If a random group of people prefers something, it certainly does not mean that you will like it too. Moreover, when it comes to a recommendation, a few trusted friends with aligned interests can easily win the battle against a group of people vouching for some other option. So, why not make a recommendation system that exploits this trust. Let's see how some techniques have tried to incorporate “trust” and “friendship” in their recommendation system and are they truly worth it?
Introduction
Social media is growing at an exponential pace and so is the data generated by it. It enables users to share information about themselves, there likes and dislikes and much more. This huge amount of data brings with itself huge amount of advantages and disadvantages. However, for recommender systems, this massive amount of data brings more good news that bad.
Nowadays the trend of checking in at places that people visit generates huge about of data that fuels this relatively newer branch of location recommender systems. The main of location recommender systems is to provide recommendations for places to visit based on social network applications where users shared their check-in details.
Who do you trust?
It is not surprising that users prefer to refer to their trusted friends over the mass population. This is the hypothesis behind the trust-based recommender systems. The use of trust-based recommender systems not only gives better recommendations but also solves many other problems that tradition recommender systems still face.
- Sparsity problem
- Cold start problem
- Trustworthiness problem
How to make rec-sys “trust” and “friendship” aware?
Recently, many types of research have been conducted which include trust factor between users to provide better recommendations. Two such trust-based recommender systems are discussed here.
[1] and [2] both incorporate trust factor in recommender systems.
The basic approach used in [1] is as follows:
- User access location probability calculation
- User trust calculation
- Find location recommendation probability
In step one, a target user is selected and say ‘N’ users exist in the system. This target user is compared to the remaining ‘N’ users and similarity is calculated. The sign in information about the user includes their unique id, location in longitude and latitude and time attribute. Pearson's coefficient was used and a matrix of target users and remaining ‘N’ users was created to measure similarity.
In step two, trust between two users is calculated using the formula
Trust (v|u) = num / totalnum
Where, num= common locations of u and v, and totalum denotes the total number of locations accessed by u.
In the last step, the probability is calculated.
In [2] a much elaborate system is used to establish trust between users in the system. Unlike the previous approach where the users may or may not have had any connection between them, this approach harnesses the power of social media and links a user to his or her close friends.
In order to do so, 5 stages of computation is done.
- Weight measurement
- Edge trust measurement
- Node trust measurement
- Edge influence measurement
- Node influence measurement
The relationship between the users is represented as a graph G=(V,E), where V is the set of users and E is the relationship between the user. Sine the trust factor between the users is not a symmetric relation multi edges are present between the nodes.
Figure 1: Weight, Trust and Influence between Nodes in the social network
Then the 5 stages are explained as follows
Edge trust is the trust friend A has on feind B, Node trust is the trust that A node has on the entire node. Edge influence is the influence that A has on B, and node influence is the influence that A has on the entre network. Weight is the correlation value between the the attributes of the two, calculated using pearson's correlation.
EdgeTrust (fri, frj) = Weight(fri, frj)/ ∑ Weight(fri,frx)
NodeTrust (fri) = ∑ Edge Trust
EdgeInfluence (fri, frj)= EdgeTrust (fri, frj) * RecRate
Where, RecRate (fri, frj) = VisitedPlaces (fri, frj)/ ∑ VisitedPlaces
Based on these calculations a recommended list is created.
Conclusion
Trust based recommendation clearly gives better recommendations both practically and intuitively. This is backed by a simple fact that we prefer recommendation of the people we trust, or at least those who share a common preferences with us. The above stated methods take into account these factors and therefore outperform the conventional recommendation system.
References:
[1] BaoLei, Wang Zhanquan, Shalin Huang, “Location Recommendation Algorithm for Online Social Networks Based on Location Trust”.College of Information Science and Engineering, East China University of Science and Technology, Shanghai 200237, China
[2] Heba M. Wagih, Hoda M. O. Mokhtar, Samy S. Ghoniemy, “Location Recommendation Based on Social Trust”. In the proceedings of 13th International Conference on Semantics, Knowledge and Grids 2017.
Image Sources:
Fig1:https://www.google.co.in/searchbiw=1484&bih=733&tbm=isch&sa=1&ei=WkuxWo65PMzcvTJ3LTACw&q=friend+recommend&oq=friend+recommend&gs_l=psyab.3...3002.9891.0.10432.0.0.0.0.0.0.0.0..0.0....0...1c.1.64.psy-ab..0.0.0....0.a2O7ZzAPWe4#imgrc=gHWhvo0IQ9LuEM:
Fig2: Heba M. Wagih, Hoda M. O. Mokhtar, Samy S. Ghoniemy, “Location Recommendation Based on Social Trust”. In the proceedings of 13th International Conference on Semantics, Knowledge and Grids 2017.
Comments
Post a Comment