Software Major Features
- Discovers rule-based knowledge from data by evolving inductive reasoning
- Builds predictive models for automated & reliable decision-making
- Includes an extensive validation process of the models, to deliver only the most robust ones
- Benchmarks vs. Logistic Regression, Random Forests, Boosted Trees, Neural Networks
- Includes all the functionalities of XTRACTIS PREDICT
- Computes Predictions in real-time by deductive reasoning
- Infers rules at very high frequency
- Computes the most optimal solutions by abductive reasoning
- Inverts models to calculate solutions best matching requests
- takes into account multi-objective requests and flexible constraints
- Includes all the functionalities of XTRACTIS PREDICT
- Oversees the quality of deployed models
- Detects changes in the process behavior
- Monitors the maintenance of the models
More about XTRACTIS Features
XTRACTIS can process data of various kinds: continuous numeric, integer numeric, and nominal variables (binary or multinomial), so quantitative or qualitative data (hedonistic evaluations, preferences, liking, expert opinions, sensory or instrumental measurements, signals, spectra, images, socio-economic data, formulations, product characteristics, genetic sequences...). Unstructured data (image, signal, spectrum, text) will require prior preprocessing via intelligible variables, i.e., which have a meaning for experts, to allow the induction of intelligible models.
The variable to predict may be a numerical variable (regression), one or more classes (multi-classification), or a risk of occurrence of an event (scoring). The variables from which it will be predicted are called "Potential Predictors".
To launch your inductive modeling, you must provide an observation set with the potential predictor's values as inputs and the variable-to-predict values as output. This reference data set must be structured in a table (rows/columns). Use all the original variables available without aggregating them and without imputation when the value is missing.
The quality of the models depends on the quality of the data. However, XTRACTIS can handle missing data, noisy data, or a small amount of data.
A robust model can also be used to detect and filter noise present in the data set.
It should be noted that:
- If the reference dataset did not contain enough exploitable information, XTRACTIS would be unable to discover a robust model. But at least you will know it very quickly and improve your dataset.
- When the output values are unavailable, XTRACTIS can detect stable clusters in the data by unsupervised learning (clustering).
Depending on the problem to solve, the modeling type can be:
- REGRESSION: to predict a continuous / integer numeric variable
- BINOMIAL / MULTINOMIAL CLASSIFICATION: to predict an N-modality variable
- SCORING: to predict the possibility of the occurrence of the studied event
In classification/scoring, and when the output values are unavailable, use the XTRACTIS Clustering feature to identify stable classes in your data. Thus, you can complete your dataset and launch the induction process.
XTRACTIS was initially invented to model benevolent human behavior: sensory perception of products, consumer liking, and subjective evaluation. It was then successfully used to detect malicious human behavior (fraud, security, cybersecurity).
Actually, its universal approach makes it possible to model any complex process or phenomenon, including natural phenomena in science... as long as a dataset is available.
No! XTRACTIS handles evenly small and big data.
As it is a holistic modeler, it considers in first intention all the available variables and possible interactions between them. Then, it automatically discards non-impacting variables and finally ranks the retained predictors by their impact, from strong to weak signals.
Yes! XTRACTIS can handle datasets that include missing data without any imputation. The reference dataset could also have noisy data –it will be detected– or data in small quantities (more columns than rows).
The induction robots need a few minutes to a few weeks to do their job, depending on the complexity of the studied process, the number of explored induction strategies, and the allocated CPU power: the higher the complexity of the problem to be solved, the higher the license power, the number of robots and parallel logic cores required. And obviously, the more you let them work in parallel, the fastest they reach their goal –finding the most robust and intelligible model– thanks to their competing and evolving inductive reasoning.
Once your project is loaded, GENERATE inductive robots explore as many strategies as necessary among the infinity of induction algorithms they have.
Firstly, each robot reasons on its own to produce a model. It systematically, objectively, and intensively evaluates the performance of its model by Cross-Validation: descriptive performance (Training), predictive performance (Validation), and real performance (Test).
