Evaluating Matching Algorithm on Marketplace Catalogues

1
Evaluating Matching Algorithm on Marketplace
Catalogues
Antonia Donà Elena Andretta - Laura
Gatti Alessandra Molani Luciano Serafini
2
Marketplaces

• A marketplace is a conceptual place where
  buyers and sellers meet in order to exchange
  products and services
• Sellers describes products and services they
  provide.
• Buyers search for products and services that feet
  their needs.

MARKETPLACE
BUYERS
SELLERS

• Problem
• How can buyers and sellers be able to
  communicate? How can buyers search for
  products/services on sellers products/services
  descriptions?

3
The communication problem

• Communication problems are actually solved using
  Standard Classification
• The marketplace manager supplies a common
  classification and sellers and buyers who want to
  join the marketplace must map their
  classifications to the standard.

• There is more than one standard classification
• standard classification are not the right
  solution to the communication problem
• our matching can be a solution!

4
Standard industrial classification

• How does standard industrial classification work?
• A standard classification is a hierarchy of a
  fixed number of levels
• Each level contains a numerical code and a
  textual description
• Products and services are linked to the hierarchy

5
From classifications to contexts

• English
• Kompass
• 71 contexts, 66.000 concepts
• Sic
• Standard Industrial Classification
• 10 contexts, 2.300 concepts
• Naics
• North American Industry Classification System
• 20 contexts , 2.300 concepts
• Ecl_at_ss
• Developed by some German companies
• 22 contexts, 15.300 concepts
• Un/Spsc
• United Nations Standard Products and Services
  Classification
• 57 contexts, 21.500 concepts
• Italian
• Kompass
• 71 contexts, 66.000 concepts

For a total of 321 Contexts!

• Main characteristics
• From 3 to 5 levels
• Concepts labels with more than 5 words

6
Just a feel..
7
Tables of mapping

• A table of mapping is a list of pairs identifying
  equivalent classes of products belonging to two
  different marketplace classifications.
• Tables of mapping are produced manually as
  solution to the problem of having so many
  different standards!

8
Tables of mapping

• English
• Sic ? Naics

• Italian
• Ateco ? Sic

Professional, Scientific, and Technical
Services
Health Care and Social Assistance
Educational Services
Service Industries
9
Testing based on tables of mapping

• Our testing methodology
• Run the algorithm on all different pairs of
  contexts (e.g. SIC.ctx1 and NAICS.ctx2)
• The output of the matching algorithm has the
  following form
• ctx1.n1 Is Equivalent ctx2.n2
• ctx1.n1 Is Less General ctx2.n2
• ctx1.n1 Is More General ctx2.n2
• ctx1.n1 Has Opposite Meaning of ctx2.n2
• Compare the output of the algorithm (ctx1.n1 Rel
  ctx2.n2) with the content of tables of mapping
  (ctx1.n1 ctx2.n2)
• ctx1.n1 ctx2.n2 AND ctx1.n1 Is Equivalent
  ctx2.n2 -gt OK
• ctx1.n1 ctx2.n2 AND ctx1.n1 Is Less General
  ctx2.n2 -gt OK
• ctx1.n1 ctx2.n2 AND ctx1.n1 Is More General
  ctx2.n2 -gt OK

10
Testing based on user satisfaction

• Our testing methodology
• We will contact a group of buyers and sellers.
• We will collect their classification structures
  (catalogue, price list ) focused on a particular
  domain, and, if possible we will ask for product
  cards to classified under the collected stucures.
• We will run the algorithm on all pairs of
  collected structures. The output will be
  interpreted as follows given two structures
  (ctx1 and ctx2), how to re-classify the product
  cards classified on ctx1 under ctx2.
• We will run the algorithm for comparing the
  collected classification structures with the
  standard ones (Sic, Kompass, Ecl_at_ss, .)
• We will ask feedback about our results.

11
Conclusions

• We have run a first step of the experimentation
  using tables of mappings, and we have use the
  results to debug our matching code.
• We are now working on bug fixing.
• Next step of experimentation will be again on
  tables of mappings, trying to collect information
  on the validity of our matching.
• In parallel we will start to collect structures
  from kompass customers.