GACFF (version 1.0)

NewKNN: Nearest Neighbors.

Description

Determining of nearest neighbors and their id to determine the number of genes in a chromosome.

Usage

NewKNN(ratings, active_user, Threshold_KNN, max_scour, min_scour)

Arguments

ratings

A rating matrix whose rows are items and columns are users.

active_user

The id of an active user as an integer greater than zero (for example active_user<-6).

Threshold_KNN

Maximum number of neighbors.

max_scour

The maximum range of ratings.

min_scour

The minimum range of ratings.

Value

An object of class "NewKNN", a list with components:

call

The call used.

sim_NewKNN

The similarities between near users and the active user that have obtained from the "NewKNN" method.

pre_NewKNN

The predicted ratings for the active user by the NewKNN method.

item_NewKNN

A set of recommended items id, obtained from the NewKNN method.

near_user

Neighbors of the active user by the NewKNN method orderly.

time_NewKNN

The elapsed time in NewKNN method.

Details

The number of neighbors for the active user determines the number of genes in the chromosome of the genetic algorithm. The fitness function is MAE which by being minimized, the similarity of the neighbor users is optimized within the processes of the genetic algorithm. The following equation is used to determine the starting points of the genetic algorithm, which are essentially approximation similarities. Using these starting points, the genetic algorithm converges faster.

$$sim_dif = (max rating - dif)/sum(ratings)$$ $$range of dif: [min rating-1,\dots, max rating-1]$$

dif is the difference in the existing ratings. For example, for a difference of 0.5, the approximate similarity is \(4.5/15\) and for a difference of 0, the similarity is \(5/15\). In this method, the number of neighbors varies for each active user, so the problem of predetermining it is solved.

The steps of this function are:

1) The rating matrix is assigned to the form of the Item-user matrix (Items in rows and users in one column).

2) The users rating differences of each item are calculated for each pair of related users.

3) For each user, the related pairwise are separated from all rows in one column.

4) If a pairwise is repeated several times, the average values of the differences are calculated. The number of neighbor users is different for each active user.

5) The rating differences are sorted in ascending order.

6) Neighbor users are selected based on lower rating differences. If the threshold for the difference is already specified, the out-of-area relationships are eliminated.

References

Koohi, H. and Kiani, K. (2017). A new method to find neighbor users that improves the performance of Collaborative Filtering. Expert Systems with Applications, vol. 83, pp.30-39.

Examples

# NOT RUN {
ratings <- matrix(c(  2,    5,  NaN,  NaN,  NaN,    4,
                    NaN,  NaN,  NaN,    1,  NaN,    5,
                    NaN,    4,    5,  NaN,    4,  NaN,
                      4,  NaN,  NaN,    5,  NaN,  NaN,
                      5,  NaN,    2,  NaN,  NaN,  NaN,
                    NaN,    1,  NaN,    4,    2,  NaN),nrow=6,byrow=TRUE)
                    
NewKNN.out  <- NewKNN (ratings, active_user=6, Threshold_KNN=4,
                       max_scour=5, min_scour=1)
# }