Then, the robots work together to compare their results and perfect their inductive reasoning strategies. And this is for as many cycles as needed to produce the most robust and least complex models.
So basically, robots automatically and collectively improve their reasoning strategies to continuously improve the robustness of the discovered models, even if the database does not change: They learn to learn better!
At the end of the procedure, XTRACTIS reveals all the generated models, ranked according to their performance and complexity.
Then the selected model is validated by humans thanks to its intelligibility and pushed to end-use. It remains unchanged during operation until MONITOR detects that it is about to expire and needs to be updated with an additional set of data.
PREDICT deductive robots trigger instantaneously the relevant rules of the model to compute the predictions in real-time. Again, over 70,000 decisions per second on a basic 8-physical core CPU, quite sufficient for online dynamic processes!
OPTIMIZE abductive robots generally need a few minutes to a few hours, depending on the complexity of the multi-objective request, the complexity of the models, and the allocated CPU power (the higher the complexity of the problem to be solved, the higher the license power, the number of robots and parallel logic cores required).
We have granted XTRACTIS the right to refuse to decide when it knows it does not know. And this makes it a wise AI, according to Socrates!
It can exercise its right to refuse in several distinct situations:
- It can advise us against using models it has discovered but has evaluated as not being robust.
- During its inductive reasoning, it may also refuse to consider reference dataset points it has estimated to be erroneous (noise).
- It can refuse to cover a particular area of the workspace because of the absence of data and/or the presence of contradictions at the borders of this area.
- When using a model in Prediction, it can sometimes refuse to predict the consequence for a new case. Several reasons can lead to a "Refusal":
- Out Of Range: the case is located very far from the boundaries of the operating space, and XTRACTIS forbids substantial extrapolation;
- Out Of Mapping: the case is located in an area not covered by the model;
- Indecision: the case is undecidable between two classes with close possibilities;
- there is an absence of consensus in the College of Virtual Experts (CEV) that must make the decision;
- a majority of the Individual Virtual Experts of the CEV refuse to make a decision;
For applications that would forbid the refusal, it is always possible to force XTRACTIS to deliver the most appropriate decision, but by authorizing it to alert the user about the low reliability of the decision it has been forced to make.
After the induction phase, XTRACTIS reveals all the generated models, ranked according to their performance and complexity.
The Performances Report delivered by XTRACTIS GENERATE synthesizes the performance indicators that endow the quality and robustness of the model: the descriptive performance (calculated on the training subset), and the predictive performance (calculated on the validation subset), which should be close to each other and confirmed by the real performance (calculated on the test subset and/or external test subset).
The performance metrics are classical statistical metrics and depend on the type of modeling and the process to model.
The Structure Report delivered by XTRACTIS GENERATE presents the structure of the model, exhibiting all the knowledge synthesized from the information included in the learning dataset: the compilation of fuzzy rules explaining the interactions between the predictors and their fuzzy relations with the variable to predict, the impact of each rule and of each predictor, the fuzzy sets qualifying the predictors, the logical operators used in the inductive reasoning...
This report corresponds to the “Ph.D. thesis” produced by the GENERATE robot that has induced the model. It thus exposes the laws of behavior of the process under study and, therefore, the entire decision-making logic of its model. Everything is transparent, understandable, and certifiable by humans!
The Prediction Report delivered by XTRACTIS explains the deductive reasoning of the PREDICT robots to calculate each decision from the model's rules for a new real or simulated case.
The decision is fully explainable and traceable, and you can understand how the system computes its fuzzy decision then final binary decision!
XTRACTIS is based on 100% proprietary and secret algorithms. It uses only in-house algorithms, except for the Benchmark functions, which use built-in and up-to-date open-source frameworks.
