[Paper Review] Mining high on-shelf utility itemsets with negative values from dynamic updated database

Anjali N. Radkar, Sayali Pawar|arXiv (Cornell University)|Jul 7, 2015

Data Mining Algorithms and Applications12 references3 citations

TL;DR

This paper proposes a dynamic algorithm for mining high on-shelf utility itemsets with negative profits in temporal databases that evolve over time. It efficiently updates results incrementally instead of re-running from scratch, significantly reducing computational overhead while accurately capturing itemset utility over time, including items with negative utility values.

ABSTRACT

Utility mining emerged to overcome the limitations of frequent itemset mining by considering the utility of an item. Utility of an item is based on user's interest or preference. Recently, temporal data mining has become a core technical data processing technique to deal with changing data. On-shelf utility mining considers on-shelf time period of item and gets the accurate utility values of itemsets in temporal database. In traditional on-shelf utility mining, profits of all items in databases are considered as positive values. However, in real applications, some items may have negative profit. Most of the traditional algorithms are used to handle static database. In practical situations, temporal databases are continually appended or updated. High on-shelf utility itemsets needs to be updated. Re-running the temporal mining algorithm every time is ineffective since it neglects previously discovered itemsets. It repeats the work done previously. In this paper, an effective algorithm is proposed to find high on-shelf utility itemsets with negative values from the dynamic updated temporal database.

Motivation & Objective

• To address the limitation of traditional on-shelf utility mining that assumes only positive profits for items.
• To enable efficient discovery of high utility itemsets in temporal databases that are continuously updated.
• To support negative utility values for items, reflecting real-world scenarios like losses or costs.
• To reduce computational cost by incrementally updating results instead of reprocessing the entire database.
• To maintain accuracy in utility estimation by considering both on-shelf duration and item-specific profit/loss values.

Proposed method

• Extends the on-shelf utility model to include negative utility values for items, reflecting costs or losses.
• Introduces an incremental update mechanism that leverages previously discovered itemsets to avoid recomputation.
• Maintains a dynamic structure to track itemset utility based on on-shelf time periods and utility values.
• Applies utility calculation that aggregates item utility over time, considering both positive and negative contributions.
• Uses a pruning strategy based on utility bounds to reduce search space during incremental updates.
• Employs a transaction-based update mechanism that processes new or modified transactions efficiently.

Experimental results

Research questions

• RQ1 How can on-shelf utility mining be extended to handle items with negative utility values?
• RQ2 What is an efficient way to maintain high on-shelf utility itemsets in a dynamically updated temporal database?
• RQ3 Can incremental updates reduce the computational cost of re-running utility mining algorithms after each database change?
• RQ4 How does the inclusion of negative utility values affect the accuracy and completeness of mined itemsets?
• RQ5 What mechanisms ensure that previously discovered itemsets are reused effectively during incremental updates?

Key findings

• The proposed algorithm achieves significant performance improvement over full reprocessing by reusing prior results and minimizing redundant computation.
• The method accurately captures itemsets with negative utility values, reflecting real-world business scenarios involving losses or costs.
• Incremental updates reduce the time complexity compared to batch processing, especially in frequently updated databases.
• The algorithm maintains high utility itemsets with precision by considering both on-shelf duration and utility sign (positive/negative).
• Empirical evaluation shows that the incremental approach scales better with increasing database size and update frequency.
• The approach supports practical deployment in real-time or near-real-time data mining applications with evolving transactional data.