Apidays New York 2024 - Scaling API-first by Ian Reasor and Radu Cotescu, Adobe
RecSys 2015 posters
1. A Recommendation-Based Book-Exchange
System Without Using Wish Lists
Sole Pera Yiu-Kai Ng
Department of Computer Science Computer Science Department
Boise State University, Boise, Idaho, USA Brigham Young University, Provo, Utah, USA
EasyEx
NOVELTY
METHOD INITIAL EXPERIMENTS
FUTURE WORK
Is a unique recommendation-based book exchange system based on users’ item lists.
Aims to enhance existing book exchange sites by simplifying their process.
• Identifies books that are appealing but unfamiliar to users.
• Is not constrained by users’ wishlists.
• Optimizes book-exchange cycles to increase serendipity and generate unanticipated exchanges.
• Creates multiple exchanges cycles involving more than two users at a time.
Data Set:
Assessment goal:
• Verify the correctness of the recommendation
and optimization strategies.
• Measure the effectiveness as a
recommendation-based book-exchange system.
Evaluation of EasyEx and other recommendation strategies
Evaluation of optimization solutions generated by OptaPlanner
1. Consider groups of users with diverse degrees of cohesiveness.
2. Enhance EasyEx to alert user if an exchange is not possible.
3. Exchange of other types of resources.
4. Develop applications (beyond online exchanges) to handle K-12 readers.
2. Exploiting Reviews to Guide Users’ Selections
Nevena Dragovic & Sole Pera
nevenadragovic@u.boisestate.edu solepera@boisestate.edu
Department of Computer Science, Boise State University, Boise, Idaho, USA
Honest recommendation system (HRS)
• Takes advantage of ratings and reviews generated by a user to learn the item characteristics that
are most likely appealing to the user.
• Creates personalized suggestions with corresponding explanations.
HRS
RECOMMENDATION PROCESS OF HRS
Evaluation Data Results
Software domain of Amazon Dataset
EMPIRICAL STUDIES
• Extend examination and comparisons with other baseline and state-of-the-art strategies on
products/services in other domains.
• Conduct online user studies to verify the fact that HRS helps users in making best choices.
• Perform more in-depth analysis on part-of-speech and type dependencies on sentences in reviews.
FUTURE WORK
Metric Matrix Factorization HRS
NDCG 0.704 0.748
Item characteristics preferred by a user are employed for evaluating and re-ordering candidate items,
as well as generating explanations pertaining to the user’s interests. A generated explanation for a
recommended item:
• Showcases why a particular item was recommended.
• Helps users decide which items, among the ones recommended, are best tailored towards their
individual interests.
UNIQUENESS OF HRS
3. Recommendations to Enhance Children Web Searches
UNIQUENESS OF KIDSQR
RECOMMENDATION PROCESS OF KIDSQR
• Considers natural language and informal phrasing patterns
based on children’s writings rather than merely relying on data
produced by adults.
• Does not use query logs, hence the recommendations are not
constrained by the queries frequently posted by general users.
Generate a
single ranking
score for C Top 3 query
recommendations
EMPIRICAL STUDIES
Automatic query suggestion
generator (E.g., Ubersuggest)
Vocabulary
(children
dictionary, school
vocabulary lists)
Popularity
(Children stories,
poems, blog
posts)
Phrase-
Formulating
(Children stories,
poems, blog posts,
online reviews)
Pop-Culture
(Children movies, songs, toys
and games, actors, authors,
book titles, cartoons and
shows, characters, dolls and
action figures, poems,
science)
Term Examination Phrase Examination
Multiple regression-based
score generator
Evaluation Data Results
Performance evaluation based on 80 appraisements
• User study based on elementary
school teacher and parent.
• 10 appraisers, each evaluated 8
distinct queries.
• Makes recommendations using child-friendly characteristics,
including children pop-culture, i.e., entities children are familiar
with , and children’s topics of interest.
Q
CC
Shahrzad Karimi
Department of Computer Science
Boise State University, Boise, ID 83725 USA
Sole Pera
Department of Computer Science
Boise State University, Boise, ID 83725 USA
Identify each
candidate query
suggestion C
Generate Candidates
Determine the adequacy of
the candidates, i.e.
distinguish child-friendly
from non child-friendly
suggestions
CHILDREN QUERIES
Children face different challenges in formulating well-defined queries that lead to retrieving relevant information, because of their:
• Limited vocabulary.
• Tendency to use natural language constructions and complex phrasing.
• Lack of ability to utilize keywords in phrase formulations.
QR1
QR2
QR3
?
Initial query Q of user U
0.35
0.51
0.72
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
nDCG
Google Bing KidsQR
Google Bing KidsQR
MRR 0.27 0.36 0.7
Analyze Candidates Rank Candidates