We understand that you may wonder how our algorithms work and how XTRACTIS results stack up against open-source AI techniques. That's why our platform includes this feature that allows you to use Logistic Regression, Random Forests, Boosted Trees, and Neural Networks on the same dataset and the same splits and with the same number of explored strategies and the same objective criteria to evaluate the performance of models from different techniques, in a single software environment.
This functionality works only with the mouse like the other XTRACTIS features and is only available for GENERATE licenses of 12+ Tflops FP64.
Yes! All functionalities are used only by mouse clicks; even the Benchmark features and the interface is the same for all applications. No need for programming or using any external framework!
Yes, for XTRACTIS or any analytic models!
Just use the ENGINE license versions of PREDICT (for predictions) and OPTIMIZE (for the most optimal solutions). Your external program can call these robots via an API (C, Java, Python).
XTRACTIS has been tested and certified on SUPERMICRO and BULL HPC servers. Certification on other server brands will be on request.
Currently, XTRACTIS runs exclusively on Windows OS. When clients confirm a strong demand, a Linux version will be developed in a few months.
Essentially by posing the problem correctly, providing compliant reference datasets, and auditing your transparent models.
You must:
- Define the studied variable and exhaustively list all the potential predictors you have.
- Use all the original variables available without aggregating them and without imputation when the value is missing.
- Do the preprocessing of unstructured data, as much as possible with intelligible variables.
- Provide the structured dataset and launch the induction process.
- Set the inductive reasoning parameters, among which the number of strategies to be launched.
- Go about other business and let your robots reason for you.
- Once REVEAL robots provide the best models, select the most robust model, i.e., the one with the best predictive performance, while checking that it remains close to its descriptive performance. These performances are confirmed by the real performance observed. With similar predictive performance, you will favor the most intelligible model: the one with the fewest predictors per rule, the fewest classes qualifying the predictors, and the fewest rules, therefore, the least complex one.
- Consult the model's structure report to audit its entire decision-making logic and understand the laws of behavior, the interactions between the variables, and the contribution and influence of each predictor and of each rule.
- Submit such a model to the regulator for certification when mandatory.
Then you use your validated/certified model or push it to end-users for:
- Prediction: let PREDICT robots compute the model output in real-time for a new scenario in a deterministic and rational way.
- Optimization: let OPTIMIZE robots find solutions that best match a non-linear multi-objective request from induced or analytical models.
XTRACTIS Licenses Plans
REVEAL 365
Ideal for small to medium complexity projectsREVEAL CDU
Ideal for large to very large complexity projectsPREDICT 365
Use models for predictive analysisOPTIMIZE 365
Use models for prescriptive analysisAll prices exclude VAT and are given as an indication.
Only a commercial proposal published by INTELLITECH will engage INTELLITECH.
Sales are governed by our legal documents, available on request.
+ High level Service & Support
Proof of Concept & Pilot Project
Test XTRACTIS on your data
Depending on the selected problem and your budget, we mobilize our resources over 1 to 6 months to show you either only the obtained performances (Proof of Concept) or the full models (Pilot Project).
In the latter case, the models become your IP/assets. You then use them with PREDICT or OPTIMIZE to enhance your processes.
Benchmarks
Compare the performance of your XTRACTIS model vs. its challengers
In a single software environment, we provide you with a comprehensive benchmark of XTRACTIS against other open-source techniques: models produced by Logistic Regression, Random Forest, Boosted Trees, and Neural Network on the same dataset, the same splits and with the same number of explored strategies and assessed with the same performance objective criteria.
Training & Workshops
Master the latest advances in AI
We teach you the fundamentals of the XTRACTIS Augmented Fuzzy Symbolic AI and other predictive Machine Learning techniques. You will also learn to use XTRACTIS and deploy your models in your business processes.
Coaching & Mentoring
Benefit from personalized support
Our Expert and Senior R&D Engineers provide personalized support. First, we check that your modeling problems are correctly posed. Then, we analyze with you the data and the results obtained.
We also assist you with your AI transformation strategy.