Reviewscan API Demo 2019
?
1 like?139 views
Data teams know that a machine learning prototype in Jupyter Notebook is still far from being a deployed product. This summer, our team won the AI challenge of Rotterdam.ai with Reviewscan, a new product that unlocks data value of online reviews with AI. Since then we have moved beyond prototyping, validating our business proposition and building an automated pipeline using GCP. In this talk, I will explore what it takes to build a ML-product by giving a demo of the Reviewscan API. I will also discuss what challenges are faced when operationalising machine learning.
Reviewscan API Demo 2019
- 1. Reviewscan API demo: from prototype to product?
- 2. Better overview of reviews Improved ?lter search by topic Improved top recommendations AISOLUTIONMEETSTHE ENDUSERSNEEDS
- 3. Problem/Solution Low conversion rate due to 39% research / just looking *based on 500 leading international travel brands by SaleCycle (UK)
- 4. Problem/Solution Importance of reviews will increase more than ever 86% read reviews before booking* 91% trust online reviews as much as personal recommendations** Reading average of 10 online reviews *TripBarometer 2017/2018 global study **BrightLocal research USA 2018
- 5. WHATISTHISTALKABOUT ? Our journey ? Prototyping ? Choosing the ML-model and applying NLP ? Challenges ? Validating Business Value Proposition ? Experiments ? Implementation ? Product deployment ? Google Cloud Backend Tooling ? Building an automated pipeline ? Lessons learned
- 6. OURJOURNEY Winning the AI-challenge Python for Data Science Rotterdam.ai at CGI Summer 2019 Lu Zheng New Business Data Engineer Elien van Riet Analytics Translator Data Scientist David Fortini Machine Learning Engineer Xu Zhang PhD Data Scientist
- 12. SENTIMENTANALYSIS ? Opinion miner ? Extract different polarity i.e. positive or negative ? Classifying the reviews based on the sentiment expressed by the reviews
- 14. BEFOREANDAFTER
- 16. OVERCOMING Review data access NLP model loosing relevant information by ? Filtering out stopwords too discriminative: Only nice thing is the nearby park. ? Bigrams and trigrams help to better understand reviews CHALLENGES
- 18. 20+ interviews + desk research Multivariate Testing Landing Pages Feedback Integration & Pricing testing
- 20. Better conversion API Keyword Extractor Dashboard Recommender system More customer value Latest NLP applications Visual insights for meetings Trained model Psychological blind spot of UX Scraped data Deployed API On top of customer trends Customer drop off Lots of untapped big data
- 21. Better overview of reviews Improved ?lter search by topic Improved top recommendations IntegratedAPI incustomerjourney
- 23. MONTHLY FEE Pro API Keyword ranking Premium Customer care Dashboard Enterprise Custom dashboard Recommender filter CUSTOMERCENTRICBUSINESSMODEL
- 25. CHOOSINGTHERIGHTPLATFORM Documentation up to date & easy to find Scaling possibility: pay as you go Clear architecture & transparant service listings and applications
- 26. ٴմ±۰ٴմ
- 27. GCP
- 29. 5LESSONSLEARNED 1. Getting started is more important than being right 2. Building a model that is not too discriminative 3. Choosing the right Cloud Platform for optimal ROI 4. Structured lean startup methodology 5. Diverse team??#$%?
- 30. DOYOUWANTMORECUSTOMERVALUE? Dont worry. We can do it for you.