International Business Machines Corporation

Skoopin

Jun Wang

Jing James Xu

Xiao Ming Ma

Si Er Han

Xue Ying Zhang

Dong Hai Yu

Lei Gao

A Peng Zhang

Wen Pei Yu

Jing Xu

Jin Wang

Ji Hui Yang

Yao Dong Liu

Jiang Bo Kang

Dakuo Wang

Yuan Feng

Ling Zhuo

Song Bo

Xin Feng Zhu

Bo Song

Chuang Gan

Bei Chen

Ke Wei Wei

Lingfei Wu

Lei Tian

Petros Zerfos

Yue Yang

Peter Daniel Kirchner

Xing Wei

Horst Cornelius Samulowitz

Wesley M Gifford

Yi Dai

Dhavalkumar C Patel

Syed Yousaf Shah

Gregory Bramble

Saket Sathe

Guang Ming Zhang

Xuan-Hong Dang

Shuang Ys Yu

Pei Pei Liang

Geng Wu Yang

Rui Wang

Yi Wang

Kai Li

Xuye Liu

Long Vu

Weijun Zhang

Qi Min

One or more computer processors calculate a cache prediction for a received inference request within an inference cache structured as a self-learning tree, wherein the inference request comprises a set of input values. The one or more computer processors responsive to the retrieved cache prediction exceeding a cache prediction threshold, transmit the cache prediction. The one or more computer processors parallel compute a model prediction for the received inference request utilizing a trained model. The one or more computer processors responsive to the retrieved model prediction exceeding a model prediction threshold, convert the trained model into a tree structure. The one or more computer processors update the inference cache with the converted train model. The one or more computer processors transmit the model prediction.

Efficient machine learning model inference

A method, a structure, and a computer system for object trail analytics. The exemplary embodiments may include obtaining time series data detailing an average speed of one or more roads within a traffic network at one or more times. The exemplary embodiments may further include extracting one or more features corresponding to the time series data, and generating one or more time series forecasting models based on the time series data and the one or more features. Additionally, the exemplary embodiments may include identifying a current location of a moving object within the traffic network, and predicting a speed of the moving object based on applying the one or more time series forecasting models to the current location.

Object trail analytics

A computer-implemented system, platform, programing product, and/or method for improving transformation selection in an ensemble machine learning (ML) model that includes: providing all base ML models of the ensemble ML model; identifying all of a plurality of Derived Fields in all the base ML models; performing a Derived Field run prediction analysis for all the Derived Fields; computing the Derived Field Importance Weight for Field (DFIW4F) and the Derived Field Importance Weight for Model (DFIW4M) for all the Derived Fields; clustering all the Derived Fields into a plurality of Derived Field clusters, wherein each Derived Field cluster is based upon the DFIW4M and the DFIW4F for the Derived Field; sorting all the Derived Field clusters by best cluster based upon DFIW4M and DFIW4F; and running the base ML models based upon the Derived Fields in the best Derived Field cluster until sufficient base ML models have been run.

System and method for efficient transformation prediction in a data analytics prediction model pipeline

One or more computer processors group a plurality of predictors contained in training data into a plurality of predictor groups. The one or more computer processors create a plurality of sample sets, wherein each sample set in the plurality of sample sets contains one or more predictors selected from a respective predictor group in the plurality of predictor groups. The one or more computer processors create a cluster model for each created sample set in the plurality of created sample sets. The one or more computer processors generate a score for a record with one or more missing values utilizing at least one created cluster model of the created cluster models and at least one created sample set of the created sample sets.

Intelligent scoring of missing data records

A computer-implemented method for anomaly detection for a time series data is provided. Aspects include receiving a time series data including a plurality of sequential data points, calculating an expected next value for the time series data based on the plurality of sequential data points, and receiving an actual next value corresponding to the time series data. Aspects also include calculating an anomaly strength estimate based on the expected next value and the actual next value, identifying one of a plurality of anomaly detection pipelines based on the anomaly strength estimate and a portrait associated with each of the plurality of anomaly detection pipelines, and obtaining an anomaly prediction by inputting the time series data and the actual next value into the one of the plurality of anomaly detection pipelines.

Anomaly detection for time series data

A computer-implemented method for treating post-modeling data includes computing, sequentially for each category of a feature, a category importance (CI) value. The CI value is based on a model accuracy change when records of a category being examined are reassigned to a remaining set of categories of the feature according to a cumulative distribution of records among the remaining set of categories of the feature, wherein the remaining set of categories include all categories of the feature, except for the category being examined. A post-modeling category is performed to merge of each category having the CI value less than a CI value threshold.

