Search & Recommendation: Birds of a Feather?


Published on

In just a little over half a century, the field of information retrieval has experienced spectacular growth and success, with IR applications such as search engines becoming a billion-dollar industry in the past decades. Recommender systems have seen an even more meteoric rise to success with wide-scale application by companies like Amazon, Facebook, and Netflix. But are search and recommendation really two different fields of research that address different problems with different sets of algorithms in papers published at distinct conferences?

In my talk, I want to argue that search and recommendation are more similar than they have been treated in the past decade. By looking more closely at the tasks and problems that search and recommendation try to solve, at the algorithms used to solve these problems and at the way their performance is evaluated, I want to show that there is no clear black and white division between the two. Instead, search and recommendation are part of a much more fluid continuum of methods and techniques for information access.

(Keynote at "Mind The Gap '14" workshop at the iConference 2014 in Berlin, Germany)

Published in: Technology
  • Be the first to comment

No Downloads
Total views
On SlideShare
From Embeds
Number of Embeds
Embeds 0
No embeds

No notes for slide

Search & Recommendation: Birds of a Feather?

  1. 1. Search & Recommendation: Birds of a Feather? Toine Bogers Aalborg University Copenhagen Copenhagen, Denmark ‘Mind the Gap ’14’ workshop @ iConference 2014, Berlin March 4, 2014
  2. 2. Outline • Introduction • Search vs. recommendation - Use case - Algorithms & evaluation - Trends • Conclusions 2
  3. 3. 3
  4. 4. Success of search engines • Search engines have had a huge impact on the information economy - Academia ‣ Vibrant & growing research community with many dedicated conferences and journals ‣ Evaluation initiatives like TREC were shown to have a great impact on the performance of Web search engines - Industry ‣ Google → ~13 billion USD in profit in 2013 4
  5. 5.
  6. 6. Success of recommender systems • Recommender systems have seen a meteoric rise to success in past two decades - Academia ‣ From specialized workshops to dedicated conference and journals - Industry ‣ Amazon → 35% sales from recommendations ‣ Netflix → 75% of what its users watch comes from recommendations ‣ Google News → recommendations generate 38% more click-through 6
  7. 7. Different perspectives? • Search and recommendation are commonly treated as different (but related) research areas • Search perspective → recommendation is a special type of search problem - Smaller research community with few dedicated venues - Dedicated sessions at major IR conferences • Recommendation perspective → field of its own - Rapidly growing research community with s - Strong industry support - Separate data sets, experimental protocol, and evaluation 7
  8. 8. ....but are they really that different? • Looking at search and recommendation in isolation can be counter-productive in many situations! • Three aspects of where both fields are (growing) close(r) - Use cases - Algorithms & evaluation - Trends 8
  9. 9. Use cases
  10. 10. Comparing use cases • What are the characteristics of the information access paradigms? - What problem are they trying to solve? - What do we know about what the user wants? - What do we know about the user? - How do we know we have solved the user’s problem? 10
  11. 11. Comparing definitions tions “A recommender system is software that provides sugges to users on which items could be of use to them.” — Ricci et al. (2011) retrieval (IR) is finding material of an “Information ation ructured nature that satisfies an inform unst need from within large collections.” — Manning et al. (2008) 11
  12. 12. Search characteristics • Information need - Explicit representation of user’s information need as a query (and occasionally a description or narrative) ‣ Typically at Taylor’s last two stages Taylor’s four stages 1. Visceral 2. Conscious 3. Formalized 4. Compromised • Knowledge about the user - User characteristics traditionally abstracted away - More focus on the user in recent years (e.g., search history) 12
  13. 13. Search characteristics • Matching criteria - Relevance ‣ Assessment of perceived topicality, pertinence, usefulness or utility of an information source by an actor or algorithm with reference to a task at a given point in time - Relevance is a multi-dimensional concept → many different flavors! ‣ Topical relevance most common interpretation ‣ Textual similarity used as a proxy for topical relevance Saracevic’s categories • Algorithmic relevance • Topical relevance • Temporal relevance • Situational relevance - See Borlund (2003) for a comprehensive overview of relevance in IR 13
  14. 14. Recommendation characteristics • Information need - Implicit representation of user’s information need as a the user’s profile ‣ Typically at Taylor’s first two stages • Knowledge about the user Taylor’s four stages 1. Visceral 2. Conscious 3. Formalized 4. Compromised - User profile representing the user’s interests - Usage patterns, past interactions with the system, requirements • Matching criteria - Interest / Usefulness 14
  15. 15. No user profile Classic IR Explicit need Web search PopularityBrowsing based methods Implicit need Information filtering ? Recommendation User profile 15
  16. 16. Search & recommendation form a continuum • Search (“Show me all books about X”) • Focused recommendation (“Show me interesting books about X!”) • Recommendation (“Show me interesting books!”) Search Focused recommendation Recommendation 16
  17. 17. LibraryThing forum topic
  18. 18. How prevalent is focused recommendation? • Is there evidence for such a continuum? - Search engines see millions of pure search requests every day - Netflix and Amazon profit immensely from pure recommendation scenarios - But how prevalent are these focused recommendation requests? • Possible explanations for underrepresentation - Perhaps we are looking in the wrong places? - Interfaces offer little support for entering complex requests 18
