Clustering and Load Balancing Optimization for Redundant Content Removal

Clustering and Load Balancing Optimization for Redundant Content Removal

Shanzhong Zhu (Ask.com)

Alexandra Potapova, Maha Alabduljalil (Univ. of California at Santa Barbara)

Xin Liu (Amazon.com)

Tao Yang (Univ. of California at Santa Barbara)

Redundant Content Removal in Search Engines

• Over 1/3 of Web pages crawled are near duplicates

• When to remove near duplicates? Offline removal

Online removal with query-based duplicate removal

Online index matching & result ranking

Duplicate removalUser

query

Final results

Offline data processing

Duplicate filtering

WebPages

Online index

Tradeoff of online vs. offline removal

Online-dominating approach

Offline-dominating approach

Impact to offline High precisionLow recall

Remove fewer duplicates

High precisionHigh recall

Remove most of duplicates

Higher offline burden

Impact to online More burden to online deduplication

Less burden to online deduplication

Impact to overall cost

Higher serving cost Lower serving cost

Challenges &issues in offline duplicate handling

• Achieve high-recall with high precision All-to-all duplicate comparison for complex/deep

pairwise analysis Expensive parallelism management &

unnecessary computation elimination• Maintain duplicate groups instead of duplicate

pairs Reduce storage requirement. Aid winner selection for duplicate removal Continuous group update is expensive.

Approximation. Error handling

Optimization for faster offline duplicate handling

• Incremental duplicate clustering and group management Approximated transitive relationship Lazy update

• Avoid unnecessary computation while balancing computation among machines Multi-dimensional partitioning Faster many-to-all duplicate comparisons

Pagepartition …

Pagepartition

Pagepartition

Pagepartition

Two-tier Architecture for Incremental Duplicate Detection

Approximation in Incremental Duplicate Group Management

• Example of incremental group merging/splitting

• Approximation Group is unchanged when updated pages are still similar to

group signatures Group splitting does not re-validate all relations

• Error of transitive relation after content update A<->B, B<-> C A<->C A <->C may not be true if content B is updated.

• Error prevention during duplicate filtering: double check similarity threshold between a winner and a

loser

Multi-dimensional page partitioning

• Objective One page is mapped to one unique partition Dissimilar pages are mapped to different partitions. Reduce unnecessary cross-partition comparisons.

• Partitioning based on document length Outperform signature-based mapping for higher

recall rates.• Multi-dimensional mapping

Improve load imbalance caused by skewed length distribution

Pages Pages Pages…

Multi-dimensional page partitioning

Dictionary

Sub-dictionary

Sub-dictionary

A=(600) A=(280,320)

1D length space

2D length space

When does Page A compare with B?

• Page length vector A= (A1, A2) , B=(B1,B2)

• Page A needs to be compared with B only if

• τ is the similarity threshold• ρ is a fixed interval enlarging factor

Implementation and Evaluations

• Implemented in Ask.com offline platform with C++ for processing billions of documents

• Impact on relevancy Continuously monitor top query results. Error rate of false removal is tiny.

• Impact on cost. Compare two approaches

– A: Online dominating. Offline removes 5% duplicates first. Most of duplicates hosted in online tier-2 machines

– B: Offline dominating.

Cost Saving with Offline Dominating Approach

• Fixed QPS target. Two-tier online index for 3-8 billion URLs.

• 8%-26% cost saving with offline dominating Less tier-2 machines due to less duplicates hosted. Online tier 1 machines can answer more queries Online messages communicated contain less duplicates

Reduction of unnecessary inter-machine communiation & comparison

Up to 87% saving when using up to 64 machines

Effectiveness of 3D mapping

• Load balance factor with upto 64 machines

• Speedup of processing throughput

Benefits of incremental computation

• Ratio of non-incremental duplicate detection time over incremental one for a 100 million dataset. Upto 24-fold speedup.

• During a crawling update, 30% of updated pages havesignatures similar to group signatures

Accuracy of distributed clustering and duplicate group management

Relative error in precision compared to a single-machine configuration

Relative error in recall

Conclusion remarks

• Budget-conscious solution with offline dominating redundant removal Up to 26% cost saving.• Approximated incremental scheme for

duplicate clustering with error handling Upto 24-fold speedup Undetected duplicates are handled online.

• 3D mapping still reduces unnecessary comparisons (upto 87%) while balancing load (3+ fold improvement)