Clustering Algorithms are promising for identifying the classes of spatial databases. In this study, an attempt has been made to obtain the model corresponding to the highest performance by tuning it for efficiency for the dataset originating from the work of Jackson et. al. This has been achieved by the means of pre-processing the data, standardising, and feature reducing.
The imputation techniques such as MICE, feature reduction methods like Principal Component Analysis, are the highlights of aforementioned stage. The algorithms employed for clustering are based of an array of diverse utilisation. From the measure of distances in k-means, to their similarity in Affinity Propagation, a vast majority of patterns have been attempted to be identified. All this yields the labels of the data corresponding to the clusters obtained.
The performances of the algorithms are then compared by using internal evaluation. A comparative analysis, to the source of the data is also done, in turn, indicating of a better model for the dataset. The labels of the data can be used further for developing a better inventory policy.
This study discusses an application of various clustering algorithms to solve Aggregation problem in order to obtain the best fitting algorithms for the dataset. Here, the "ground truth" is not known, algorithms are compared based only on the internal evaluation, making this study a descriptive clustering analysis.
Since SKU’s groups should take into account all the attributes having a sufficient impact on the certain inventory operation, features such as information beyond inventory cost and volume used in classical ABC analysis are included. All the necessary calculations and data transformation, are performed by using Python 3.10 and extensively a free machine learning library "scikit-learn" as a core.
The aim of this study is to obtain a better performing algorithm corresponding to the data obtained.
The objectives of the study are as follows:
- Implementing various clustering algorithms with varying approaches to obtain a list of labelled data linked to clusters
- Utilizing better programming practices to achieve consistency of solutions, by reducing redundancy
- Adapting better practices and algorithms prior to clustering implementation to refine the data for analysis,
- Comparing the performances of the clusters obtained in this study, by using internal evaluation methods,
- Comparing the performances of the algorithms with the benchmarks set by the study of Jackson et al.
- Deploying the model made using the framework, with a satisfactory level of robustness, in order to ensure stability of outcomes, even after multiple iterations
This work is heavily influenced by the work of:
Jackson, Ilya & Avdeikins, Aleksandrs & Tolujew, Juri. (2019). "Unsupervised Learning-Based Stock Keeping Units Segmentation: Selected Papers from the 18th International Conference on Reliability and Statistics in Transportation and Communication", RelStat’18, 17-20 October 2018, Riga, Latvia. 10.1007/978-3-030-12450-2_58.