IC-SPLC-v2: A Supply Chain Structure Neo4j Data

Abstract

This dataset, IC-SPLC (Integrated Circuit Supply Chain) version 2, is a year-labelled directed supply-chain network covering the period 2013 to 2025, released in anonymized form. It was assembled to address a specific research hypothesis: that publicly available web text, read and structured by a large language model (LLM) workflow, can yield a usable map of supply-chain relationships without depending on the disclosure channels that existing commercial databases rely on. The released data form a directed property graph. It contains only one node label EntityObj representing firms and other organizations, and directed edges of one type, SupplyProductTo, whose direction is from supplier to customer. Every edge carries the year in which the relationship was observed. Nodes carry a country field, a first-level industry classification, a second-level industry classification with alternative candidate values in other_possible_industry_2, a three-level product-category classification available, and a 128-dimensional representation vector embedding the information of firm descriptions. All categorical values are anonymized letter, and the second-level industry codes and third-level product-category codes are shared between the primary and the alternative-candidate fields, so that identical codes in those two pairs of fields refer to identical values. In other cases, same letter does not mean the same value. The pipeline was evaluated against human annotation. Entity recognition reaches an accuracy of 0.9297, and relation extraction reaches a precision of 0.4524 and a recall of 0.6214. The re-verification stage is what makes the network usable, since it raises relation precision from 45.2 percent to 84.3 percent at a recall of 85.8 percent; the system as a whole therefore attains a precision of about 81.6 percent and a recall of about 53.3 percent. Finally, the version deposited here is anonymized. Firm names and all identifying fields have been removed, the values of the categorical attributes have been replaced by frequency-ranked codes whose mapping to the original values is not published, rare and missing values have been turned into blanks, the representation vectors have been randomly permuted along their dimensions, a transformation that preserves all similarities and distances between nodes while removing the meaning of individual dimensions. The value of the dataset therefore lies in network structure and attribute distributions rather than in firm-level profiles. The data need to be loaded to Neo4j before using. We recommend Neo4j community version 5.19.0 since it is where we produced it.