- 快召唤伙伴们来围观吧
- 微博 QQ QQ空间 贴吧
- 视频嵌入链接 文档嵌入链接
- 复制
- 微信扫一扫分享
- 已成功复制到剪贴板
Automating Loss Prevention Using NLP with FastAI on Azure Databricks
PetSmart, with over 1,600 stores in North America, is the largest specialty pet retailer of services and solutions for the lifetime needs of pets. The Advanced Analytics Group is a small team of highly business-oriented strategy and data science professionals that uses various data and modeling methodologies to generate breakthrough insights for various business units throughout the company to deliver top- and bottom-line growth for PetSmart.
As a retailer, PetSmart has many years of transaction data, loss prevention store reports, customer feedback, labor schedules, supply chain, and other data. Loss prevention deals with reduction of preventable losses whether it be from, theft, fraud, vandalism, waste, abuse, incidents, accidents, or misconduct.
Store leaders at PetSmart locations submit free text reports to the Loss Prevention team of investigators which must be prioritized for further resolution. Most reports are of low priority and are reported as a matter of policy fulfillment but some require further investigation by this team. However, the team must still read each report in order to filter out low priority reports and then spend time investigating the higher priority reports. The Advanced Analytics Group was asked if we could help automatically prioritize these reports. Developing a prioritization system with performance high enough to automatically prioritize would require near-human performance. To achieve that level (96% accuracy) we utilized FastAI’s ULMFiT NLP classifier.
FastAI is not natively supported on Azure Databricks so setup required special configuration. Azure Databricks newly released ML Beta and GPU clusters were instrumental in enabling the setup. Other challenges included actually extracting the data from a legacy reporting system. Without the flexibility Azure Databricks provides, the iterations, training, and eventual operationalization of the model would have taken much longer and at a greater cost.
展开查看详情
1 .WIFI SSID:Spark+AISummit | Password: UnifiedDataAnalytics
2 .Automating Loss Prevention Using NLP with FastAI on Azure Databricks Mike Vedomske, PhD, PetSmart #UnifiedDataAnalytics #SparkAISummit
3 .Who are we? • 1,600 stores in North America • Largest specialty pet retailer Our team: Advanced Analytics Group A small team of business-oriented strategy and data science professionals #UnifiedDataAnalytics #SparkAISummit 3
4 .The Use Case Loss prevention Reduction of preventable losses whether it be from, theft, fraud, vandalism, waste, abuse, incidents, accidents, or misconduct. #UnifiedDataAnalytics #SparkAISummit 4
5 .The Use Case • The Ask: Can we automate this please? – If we can auto-classify low-priority reports so investigators can spend time on higher-priority – I.e., actually investigate – Improve daily work experience and results of current investigators, reduce need for more hires The (Implied) Need: Near-Human Performance #UnifiedDataAnalytics #SparkAISummit 5
6 .The Data • 13 years or ~500k store reports • Imbalanced data: most were low-priority • Free text, frequent incorrect spelling and grammar, unconventional abbreviations • Legacy system was a mess, broken since 2001: – Field leakage – Some behavior auto-triggered high priority #UnifiedDataAnalytics #SparkAISummit 6
7 .The Old Workflow Store Investigations Team Field Store inputs Investigators Severity is report into access CMS to assigned through Appropriate Case review manual review of Resolution Management incidents “Team” queue System (CMS) Critical incidents sent to field team 72% incidents are classified as Potential for “Team” for manual review automation #UnifiedDataAnalytics #SparkAISummit 7
