Complex Network Modeling of Supply & Demand Data: An Application Case in the Plastics Recycling Industry
Stefan Bloemheuvel, Jurgen van den Hoogen, Martin AtzmuellerThe supply chains of raw materials have been studied by several researchers in the last decade. Such chains can often be modeled as complex networks, defined as all stages involved in producing and delivering a product from supplier to customer. Historically, such distribution networks are considered as a series of steps. However, by viewing supply chains as interconnected networks of suppliers, higher-order connections between companies can then arise. For example, network theory was used to map product platform risk by analysing their underlying supply chains, while others modelled supply networks of car-parts for well-known automotive brands.
However, such networks only model from the supplier's point of view. In cases such as the plastics recycling industry, where batches of recycled waste are either demanded or offered, the notion of a buyer-seller marketplace is more appropriate. Nevertheless, due to problematic data collection, research into the supply and demand of recycled plastic is lacking. The sketched problem originates from the absence of information on the flow of materials, since origin, current location and the chemical composition of recycled plastic material (rPM's) supply and demand are not tracked.
In this study, we will collect data from companies in the plastics recycling industry (region = North-West Europe) that contribute to the Di-Plast project. Companies will provide material specifications from both the buyer and seller point of view. We model the companies and their individual batches of plastics as nodes in a network. An edge between two batches is predicted based on the similarity of node-attributes.
Such batches can hold information about the chemical composition of the material, such as: weight, type of plastic (e.g., PET or PP), viscosity, melt temperature and shear rate. With aforementioned variables, possible supply-demand relationships can be predicted. The more similar two batches are, the easier a company could incorporate (with minor alterations, e.g., increasing the viscosity or lowering the melt temperature) the plastic into their production line. Due to the mixed-variable types, the Gower distance was used to measure the similarity of the plastic bags, which combines Euclidean distance and Levenshtein distance for measuring similarities between numerical values and strings.
The network that is then created which is represented by an attributed graph is novel in several ways. First, the initial edges between the nodes and the batches are Query-Based, since they originate from the demand and supply of companies. Second, the predicted edges are not based on structural properties of the graph, but instead on the node attributes. Using node attributes to form the initial edges in the graph helps with combating the cold-start problem. Lastly, the communities that evolve from the network could be used to calculate standards for material properties and could help to automate supply and demand for the plastics recycling industry. Such improvements can potentially help companies to prevent paying for their waste disposal.