recommendation system based on similarity of users and article popularity

• pandas, matplotlib, numpy with python3.6

• given a list of user-article interaction data/user-item-interactions.csv, how we recommend relevance articles based on interaction of other similar users or the popularity of the article itself.
• recommend articles the new users (cold-start)

• Rank similarity based dot product of two users, a higher product shows a better similarity,
• Rank popularity of articles based on the total interaction on platform
• Use Single Value Decomposition from numpy.linalg.svd to split a user-item matrix to three matrices: U (users), S(sigma, latent feature) and VT (V transpose for latent feature and articles)
• SVD only works with a dense matrix, no missing value, otherwise, we need to approximate U, S, VT using FunkSVD algorithm

• the accuracy is high (>99%) but it is more likely as the antifacts of memorization, in which the matrics from SVD is optimized for a single use. Note that, the value in matrix is mostly zeros or imbalanced class so the high accuracy prediction is biased.
• the noise when introducing more latent feature reduces (slightly) accuracy of testing data

• the framework of notebooks is prepared by Udacity.com
• StackOverflow