Weighted Instance Based Learner (WIBL) for user profiling

Abstract

With an increase in web-based products and services, user profiling has created opportunities for both businesses and other organizations to provide a channel for user awareness as well as to achieve high user satisfaction. Apart from traditional collaborative and content-based methods, a number of classification and clustering algorithms have been used for user profiling. Instance Based Learner (IBL) classifier is a comprehensive form of the Nearest Neighbour (NN) algorithm and it is suitable for user profiling as users with similar profiles are likely to share similar personal interests and preferences. In IBL every attribute has an equal effect on the classification regardless of their relevance.In this paper, we proposed a weighted classification method, namely Weighted Instance Based Learner (WIBL), to build and handle user profiles. With WIBL, we introduce Per Category Feature (PCF) method to IBL in order to distinguish the effect of attributes on classification. PCF is an attribute weighting method and it assigns weights to attributes using conditional probabilities. The direct use of this method with IBL is not possible. Hence, two possible solutions were also proposed to address this problem. This study is aimed to test the performance of WIBL for user profiling.To validate the performance of WIBL, a series of computer simulations were carried out. These simulations were conducted using a large user profile database that includes 5000 training and 1000 test instances (users). Here, each user is represented with three sets of profile information; demographic, interest and preference data. The results illustrate that WIBL with PCF methods performs better than IBL on user profiling by reducing the error up to 28% on the selected dataset.

User Profiling - A Short Review

“…In summary it can be said that weighted multi-dimensional user profiling could be the new profiling method for the future service personalization. Although in [48] authors proposed WIBL for this purpose, there are other potential feature weighting algorithms that could be used with IBL to achieve much better accuracy with multi-dimensional user profiles.…”

“…Moreover, Locally Weighted Learning (LWL), RepTree, Decision Table and SVMReg classifiers were compared for classification. Previous works [44], [45], [46] and [47] have been the first in the literature to present the comparison of the classification and clustering accuracy performance of different algorithms with user profiles. In [45] NB, Instance Based Learner (IBL), Bayesian Network (BN) and Lazy Bayesian Rules (LBR) classifiers were compared using a user profile dataset.…”

“…Classification and Regression Tree (SimpleCART), NBTree, Id3, J48 -a version of C4.5-and Sequential Minimal Optimization (SMO)) were included and compared with large user profile data. Following this, in [48] a variant of the IBL algorithm, namely Weighted Instance Based Learner (WIBL), for the user profiling has been proposed. The WIBL uses a modified version of the PerCategory Feature Weighting algorithm to assign weights to the features during the user profiling process.…”

“…The WIBL uses a modified version of the PerCategory Feature Weighting algorithm to assign weights to the features during the user profiling process. In [47] comparison of the performance of WIBL with other well known clustering algorithms have been done. In this study simulation results showed that WIBL outperforms the well known clustering algorithms.…”

Testing and Analysis of Activities of Daily Living Data with Machine Learning Algorithms

“…The closest training instance predicted to has the same class label with the test instance [5] [6] [7]. In case of more than one training instance qualified as the closest, the class label of the first one is assigned to be the class label of the test instance [8].…”