Maximum Coverage Method Modification with Timeliness in Non-Personalized Recommendation for Pure Cold-Start Users

• Main Author: Alkaf Bahraini Saputra, Novan
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/39693
• Institution: Institut Teknologi Bandung

Humans have limitations in filtering the information they receive, both energy and time, so that a search is tiring. In a system that has thousands or even millions of items, of course a system is needed that can provide a recommendation that is in accordance with the interests of users, the system is called the recommendation systems . The recommendation system functions to filter items and produce a list of recommendations that might have high possibilities of interest by users. The recommendations generated by the system are called personalized recommendations, where the system will take into account each system user data to produce recommendations that are relevant and on target. However, this system is not possible to investigate the interests of someone who has not been part of the system. Such users are called pure cold-start users. This user cannot be accessed due to various things such as the use of incognito mode when browsing a system, the difficulty of getting IP and cookies, and the high cost of knowledge and energy needed. Though such users are potential users, the list of recommendations provided by the system plays a role in the acquisition of cold-start users becoming users of the system. For this reason, another approach to the problem is non-personalized recommendations. Non-personalized recommendations will provide a global list of recommendations so everyone have same recommendation with aggregation and calculation results for users who already exist in the system. The aim of the global recommendation is to accommodate as much as possible the aspirations of the majority of users on the system to produce recommendations that also match the majority of cold-start pure users. In looking at the quality of recommendations, not only is seen based on the level of accuracy, but also seen from the utility of the recommendations produced, the influence of a diverse list of recommendations and the reach of cold-start pure users. To reach as much as possible users is important because the goal of non-personalized recommendations is to provide global recommendations with a wide range. To reach the maximum, the maximum coverage method is used. The aim is to produce recommendations that contain items that can reach all users. This method only covers items with maximum range. However, there is a possibility that this is not done because the bias in the user is consumption of items based on relevance and time so that timeliness metrics are used to complete the method. This metric is used to produce maximum coverage with items that have better timeliness than other items. The study was conducted on the 100k and 1M movielens datasets using maximum-coverage-timeliness in 2 scenarios, namely scenario 1 using all item data as input while scenario 2 cutting the number of items based on user coverage of items that are assumed to be popular. In the dataset movielens 100k, MaxCovTL_100 from scenario 2 is a method that becomes a reference by having a usability value better than other methods on top-15 and top-20 and has the best range value in the top-15. However, after conducting a significance test with the Wilcoxon test, the increase that occurred was not significant. Whereas in the movielens 1M dataset, the modified method in this study has the same value as the maximum coverage method. The addition of timeliness does not go well because the items with the most coverage are owned by items that have long been circulating, while newer items have less coverage. As for the other disadvantages, the modified method cannot reach users who like less popular items. With these results, it can be concluded that the addition of timeliness in the maximum coverage method cannot improve the quality of recommendations in terms of utility, diversity and user conquered.