Abstract
Given the importance of recommender systems, a detailed understanding of the collaborative filtering systems is imperative. The standard formulation of most collaborative recommender algorithms is either based on prediction of a missing user-item rating or it tends to derive topmost k, users, or items. Collaborative methods could be further divided into user-based filtering and item-based filtering. Regression-based approach can help in combining the advantages of both user-based and content-based methods. Machine learning-based models on the other hand offer further generalization of above models and formally segregate the data modeling, training, and prediction phases. Latent factor models further the same idea by formally incorporating the dimensionality reduction concept and have been found to be very effective. The article is likely to be well received by the academics, especially the doctoral students/researchers in the field of recommender systems as well as the practitioners either utilizing or trying to procure recommendation systems for their organizations.
TopNeighbourhood Based Collaborative Filtering
Neighbourhood based collaborative filtering algorithms are amongst the earliest algorithms for the purpose of developing recommendations. The foundational idea of these algorithms is that similar consumers are likely to rate on similar lines and at the same time similar items have a higher probability of achieving similar ratings. As has been discussed before, the process of computing similarity is a primary activity among these algorithms. Let us discuss the steps involved in the execution of the neighborhood-based collaborative filtering algorithm and simultaneously develop an understanding about these algorithms. For assessing similarity between users we can utilize various similarity indices such as raw cosine similarity, adjusted cosine similarity or the Pearson’s coefficient. Pearson’s coefficient is considered better for the algorithm as it has a bias-adjustment effect as compared to other cosine indices. Another variation to similarity index is achieved by using a discount factor in cases when the common ratings are below a threshold value.
Key Terms in this Chapter
Model-Based Methods: As against neighborhood based methods, model based methods are formal supervised or unsupervised learning methods where a model has been established for prediction of test case.
Neighbourhood Based Methods: These algorithms are based on the idea that similar consumers are likely to rate on similar lines and at the same time similar items have a higher probability of achieving similar ratings.
Latent Factor Methods: In the latent factor method, the user-rating matrix is transformed into a lower-dimensional space by using either principal component analysis (PCA) or singular value decomposition (SVD).
Singular Value Decomposition: SVD involves factorization of a matrix into two matrices U and L, such that requires U and L to be orthogonal.