Key category identification and visualization

An approach to time-series data point anomaly detection may be presented. Data point anomalies in time-series data can cause a cascade of incorrect predictions in a time-series data prediction model. Presented herein may be an approach to decompose a time-series training data set into elementary components, such as seasonal, trend and residual. The approach may determine one or more confidence intervals for elementary components of data points including level shift, variance, and outlier. From these confidence intervals, new data points can be analyzed and identified as anomaly data points. The approach may also prevent anomaly data points from being incorporated into a time series data prediction model, reducing prediction error in the prediction model.

Optimization of time-series anomaly detection

In an approach for post-modeling data visualization and analysis, a processor presents a first visualization of a training dataset in a first plot. Responsive to receiving a selection of a data group of the training dataset to analyze, a processor identifies three or fewer key model features of the data group of the training dataset. A processor ascertains a representative record of each key model feature of the three or fewer key model features using a Local Interpretable Model-Agnostic Explanation technique. A processor presents a second visualization of the three or fewer key model features and the representative record of each key model feature in a second plot.

Visualize data and significant records based on relationship with the model

A method, computer system, and a computer program product are provided for post-modeling feature evaluation. In one embodiment, at least at least one post model visual output and associated data is obtained that at least includes an individual conditional expectation (ICE) plot and a partial dependence (PDP) plot. Using the associated data and the plots, a Feature Importance (PI) plot is provided. A plurality of features is then determined for each PI, PDP and ICE plots to calculate at least one Interesting Value for each plot. An overall score is also calculated for each plurality of features based on the associated Interesting Values for each PDP, ICE and PI plots. At least one top feature is selected based on said scores. A final plot is then generated at least reflecting the top feature. The final plot combines the PI, PDP and ICE plots together.

Visual representation using post modeling feature evaluation

Systems and computer-implemented methods select a subset of methods to generate data schemas for input data from a list of methods for generating data schemas, based on output of a regression model; generate a candidate schema for each method in the subset of methods to generate data schemas; and generate a master data schema for the input data by merging the candidate schema for each method in the subset of methods to generate data schemas, utilizing predetermined rules.

Automated determination of accurate data schema

Disclosed are a computer-implemented method, a system and a computer program product for model exploration. Model feature importance of each model of a plurality of models can be obtained, the plurality of models can be grouped into a plurality of model clusters based on the model feature importance of each model, and the model feature importance can be presented by box-plot or confidence interval.

Feature importance based model optimization

A method for forecasting server demand includes collecting a historical number of scoring requests from a network using a serverless architecture. A scoring request capacity per server is determined using the historical number of scoring requests. A prediction model predicts a first future value of scoring requests for a first future time span. A current number of servers in a pool of servers handling the scoring requests. Using the prediction model, a determination of whether the current number of servers is capable of handling the first future value of scoring requests for the first future time span is made. Upon determining that the current number of servers is incapable of handling the first future value of scoring requests, one or more additional servers are warmed up. The warmed-up additional servers are added to the pool of servers prior to an arrival of the first future time span.

Efficient serverless method and system of serving artificial intelligence models

Mechanisms are provided for intelligently identifying an execution environment to execute a computing job. An execution time of the computing job in each execution environment of a plurality of execution environments is predicted by applying a set of existing machine learning models matching execution context information and key parameters of the computing job and execution environment information of the execution environment. The predicted execution time of the machine learning models is aggregated. The aggregated predicted execution times of the computing job are summarized for the plurality of execution environments. Responsive to a selection of an execution environment from the plurality of execution environments based on the summary of the aggregated predicted execution times of the computing job, the computing job is executed in the selected execution environment. Related data during the execution of the computing job in the selected execution environment is collected.

Intelligent identification of an execution environment

The present disclosure relates to privacy protection in a search process. According to a method, a target emotion vector is extracted from a search interaction, the target emotion vector representing emotional information in the search interaction. Respective emotion distances between the target emotion vector and respective emotion vectors associated with a plurality of text clusters are determined. The plurality of text clusters is clustered from a dictionary of text elements. A first number of text clusters are selected from the plurality of text clusters based on the determined respective emotion distances. The first number of text clusters have emotion distances larger than at least one unselected text cluster among the plurality of text clusters. A plurality of confused search interactions are constructed for the search interaction based on the first number of text clusters, and the plurality of confused search interactions are performed.

Privacy protection in a search process

A computer-implemented method, system and computer program product for recommending join operations of relational data among different tables. Relationships between one or more pairs of columns of relational data for one or more pairs of tables are identified by determining the semantic similarity between each pair of columns of relational data. The data content join converge for each of the identified relationships in connection with joining the tables involved in such identified relationships is determined. Furthermore, a join strength (indication of the degree that the relational data in the paired columns match) is calculated for each of the identified relationships based on the semantic similarity and the data content join coverage for such identified relationships. Based on the calculated join strength as well as other factors, a combination optimization algorithm identifies the best join combinations of relational data of the tables among a set of tables.

Recommending join operations of relational data among tables based on optimization model

Growing community of inventors

Xi'an, China

Jun Wang