S8768- Contextual Product Search With Vectorized Part Descriptions - PowerPoint PPT Presentation

S8768- Contextual Product Search With Vectorized Part Descriptions Danny Godbout | Data Scientist

Application Engineering Part Search

Which Vehicles Were Built With Aerodynamic Fairings?

N33-1068-199AC3 FAIRING, ROOF J83-1001-193AD3 LIGHT , ROOF S78-1201-3117 SEAT , DRIVER 15-1068-199AC3 FAIRING, ROOF 15-6023-5J22K FRNG, ROOF BRACKET N33-1201-3117 FRNG, FUEL TANK AACS16034-033123 SEAT , FASTENER M12

N33 N33-1068-199AC3 FAIRING, ROOF J83-1001-193AD3 LIGHT , ROOF S78-1201-3117 SEAT , DRIVER 15-1068-199AC3 FAIRING, ROOF 15-6023-5J22K FRNG, ROOF BRACKET N33-1201-3117 FRNG, FUEL TANK AACS16034-033123 SEAT , FASTENER M12

FAIRING N33-1068-199AC3 FAIRING, ROOF J83-1001-193AD3 LIGHT , ROOF S78-1201-3117 SEAT , DRIVER 15-1068-199AC3 FAIRING, ROOF 15-6023-5J22K FRNG, ROOF BRACKET N33-1201-3117 FRNG, FUEL TANK AACS16034-033123 SEAT , FASTENER M12

FAIRING, FRNG, N33, … N33-1068-199AC3 FAIRING, ROOF J83-1001-193AD3 LIGHT , ROOF S78-1201-3117 SEAT , DRIVER 15-1068-199AC3 FAIRING, ROOF 15-6023-5J22K FRNG, ROOF BRACKET N33-1201-3117 FRNG, FUEL TANK AACS16034-033123 SEAT , FASTENER M12

FAIRING, FRNG, N33, … N33-1068-199AC3 FAIRING, ROOF J83-1001-193AD3 LIGHT , ROOF S78-1201-3117 SEAT , DRIVER 15-1068-199AC3 FAIRING, ROOF 15-6023-5J22K FRNG, ROOF BRACKET N33-1201-3117 FRNG, FUEL TANK AACS16034-033123 SEAT , FASTENER M12

Approach Interpret vehicle options like expert reading BoM

Pre- Define Cosine Vectorize Process Kernels Distance

Pre-Processing Pre- Process N33-1078-03123 | FAIRNG, TRUCK ROOF 45” CUSTOMERASPECIAL Base P/N nthreethree FAIRNG, TRUCK ROOF 45” CUSTOMERASPECIAL Vectorize Lowercase nthreethree fairng, truck roof 45” customeraspecial Remove Non-Alpha nthreethree fairng, truck roof 45” customeraspecial Define Kernels Jaro-Winkler Match nthreethree fairing truck roof customeraspecial Re-Sample nthreethree fairing truck roof customeraspecial Cosine Distance nthreethree fairing truck

Vectorization- Word2Vec Pre- Process Vectorize Define Kernels Cosine (�) 𝐾 ��� = log(𝑅 � 𝐸 = 1 | 𝑢ℎ𝑓, 𝑛𝑏𝑢 ) + log (𝑅 � 𝐸 = 0 𝑨𝑓𝑐𝑠𝑏, 𝑛𝑏𝑢)) Distance Image Source: “Vector Representations of Words | TensorFlow.” TensorFlow , www.tensorflow.org/tutorials/word2vec.

BlazingText – GPU Acceleration Pre- Process Vectorize Define Kernels Cosine Distance Source: Gupta S., Khare V. “BlazingText: Scaling and Accelerating Word2Vec using Multiple GPUs” MLHPC’17: Machine Learning in HPC Environments , November 12–17, 2017, Denver, CO, USA

Word Vectors Pre- Process Vectorize Define Kernels Cosine Distance

Part Vectors Pre- Process Vectorize Define Kernels Cosine Distance

Kernels Pre- Process Aerodynamic Vectorize Define Attachment Kernels Cosine Distance

Description Matrix ● Search Kernels = Term- Kernel Similarity Pre- Process Roof Vectorize Fairing Define Kernels Bracket Max Pool(Kernel Similarity) = Description Vector Cosine Distance

Part Search Pre- Process Part Vectors Vectorize Define 𝐵 � 𝐶 𝑡𝑗𝑛𝑗𝑚𝑏𝑠𝑗𝑢𝑧 = cos 𝜄 = Kernels 𝐵 | 𝐶 | Search Term: FAIRING AND NOT BRACKET Cosine Distance

Results Match quality, speed

Measures Kappa Sensitivity Fall-Out (TPR) (FPR) Naïve Search 0.28 0.30 0.07 Raw Word2Vec 0.61 0.88 0.21 Kernel Search 1.00 1.00 0.00

Recap  Word Vectors applied to domain-specific vocabulary  Improves search reliability in multi-generational part catalog  GPU acceleration: More iterations in a given timeframe