Query-time Nonparametric Regression with Temporally Bounded Models - Patrick Heck, Needham Software & David Smiley, D W Smiley LLC
Download as PPTX, PDF1 like175 views
This document summarizes a presentation about query-time nonparametric regression and time routed aliases in Solr. It discusses how nonparametric multiplicative regression was used to continuously predict user interests for an online career coaching system based on click-through data. It also describes how time routed aliases in Solr provide a built-in way to implement time-partitioned indexing of timestamped data across multiple collections while automatically adding and removing collections over time.
Query-time Nonparametric Regression with Temporally Bounded Models - Patrick Heck, Needham Software & David Smiley, D W Smiley LLC
- 1. Query-time Nonparametric Regression with Temporally Bounded Models Gus Heck & David Smiley #Activate18 #ActivateSearch
- 2. Agenda • Presenters • Query time Nonparametric Regression • Demo – Suggesting tagged images • Time Routed Aliases • Demo – Creating a Time Routed Alias
- 3. Presenters Patrick (Gus) Heck • Solr Contributor since 2013 • Consulting since 2012 • Enterprise Search since 2010 • Apache Ant contributor 2003-2004 • Web Applications since 2003 David Smiley • Lucene/Solr Committer (PMC) • Consulting • Author of first book on Solr • Presentations & Training
- 4. In the beginning.. Dave Mackey - Vision for an online coaching platform leveraging machine learning technology to help companies help their middle managers. Engaged me first as Chief Architect and later as CTO to bring this vision to reality. It worked, but was not funded...
- 5. Brief Overview of Simply Coached Simply Coached provided online career coaching system Key Goal was connecting users with relevant curated content Relevance was determined by • Identifying topics the user is interested in by observing and learning from click-throughs • Several other customized metrics
- 6. Learning What Interests the User Goals: • Suggest things of interest, things the user is willing to learn • Make predictions based on the user’s behavior • Unique prediction per user • Predictions based on entire user base • Avoid calcification of the predictive model, old data needs to be sunset regularly.
- 7. Modeling Candidate - Neural Nets Simply Coached started right at the dawn of the “3rd wave” of neural networks Limitations of Neural Networks • They don’t adapt well to a changing problem • We would need to retrain predictive network regularly • Require massive data for training
- 8. Modeling Candidate - Regression Not as “cool” as deep learning, but more suitable • Can converge faster with less data • But, most techniques are assuming that the residuals are normally distributed • Many of our features will be binary • Normality is right out the window I decided to look for a Non-Parametric Regression technique
- 9. Non-Parametric Multiplicative Regression • Can predict continuous variable • Good for sorting/ranking • Kernel smoother, presence/absence kernels already well known • Non-parametric Nice discrete sub-calculations... Used in Habitat Ecology for predicting habitat suitability https://ir.library.oregonstate.edu/concern/defaults/z029p910z
- 10. NPMR Equations Performance across multiple factors (features) Local Mean estimator Example kernels continuous or presence/absence: Key realization: the products and sums for i,j can pre precalculated: I called these pre-calculated portions “Partials” and added temporal metadata
- 11. But wait... What are we predicting? Needed a continuous statistic to predict Thesis: Given a good teaser etc., the user will click more rapidly on more interesting articles. Suggest documents in order of predicted reaction time
- 12. Demo Three users: • Sporty Black likes black sports cars • Cooper Blue likes blue coupes • Racy Yeller likes yellow formula cars System was trained by presenting ~20 sets of 9 randomly selected cars and clicking with varying speed Images copyright Andrew Mauldin, used with permission
- 13. Streaming Expression! As mentioned, NPMR prediction was only part of the overall product It occupies the top line of Boxes A1 and B1
