Recommendation Systems Main problems of RS Proposed solution Results and evaluation Conclusions Questions
Alleviating cold-user start problem with users’social network data in recommendation systems
Eduardo Castillejo Aitor Almeida Diego Lopez-de-Ipina
DeustoTech - Deusto Institute of Technology, University of Deustohttp://www.morelab.deusto.es
Preference Learning: Problems and Applications in ArtificialIntelligence, 2012
Recommendation Systems Main problems of RS Proposed solution Results and evaluation Conclusions Questions
Index
1 Recommendation SystemsHow do they workWhat are they
2 Main problems of RSKnown problems
3 Proposed solutionFoursquareEigenvector centralityExample and analysis
4 Results and evaluationEvaluationResults
5 ConclusionsConclusionsFuture work
6 Questions
Recommendation Systems Main problems of RS Proposed solution Results and evaluation Conclusions Questions
How do they work
Amazon
Recommendation Systems Main problems of RS Proposed solution Results and evaluation Conclusions Questions
How do they work
Youtube
Recommendation Systems Main problems of RS Proposed solution Results and evaluation Conclusions Questions
What are they
Commonly built under a web-based platform they gatherinformation about every entity which takes part in ane-commerce interaction process to make recommendationsto the users increasing the benefits of the e-commercecompany.
They use algorithms which base their recommendations inexplicit and implicit data from the users (ratings,purchases, previous searches...).
Recommendation Systems Main problems of RS Proposed solution Results and evaluation Conclusions Questions
Known problems
Recommendation systems are a good tool to suggest items tousers based in their own interaction with the system, but they alsohave some intrinsic problems which are difficult to solve:
Sparsity: it occurs when available data are insufficient foridentifying similar users (neighbors) and it is a major issuethat limits the quality of recommendations and theapplicability of collaborative filtering (CF).
Scalability: CF requires computations that are veryexpensive and grow polynomially with the number of usersand items in a database. Therefore, in order to bringrecommendation algorithms effectively on the web, andsucceed in providing recommendations with high accuracy andacceptable performance, sophisticated data structures andadvanced, scalable architectures are required.
Cold-start
Recommendation Systems Main problems of RS Proposed solution Results and evaluation Conclusions Questions
Known problems
Recommendation systems are a good tool to suggest items tousers based in their own interaction with the system, but they alsohave some intrinsic problems which are difficult to solve:
Sparsity: it occurs when available data are insufficient foridentifying similar users (neighbors) and it is a major issuethat limits the quality of recommendations and theapplicability of collaborative filtering (CF).
Scalability: CF requires computations that are veryexpensive and grow polynomially with the number of usersand items in a database. Therefore, in order to bringrecommendation algorithms effectively on the web, andsucceed in providing recommendations with high accuracy andacceptable performance, sophisticated data structures andadvanced, scalable architectures are required.
Cold-start
Recommendation Systems Main problems of RS Proposed solution Results and evaluation Conclusions Questions
Known problems
The cold-start problem arises when a new entity enters thesystem for the first time. In this situation the recommendationengine can not predict suggestions because of the lack ofinformation about the current entity. It usually includes 3 entities:
Items
Users
Systems
Recommendation Systems Main problems of RS Proposed solution Results and evaluation Conclusions Questions
Known problems
Amazon
Recommendations related with Kindle, watches special pricesand laptops... ¿?
Recommendation Systems Main problems of RS Proposed solution Results and evaluation Conclusions Questions
Known problems
Youtube
Recommendations about videos of people we don’t actuallyknow... ¿?
Recommendation Systems Main problems of RS Proposed solution Results and evaluation Conclusions Questions
Known problems
We focus our research in alleviating the so called cold-userproblem by collecting information about the user digging intheir social network interactions.
But... how?
Recommendation Systems Main problems of RS Proposed solution Results and evaluation Conclusions Questions
Known problems
We focus our research in alleviating the so called cold-userproblem by collecting information about the user digging intheir social network interactions.
But... how?
Recommendation Systems Main problems of RS Proposed solution Results and evaluation Conclusions Questions
Foursquare
Foursquare is a location-based social networking websitewhich allows users to ”check in” at venues using theirsmartphones.
Thanks to its API developers can request some user data (e.g.location, friends, last check-ins, etc.).
Developed Android app.
