Understanding Voice of Members via Text Mining – How Linkedin Built a Text Analytics Engine at Scale
0 likes628 views
This document summarizes how LinkedIn built a text analytics platform to understand member feedback at scale. It discusses how LinkedIn used techniques like text classification, topic modeling, and machine learning to analyze large amounts of unstructured social data from members in order to gain insights into trends, sentiments, and values. The platform was designed for scalability, availability, and ease of use. It provides centralized insights that have improved analytics efficiency and driven business impacts.
Understanding Voice of Members via Text Mining – How Linkedin Built a Text Analytics Engine at Scale
- 1. Understanding Voice of Members via Text Mining – How Linkedin built a text analytics platform at scale Chi-Yi Kuan Weidong Zhang Tiger Zhang
- 2. Who are we? www.linkedin.com/in/chiyikuan Chi-Yi Kuan www.linkedin.com/in/weidongzhang1 Weidong Zhang Tiger Zhang www.linkedin.com/in/tigerzhang • Director, Analytics at Linkedin • Big data evangelist and practitioner • Manager, Analytics Platform & Apps at Linkedin • Build big data and analytics products • Sr. Staff, Analytics at Linkedin • Text mining scientist and big data enthusiast Strata + Hadoop World, 12/8/2016
- 3. Strata + Hadoop World, 12/8/2016
- 4. KnowledgeSchoolsSkillsJobsCompaniesMembers 467M 7M 6M 3B 27k 200k Endorsements Daily posts Strata + Hadoop World, 12/8/2016
- 5. 467M 2B Billions LinkedIn Big Data Strata + Hadoop World, 12/8/2016
- 6. Strata + Hadoop World, 12/8/2016 467+ million members = a lot of data
- 7. Voices: drive actionable intelligence from member voices… What’s trending Products Home Page Mobile Inbox Sentiments Value Props Hire Market Sell Relevance filtering Classification Topic mining Identify content that is relevant to Linkedin brand and products/ services Structuralize unstructured textual data into well-defined categories Find most significant topics and stories in a certain time window Strata + Hadoop World, 12/8/2016
- 8. …creating impact across business metrics Developed game-changing solutions to drive Voice of Member impact Improved analytics efficiency with unstructured data by 20X Drove end-to-end technological integration on big data and embedding NLP solutions Piloting operational solutions to scale advanced analytics impact for broader organization Strata + Hadoop World, 12/8/2016
- 9. LinkedIn Hadoop Ecosystem HDFS Map-Reduce Tez Spark Pig Hive Scalding YARN AZKABAN Strata + Hadoop World, 12/8/2016
- 10. Design Principles for Voices Platform Scalability Availability Easy to Use Process Platform Data Systems Application Framework Kafka, Hadoop Spark Gobblin Elasticsearch NoSQL Phoenix Elasticsearch Highcharts Strata + Hadoop World, 12/8/2016
- 11. E2E Voices Platform Architecture Strata + Hadoop World, 12/8/2016
- 12. Data Processing at Scale – with Generic ETL Strata + Hadoop World, 12/8/2016
- 13. Smart IDs – for Viral Mentions with Threading Strata + Hadoop World, 12/8/2016
- 14. High Availability – through Heterogeneous Data Strata + Hadoop World, 12/8/2016
- 15. Machine learning based analytic engine to surface insights to everyday business users Customized Feeds Central navigation Trending insights Social analytics & topic mining Deep dives Sentiment solutions Strata + Hadoop World, 12/8/2016
- 16. Text mining is a crowded space Strata + Hadoop World, 12/8/2016
- 17. Our solution targets unique use cases for LinkedIn Member info • Identity • Behavior • Social Social data Customer feedback • Customer service • Group updates • Network updates Survey results What’s trending Products Sentiments Value Propositions PYMK Group Home Page Mobile Inbox Identity Network Hire Market Sell Relevance solution Topic mining Text Classification Strata + Hadoop World, 12/8/2016
- 18. ▪ Product insights, launches, and events ▪ Horizontal themes ▪ PR and marketing campaigns ▪ Brand and value ▪ LinkedIn’s strategy, financial performance, international etc. Relevant: Non-relevant: ▪ Status update, e.g. "I posted something on Linkedin"; ▪ Social mentions, e.g. "Please connect with me on Linkedin" or "Follow me on Linkedin"; ▪ Self promoting materials, e.g. “share on LinkedIn” ▪ SPAMs 1) Focusing on relevant data Strata + Hadoop World, 12/8/2016
- 19. Keyword based approach Relevance prediction power Rules 56% Whitelist Blacklist 10% 60% 6% 19% 35% Strata + Hadoop World, 12/8/2016
- 20. Generic text classification framework ▪ Feature generation ▪ Feature selection ▪ Machine learning algorithms: – Naïve Bayes (NB) – Logistic Regression (LR) – Support Vector Machines (SVM) (LibLinear) ▪ Cross-validation and evaluation Applications ▪ LinkedIn relevance ▪ Sentiment analysis ▪ Product categorization ▪ Value proposition classification 2) Leveraging text classification engine Strata + Hadoop World, 12/8/2016
- 21. Machine learning approach increases overall relevance by 40% Relevance prediction power Rules 56% Whitelist Blacklist 6% 19% 40% 100% SVM 35% SVM: great gain in balancing precision and recall Strata + Hadoop World, 12/8/2016
- 22. 3) Enabling topic mining HIGH SPARK Description POS pattern matching Part-of-speech (POS) tagging (Stanford CoreNLP) This is great. … … Topic pruning - Stemming - removing stop words - merging synonyms - clustering (optional) **** ing ****** s = = Topic ranking: TF-IDF weighting and DF ranking Strata + Hadoop World, 12/8/2016
- 23. Trending Insights – identify organic trending topics Didi and Kuaidi merger Product release Strata + Hadoop World, 12/8/2016
- 24. LinkedIn’s customer support has evolved into an intelligence platform… Scaling to have a broader impact across LinkedIn ▪ GCO cases ▪ Issue resolution ▪ Support focused ▪ Internal data (GCO, surveys, site feedback) ▪ App review ▪ LI.com ▪ Social data ▪ Product insight ▪ Member insight ▪ Launch tracking ▪ Social sentiment ▪ Brand tracking ▪ Viral mentions Reactive Multi-channel Intelligent Predictive Support Feedback Insights Anticipation Strata + Hadoop World, 12/8/2016
- 25. …breaks down into sentiment and drivers… 4 (For LI data ) deep dive into MLC segmentation… 6 …geographic locations… 5 …and audience segmentation… 7 …generates automatic reporting, alerts and escalations… 8 …and close the feedback loop with support and PR solutions 9 This is what the future could look like From the first time we pick up an isolated comment… 1 Machine determines if there is significant reach… 2 …and whether it is a trending topic… 3 Strata + Hadoop World, 12/8/2016
- 26. Best customer experience starts from understanding Voices of members! Thank You!
- 27. Engineering blogs for Voices Strata + Hadoop World, 12/8/2016 Part I. Voices: a Text Analytics Platform for Understanding Member Feedback Part II. Technical Details for Topic Mining
- 28. References 1. LibLinear: a library for large linear classification, available at https://www.csie.ntu.edu.tw/~cjlin/liblinear/ 2. LingPipe: a Java-based toolkit for processing text using computational linguistics, available at http://alias-i.com/lingpipe/ 3. NLTK: a leading platform for building Python programs to work with human language data, available at http://www.nltk.org/ 4. Stanford CoreNLP: an open source project lead by Stanford NLP group, available at http://nlp.stanford.edu/software/