How to categorise 100K search queries in 15 minutes - MeasureFest
- 1. How to categorise 100,000 queries in 15 minutes Richard Lawrence Rise at Seven @richlawre
- 2. WHAT YOU WILL LEARN
- 3. How to grab thousands of queries
- 4. How to categorise them quickly
- 5. How to train a text classifier
- 6. How to make a tool anyone can use to do all this
- 7. How to gain useful insight
- 8. Loads of code examples throughout!
- 10. GRABBING THE QUERIES ● Set up the API: bit.ly/gsc-api- setup USING PYTHON
- 11. GRABBING THE QUERIES ● Use this Colab notebook: USING PYTHON bit.ly/gsc-api-python
- 12. GRABBING THE QUERIES WHAT YOU GET ● List of queries and associated metrics
- 13. ● Use this dashboard: USING DATA STUDIO bit.ly/aleyda-gsc- dash GRABBING THE QUERIES
- 16. CATEGORISING QUERIES THREE POTENTIAL WAYS Keyword mentions Topic similarity Intent
- 18. BY KEYWORD MENTIONS USING AN APRIORI ALGORITHM Creates categories for commonly mentioned keywords QUERY KEYWORD MENTION CATEGORY BUY SHIRTS SHIRTS SHIRTS IDEAS SHIRTS SALE
- 19. BY KEYWORD MENTIONS USING AN APRIORI ALGORITHM Specify number of keywords needed to create a category QUERY KEYWORD MENTION CATEGORY BUY SHIRTS SHIRTS SHIRTS IDEAS SHIRTS SALE
- 20. BY KEYWORD MENTIONS USING AN APRIORI ALGORITHM Specify around 1,000 for 100K queries QUERY KEYWORD MENTION CATEGORY BUY SHIRTS SHIRTS SHIRTS IDEAS SHIRTS SALE
- 21. BY KEYWORD MENTIONS A FULL EXPLANATION bit.ly/apriori- explanation
- 22. bit.ly/apriori-colab ● Use this Colab notebook: GETTING HANDS ON BY KEYWORD MENTIONS
- 24. EXAMPLE TOPIC CLUSTERS BY TOPIC SIMILARITY Cats Kittens Feline Dogs Puppies Canine
- 25. THE PROCESS BY TOPIC SIMILARITY Turn words into numbers plotted on a graph by meaning Find clusters of words on graph to represent topics Reduce dimensions of graph
- 26. TURNING WORDS INTO NUMBERS Example embeddings: ● BERT ● Word2vec ● GloVe BY TOPIC SIMILARITY
- 27. REDUCING DIMENSIONALITY Example algorithms: ● PCA ● SVD ● UMAP BY TOPIC SIMILARITY
- 28. CLUSTERING INTO TOPICS Example algorithms: ● K-Means ● BIRCH ● HDBSCAN BY TOPIC SIMILARITY
- 29. GETTING HANDS ON BY TOPIC SIMILARITY ● Use this Colab notebook: bit.ly/topic-similarity
- 30. USING BERTOPIC FOR ALL ● Embeddings: BERT ● Reduce dimensions: UMAP ● Clustering: HDBSCAN ● Also summarises topics! BY TOPIC SIMILARITY
- 31. GETTING HANDS ON BY TOPIC SIMILARITY ● Use this Colab notebook: bit.ly/bertopic
- 32. BY INTENT
- 33. BY INTENT SOME EXAMPLE INTENT CLUSTERS Definition Intro Explain Setup Create Make
- 34. BY INTENT THE PROCESS QUERY KEYWORD MENTION CATEGORY BUY SHIRTS SHIRTS SHIRTS SALE SHIRTS IDEAS Categorise by keyword mention with the apriori algorithm
- 35. BY INTENT THE PROCESS QUERY KEYWORD MENTION CATEGORY TF-IDF WORD BUY SHIRTS SHIRTS BUY SHIRTS SALE SALE SHIRTS IDEAS IDEAS Use TF-IDF to find rare words at the category level
- 36. BY INTENT THE PROCESS QUERY KEYWORD MENTION CATEGORY TF-IDF WORD CLUSTER BUY SHIRTS SHIRTS BUY PURCHASE SHIRTS SALE SALE SHIRTS IDEAS IDEAS INSPIRATION Cluster TF- IDF words by similarity
- 37. GETTING HANDS ON ● Use this Colab notebook: bit.ly/intent-cluster BY INTENT
- 39. TRAINING A CLASSIFIER ● Teach with labelled data & it then labels new data HOW IT WORKS
- 40. Yes No ● Keyword mentions ● If you make any manual changes to labels ● Similarity (built in) ● Intent (built in) YOU MIGHT NOT NEED ONE TRAINING A CLASSIFIER
- 41. TRAINING A CLASSIFIER WHAT YOU CAN USE IT FOR ● Classifying new Search Console data
- 42. TRAINING A CLASSIFIER WHAT YOU CAN USE IT FOR ● Classifying new keyword research QUERY CLASSIFICATION SHIRT PRICES PURCHASE SHIRT SIZES GUIDANCE SHIRT COLOURS INSPIRATION
- 43. Train neural network with training set Use test set to evaluate performance Turn words into word embeddings Split labelled data into test & training sets TRAINING A CLASSIFIER THE PROCESS
- 44. USING BERT ● You use its word embeddings to better understand the labels TRAINING A CLASSIFIER
- 45. Pros Cons ● Accurate (fine tune) ● Can use smaller dataset ● Requires GPUs ● Takes a bit longer USING BERT TRAINING A CLASSIFIER
- 46. ● Use this Colab notebook: GETTING HANDS ON bit.ly/bert-classifier TRAINING A CLASSIFIER
- 47. REMEMBER TO USE GPUS FOR BERT! Runtime > Change runtime type TRAINING A CLASSIFIER
- 48. Pros Cons ● No GPUs required ● Bit quicker USING FASTTEXT ● Less accurate ● Requires more data TRAINING A CLASSIFIER
- 49. GETTING HANDS ON ● Use this Colab notebook: bit.ly/fasttext-classifier TRAINING A CLASSIFIER
- 50. MAKING A TOOL ANYONE CAN USE
- 51. MAKING A TOOL ● Easily put a front end on your code with Python INTRODUCING STREAMLIT
- 52. MAKING A TOOL ● Get started here: GET STARTED WITH STREAMLIT bit.ly/start-streamlit
- 56. EXAMPLE INSIGHT
- 57. EXAMPLE INSIGHT
- 58. SUMMARY
- 59. Use the GSC API or GDS to grab thousands of queries
- 60. Categorise them by keyword mention, topic similarity or intent with Python
- 61. Train a classifier to label future data (but you might not need to)
- 62. Use Streamlit to build a tool & allow anyone to follow this workflow
- 63. Gain insight on how people search at different stages of their journey
- 65. THANKS