One of the first methods of collaborative filtering is memory-based technique, which is also known as neighborhood memory based collaborative filtering algorithms (Aggarwal, 2016), Memory-based collaborative filtering uses rating of the user-item mixtures to generate predication, each user with similar taste is a part of a similar group of users. If to find out the users that shares the same taste of a new user or so-called neighbors, then recommendations can be given to the new user that is likely will be matching his/her taste.
Items that have been rated previously by the user, acts as a vital part in looking for a neighbor that shares same taste with him/her (Shang & Zhao, 2010).Once a neighbor of a user is detected, various algorithms can be applied and are able combine the interests of neighbors to produce recommendations.
There are two types of neighborhoods:
· User-based nearest neighbor collaborative filtering
If there are two user(A) and user(B), have given similar rates in the past for an item and a new item is introduced that was liked by user(A) then the probability that user(B) will like the new item is high, the main idea is to determine users, who are similar to the user (B), and recommend ratings for the ignored or undetected ratings of user(B) by calculating weighted averages of the ratings of this peer group or so-called neighbors (Ricci, Rokach, & Shapira, 2001).
In the process of User-based neighbor collaborative filtering, two tasks are being implemented so that recommendations for an active user can be generated (Thorat, Goudar, & Barve, 2015)
(a) finding out the user nearest neighbors that is like the active user by applying the selected similarity measurements.
(b) After identifying nearest neighbors similar to the active user, the calculation process is performed in order to calculate the prediction of an item to an active user, one of the following aggregation methods is repeatedly used: the average, the weighted sum and the adjusted weighted aggregation (Thorat, Goudar, & Barve, 2015) .
User-based collaborative filtering figures the closeness and the similarity between the users by looking at their ratings on the same item, it also calculates the predicted rating for the item by an active user and consider it as a weighted average of the ratings of the item by similar users to the active user where weights are the similarities of these users with the target item (Ojokoh, Yetunde Oluwatoyin, & Isinkaye, 2015).
· Item-based collaborative filtering technique
Unlike user-based collaborative filtering, item-based method focusses on the item not the user, it uses methods to calculate predictions by analysing similarities between the items not the users, it constructs an item similarities models that recover pervious rated items by an active user from an item-user matrix, it also estimates the similarity of the pervious items to the target item and the estimation process is implemented by using a weighted average rating of an active user of similar items (Ojokoh, Yetunde Oluwatoyin, & Isinkaye, 2015). For example, User user’s(A) ratings on drama movies such as Braveheart and Gladiator can be used to guess his future rating on King Arthur (Aggarwal, 2016).
Many authors have discussed the various similarity methods are used in memory based collaborative filtering, Ojokoh et. al, 2015 mentioned that the most two common similarity metric approaches to determine the closeness between two documents in memory based collaborative filtering are the correlation-based and the cosine-based (Ojokoh, Yetunde Oluwatoyin, & Isinkaye, 2015). Pearson correlation coefficient is used to find closeness between user(A) and user(B), the method was originally presented in the context of GroupLens project (Vozalis & G. Margaritis, 2003)
Cosine similarity is a different approach that is commonly used in the areas of texts mining to compare texts documents, Cosine similarity is calculated using all users who have rated two items, the only difference from the other measure is that adjusting the average is executed with respect to the user, not to the item (Sen, Herlocker, Frankowski, & Schafer, 2007).
According to Aggarwal et al, 2016 memory-based collaborative filtering algorithms are simple to employ and to use, they also give recommendations results that are often easy to express. however, memory-based algorithms don’t function efficiently with sparse ratings matrices. For instance, it might be tricky to find enough users that have the same taste as user(A), who gave rating for the movie Gladiator and in that case, it’s not going to be easy to predict user’s(A) rating for the movie Gladiator (Aggarwal, 2016).
Sarwar et. al, 2001 discussed the importance of model-based collaborative filtering algorithms and mentioned that it the methods of model-based gives item suggestion by creating a model of user ratings. Model-based methods fellows a probabilistic strategy that have the capability to produce predictive models, based on the past user behavior and actions using machine learning algorithms such as Bayesian network, clustering, and rule based-approaches (Sarwar, Karypis, Konstan, & Riedl, 2001),this approach uses the previous user’s ratings to predict a model to enhance the performance of Collaborative filtering system. The process of creating models are done by data mining techniques, which recommend a set of items rapidly that uses pre-computed model, the techniques includes different example such as Regression, Clustering and decision trees (Ojokoh, Yetunde Oluwatoyin, & Isinkaye, 2015),model-based methods examine the user-item matrix and relation to find out If there is a connection between the items, it uses techniques of learning algorithms that solve the problem of sparsity that is associated with recommendation systems.
Ricci et al.2001, mentioned that neighbourhood memory-based technique stored ratings are used directly for predication but model-based methods utilize the ratings to produce predictive models, the main idea is to analyse the user-item interactions with factor representing characteristics of the items and the users such as the user taste and the item category and classification, using the data the model can be trained and utilized to anticipate user’s future ratings for new items (Ricci, Rokach, & Shapira, 2001). The most common approaches that are used in mode-based for the task of recommending new items includes Bayesian network model that creates a probabilistic strategy for collaborative filtering problem, clustering model deals with collaborative filtering techniques as a categorization problem and proceed by clustering similar users in same category and calculating the probability that a specific user is in a specific category, the rule-based methods applies association rule discovery algorithms to find connection between co-purchased items and then produces item suggestion based on the strength of the association between items (Sarwar, Karypis, Konstan, & Riedl, 2001).