Given a directed social graph, we have to predict missing links to recommend frienns/connections/followers(Link Prediction on graph).
Dataset from facebook's recruiting challenge on kaggle: https://www.kaggle.com/c/FacebookRecruiting Data contains two columns: source and destination edge pairs in the directed graph.
- Data columns(total 2 columns):
- source_node int64
- destination_node int64
-
Map this to a binary classification task with 0 implying an absence of an edge and 1 implying the presence as y_i's
-
Now, we need to featurize a pair of vertices (ui, uj), such that these features can help us predict the presence/absence of an edge. A simple feature could be number of common-friends to u_i and u_j which is highly indicative of an edge between u_i and u_j.
-
Reference papers and link:
-
Assignments: Add another feature called Preferential Attachment with followers and followees data of vertex. You can check about Preferential Attachment in below pdf: http://be.amazd.com/link-prediction/
- No low_latency requirements
- Predicting the probability of a link is useful so as to recommend the highest probability links to a user
- We got to suggest connections which are most likely to be correct and we should try and not miss out any connections
- Both precision and recall are important, hence F1 score is good choice
- Confusion matrix
- Accuracy can also be checked