Compound structural identification for non-targeted screening of organic molecules in complex mixtures is commonly carried out using liquid chromatography coupled to tandem mass spectrometry (UHPLC-HRMS/MS and related techniques). Instrumental developments in recent years have increased the quality and quantity of data available; however, using current data analysis methods, structures can be assigned to only a small fraction of compounds present in typical mixtures. We present a new data analysis pipeline, "MSEI", that harnesses data science methodologies to improve structural identification capabilities from tandem mass spectrometry data. In particular, feature vectors for fingerprint calculation are found directly from tandem mass spectra, strongly reducing computational costs, and fingerprint comparison uses an optimised methodology accounting for uncertainty to improve distinction between matching and non-matching compounds. MSEI builds on the identification of a small number of compounds through current state-of-the-art data analysis on UHPLC-HRMS/MS measurements and uses targeted training and tailored molecular fingerprints to focus identification to a particular molecular space of interest. Initial compound identifications are used as training data for a set of random forests which directly predict a custom 75-digit molecular fingerprint from a vectorised MS/MS spectrum. Kendrick mass defects (KMDs) for peaks as well as "lost" fragments removed during fragmentation were found to be useful information for fingerprint prediction. Fingerprints are then compared to potential matches from the PubChem structural database using Euclidean distance, with fingerprint digit weights determined using an SVM to maximise distance between matching and non-matching compounds. Potential matches are additionally filtered for hydrophobicity based on measured retention time, using a newly developed machine learning method for retention time prediction. MSEI was able to correctly assign > 50% of structures in a test dataset and showed > 10% better performance than current state-of-the-art methods, while using an order of magnitude less computational power and a fraction of the training data.