Appendix

1
Appendix
2
Main Variables in Adaptive Information Retrieval

• Query
• User
• Time of request
• Information sources

3
1st-Order Markov Model

• Current action depends on the history H(U) only
  through the last observed action A
• ?0,k is the probability of observing H(U) as the
  first action in the history with mixture
  component k,
• ?(h-gth1),k is the probability of observing a
  transition from action h to h1 in the history
  for that component. Note that Markov model only
  depends on the bigrams.

4
Maximum Entropy Model Definition

• Maximum entropy objective function leads to the
  following model

• Fs(A,H(U)) are the feature indicator functions. S
  is the total number of bigramstriggers for
  action A.
• s are maxent model parameters for component k.
• If s corresponds to bigram (a,b) then
  Fs(A,H(U))1 iff Aa and Aprevb, and
  Fs(A,H(U))0 otherwise.
• Z?,k(H(U)) is a normalization constant.

5
Experimental Setup

• 3 months of CiteSeer data was gathered and
  sessionized based on the inactivity time
  criterion (10 minutes)
• Robots were identified and eliminated based on
  the number of CiteSeer accesses
• Queries were extracted from the sessionized data
• The type of queries were identified by the
  following procedure
• 1st Query in a session is always a new query
• Current query Qcurrent is a new query if it
  doesnt have any common terms with the previous
  query Qprev (excluding stop words and logical
  operators)
• Qcurrent is an old query if its exactly the same
  as Qprev
• Qcurrent is a modified query if it has at least
  one common term with Qprev
• Qcurrent is a document query if the documents
  were searched, and citation query if the
  citations were being searched.

6
CiteSeer Users Document Request Patterns

• Document request logs are valuable because they
• bear contextual information
• facilitate finding similar users, similar
  documents
• indicate users accustomed search actions.
• How frequently does the user request new
  documents?
• Does (s)he ever request documents (s)he has seen
  in previous sessions?
• When the user views a new document is (s)he done
  with it in 1 request, or does (s)he request the
  new document frequently in that session?

7
Document Request Patterns - Motivation

• The object of this study is the 3rd point which
  we believe could play and important role in
  recommendation strategies and user interface
• If the user hardly requests documents viewed in
  old sessions should we bother recommending them
  at all?
• If the user searches citations mostly should we
  make that users default search citation
  search?
• If the user visits the new found documents in
  that session frequently, should we have links to
  those documents on every page during that
  session?
• If the user requests old documents more
  frequently what is the reason? Are we showing the
  same results? Should the new documents have a
  higher priority in recommendations to this user?

8
Document Request Patterns

• We identify the types of sessions rather than
  just identifying only each document request type.
  Its straightforward to cluster the users once we
  identify the sessions.
• Document Request Types
• New Document Users first request of this
  document
• This Session Document Document was requested
  earlier in this session
• Old Document User had requested this document in
  previous sessions
• Session Types
• New Documents Session All documents requested
  are new documents
• Old Documents Session All documents requested in
  this session were visited in previous sessions
• New And This Session Documents Session Documents
  requested were never viewed in previous sessions,
  but at least one of the new documents were
  requested more than once in this session
• New And Old Documents Session Documents were
  either old or new
• All Document Types Session This session has at
  least two requests to a new document and at least
  one to an old document.

9
Session Types - Example

• Input
• ltsessionID userID documentID1 documentID2 gt
• 1 uid1 475572 475572
• 2 uid1 102844 102844
• 3 uid1 23039 23039 263905 263905 430232 34069
  34069 123882 87389 87389 470739 89795 89795
  317832 18670 506897 94170 506428 506428 501759
  286736 286736 99777 99777 40383 391244
• 4 uid1 209066 209066
• 5 uid1 84571
• 6 uid1 496615 496615
• 7 uid1 525188 501003
• 8 uid1 501621 501003
• 9 uid1 186821 186821
• 10 uid1 88085 88085
• 11 uid1 15714 15714
• 12 uid1 25344 25344
• Output
• ltuid session1Type session2Type gt
• uid1 New New New New NewANDOld New New New New
• (first 3 sessions were excluded therefore
  sessions 4 through 12 are in the final output)

10
Document Request Behavior - Results
Only New Doc Requesters
11
Conclusions Future Work

• Users differ in their use of CiteSeer in every
  way
• Some view only new documents, some hardly ever do
  so, whereas some like to spend more time on new
  found documents
• Some search only documents, some search citations
  mostly, and some relentlessly try to modify their
  queries to find the right documents.
• This study has shown that by looking at the query
  formulations and the type of the documents
  requested we can identify bottlenecks for
  different user groups
• Our goal is to use the proposed methods to
  provide personalized recommendations to CiteSeer
  users, as well as automatically adapting the
  design of the user interface to better suit their
  needs