Classifying Search Queries Without User Click Data
summary
Claims a pipeline that classifies queries using LSTM and convolutional nets plus XGBoost and word2vec features derived from SER titles and Wikipedia categories, avoiding no user click data. Trains on ~1 million queries, maps queries and ~30,000 categories into embeddings, and reports ~62% validation accuracy for the large LSTM while noting one epoch takes ~2 hours on a GPU. Also builds a fast rule+model stack that labels millions per hour and outperforms traditional log‑based heuristics.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 1:05Classify queries without logs
- 3:57Business need for classification
- 5:00Query category types
- 7:38Rule-based classification
- 8:43Data and features needed
- 10:10Word2vec for query embeddings
- 12:40Boosting with CNN model
- 13:46LSTM for Wikipedia categories
- 15:52LSTM model limits
- 17:18User intent and buying cycles
- 20:27Convolutional neural network for classification
- 22:14Model architecture details
- 24:51Results and examples
- 26:35Conclusions and future work
- 27:23Q&A
related talks
43mSearch Logs + Machine Learning = Auto-Tagged Inventory
31mSponsored Session: From Clicks to Rankings: Building a Learning-to-Rank Pipeline with…
47mLearning to rank search results
31mWhen humans teach machines
34mIs That a Duplicate Quora Question?
22mImplementing a Text Classifier based on Lucene & LIBSVM
01h 30mBuilding a Search Engine
26mAdvancing Multi-Modal Search Capabilities in Search Pipeline