AI is at the heart of the Quantexa Decision Intelligence Platform, helping users unify data and create context to make accurate and reliable decisions.
Machine Learning (ML) and AI are applied in four distinct areas within the Quantexa Platform:
Composite AI within Quantexa
The Quantexa Platform uses multiple AI techniques to solve a wide range of business problems more efficiently. These include AI-driven production code, models, and development tools to help users deploy AI appropriate to their use case. Quantexa’s Composite AI Stack combines subject matter expertise and domain knowledge with various machine learning, Natural Language Processing (NLP), and deep learning techniques.
The diagram below summarizes AI techniques available across Quantexa-based workflows, with flexibility and modularity, to help identify risks and opportunities.
Key Platform Components Driving Composite AI
Entity Resolution with Machine Learning
AI models enable users to identify opportunities to optimize their Entity Resolution configurations to drive more accurate entities
Overlinking Root Cause Analyzer (ORCA)
This is for analysts to determine the root cause of any potential overlinking issues during Entity Resolution. ORCA uses advanced statistical techniques to approximate match-rate distributions across data sources. These distributions are fed into sophisticated anomaly detection algorithms to:
- Reveal bad input data (e.g. default values) adversely impacting Entity Resolution.
- Indicate which compounds should be tweaked by adding fine-tuned exclusions to improve their performance.
- Provide interactive graphical outputs such as heat maps that quickly alert users to problem areas of their data.
Read more: Entity Quality ORCA (login required)
Quantexa’s Pre-Built AI Models for Risk Identification
- Shell Company and Mini Umbrella Company Detection (login required)
- Calculation of Ultimate Beneficiary Owners (login required)
- Identification of SMEs (Small Medium Enterprises) (login required)
- Identification of related high-risk entities through multi-hop risk detection
The QKnowledge Graph Python Library
The QKnowledge Graph Python Library includes a suite of tools to work with large Entity-based networks in Python. Some of the features are:
Large-scale analytics for Quantexa Networks, notebook-friendly visualizations, and extensible by data scientists:
- Enables easy application of graph analytics to Quantexa networks (raw bi-partite, heterogenous graphs are unsuitable for most algorithms)
- Provides multiple perspectives on the same graph data, utilizing MetaGraphs
- Enables efficient application of global algorithms, e.g. PageRank
Fast & easy neighborhood creation:
- Cypher-like path-based queries
- Traverse billions of edges in minutes
- Efficiently analyse highly linked nodes through techniques such as sampling and filtering
State-of-the-art Graph Learning at scale, facilitating rapid iteration for exploration and modelling:
- Seamless interfaces to Quantexa platform for data ingest and scoring (Quantexa Batch Resolver and Assess)
- Open interface to leading open-source graph libraries
- Graph learning with libraries such as PyG and Graph Neural Networks (GNNs)
Read more: QKnowledgeGraph (login required)
Customisable Scoring with Quantexa Assess
Quantexa Assess is a framework to enable the production of complex network and Entity-based detection logic, with scores working as business rules, triggers or features within ML.
Data Science teams can build/maintain their own models as needed with Assess, using Scala and Python interfaces to access the widest set of ML algorithms, and providing version-controlled parameterization to allow business users to run and maintain use case models. The framework:
- Enables write-once scores, servicing both batch and dynamic infrastructures
- Manages score dependencies (DAG) and facilitates rapid iteration of individual scores
- Supports rich score descriptions & node highlighting in UI
- Provides flexible Alerting and Re-alerting logic
- Supports config deployable prebuilt scoring models for use with multiple data models
- Records all scoring output for audit and model retraining purposes
Assess understands entities and networks and enables multiple scoring levels:
The QPython Library
The QPython Library allows data scientists and Python developers to Pythonically interact with Quantexa, for example:
For ad hoc analytics
- Explore data and test hypotheses
- Allows Data Scientists to discover additional insights over existing detection logic
- Directly share comprehensive findings with Business Users within the UI, raising 'Tasks’ containing networks and detailed scoring data.
As a development tool
- Productionize Python detection logic within Assess – combining Python with Scala as desired.
- Combine with the full range of out-of-the-box scores and Assess scoring features to raise alerts including custom Python scores and logic.
For tuning and improvement
- Analyze scoring outcomes utilizing Python tooling
- Quantexa Tools compare both alerts and scoring outcomes
- Simulate threshold changes on historical data.
- Utilize Python-based ML to optimize changes
- Record and evidence changes with notebook functionality
Product Documentation for Composite AI Components within Project Example
Note: All Documentation site links require login.
- Rule based standardization, cleansing and parsing, NLP based parsing models:
- Automated field identification:
- Transparent Entity Resolution configuration, optimised configuration:
- Entity Quality Scoring Models:
- Graph Script network definitions:
- Use case specific risk indicators and mitigators, Expert Scorecard, Alerting and re-alerting:
- ML based risk detection, Model optimisation
- Model monitoring and tuning
Visualization and Exploration of Risk
- Top-down thematic filtering (Explorer)
- Temporal and geospatial based presentation of risk
- Contextual description & visualization
- Graph Perspectives / Model based Perspectives (e.g., Entity to Entity in investigations)
Further Resources from the Community
Announcements
Articles & Discussions
Further Resources from the Documentation Site
