International Business Machines Corporation

Ravi Chandra Chamarthy

Manish Anand Bhide

Madhavi Katari

Arunkumar Kalpathi Suryanarayanan

Prateek Goyal

Prashant Pandurang Mundhe

Remus Lazar

Swapna Somineni

Ruchir Puri

Kishore Patel

Trent A Gray-Donald

One or more systems, devices, computer program products and/or computer-implemented methods of use provided herein relate to a process to facilitate abnormal document self-discovery. A system can comprise a memory that stores computer executable components, and a processor that executes the computer executable components stored in the memory, wherein the computer executable components can comprise a fairness component, an identification component, a removal component, and an evaluation component. The fairness component can receive a dataset for a tree-based model and calculates a first fairness, and the identification component can identify root to leaf paths in the tree-based model for one or more records of the dataset and one or more corresponding perturbed records. The removal component can remove at least one record of the one or more records having similar root to leaf paths to the other records of the one or more records.

System and methods for evaluating the sensitivity and fairness bounds of tree-based models

A machine learning model data quality improvement detection tool is provided for identifying an accurate reference group and an accurate monitored group of a machine learning model. The tool monitors a behavior of the machine learning model for a predetermined time frame. The tool compares a determined fairness metric a pre-defined fairness threshold. Responsive to the fairness metric failing to meet the pre-defined fairness threshold, the tool modifies the monitored group to include a first portion of the reference group. The tool compares a newly determined fairness metric to the pre-defined fairness threshold. Responsive to the newly determined fairness metric meeting the pre-defined fairness threshold, the tool identifies the modified monitored group including the first portion of the user-defined reference group as a new monitored group and the modified reference group without the first portion of the user-defined reference group as a new reference group.

Decreasing error in a machine learning model based on identifying reference and monitored groups of the machine learning model

A computer-implemented method includes: receiving, by a computing device and from a user device, a request to validate an application program interface (API); validating, by the computing device, the API by performing a fetch analysis using different user profiles; returning, by the computing device and to the user device, a result of the fetch analysis; validating, by the computing device, the API by performing an insert/update analysis using the different user profiles; and returning, by the computing device and to the user device, a result of the insert/update analysis.

Data governance policy recommendations for application program interfaces

An example operation may include one or more of storing a batch scoring engine and an application programming interface (API) for the batch scoring engine, receiving a trigger to perform a batch prediction process, reading input data from a source data store and executing, via the batch scoring engine, one or more predictive models on the input data to generate a predictive output and metadata associated with the predictive output, storing the predictive output and the metadata in a target data store, and updating the API with a location of the predictive output within the target data store and a location of the metadata within the target data store.

API management for batch processing

A system and related method score a fairness of an outcome model. The method comprises receiving a set of original transaction records (OTRs), and selecting an OTR subset of the OTRs according to a subset selection criteria in order to reduce a number of OTRs to send to outcome model. For each OTR in the subset a perturbed transaction record (PTR) is created based on the OTR that includes changing at least one attribute in the PTR from the OTR, sending the OTR and the PTR to the outcome model, receiving an OTR outcome and a PTR outcome from the outcome model, and determining a record bias score for the OTR outcome and the PTR outcome respectively that indicates bias in the respective outcome. The OTR and the PTR bias score are stored in a bias determination system (BDS) database.

Batch scoring model fairness

Early indications of application programming interface (API) usage are identified by correlation to particular issues with the API including singular and mutual consistency, completeness, accuracy, and staleness. Analysis of API input and output along with data type and formatting information facilitates identification of the API issues. Establishing a correlation between API usage and issues supports early detection of potential usage reduction on a case-by-case level. Corrective action to resolve identified issues may be performed in a timely manner to maintain usage levels.

Correlating usage to issues in application programming interfaces

One example of a system comprises using a processor for identifying a model to be validated that is stored in a repository; automatically computing and recording one or more model metrics for the model to be validated in a tamper-proof manner; comparing the computed tamper-proof metrics with one or more encoded rules and policies to determine if the model to be validated complies with the one or more encoded rules and policies; and outputting a notification to a device indicating a validation status of the model to be validated based on the comparison of the computed tamper-proof metrics with the one or more encoded rules and policies.

AI governance using tamper proof model metrics

Artificial intelligence (AI) model monitoring and ranking includes obtaining metric values indicative of performance of AI model deployments, the metric values including respective metric values measured across metrics, determining violation statuses of the metrics for each of the AI model deployments, the violation statuses indicating, for each AI model deployment, which of the metrics are violated by the AI model deployment as reflected by respective metric values for that AI model deployment, ranking the AI model deployments against each other according to a ranking model and based on the determined violation statuses for each of the AI model deployments, and providing a rank of at least some of the AI model deployments to a user.

Artificial intelligence model monitoring and ranking

A processor may receive an original dataset. The processor may segment, automatically, the original dataset into a plurality of data groups. The plurality of data groups may include a model training dataset and a holdout dataset. The processor may generate a model with the model training dataset. The processor may validate the model with the holdout dataset.

Validation of AI models using holdout sets

A system includes a memory having instructions therein and at least one processor in communication with the memory. The at least one processor is configured to execute the instructions to run a machine learning base model on input data to generate base model prediction data and run a machine learning error prediction model on the input data to generate error prediction data. The at least one processor is configured to execute the instructions to generate predicted correct base model prediction data based on the base model prediction data and the error prediction data. The at least one processor is configured to execute the instructions to generate confusion values data based on the base model prediction data and the predicted correct base model prediction data. The at least one processor is also configured to execute the instructions to generate base model accuracy fairness metrics data based on the confusion values data.

Machine learning model accuracy fairness

One example of a method comprises identifying a model to be validated that is stored in a repository; automatically computing and recording one or more model metrics for the model to be validated in a tamper-proof manner; comparing the computed tamper-proof metrics with one or more encoded rules and policies to determine if the model to be validated complies with the one or more encoded rules and policies; and outputting a notification to a device indicating a validation status of the model to be validated based on the comparison of the computed tamper-proof metrics with the one or more encoded rules and policies.

An approach is provided for optimizing data fetching. A query employing a method to fetch data from a JSON document is received. An amount of time required to execute the query and a number of nested layers in a traversal of the JSON document required to fetch the data are determined. Based on the amount of time and the number of nested layers, a cost associated with an execution of the query is calculated. The cost is determined to exceed a threshold value. Responsive to the determination that the cost exceeds the threshold value and using historical query patterns and historical query execution times, a schema of the JSON document is re-designed. The data is fetched from the JSON document using the re-designed schema.

Optimizing JSON document usage

Systems and methods for detecting indirect bias in machine learning models are provided. A computer-implemented method includes: receiving, by a computer device, a user request to detect transitive bias in a machine learning model; determining, by the computer device, correlations of attributes of neighboring data not included in a dataset of the machine learning model; ranking, by the computer device, the attributes based on the determined correlations; and returning, by the computer device, a list of the ranked attributes to a user that generated the user request.

Automatic correction of indirect bias in machine learning models

A method, computer program product, and system for pre-parsing a large XML document to enable faster processing, comprising providing an XML document that is idle or is being generated, pre-parsing the XML document to generate one or more indices associated with the XML document, storing the indices in the XML document, and storing the XML document in a storage area. At the time of processing the XML document, an XML consumer is able to process the XML document more quickly than normal by using the one or more indices.

Growing community of inventors

Hyderabad Telangana, India

Ravi Chandra Chamarthy