An evolutionary approach for automating the selection of optimum Algorithm in Collaborative Filtering Recommender Systems
mojdeh robati anaraki, nooshin riahi · Archives de l'Institut d'Hessarek. · 2023
The expansion of online resources over the last few decades, has made the internet the main source for multimedia information, such as movies, books, music, etc.When the amount of choices becomes overwhelmingly large, the need for a recommender system arises.Recommender systems can be definedas software programs to suggest the most appropriate and closest items to the user's taste.They act as a counselor, behaving in such a way to guide people through the discovery of products of interest.Nowadays, there are numerous recommendation methods used to implement a recommender system.These methods, generally, fall into three main categories. 1) Content-based techniques that recommend items similar to the user's highest rated items.2) Collaborative filtering techniques, which use similarity between users or items according to the user's rating pattern, for generating recommendations.These recommendations are based on the preferences of users who have similar tastes to the target user.3) Hybrid techniques that combine these two techniques to address the shortcomings of each one.Collaborative filtering methods are categorized as 1) Model-based methods, in which they use the user's rating to create a model for recommendations, and 2) Memory-based (Neighborhood-based) methods, which find the most similar users to the target user (neighbors) to create recommendations for the user.Various similarity functions and metrics have been used to create the model or compute the similarity in collaborative filtering methods.When using collaborative filtering methods, one of the challenges encountered, is the cold start problem.This occurs because in e-commerce systems, most items have only been rated by a few users.Because there are not enough ratings, it becomes difficult to calculate the similarity between users or items.There are several approaches to tackle the cold start problem in collaborative filtering systems.When building a recommender system, it can be challenging to manually select a method from the recommendation methods, since there are many methods to choose from.The best method to generate the most relevant recommendations, may vary depending on the available data of users and items, since each approach has its own particularities and depends on the context.Furthermore, because the data in online systems are constantly changing, with a fundamental change in the data, the best method for the system might also change.Therefore, there is a need for a technique to automate the selection of recommendation technique for the system.This paper proposes a method that allows us to choose the best combination of memory-bsed collaborative filtering algorithms that, when used on selected data, will make the optimal recommendations in a limited amount of time.Among the main methods to create recommendations, collaborative filtering methods have better results because they rely on quality data provided by users, compared to content-based methods that rely on item information.Therefore, collaborative filtering methods have been selected as the base methods for the purposed approach, and since the goal is to create a system that can choose the optimum recommendation algorithm based on general data statistics, such as the ratio of items to users, if the data changes fundamentally, we need to be able to rerun the algorithm to select a new combination of methods.Since the selected methods for this system need to be run in a relatively short time, only the memory-based methods in collaborative filtering methods have been selected, because they do not need a training step, also they can adapt themseleve to changes in the data, faster.To address the cold start issue, in addition to computing the similarity between users, inferred trust is also computed to increase the amount of available information about relations between users' interests.Since we the time is limitted, we need a technique to find the best combination of methods without testing each combination.Therefore, we used the genetic algorithm that reflects the process of natural selection, where the fittest individuals are selected for reproduction in order to produce offspring of the next generation.This process is continued to iterate and eventually, a generation with the fittest individuals will be found within a limited time.This approach, ultimately, proposes a combination of collaborative filtering techniques for each data set.This will automate the selection of the optimum algorithm for a recommender system, and the resulting combination, in addition to considering time limits, will have an acceptable precision for making recommendations.The proposed approach, uses a genetic algorithm for ranking aggregation of memory-based collaborative filtering methods, selects the most relevant recommendations generated by different similarity techniques to create a Top-N recommender system.This approach has been compared to individual memory-based collaborative filtering methods and other similar methods.The experiments were performed using 1M MovieLens and 100k MovieLens and HetRec2011 data sets.The results show that the proposed approach in this paper, is performmed better and it has a higher precision in generating recommendations for users, than the other similar algorithms.