Temporary rules of retail product sales time series based on the matrix profile

Abstract

Correlation analysis in the retail industry mainly involves market basket analysis. This kind of correlation analysis of retail product sales does not reflect the information regarding the time or quantity of the sales. Product sales datasets contain rich information about the correlations between different products at different times. The co-occurrence of similar sales subsequences reveal that product sales are correlated in a specific time period. Therefore, searching for similar co-occurrence patterns can help analyze the temporary correlations between products. The search for similar subsequences can be viewed as motif discovery in time-series datasets. In the field of motif discovery, the matrix profile (MP) provides an overwhelming advantage in detecting motifs. In this study, our aim is to discover motifs using MP, and hence, analyze the temporary sales correlations between products. The results of our numerical experiments indicate what products customers will purchase at what time. As opposed to strong association rules, we name the correlation rules in this work as temporary rules (TRs). Our results also show that customers’ preferences are not stable and change with time. In the retail industry, TRs can help business owners make suitable product promotions at appropriate times. Moreover, our analysis demonstrates that TRs can extract more interesting information and patterns than mining with association rules.