grapeVin
- 1. GrapeVin Anasuya Das Insight Data Science
- 2. Can wine recommendations be crowd sourced?
- 3. Can wine recommendations be crowd sourced? 130,000 reviews 82,342 unique wines 655 unique users Images from:
- 4. Can wine recommendations be crowd sourced? http://GrapeVin.us 130,000 reviews 82,342 unique wines 655 unique users Images from:
- 6. Content based recommender Deep ruby color leads to a nose of currant and cedar chest. The complex body contains plum,black cherry and just the right amount of oak.
- 7. Content based recommender Deep ruby color leads to a nose of currant and cedar chest. The complex body contains plum,black cherry and just the right amount of oak. Get most similar wines in latent semantic space
- 8. Content based recommender Deep ruby color leads to a nose of currant and cedar chest. The complex body contains plum,black cherry and just the right amount of oak. Deep ruby color, currant, dark cherry, vanilla, cedar, full bodied, oak Get most similar wines in latent semantic space
- 9. 2/8/2015 mpld3 plot -0.6 -0.4 -0.2 -0.0 0.2 0.4 0.6 0.8 -0.6 -0.4 -0.2 -0.0 0.2 0.4 0.6 0.8 Pinot noir Champagne component3 component 2 Bordeaux Wine reviews cluster by the type of wine
- 10. Preserved're(notopic'organiza(on'in'early'visual'cortex'followin 'implica(ons'for'visual'func(on?' How*does*V1*damage*affect*the*re;notopic*organiza;on*of*spared* visual*cortex?**How*is*this*related*to*spared*visual*func;on?* V2d' V2v' V3v' V4' V3a' hMT+' V3d' pole' calcarine' V1'LO1/LO2' V2d' V2v' V3v' V4' V3a' hMT+' V3d' pole'calcarine' V1' LO1/LO2' V2d' V2v' V3a' hMT+' V3d' pole' calcarine' V1'LO1/LO2' Background& Anasuya&Das1,2,&Elisha&P.&Merriam3,&David&J.&Heeger3,&Krystel&R.& 1Flaum&Eye&Ins+tute&and&2Centre&for&Visual&Science,&University&of&Rochester,&3Center&for&N Damage&to&the&primary&visual&cortex&(V1)&or&its&afferents&produces&a&severe& loss& of& vision& in& the& contralateral& visual& hemifield& (VF),& called& cor+cal& blindness& (CB).& Studies& in& nonGhuman& primates& with& lesions& to& V1& observe& reduced& but& organized& re+notopic& ac+vity& in& the& lesioned& visual& cortex& (Schmid,&2009&&&2010).&These&and&other&studies&suggest&that&residual&visual& processing&is&mediated&by&either&spared&V1&or&extraGgeniculoGcalcarine&input& to&extrastriate&visual&areas&(reviewed&in&Das&and&Huxlin,&2010).&Human&fMRI& studies&of&CB&have&examined&single&subjects,&and&the&lesion&characteris+cs&of& these&subjects&have&varied&across&studies.&The&re+notopic&organiza+on&of&the& damaged&visual&cortex&in&humans&is¬&well&characterized.& Group*1:*Re;notopic*organiza;on*is*preserved*around*lesion** But$extrastriate$cortex$has$greater$representa1on$of$central$field$than$spared$V1(CB8)$& Intact&hemisphere&Damaged&hemisphere& Damaged&hemisphere& Ques;ons& V2d' V2v' V3v' V4' V3a' hMT+' V3d' pole'calcarine' V1' LO1/LO2' The story so far … y*visual*cortex*is*
- 11. Questions?
- 13. Cross validation: How accurately can the star rating of a recommended wine be predicted? red,spicy, oak Train classifier Get similar wines Predict ratings
- 14. Cross validation: How accurately can the star rating of a recommended wine be predicted? Train a Naive-Bayes to learn the user preference of each user. Test on the reviews written by other user for recommended wines
- 15. Cross validation: How accurately can the star rating of a recommended wine be predicted? 2/8/2015 mpld3 plot 1 2 3 4 5 0 10 20 30 40 50 60 Rating prediction accuracy Numberofusers 0.2 0.4 0.6 0.8 1 Train a Naive-Bayes to learn the user preference of each user. Test on the reviews written by other user for recommended wines
- 16. A L G O R I T H M Remove stop words Stem using wordNet Synsets to defeat adverbs Detect language and filter Unicode, HTML scrubbing Cosine similarity in lower dimensional space reviews = n Truncated SVD, a.k.a Latent Semantic Indexing on 132,000 reviews x 20,000 words TfIdf - m columns x11 x12 …. x21 x22 …. x31 x32 …. ~ n x r components ~ x11 x12 …. x21 x22 …. x31 x32 …. Recommend top 10 most similar and highest rated wines
- 17. 2/8/2015 mpld3 plot 1 2 3 4 5 0 10 20 30 40 50 60 Crossvalidation: What is the probability of recommending a wine that is already reviewed P of recommending already reviewed wine Numberofusers 0.2 0.4 0.6 0.8 1 Data: SuperUsers: Top 100 users with the most reviews Wines: Top 100 wines reviewed by SuperUsers Method: For each wine reviewed by SuperUser recommend 20 most similar wines based on remaining 99 users
- 18. 1. convert to lower case 2. remove stop words and punctuation and html code 3. deal with broken unicode characters and replace with plain text 4. detect language and only include reviews in english 5. lemmatize using wordNet- works only on nouns and adjectives 6. do synset and pertainyms to convert adverbs to adjectives 7. use bigrams 8. tokenize using term frequence- inverse document frequency Text processing steps
- 19. 1/29/2015 mpld3 plot 500 1,000 1,500 2,000 2,500 3,000 20 30 40 50 60 70 80 90 Explainedvariance(%) number of components Selecting k components
- 20. 1. Incorporate keyword search 2. Scale up and increase inventory 3. Scrape wine pricing information and local availability 4. Analyze tasting notes by vineyard or geographical region - does soil and climate really impact how wines taste Future directions