Article Preview
TopIntroduction
Frequent Item-sets Mining(FIM) (Agrawal, Imielinski, et al., (1993)) (Rakesh Agrawal, Ramakrishnan Srikant. (1994)) is a fundamental data mining task, which discovers the high frequent item-sets from the transactional dataset. To full fill various requirements specific to bounded sub-areas, FIM extended tasks (High Utility Item-set Mining (HUIM)(Gan, Lin, et al.,), Sequential Pattern Mining (SPM)(Fournier Viger, Philippe, et al., (2017)), Rare Item-set Mining (RIM)(Borah, Nath, (2019)),..etc.) came into existence. The advantage of HUIM over FIM is, HUIM considers the utility (profit & quantity) of items whereas FIM considers only quantity. Again to full fill the various requirements specific to applications, HUIM extended as HUIM with Positive and Negative utility values(HUIMPN) (Singh, Singh, et al., (2019)), top-k HUIM (Tseng, Vincent, et al., (2015)), High Utility Sequential Item-set Mining(HUSIM)(Truong-Chi, Fournier-Viger, (2019)),...etc. The advantage of HUIMPN over traditional HUIM is, HUIMPN compatible with the positive and negative utility values of items whereas HUIM compatible with only positive utility values.
In general, retail markets provide various discounts on the tag price of items. In this process, few items may make a negative profit. But, overall profit could be positive. Suppose, an item 'Y' is free along with a purchase of item 'X'. Here, 'Y' makes a negative profit and 'X' makes a positive profit. But, the combined profit of X, Y could be positive. Hence, the algorithms must compatible with the data sets having positive and negative profits are needed.
The set of items purchased by a customer in one visit is called a transaction. A stream of transactions comes into the retail market database every time due to the bulk number of customers visit. There could be different kinds of discount strategies applied to different items. Some of them may be changed from time to time and some of them may hold for a long time. To analyze the sales trend (profits, customer interests) dynamically, techniques which process the data stream are required. As data comes in fly continuously, some technique is required to hold it. Many techniques are available for this purpose, sliding window one such technique (Lee, Jin, et al., (2014)).
A window is nothing but a set of bathes, and a batch is nothing but a set of transactions. The size of the window and batch is set by the user. Suppose, the window and batch sizes are set to 10 and 1000 respectively then, B1 (batch-1) possess transactions T1 to T1000, and B2 (batch-2) possess transactions T1001 to T2000 likewise batches filled with the transactions. In sliding window technique, W1 (window-1) possess batches B1 to B10, and W2 (window-2) posses batches B2 to B11 likewise windows are filled with batches. A batch is like a bottom closed bucket. It is filled with the sequence of transactions up its capacity, and further coming transactions are placed in the next batch. A sliding window is like a controlled leaky bucket. For the first time window is completely filled with the batches and sent for mining. After that, if a new batch comes, the oldest batch is dropped, and the newest batch is added to the window called window sliding. For every window slide, it is sent for mining. The advantage of this technique is very recent data is used to mine.
To give better insights to retail markets, HUIM algorithms must be capable to full fill the following requirements. i. They must apply for the datasets having positive and negative utility values. ii. They need to process the data stream. Most of the available HUIM algorithms () may not full fill the specified requirements. The proposed Extended Global Utility Item-sets tree (EGUI-tree) algorithm is well suitable for the given environment. The rest of the paper is organized as follows. Section 2 introduces Preliminaries, problem statement, and the relevant literature. Section 3 presents the proposed algorithm. Section 4 Contains experimental results. Finally, Section 5 concludes the paper.