Hierarchical clustering in Python and beyond
Download as PPTX, PDF18 likes10,781 views
The document discusses hierarchical clustering techniques in Python, covering theory, practical applications, and relevant methodologies for data analysis. It highlights the importance of various clustering methods, metrics, and considerations for achieving efficient results, particularly in relation to search keyword analysis. Additionally, it emphasizes the role of visualization tools and the necessity of preprocessing data for effective clustering outcomes.
Hierarchical clustering in Python and beyond
- 1. Hierarchical clustering in Python & elsewhere For @PyDataConf London, June 2015, by Frank Kelly Data Scientist, Engineer @analyticsseo @norhustla
- 2. Hierarchical Clustering Theory Practice Visualisation Origins & definitions Methods & considerations Hierachical theory Metrics & performance My use case Python libraries Example Static Interactive Further ideas All opinions expressed are my own
- 3. Who am I? All opinions expressed are my own
- 5. Clustering is an unsupervised learning problem "SLINK-Gaussian-data" by Chire - Own work. Licensed under CC BY-SA 3.0 via Wikimedia Commons - https://commons.wikimedia.org/wiki/File:SLINK-Gaussian-data.svg#/media/File:SLINK-Gaussian-data.svg based on some notion of similarity. whereby we aim to group subsets of entities with one another
- 7. Diverse applications Attribution: stack overflow, wikipedia, scikit-learn.org, http://www.poolparty.biz/
- 8. Two main purposes Exploratory analysis – standalone tool (Data mining) As a component of a supervised learning pipeline (in which distinct classifiers or regression models are trained for each cluster). (Machine Learning)
- 9. Clustering considerations Partitioning criteria (single / multi level) Separation Exclusive / non- exclusive Clustering space (Full-space / sub-space) Similarity measure (distance / connectivity)
- 10. Use case: search keywords RD P P P KW KW KW KW KW CP CP KW KW KW The competition! KW KW CP CD You Opportunity! CD = Competing domains CP = Competitor’s pages RD = Ranking domain P = Your page KW = Keyword
- 11. ….x 100,000 !!
- 12. Use case: search keywords KW…so we have found 100,000 new ‘s – now what? How do we summarise and present these to a client?
- 13. Clients’ questions… • Do search categories in general align with my website structure? • Which categories of opportunity keywords have the highest search volume, bring the most visitors, revenue etc.? • Which keywords are not relevant?
- 15. Requirements • Need: visual insights; structure • Allow targeting of problem in hand • May develop into a semi- supervised solution
- 16. • High-dimensional and sparse data set • Values correspond to word frequencies • Recommended methods include: hierarchical clustering, Kmeans with an appropriate distance measure, topic modelling (LDA, LSI), co-clustering Options for text clustering?
- 18. 2 types Agglomerative Divisive Deterministic algorithms! Attribution: Wikipedia
- 19. Agglomerative Start with many “singleton” clusters … Merge 2 at a time continuously … Build a hierarchy Divisive Start with a huge “macro” cluster … Iteratively split into 2 groups … Build a hierarchy
- 20. Agglomerative method: Linkage types • Single (similarity between most similar – based on nearest neighbour - two elements) • Complete (similarity between most dissimilar two elements) Attribution: https://www.coursera.org/course/clusteranalysis
- 23. Agglomerative method: Linkage types Average link ( avg. of similarity between all inter-cluster pairs ) Computationally expensive (Na*Nb) Trick: Centroid link (similarity between centroid of two clusters) Attribution: https://www.coursera.org/course/clusteranalysis
- 24. Ward’s criterion • Minimise a function: total in-cluster variance • As defined by, e.g.: • Once merged, then the SSE will increase (cluster becomes bigger) by: https://en.wikipedia.org/wiki/Ward's_method
- 25. Divisive clustering • Top-down approach • Criterion to split: Ward’s criterion • Handling noise: Use a threshold to determine the termination criteria Attribution: https://www.coursera.org/course/clusteranalysis
- 26. Similarity measures This will certainly influence the shape of the clusters! • Numerical: Use a variation of the Manhattan distance (e.g. City block, Euclidean) • Binary: Manhattan, Jaccard co-efficient, Hamming • Text: Cosine similarity.