8 .The Interim Workflow Store AI Classifying Store inputs Model Probabilities report into Provides Displayed in Case Probability of Separate Management High/Medium Dashboard System (CMS) Severity Field Investigations Team Investigators Severity is access CMS to Appropriate assigned through review incidents Resolution manual review of in order of Critical incidents “Team” queue sent to field team probability #UnifiedDataAnalytics #SparkAISummit 8
9 .The Future-State Workflow Store AI Automation Store inputs Model Low Priority report into Provides Reports Recorded Case Probability of as Such and 72% Management High/Medium Automatically workload System (CMS) Severity Closed reduction! Field Investigations Team Final High or Med Severity is Investigators Appropriate assigned through Only Investigate Resolution manual review of High/Med Cases Critical incidents sent to field team Hi/Med queue #UnifiedDataAnalytics #SparkAISummit 9
10 .Early Modeling Attempts • TF-IDF, word counts, n-gram TF-IDF • Classifiers tried: – Logistic Regression (81% AUC) – Random Forest (64% AUC) ß ¯\_(ツ)_/¯ – XGBoost (75% multiclass accuracy) • Performance was…fine, but not good enough #UnifiedDataAnalytics #SparkAISummit 10
11 .FastAI:NLP A (Super) Brief Intro • Who: Jeremy Howard & Sebastian Ruder • What: “an effective transfer learning method that can be applied to any task in NLP, and…techniques that are key for fine-tuning a language model.” • What: “Our method significantly outperforms the state-of-the-art on six text classification tasks, reducing the error by 18-24% on the majority of datasets. Furthermore, with only 100 labeled examples, it matches the performance of training from scratch on 100x more data.” #UnifiedDataAnalytics #SparkAISummit 11
12 . FastAI:NLP A (Super) Brief Intro • 3-layer LSTM • 1150 unit hidden layer • 400 unit embedding Source: arXiv:1801.06146 Pretrained Fine-Tune Fine-Tune Language on Your as Model Text Classifier #UnifiedDataAnalytics #SparkAISummit 12
13 .FastAI:NLP A (Super) Brief Intro Pretrained Fine-Tune Fine-Tune Language on Your as Model Text Classifier 1. “carefully control how fast our model learns” 2. “update the pre-trained model so that it does not forget what it has previously learned” 3. “we can make training text classification models for languages other than English a lot easier as all we need is access to a Wikipedia, which is currently available for 301 languages” #UnifiedDataAnalytics #SparkAISummit 13
14 .FastAI on “the big Mac” The Original Setup • Mac Pro, 64 GB RAM, 2.7 GHz 12- Core Intel Xeon E5 • NO NVIDIA GPUs • Jupyter notebook 7 days for Language Model Tuning…we need something else The Big Mac #UnifiedDataAnalytics #SparkAISummit 14
15 .FastAI on Azure Databricks • The Setup – FastAI 0.7.0 "This is an alpha version.” – Needed Python 3.6+ and a bunch of dependencies • Providence Provided! – 5.0 ML (Beta), now deprecated – Driver: 4 Nvidia GPUs, 224 GB RAM – 1 worker: 2 GPUs, 112 GB RAM #UnifiedDataAnalytics #SparkAISummit 15
16 .FastAI Performance • Initial Model Performance – Forward only model: 92% accuracy – Bidirectional model: 93% accuracy • Language Model fine tuning: ~24 hours • Classifier Training: ~24 hours #UnifiedDataAnalytics #SparkAISummit 16
17 .Squeezing Out a Bit More Juice • Train LM on all data 2005-Present • Train Classifier on 2016+ data • Final Classifier Performance: 96% accuracy…yeah, we can use this Pretrained Fine-Tune Fine-Tune Language on Your as Model Text Classifier #UnifiedDataAnalytics #SparkAISummit 17
18 .FastAI in Production • Score hourly (streaming in the future?) • Retrain monthly, adds latest reports • Moving it from dev to production was super easy on Azure Databricks: 1. Clone 2. Configure Cluster 3. Create Job #UnifiedDataAnalytics #SparkAISummit 18
19 .Final Thoughts • Took some work to get it up and running • Performance is spectacular • Azure Databricks was most cost effective #UnifiedDataAnalytics #SparkAISummit 19
20 .DON’T FORGET TO RATE AND REVIEW THE SESSIONS SEARCH SPARK + AI SUMMIT