Recommendation Systems Main problems of RS Proposed solution Results and evaluation Conclusions Questions
Foursquare
Using the herenow API endpoint we get the previouscheck-ins done by other users at the current checked in venue.
current user
user 1user 2
user 3
user 4
LegendFriend user
Unknown user
1 hour interval
2 hours interval
3 hours interval
Users' check-ins time stampcurrent user
user 1
user 2
user 3
1:00 pm
2:00 pm
3:00 pm
1:00 pm
user 4 4:00 pm
Recommendation Systems Main problems of RS Proposed solution Results and evaluation Conclusions Questions
Eigenvector centrality
Once we have completed the matrix, we apply eigenvectorcentrality.
Since degree centrality gives a simple count of the number ofties a node has, eigenvector centrality acknowledges thatnot all connections are equal.
Denoting the centrality of a node i by xi , then it is possibleto make xi proportional to the average of the centralities of i’snetwork neighbours:
where λ is a constant. This equation can be also rewrittendefining the vector of centralities x = (x1, x2, ...):
Recommendation Systems Main problems of RS Proposed solution Results and evaluation Conclusions Questions
Example and analysis
Applying eigenvector centrality to our previous matrix A”
we obtain that the eigenvalue λ = 12,502, and its eigenvector:
Recommendation Systems Main problems of RS Proposed solution Results and evaluation Conclusions Questions
Example and analysis
The first value of the vector e1 is related to the node 0 (thecurrent user), so its value has to be ignored (in this case thehighest value corresponds with the current user).
The second highest value corresponds with the node 1. Thatmeans that his recommendations would be more ”pleasing” tothe user.
Recommendation Systems Main problems of RS Proposed solution Results and evaluation Conclusions Questions
Evaluation
Amazon default recommendations VS. our Amazoncategories estimation: We presented to our test usersthe default recommendations that Amazon.com for newusers, and another list with our Amazon categoriesrecommendations computed with our solution.
Once our users compared both lists, they fulfilled aquestionnaire to capture their satisfaction level with theobtained results.
Amazon default recommendations: Kindle related products,clothing trends, products being seen by other customers, bestwatches prices, laptops best prices, top seller books.
Amazon default categories: Home, garden and tools,clothing, shoes and jewelry; books; electronics and computers;automotive and industrials; movies, music and games; grocery,health and beauty; toys, kids and baby; sports and outdoors.
Recommendation Systems Main problems of RS Proposed solution Results and evaluation Conclusions Questions
Results
Recommendation Systems Main problems of RS Proposed solution Results and evaluation Conclusions Questions
Conclusions
This paper explores the possibility of using relevant data fromusers’ social network to alleviate the cold-user problems in arecommender system domain. The proposed solution extractsthe most valuable node in the graph generated by check in avenue with an Android application using the Foursquare API.By obtaining the recommendations to this node we estimatethe probability of some categories to be similar to userstastes...
... but we suffered several limitations:
Few users and data...Near venues are not checked in enough...Sparsity.
Recommendation Systems Main problems of RS Proposed solution Results and evaluation Conclusions Questions
Conclusions
This paper explores the possibility of using relevant data fromusers’ social network to alleviate the cold-user problems in arecommender system domain. The proposed solution extractsthe most valuable node in the graph generated by check in avenue with an Android application using the Foursquare API.By obtaining the recommendations to this node we estimatethe probability of some categories to be similar to userstastes...
... but we suffered several limitations:
Few users and data...Near venues are not checked in enough...
Sparsity.
Recommendation Systems Main problems of RS Proposed solution Results and evaluation Conclusions Questions
Conclusions
This paper explores the possibility of using relevant data fromusers’ social network to alleviate the cold-user problems in arecommender system domain. The proposed solution extractsthe most valuable node in the graph generated by check in avenue with an Android application using the Foursquare API.By obtaining the recommendations to this node we estimatethe probability of some categories to be similar to userstastes...
... but we suffered several limitations:
Few users and data...Near venues are not checked in enough...Sparsity.
Recommendation Systems Main problems of RS Proposed solution Results and evaluation Conclusions Questions
Future work
Include data not only from Foursquare
Combine different social network analysis metrics
Take into account more than the most valuable node fordoing recommendations.
Store the obtained matrices for each venue and update themwith every check-in.
Test the solution among a higher number of users.
Recommendation Systems Main problems of RS Proposed solution Results and evaluation Conclusions Questions
Thank you [email protected]
Preference Learning: Problems and Applications inArtificial Intelligence, 2012