- 27. Cosine similarity Represent a document by a bag of terms Record the frequency of a particular term (word/ topic/ phrase) If d1 and d2 are two term vectors, …can thus calculate the similarity between them Attribution: https://www.coursera.org/course/clusteranalysis
- 29. Gather word documents = keyword phrases
- 30. Aggregate search words with URL “words”
- 31. Text clustering: preparations • Add features where possible o I added URL words to my word set • Stem words o Choose the right stemmer – too severe can be bad • Stop words o NLTK tokeniser o Scikit learn TF-IDF tokeniser • Low frequency cut-off o 2 => words appearing less than twice in whole corpus • High frequency cut-off o 0.5 => words that appear in more than 50% of documents • N-grams o Single words, bi-grams, tri-grams • Beware of foreign languages o Separate datasets if possible
- 32. Text preparation
- 33. Dimensionality • Get a sparse matrix o Mostly zeros • Reduce the number of dimensions o PCA o Spectral clustering • The “curse” of dimensionality
- 37. The dendrogram
- 38. Assess the quality of your clusters • Internal: Purity, completeness & homogeneity • External: Adjusted Rand index, Normalised Information index
- 39. Topic labelling
- 41. Life on the inside: Elasticsearch • Why not perform pre-processing and clustering inside elasticsearch? • Document store • TF-IDF and other • Stop words • Language specific analysers
- 42. Elasticsearch - try it ! - • https://www.elastic.co/ • NoSQL document store • Aggregations and stats • Fast, distributed • Quick to set up
- 43. Document storage in ES
- 44. Lingo 3G algorithm • Lingo 3G: Hierarchical clustering off-the-shelf • Built-in part of speech (POS) • User-defined word/synonym/label dictionaries • Built-in stemmer / word inflection database • Multi-lingual support, advanced tuning • Commercial: costs attached http://download.carrotsearch.com/lingo3g/manual/#section.es http://project.carrot2.org/algorithms.html
- 45. Elasticsearch with clustering – Utopia? Carrot2’s Lingo3G in action : http://search.carrot2.org/stable/search Foamtree visualisation example Visualisation of hierarchical structure possible for large datasets via “lazy loading” http://get.carrotsearch.com/foamtree/demo/demos/large.html
- 46. Limitations of hierarchical clustering • Can’t undo what’s done (divisive method, work on sub clusters, cannot re-merge). Even true for agglomerative (once merged will never split it again) • Every split or merge must be refined • Methods may not scale well, checking all possible pairs, complexity goes high There are extensions: BIRCH, CURE and CHAMELEON
- 47. Thank you! A decent introductory course to clustering; https://www.coursera.org/course/clusteranalysis Hierarchical (agglomerative) clustering in Python: http://scikit- learn.org/stable/modules/generated/sklearn.cluster.AgglomerativeClustering.html Recent (ish) relevant Kaggle challenge: https://www.kaggle.com/c/lshtc Visualisation: http://carrotsearch.com/foamtree-overview Clustering elsewhere (Lingo, Lingo3G) with Carrot2:http://download.carrotsearch.com/ Elasticsearch: https://www.elastic.co/ Analytics SEO: http://www.analyticsseo.com/ Me: @norhustla / [email protected] Attribution: http://wynway.com/
- 48. Extra slide: Why work inside the database? 1. Sharing data (management of) Support concurrent access by multiple readers and writers 2. Data Model Enforcement Make sure all applications see clean, organised data 3. Scale Work with datasets too large to fit in memory (over a certain size, need specialised algorithms to deal with the data -> bottleneck) The database organises and exposes algorithms for you conveniently 4. Flexibility Use the data in new, unanticipated ways -> anticipate a broad set of ways of accessing the data