- 14. The actual demo expression
- 15. Streaming Expressions The final sorted list of documents can be calculated The bold portion is the final summation
- 16. Continuously running, recalculating sums/products once a minute Included traditional document indexing (article scanner) and also streaming expression update() (send_per_user_sum_update) By end, running continuously for months at a time unattended. http://www.jesterj.org for more info about JesterJ Indexing With JesterJ
- 17. Adapting and Scaling • Adapt at query time with filters, could also filter in other ways with additional metadata • Four Dimensions: 1. Content to recommend - constant set 2. Users - per user pricing to the rescue 3. Activity - price for average activity level 4. Time - data accumulates over time Time Routed Aliases are a solution to #4
- 18. Time Routed Aliases David Smiley
- 19. Have a Lot of Timestamped Data? And you need keyword search and/or analytics (faceting/aggregations) capabilities. Examples: • Logs • Sensor data (IoT) • Social media posts Characteristics: tons of docs, continuously flowing in, limited retention
- 20. Strategies Hash Partitioned: • One collection, hash routed shards (built-in) Time Partitioned: • One collection, time routed shards (DIY) • Time partitioned collections (DIY) • Time partitioned collections via TRAs (built-in) “Partitioned” = “Routed” = “Organized”
- 21. One Collection, Hash Routed Shards Hash on ID, even distribution router.name=compositeId + Easy (default) + High write throughput - Deleted dead-weight - Poor realtime search - Queries execute everywhere - Inflexible sizing - … thus Expensive (uniform hardware requirements) Deleted docs Live docs See DocExpirationUpdateProcessorFactory shard1 shard4 shard7 2 5 8 9
- 22. Time Partitioned Data Generally better... • New data always goes to most recent partition • Variable or equal sized (depends on approach) • Addresses all negatives of hash routing • Write throughput can be addressed by sharding within partition • Opportunities to “optimize” aged indexes • Flexibility in assigning partitions to better or cheaper hardware • … all leads to cost savings How?... 2017-01 2017-02 2017-03 2017-04 2017-05 2017-06 2017-07 2017-08 2017-09
- 23. Time Partitioned Data Implementation Strategies: • One Collection, router.name=implicit • See this sample code: (DIY) https://github.com/cga-harvard/hhypermap- bop/tree/master/bop-core/solr- plugins/src/main/java/edu/harvard/gis/hhypermap/bop/solrplugins by me • Multiple Collections • See this blog: (DIY) http://blog.cloudera.com/blog/2013/10/collection-aliasing-near-real-time-search-for- really-big-data/ by Mark Miller • TRAs… (built-in) 2017-01 2017-02 2017-03 2017-04 2017-05 2017-06 2017-07 2017-08 2017-09
- 24. About Collection Aliases SolrCloud supports collection aliases Aliases point to one or more collections Ex: Alias “tra-demo” → 2017-09, 2017-08, 2017-07, ... 2017-01 2017-02 2017-03 2017-04 2017-05 2017-06 2017-07 2017-08 2017-09
- 25. Time Routed Aliases Aliases have new tricks up their sleeves… • Aliases now have metadata (API to read & edit) • Can create a “time routed alias” w/ first collection • Collections in a TRA can • Route update requests to the correct collections • Adds/deletes collections automatically • data driven • just-in-time or preemptively Mutable configuration, mostly New in Solr 7.3
- 26. TRA Creation (using V2 API) curl http://localhost:8983/api/c -H 'Content-type:application/json' -d '{ "create-alias":{ "name": "tra-demo", "router": { "name": "time", "field": "evt_dt", "start": "2018-01-01T00:00:00Z", "interval": "+1DAY", "autoDeleteAge": "-2DAY" }, "create-collection": { "config": "_default", "numShards": 2, "maxShardsPerNode": 2 } } }'
- 27. TRAs, the fine print TODOs • Size capped (e.g. 5M docs / collection) • Query routing to subset of collections • Better “auto-scaling” tie-ins • “optimize” of older collections • TRA deletion, ease of use TRAs may not be for everyone • Partitioning is strict and must be adjacent (no gaps) • Doesn’t work with CDCR
- 28. Thank you! Gus Heck & David Smiley #Activate18 #ActivateSearch