Identification of Accident Path Elements using Supervised Learning: Observation of Diminishing Marginal Accuracy while using Cosine Similarity

With increasing outsourcing of project activities to contractors in process industries, ensuring contractor safety is becoming one of the top priorities of organizations. This study aims at identifying the major Hazardous Elements (HE) and Initiating Events (IE) of incidents occurring at a steel plant in India with respect to activities and safety incidents involving contractors. Accordingly, a supervised text classification model is used. First a training dataset is manually annotated to identify the HEs and IEs and then HE and IE are annotated for a test document based on cosine similarity values calculated between the test document and the training documents. Accordingly, 11 unique HEs and IEs have been identified in contractor safety dataset. This study also attempts to identify the optimal way to use cosine similarity to classify a test dataset. First, for each test document, top N training documents in decreasing order of cosine similarity values are considered for classification and the HE and IE as displayed by majority of those training documents are taken as the HE and IE of the test document. The percentage of accurately classified test documents is then calculated. In the next step, the study monitors the change in accuracy percentage as N is stepwise increased. The results show that on increasing N, the accuracy increases but the marginal increase starts diminishing after a threshold.