Dynamic k Neighbor Selection for Collaborative Filtering

Abstract

Collaborative filtering is a commonly used method to reduce information overload. It is widely used in recommendation systems due to its simplicity. In traditional collaborative filtering, recommendations are produced based on similarities among users/items. In this approach, the most correlated k neighbors are determined, and a prediction is computed for each user/item by utilizing this neighborhood. During recommendation process, a predefined k value as a number of neighbors is usedfor prediction processes. In this paper, we analyze the effect of selecting different k values for each user or item. For this purpose, we generate a model that determines k values for each user or item at the off-line time. Empirical outcomes on movie based dataset show that using the dynamic k values during the k-nn algorithm leads to more favorable recommendations compared to a constant k value.

Improving Prediction Performance of Dynamic Neighbor Selection in User-Based Collaborative Filtering

“…The first result shows the classical PCC with users who have rated the target item while the second result depicts the modified PCC with the significance-weight factor. The third result shows our proposed method's result and allows the system to decrease similarity k=50 in [23] k=dynamic in [23] Coverage for FilmTrust weights between users if the co-rated items are smaller than the threshold and choose dynamic k values. Figure 4 illustrates that using dynamic k value for each user with the significance-weight factor increases predictions of accuracy.…”

“…We conduct a new experiment in order to measure the effect of nominating users who rated the target item as possible neighbors in dynamic k neighbor selection. Recall that dynamic k neighbor selection nominates all users whose correlation/similarity can be calculated [23]. Each user in the dataset is selected as an active user and the rest form our training data.…”

“…Throughout our experiments, we test if the accuracy and the coverage results get better than the existing methods [14], [23]. We perform different experiments to test our intuition and demonstrate how dynamic k values and the significance-weighting factor contribute to the results.…”

“…In the online part, we randomly select five items among rated items by the active user as in [23] and estimate predictions for those items using dynamic k values and predefined significance weight value. We repeat this experiment 100 times to obtain average prediction results.…”

“…However, allowing users to select a dynamic k value increases the accuracy of the prediction [23]. Similar to traditional k-nn approaches, an active user, who is looking for a prediction, selects the best k neighbors among all users whose similarities can be computed in the dynamic k neighbors algorithm [23]. This may lead a poor coverage compared to accuracy [23].…”