increase the engine’s throughput.

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.

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.

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.

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 don’t depend on the size of the cache. O(log K). O(N).

Is There Only Caching? No!!!! There’s also PREFETCHING! What’s Prefetching: Anticipating users query by exploiting query stream properties Uhuuuu! Sounds like kind of “Usage Mining”! For instance let’s have a look at the probability of subsequent page requests. Prefetching factor p is the number of pages prefetched.

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.

Adaptive Prefetching

Theoretical Bounds

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.

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!

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

Two-level Caching Cache of Query Results Cache of Inverted Lists Both

Throughput

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 14 th international Conference on World Wide Web (Chiba, Japan, May 10 - 14, 2005). WWW '05. ACM Press, New York, NY, 257 -266.

Probability Driven Caching Lempel, R. and Moran, S. 2003. Predictive caching and prefetching of query results in search engines. In Proceedings of the 12 th 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!

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.

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 fstatic*N. Where 0< fstatic < 1 is the fraction of the total entries (N) of the cache devoted to the Static Set. Adaptive Prefetching is adopted.

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

SDC Performance Linux PC: 2 GHz Pentium Xeon - 1 GB RAM Single process. fstatic = 0. 5. No prefetching.

SDC Hit-Ratio

Why Static Set Helps?

Concurrent Caching

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.

Daily Patterns