Productionizing an unsupervised machine learning model to...
summary
Claims a usable pipeline that clusters open-text feedback so analysts avoid reading ~5,000 items a month, built with open-source Python and deployed in Docker behind an HTTPS endpoint. Evaluation uses a SentEval-style suite on downstream tasks and surprisingly TF‑IDF + truncated SVD + hierarchical clustering beats heavier models in stability and explainability; BERT shows stronger embeddings but inconsistent clusters in practice. Production tradeoffs include per-request retraining, DB and request-size limits, multilingual stopword handling, and a six-month MVP that customers accept.
AI-written summary · about
chapters (11)
- 0:00Feedback insights project
- 1:08Customer feedback analysis
- 3:38Feedback cloud solution
- 5:25Unsupervised ML model
- 9:21Evaluating embeddings with SentEval
- 12:36TF-IDF for production
- 15:05Demo: feedback cloud notebook
- 17:42Model to production pipeline
- 20:33Production challenges
- 23:21Business value reflections
- 23:42Q&A
related talks
18mBeyond Inference: Bringing ML into Production
25mContinuous ML Applications in Production
35mGetting Machine Learning Models Ready For Production Using Python
31mWhen humans teach machines
45mTransformer based clustering: Identifying product clusters f...
01h 08mGoing multi-modal
37mAutomated Machine Learning: Building a Pipeline
21mFrom 0 to a crowd sourced neural network in a jiffy