Skip to main content
Uber logo

Schedule rides in advance

Reserve a rideReserve a ride

Schedule rides in advance

Reserve a rideReserve a ride
AI

Applying Customer Feedback: How NLP & Deep Learning Improve Uber’s Maps

October 22, 2018 / Global
Featured image for Applying Customer Feedback: How NLP & Deep Learning Improve Uber’s Maps
Figure 1: Our machine learning algorithm looks at the contact type for tickets, expressed as a UUID, and the textual content before applying Logistic Regression on the concatenated word vector and one hot type vector.
Figure 2: After acknowledging the limitations of the first version of our algorithm, we considered how to apply WordCNN and LSTM to the problem of ticket classification.
EmbeddingsModelAUC ROCAUC PRrecall @ precision = 0.5
TypeTrainable
randomYesLSTM0.7940.5190.528
No0.6280.2970.062
Word2VecYes0.7910.5020.533
No0.7960.5270.538
GloVeYes0.7850.5210.568
No0.7730.4940.519
randomYesWordCNN0.8400.6120.669
No0.8220.5960.608
Word2VecYes0.8490.6200.688
No0.8420.6150.646
GloVeYes0.8360.6010.642
No0.8310.5880.627
randomNoLR0.6360.3350.124
Word2VecNo0.8100.5260.536
GloVeNo0.7880.4860.435
Figure 3: Plotting our model performance, it is clear that WordCNN delivers the best results.
Synonyms of pickupCosine DistanceSynonyms of airportCosine Distance
dropoff0.605international0.340
ministry0.609terminal0.360
37200.616delta0.415
destined0.616mco0.424
pinned0.616jetblue0.432
Synonyms of stCosine DistanceSynonyms of buildingCosine Distance
ave0.183alley0.283
pl0.244apartment0.525
avenue0.277condo0.530
broadway0.299complex0.537
Synonyms of highwayCosine DistanceSynonyms of 4thCosine Distance
freeway0.1755th0.226
interstate0.2486th0.232
ramp0.3039th0.240
Synonyms of northCosine DistanceSynonyms of mapsCosine Distance
south0.223waze0.274
west0.383google0.292
east0.385navigation0.444
mockingbird0.408navigator0.475
Synonyms of sfCosine Distance
oakland0.285
francisco0.333
berkeley0.350
san0.395
sfo0.436
Figure 4: Both version 1, using Logistic Regression, and version 2, using WordCNN, of our algorithm are implemented as an end-to-end Spark pipeline.
Figure 5: We wrap the TensorFlow WordCNN model as a Spark pipeline.
Figure 6: If we can associate a map entity with a trip ID in our database, we can join it with positive tickets and aggregate the tickets, creating higher confidence results.
Chun-Chen Kuo

Chun-Chen Kuo

Chun-Chen Kuo has been a software engineer at Uber since early 2017, and was an intern in the summer of 2016. His engineering and research interests center around machine learning.

Livia Yanez

Livia Yanez

Livia Yanez has been a data scientist at Uber since early 2017. She is passionate about applying machine learning to a wide variety of problem spaces, and to use a data-driven approach to identify how we can improve maps at Uber.

Jeffrey Yun

Jeffrey Yun

Jeffrey Yun was a software engineering intern on the Maps team in the summer of 2018. He is currently pursuing his B.S. in Computer Science and Mathematics at UCLA. His engineering interests center around algorithm design, artificial intelligence, and machine learning.

Posted by Chun-Chen Kuo, Livia Yanez, Jeffrey Yun

Category: