Use a Knowledge Graph to Fill Missing Product Attributes

A retailer has 10,000,000 products. About 35 percent lack the "material" attribute and 20 percent lack "compatible with", which hurts search filters. Product descriptions are messy text. A teammate proposes training TransE on the product knowledge graph to predict the missing tails.

1. Write the triple formulation and the TransE score, and evaluate a tiny example by hand.
2. Say what each of these gives you: TransE completion, a text classifier, an LLM extraction pipeline. Which would you try first?
3. Describe a correct evaluation protocol and compute MRR for ranks (1, 2, 4).
4. Name the data-quality risks of the graph and how you would limit them.