W10 Web Search & Recommender System Flashcards
what is the CPC model?
the cost per click model: advertisers pay the search enginge and get clicks in return (goal: induce transaction)
what is anchor text?
a descriptive text of the URL that the hyperlink points to
This information can be used to bridge the vocabulary gap between query and document: anchor texts may contain query terms that are not in the document itself
2 intuitions about hyperlinks
- the anchor text pointing to page B is a good description of page B (textual information)
- the hyperlink from A to B represents endorsement of page B, by the creator of page A (quality signal)
both signals contain noise
what is PageRank?
technique for link analysis that assigns to every node in
the web graph a numerical score between 0 and 1
main intuition: pages visited more frequently in a random walk on the web are the more important pages
3 PageRank intuitions
- incoming link counts are an important signal: a page is useful if it is cited often
- indirect citations also count: if important pages are pointing to a page, the page must be important
- smoothing: mooth citations with some random step to accommodate potential citations that have not yet been observed
how to estimate PageRank scores?
- start at a random page
- jump to another page: with probability alpha to a random page, with probability 1 - alpha to any outgoing link
- repeat step 2 until convergence of the scores
- final score is the probability that the surfer reaches the page
why do we want query result diversification?
queries are often short and ambiguous: we don’t know what the user wants
if we take query-document similarity as the most important
ranking criterion, there might be a lot of redundance in the top-ranked results
what is diversification?
do not consider the relevance of each document in isolation,
but consider how relevant the document is in light of:
1. the multiple possible information needs underlying the query
2. the other retrieved documents
goals: maximum coverage and minimum redundancy
MMR
Maximal Marginal Relevance: score a relevant document as the document’s estimated relevance with respect to the query, discounted by the document’s maximum similarity with respect to the already selected documents in D_q
f_MMR(q,d,D_q) = lambda*f1(q,d) - (1 - lambda) * max f2(d,d_j)
f1(q,d) = relevance of d to q
f2(d,d_j) = similarity of d_j to d
what is the purpose of recommendation?
economic: the more relevant the suggestion, the more consumption
user: prevent overchoice / information overload
what is the recommendation task?
- create a ranked personalized list of items, taking the context, situation, and information need into consideration
- if the user interacts with a recommended item, the system was successful
often based on previous interaction between users and items
5 goals of recommender systems
*relevance: users are more likely to consume items they find interesting
*novelty: the recommended item is something that the user has not seen in the past
*serendipity: the items recommended are somewhat unexpected/surprising (‘lucky discovery’)
*diversity: when the recommended list contains items of different types, it is more likely that the user likes one of these items
*explainability: does the user understand the recommendations
relevance signals: how to know what to recommend?
explicit feedback: like, buy, positive review
implicit feedback: browse, click, watch/listen
what are the 3 types of recommender systems models?
collaborative filtering: use user-item interactions (ratings or buying behaviour in ratings matrix)
content-based recommender systems: use attribute information about users and the items
knowledge-based recommender systems: recommendations based on explicitly specified user requirements (explicit profiles, demographics)
content-based methods
do no use data from other users, but item descriptions combined with user’s ratings
user and item embeddings in the same space
use the user profile as a query to retrieve the most relevant items
useful for new items, but not for new users => knowledge-based models or recommend the most popular items
evaluation of recommender systems
- offline evaluation with benchmarks:
- RMSE for estimated rankings
- rank correlation between RecSyst and ground truth
- ranking metrics for relevance
- user studies
- online evaluation (A/B testing)
disadvantages of offline evaluation
they do not measure the actual user response (the data and the users might evolve over time)
prediction accuracy does not capture important characteristics of recommendations, such as serendipity and novelty
disadvantages of offline evaluation
they do not measure the actual user response (the data and the users might evolve over time)
prediction accuracy does not capture important characteristics of recommendations, such as serendipity and novelty
filter bubbles / rabbit holes
if users get recommended information that is most attractive to them they may get more and more of the same
- more problematic in new recommender systems and social media than in entertainment systems
- important to aim for diversity in recommendations
