data-mining.philippe-fournier-viger.comThe Data Mining Blog | A blog by Philippe Fournier-Viger about data mining, data science, machine le

The Data Mining Blog A blog by Philippe Fournier-Viger about data mining, data science, machine learning, big data… Skip to content Home About the author ← Older posts Approximate Algorithms for High Utility Itemset Mining Posted on 2021-05-06 by Philippe Fournier-Viger On this blog, I have previously given an introduction to a popular data mining task called high utility itemset mining . Put simply, this task aims at finding all the sets of values (items) that have a high importance in a database, where the importance is evaluated using a numeric function. That sounds complicated? But it is not. A simple application is for example to analyze a database of customer transaction to find the sets of products that people buy together and yield a lot of money (values = purchased products, utility = profit). Finding such high utility patterns can then be used to understand the customer behavior and take business decisions. There are also many other applications. High utility itemset mining is an interesting problem for computer science researchers because it is hard. There are often millions of ways of combining values (items) together in a database. Thus, an efficient algorithm for high utility itemset mining must search to find the solution (the set of high utility itemsets) while ideally avoid exploring all the possibilities. To efficiently find a solution to a high utility itemset mining problem (task), several efficient algorithms have been designed such as UP-Growth, FHM, HUI-Miner, EFIM , and ULB-Miner . These algorithms are complete algorithms because they guarantee finding the solution (all high utility itemsets) However, these algorithms can still have very long execution times on some databases depending on the size of the data, the algorithm’s parameters, and the characteristics of the data. For this reason, a research direction in recent years has been to also design some approximate algorithms for high utility itemset mining . These algorithms do not guarantee to find the complete solution but try to be faster. Thus, there is a trade-off between speed and completness of the results. Most approximate algorithms for high utility itemset mining are based on optimization algorithms such as those for particle-swarm optimization, genetic algorithms , the bat algorithm , and bee swarm optimization . Recently, my team proposed a new paper in that direction to appear in 2021, where we designed two new approximate algorithms, named HUIM-HC and HUIM-SA , respectively based on Hill Climbing and Simulated Annealing . The PDF of the paper is below: Nawaz, M.S., Fournier-Viger, P., Yun, U., Wu, Y., Song, W. (2021). Mining High Utility Itemsets with Hill Climbing and Simulated Annealing . ACM Transactions on Management Information Systems (to appear) In that paper, we compare with many state-of-the art approximate algorithms for this problem (HUIF-GA, HUIM-BPSO, HUIM-BA, HUIF-PSO- HUIM-BPSOS and HUIM-GA) and observed that HUIM-HC all algorithms on the tested datasets. For example, see some pictures from some runtime experiments below on 6 datasets: In this picture, it can be observed that HUIM-SA and HUIM-HC have excellent performance. In a) b) c) d), e), f) HUIM-HC is the fastest, while HUIM-SA is second best on most datasets (except Foodmart). In another experiment in the paper it is shown that although HUIM-SA is usually much faster than previous algorithms, it can find about the same number of high utility itemsets, while HUIM-HC usually find a bit less. If you are interested by this research area, there are several possibilities for that. A good starting point to save time is to read the above paper and also you can find the source code of all the above algorithms and datasets in the SPMF data mining library . By using that source code, you do not need to implement these algorithms again and can compare with them. By the way, the source code of HUIM-HC and HUIM-SA will be included in SPMF next week (as I still need to finish the integration). Hope that this blog post has been interesting! I did not write so much on the blog recently because I have been very busy and some unexpected events occurred. But now I have more free time and I will start again to write more on the blog. If you have any comments or questions, please write a comment below. — Philippe Fournier-Viger is a professor of Computer Science and also the founder of the open-source data mining software SPMF, offering more than 120 data mining algorithms. Posted in Pattern Mining , Utility Mining | Tagged bat algorithm , Bpso , genetic algorithm , hill climbing , HUIM-GA , itemset mining , pso , simulated annealing , utility mining | Leave a comment UDML 2021 @ ICDM 2021 Posted on 2021-04-15 by Philippe Fournier-Viger Hi all, This is to let you know that the UDML workshop on utility driven mining and learning is back again this year at ICDM , for the fourth edition. The topic of this workshop is the concept of utility in data mining and machine learning. This includes various topics such as: Utility pattern mining Game-theoretic multiagent system Utility-based decision-making, planning and negotiation Models for utility optimizations and maximization All accepted papers will be included in the IEEE ICDM Workshop proceedings, which are EI indexed. The deadline for submiting papers is the 3rd September 2021 . For more details, this the website of the workshop: http://www.philippe-fournier-viger.com/utility_mining_workshop_2021/ — Philippe Fournier-Viger is a full professor working in China and founder of the SPMF open source data mining software . Posted in Big data , cfp , Conference , Data Mining , Utility Mining | Tagged cfp , data mining , icdm , icdm conference , machine learning , udml , utility mining , workshop | Leave a comment MLiSE 2021 @ PKDD 2021 – a new workshop! Posted on 2021-04-14 by Philippe Fournier-Viger I am glad to announce that I am co-organizing a new workshop called MLiSE 2021 ( 1st international workshop on Machine Learning in Software Engineering ), held in conjunction with the ECML PKDD 2021 conference . Briefly, the aim of this workshop is to bring together the data mining and machine learning (ML) community with the software engineering (SE) community. On one hand, there is an increasing demand and interest in Software Engineering (SE) to improve quality, reliability, cost-effectiveness and the ability to solve complex problems, which has led researchers to explore the potential and applicability of ML in SE . For example, some emerging applications of ML for SE are source code generation from requirements, automatically proving the correctness of software specifications, providing intelligent assistance to developers, and automating the software development process with planning and resource management. On the other hand, SE techniques and methodologies can be used to improve the ML process ( SE for ML ). The deadline for submiting papers is the 23rd June 2021 , and the format is 15 pages according to the Springer LNCS format. All papers are welcome that are related to data mining, machine learning and software engineering. These papers can be more theoretical or applied, and from academia or the industry. If you are interested to submit but are not sure if the paper is relevant, feel free to send me an e-mail. The papers will be published on the MLiSE 2021 website. Moreover, a Springer book and special journal issue are being planned (to be confirmed). Hope that this is interesting and that I will see your paper submissions in MLiSE 2021 soon:-) — Philippe Fournier-Viger is a full professor working in China and founder of the SPMF open source data mining software . Posted in artificial intelligence , Big data , cfp , Conference , Machine Learning | Tagged artificial intelligence , cfp , data , data mining , machine learning , mlise , pkdd , pkdd 2021 , software engineering , workshop | Leave a comment Mining Episode Rules (video) Posted on 2021-04-11 by Philipp...