Challenges Faced

Understanding and coming up with valid combinations of Permutation Factors

Decision between Pseudo-User profile and merging recommendation lists of individual users in our recommendation design: We chose to go with the second approach because Pseudo-User profile might ignore individual dissimilarities and burn down the rating to a similarity based value based on all-user rating. This hampers optimality criterion for Satisfaction of individual users.

Matrix-Factorization goes to NAN values at epoch zero itself leading to failure of the sparsity handling algorithm: We chose to incorporate Learning Rate schedulers to overcome this issue.

Since our data corresponds to maximum dissimilar groups, initially we tried to take social and expertise descriptor bins as statistical thresholds instead of absolute values. When it did not give good accuracy, we reverted back to absolute threshold values as per state-of-the-art. (We tried this because w1, w2 values as per existing techniques were not optimal)