Abstract
Product development in the food industry is iterative and relies largely on implicit expertise. An Austrian food-tech company specialising in baked goods documents every
development cycle, yet this archive is not used analytically. This thesis examines what
can be extracted from such records, drawing on 222 project proposals, 1492 trial proto
cols and 1939 recipes covering white bread and pastries from 2010 to 2015, and treats
their quality as an object of study rather than a precondition.
Scored against an established five-dimension data quality framework, the free-text project
goals reached 99 of 100 and the protocol and recipe data 77.76, with the deductions
concentrated in Usability, where no column documentation or metadata exists, and in
Reliability, where recipe identifiers resolve ambiguously. Two pipelines were then imple
mented. The first represents the project goals as TF-IDF matrices and transformer-based
sentence embeddings, reduces them via truncated SVD, NMF and PCA, and clusters
them into the archetypes replication and improvement against 215 manual labels. The
second treats each ingredient in turn as a regression target, fits Partial Least Squares
and XGBoost estimators across four data subsets and two amount representations, and
converts permutation feature importances and partial-dependence directions into ingre
dient interaction networks, assessed by cross-model agreement, by bivariate correlation,
and by a senior domain expert.
Clustering did not recover the archetypes. Across 80 configurations the highest agree
ment reached an Adjusted Rand Index of 0.53 at a Silhouette Score near zero, because
replication is expressed through a compact and stable vocabulary whereas improvement
is defined largely by its absence. Interaction inference fared better: 86.25% of the 80
top-ranked interactions agreed in sign with the empirical data and 72.5% did so signif
icantly, yet the expert rated only 28 of 67 distinct ingredient pairs as technologically
plausible, and the linear estimator matched or exceeded the non-linear one throughout.
Data quality proved diagnostic for both outcomes. Records that document a process
rather than probe it do not permit co-variation arising from standard operating pro
cedure to be separated from functional interaction, so the networks serve to prioritise
candidates for experimental follow-up rather than as a validated map of formulation
behaviour.
| Date of Award | 2026 |
|---|---|
| Original language | English (American) |
| Awarding Institution |
|
| Supervisor | Stephan Winkler (Supervisor) |
Studyprogram
- Data Science and Engineering
Cite this
- Standard