Sitemap
The Airbnb Tech Blog

Creative engineers and data scientists building a world where you can belong anywhere. http://airbnb.io

Discovering and Classifying In-app Message Intent at Airbnb

11 min readJan 22, 2019

--

Get embraced by plenty of natural light, brick, and plant in our new office in downtown Seattle.
Figure 1: A concept that illustrates a guest asking a host for dining recommendations nearby.

Identifying Message Intent

Intent Discovery

Figure 2: A graphical model representation of LDA by David Blei et al. along with the joint probabilities of the observed (shaded nodes) and hidden (unshaded nodes) units
Figure 3: A 2D visualization of inter-topic distances calculated based on topic-term distribution and projected via principal component analysis (PCA). The size of the circle is determined by the prevalence of the topic.

Labeling: From Unsupervised to Supervised

Intent Classification with CNN

Figure 4: Illustration of a CNN architecture for sentence classification from Ye Zhang et al.
Most similar words for “house” generated by word2vec models trained without / with extra preprocessing steps
Table 1: Comparison on overall accuracy between Phase-1&2, Phase-1 Only, and Predict by Label Distribution. Pre-trip: before a trip starts. On-trip: during a trip.
Figure 5: The normalized confusion matrix for the on-trip model results
Table 2: Example categories that are well predicted
Table 3: Example categories that are not so well predicted

Productionization — Online Serving

Figure 6: The offline training & online serving workflow of Phase II.

Applications

Conclusion

Acknowledgement

--

--

The Airbnb Tech Blog
The Airbnb Tech Blog

Published in The Airbnb Tech Blog

Creative engineers and data scientists building a world where you can belong anywhere. http://airbnb.io

Michelle (Guqian) Du
Michelle (Guqian) Du

Written by Michelle (Guqian) Du

Data Science & Machine Learning @ Airbnb

Responses (6)