Browse Prior Art Database

[CUE_Interns-Recommendations-03] Similarity-based people recommendation algorithms for social networking sites

IP.com Disclosure Number: IPCOM000190147D
Original Publication Date: 2009-Nov-18
Included in the Prior Art Database: 2009-Nov-18
Document File: 3 page(s) / 94K

Publishing Venue

IBM

Abstract

Disclosed are similarity-based people recommendation algorithms for social networking sites

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 43% of the total text.

Page 1 of 3

[

[[[CUECUECUE

CUE

_

___Interns

InternsInterns

Interns-

---RecommendationsRecommendationsRecommendations

Recommendations -

---030303


03]]]] SimilaritySimilaritySimilarity

Similarity -

---based people recommendationbased people recommendationbased people recommendation

based people recommendation

algorithms for social networking sites

algorithms for social networking sitesalgorithms for social networking sites

There exist a growing number of social networking sites available today. As the number of users on these sites also grows, it becomes difficult for an individual user to find other users they may know or would like to network with.

One possible way to help users find people on sites with a large population is to proactively recommend people to a given user. Some algorithms tend to recommend people a user already knows or is familiar with. Those algorithms help a user complete their existing social network online. Other algorithms recommend mainly unknown people based on similarity, for example, same job, skills, keyword match etc. In a professional setting (like Beehive), those algorithms are important to support reaching out to new people who can help a user in their job and career.

Note

that the "familiarity"-based approach will run out of recommendations once a user has completed their social network. "Similarity" approaches can constantly help expand your existing network with new contacts. However, similarity-based people recommenders are not yet very common today on social networking sites.

Facebook has recently added a "People you may know" page. They use a simple "Friend of a Friend" algorithm which recommends users based on the number of mutual friends you have in common. This algorithm recommends known and unknown people but recommendations are limited to one hop away from the user. The Sonar-Fringe people recommender in *IBM is similar to the Facebook recommender in that it recommends mostly "familiar" people. However, Sonar includes multiple data sources to compute familiarity.
A. The first algorithm works as follows

(1) Creating bag-of-words representation

First create a bag-of-words representation of each user, using all textual content that can be associated with the user. To improve the robustness of the word list and ensure the words are meaningful, use certain rule-based procedures to combine words and/or collapse the word list. Such procedures can include, but are not limited to:

1. A stemmer, such as the Porter stemmer, which reduces all words into their roots.

2. Combine words with short edit distances from each other.

3. Remove all words in a customized stop word list.

4. Remove all words whose word count is lower or higher than certain thresholds.

All remaining words associated with a user

u

are then used to create a word vector u

V =( )

( 1

vu

w

,…, )

( m

vu w ) to describe

u

, where

m

                     is the total number of distinct words used in all included texts after the a...