Despite the high number of prioritization techniques in the literature, prioritizing customers is still one of the most difficult tasks for marketers. A prioritized customer list consists of a selected subset of customers that should receive more marketing efforts than other customers (Figueiredo et al., 2005). The prioritization problem arises in numerous areas such as financial services, telecommunications service providers, retail trade and public utilities. In prioritized contact centers the prioritization technique filters an incoming stream of contacts prioritizing the high value hot contacts in order to maximize revenue and improve customer satisfaction (Figueiredo et al., 2005).
In prioritization problems a prioritized customer list is a subset of a set of customers that should receive more marketing efforts than other customers. To solve this problem, one administration criterion is needed to be defined. A prioritized list may have two different objectives: maximizing profit or minimizing costs.
In order to build a prioritized list from its components several techniques can be used. These techniques are categorized by three main groups:
Using this criteria, various prioritization techniques can be classified.
Among prioritization techniques a prioritized list for planned marketing efforts is perhaps the most popular (Arya & Dutta, 2003). This prioritization technique has been used when prioritizing customers based on their time of making the next purchase, in order to maximize the expected profit (Lenkeit et al., 2002) or when prioritizing wireless subscribers in order to minimize churn (Garcia-Burgos & Macdonald, 2010). Among other techniques, one can cite: data envelopment analysis (DEA), linear programming with objective function by using variables’ total revenue and number of advertisement views (Liu et al., 2013; Shimojima et al., 2001), machine learning models such as Naïve Bayes, linear SVMs and random forests in prioritizing advertisements in a digital ad campaign (Sun et al., 2016). Another method is the prioritization of patients in a clinical trial based on drug efficacy having an outcome probability to survive one year after treatment. This prioritization technique was also used in order to find the best partitioning strategy for next patient enrollment into clinical trials by minimizing the overall expected cost (Lenkeit et al., 2002). In all these cases, prioritization techniques aim at finding an optimal solution that maximizes or minimizes either utility function or objective function.
In this article, two prioritization techniques are presented. First, data prioritization is proposed with a focus on the minimum amount of data. The second prioritization technique focuses on prioritization by means of prioritized complete datasets.
RICE PRIORITIZATION
The Rice prioritization method is an inductive prioritization technique for prioritizing variables used in regression models (Raoul and Zamar, 2015). The method uses the following ranking: the variable with highest predictive power gets a rank of 1; then consecutive ranks are assigned until all variables get a rank. This prioritization technique works well when there is no clear-cut priority between two competing variables and it can be extended to more than one input variable. It was found to outperform existing prioritization methods such as stochastic gradient boosting (Friedman et al., 2009) or Bayesian optimization (Salvatier et al., 2011) which are popularly used to select features for data-driven modeling.
The Rice Prioritization method described above was developed by Raoul et al. in 2015, but it is worth mentioning that prioritizing variables for the sake of performing a regression model (or other machine learning techniques) has been addressed before. The CRISPROT prioritization tool (http://mendel.stanford.edu/SidowLab/crisprot/) performs rule based prioritization of drug targets to produce prioritized lists, similarly variable prioritization via linear programming (VPLP), random forests (RF), and support vector machines (SVM) has also been explored before. However these methods use different priorities for prediction, i.e., prioritizing genes that are most differentially expressed in an experiment, or prioritizing genes that have the highest correlation to a given category (such as cancer).
Performance comparison of lead candidate approaches for large-scale phenotype prediction, 2017 – A benchmark on three publicly available sets of human gene expression and metabolic data is reported in this paper with the aim to assess the performance of a set of input parameters that might be used by lead candidate approaches for phenome-wide association studies.
The prioritization method described in this article is unique because it uses prioritized lists of variables with known phenotypic impacts on traits of interest, allowing for prioritization of variants in causal regulatory regions. This allows one to train models more efficiently when prioritizing only certain key variables without sacrificing too much accuracy.
Rice prioritization can be performed using various software programs such as WebGestalt and gene prioritization tools available through cloud computing platforms using Amazon Web Services.
Rice score prioritization is a prioritization method used for prioritizing genetic variants with unknown causal effects on traits of interest.
The method uses pre-computed prioritized lists of factors with known phenotypic impacts to prioritize causative variants and estimate the proportional contribution of every single variant to variation in traits of interest.
In addition, rice prioritization will predict the effect of each causal variant into a prioritized list, allowing for model training without sacrificing too much accuracy. In contrast, other methods such as rice score prioritization do not predict the relative weights or relative contribution of specific variants but instead predict their statistical significance.