  19. 19. INEX Social Book Search track • Track running at INEX from 2011-2014 on book search - Amazon/LibraryThing collection ‣ 2.8 million book metadata records ‣ Mix of metadata from Amazon, Librarything, Library of Congress, and British Library - Realistic book requests & information needs from LibraryThing fora ‣ Highly varied set of requests that touch upon topics, genres, authors, engagment, reading level, personal preferences, etc. 19
  20. 20. Topic title Annotated LT topic Narrative Group name Recommended books 20
  21. 21. INEX Social Book Search track • Track running at INEX from 2011-2014 on book search - Amazon/LibraryThing collection ‣ 2.8 million book metadata records ‣ Mix of metadata from Amazon, Librarything, Library of Congress, and British Library - Realistic book requests & information needs from LibraryThing fora ‣ Highly varied set of requests that touch upon topics, genres, authors, engagment, reading level, personal preferences, etc. ‣ Collected & annotated 944 book requests from the LibraryThing fora - Relevance judgments ‣ Member suggestions (Suggestions made by other Librarything members) ‣ Reading behavior (Has the original requester added any suggestions afterwards?) 21
  22. 22. Relevance aspects of book requests • Eight LIS students annotated all requests on relevance aspects Relevance aspects % Accessibility 16 Content 74 Engagement 23 Familiarity 36 Known-item 21 Metadata 28 Novelty 4 Socio-cultural 14 0 10 20 30 40 50 60 70 80 22
  23. 23. Continuum of search & recommendation • How common are the different types of information needs? Familiarity No familiarity Content Focused recommendation (260 requests) Search (338 topics) No content Recommendation (66 topics) Context (78 topics) Sign up at https://inex.mmci.uni-saarland .de/tracks/books/! 23
  24. 24. Not just true for the book domain! 24
  25. 25. What’s next? • Focused recommendation deserves more attention! - Combines aspects of search and recommendation • Open questions - How can we best address focused recommendation requests? ‣ Likely to require a combination of both search and recommendation approaches ‣ Early indications from INEX track that a combination indeed works best - How can we support expressing these complex needs through the UI? 25
  26. 26. Algorithms & evaluation
  27. 27. Algorithms & evaluation • Past decade has seen combination & mutual inspiration - Both fields have borrowed techniques & metrics from each other - Dedicated workshops & events ‣ CARR 2011-2014 ‣ BARR 2013 ‣ Mind The Gap 2014 27
  28. 28. Recommender systems → IR • Collaborative filtering - Automates the process of word-of-mouth recommendations by looking for unseen items among other users with similar interests • Used in IR for - Collaborative search ‣ I-SPY search engine by Smyth et al. (2004) - Query suggestion - Improving 'More like this' functionality 28
  29. 29. IR → Recommender systems • Recommender systems has borrowed from many different fields - Artificial Intelligence (ML, CBR), IR, Natural Language Processing • Inspiration from IR - Algorithms ‣ TF·IDF weighting scheme for CF (Breese et al., 1998) ‣ Query expansion for recommender systems (Formosa et al., 2013) ‣ Probability ranking principle in recommender systems (Wang et al., 2006) ‣ Language modeling for recommender systems (Bellojin et al., 2013) - Evaluation ‣ Increasing use of nDCG (and MAP) as metrics for ranked list recommendation 29
  30. 30. Trends
  31. 31. Context • Incorporating contextual information into the search/ recommendation process • Search - IRiX workshop (2004-2005) - CARR workshop (2011-2014) - TREC Contextual Suggestion (2012-2013) • Recommendation - CARS workshop (2009-2012) - CAMRA workshop (2010-2011) - CARR workshop (2011-2014) 31
  32. 32. Diversity • Ensuring a diverse range of relevant results/recommendations • Search - IDR workshop (2009) - DDR workshop (2011-2012) - Many publications addressing diversity in search results • Recommendation - DiveRS workshop (2011) - Many publications addressing diversity in recommender systems 32
  33. 33. Privacy • Protecting user privacy when generating results/recommendations or releasing data sets - Hot topic in the aftermath of release of AOL and Netflix data sets - Many papers on how to (de-)anonymize of recommendation data sets and search logs • Search - PIR workshop (2014) • Recommendation - RESSON workshop (2013) 33
  34. 34. Conclusions
  35. 35. Conclusions • Search & recommendation form an information access continuum - Pure search & recommendation needs are addressed well by the respective research fields - But many other information needs fall through the cracks! ‣ Need to look at the whole range of information needs ‣ Both in terms of algorithms and interface design • Search & recommendation are already moving closer together - Exchange of algorithms & techniques - Shared evaluation metrics - Similar research trends • A continuum of requests requires a continuum of solutions! 35
  36. 36. Questions? Comments? Suggestions? 36
  37. 37. Backup slides 37
  38. 38. Example requests CONTEXT REQUEST I've just finished my undergraduate work, and as I float into the real world, I find myself missing books-and recommendations for books--in a serious way. So, those of you in a similar state (and those of you who simply love reading, and sharing): have any books that you find essential for living? I'll post what I've been reading, and you can as well...I'm especially interested in books that are a little older, a little less known, and more prone to flying under the radar. I read almost everything as well, a sentiment I'm sure most of you are familiar with. SEARCH REQUEST looking for heroine oriented love triangle romances, any recommendations appreciated. RECOMMENDATION REQUEST Just read and reviewed Moon in the Water: Reflections on an Aging Parent. I wonder if other early readers have recommendations for similar pieces...this makes me want to go back and read The Summer of the Great-Grandmother by Madeline L'Engle. I glossed through it the first time, but now that I am closer to that stage of life I wonder if it will have more meaning. 38