Interestingness of interestingness measures
Submitted by Simrat Hanspal (@simrathanspal) on Sunday, 30 April 2017
Section: Full talk for data engineering track Technical level: Advanced
Analysis of relationship between entities is at the heart of data mining problems. There are many metrics used for association mining like support, confidence, lift, mutual information etc. However many of these measures provide conflicting results about the interestingness of the association. Therefore it becomes very important to understand how to evaluate metrics for an application.
A fundamental step in pattern mining of transactional datasets is the extraction of frequent and interesting itemsets - a set of entities connected by the frequently occurring relationships between them. For instance, identifying from housing purchase data the correlations to age and income groups they belong to can lead to explainable relationships between the different data points which in turn leads to knowledge discovery. This kind of analysis is often confounded by the presence of spurious correlations and data sparsity especially in e-commerce where much of the traffic is often directed to a small percentage of the catalog.
Data mining is the process of discovering previously unknown interesting relationships which can be used to increase customer engagement and boost the sales. One very common and interesting example of this would be the discovery of strong link between diapers and beers from transactional data of walmart. This association was definitely not intuitive because you would expect customers to buy other baby products along with diapers. It was then suspected that these purchases were made by fathers baby sitting on weekends. Even Though this was found to be a strong correlation we must not forget that correlation doesn’t imply causation. A strong correlation is an interesting insight from user behaviour which can be fed back to increase customer interaction.
The goal of this talk is to set the stage for mining of associations, present commonly used techniques, followed by objective measures of interestingness of associations and finally, present an analysis on real customer datasets. In the case of sites attracting large amounts of customer traffic the main challenge is to make this process computationally effective.
We will pay special attention to low-traffic situations so that we can mine for interesting patterns without requiring large amount of data to support it. One way in which we achieve this is to build higher-level abstractions for entities such as the brand of the product rather than the product itself. In addition to this domain knowledge and transitive connections can be used to help with the cold-start problem.
Brief outline of the talk -
What is association mining
Motivation - Market basket analysis
Association mining: Basic concepts
Frequent ItemSet Generation
- Apriori algorithm
- FP growth algorithm
Evaluation of association patterns
Measures of interestingness
Case studies to compare interestingness measures
Challenges of mining sparse dataset
Unearthing unseen gems
Motivation for mining transitive or indirect association
Simrat is a Data Scientist, Engineering Ninja and Inspector Gadget at Mad Street Den. She builds data platforms and models to make sense of user and product data in e-commerce online retail.