High utility itemset mining using genetic algorithm assimilated with off policy reinforcement learning to adaptively calibrate crossover operation. (14th November 2021)

• Record Type: Journal Article
• Authors: Logeswaran, K; Suresh, P
• Is Part Of: Computational intelligence. Volume 38:Number 5(2022)
• Page Start: 1596
• Page End: 1615
• Publication Date: 2021-11-14
• Subjects: artificial intelligence -- evolutionary computation -- high utility item set -- intelligent communication model -- off policy reinforcement learning; Artificial intelligence -- Periodicals; Computational linguistics -- Periodicals; 006.3
• DOI: 10.1111/coin.12490
• Languages: English
• ISSNs: 0824-7935

Abstract

Abstract: Mining high utility itemsets (HUI) is a current thrust field in data mining that has received numerous methodologies for addressing it effectively. The difficulty with HUI is to locate a number of items that have a high degree of utility in comparison to other different sets in a transaction database. Traditional accurate HUIM algorithms usually have to solve the exponential problem of big search spaces when the size or number of different items in the database is quite vast. Evolutionary computation (EC)‐based algorithms have been offered as an alternate and successful technique to solving HUIM issues since they may generate a collection of approximately optimum solutions in a short amount of time. Many genetic algorithm (GA)‐based approaches have been developed in recent years to efficiently mine HUI from transaction databases. The selection technique, crossover probability, mutation probability, and finishing criteria of the genetic algorithm have an greater impact on generating a reasonably decent solution and the processing time. Particularly crossover is a convergence operation which is intended to pull the population toward a local minimum/maximum. During HUIM using GA both low and high crossover rate will have the problem of decreasing the quality of intermediate itemset and take longer time to converge to some optima and vice versa. This problem can be solved by adjusting the crossover rate adaptively depending on environmental inputs. The proposed approach describe a hybrid system that employs a reinforcement learning (RL) agent to adaptively calibrate the crossover operation to increase the performance of a genetic algorithm. To estimate state‐action utility values, the RL agent employs Q‐learning, which it then employs to execute high‐level adaptive control over the crossover operation in the genetic algorithm. To evaluate the performance of the proposed methodology, extensive experiments were conducted on a four benchmark datasets and compared with three state‐of‐art EC approaches HUPEUMU ‐GRAM, Bio‐HUIF‐GA, HUIM‐BPSO, and one exact approach HUP‐Miner . The result analysis witnessed that proposed approach outperforms EC approaches in terms of execution time, discovered HUIs and convergence.