PaperlayerView exact revision on arXiv

Source-linked AI summary

ATMSeer: Increasing Transparency and Controllability in Automated Machine Learning

Qianwen Wang, Yao Ming, Zhihua Jin, Qiaomu Shen, Dongyu Liu, Micah J. Smith, Kalyan Veeramachaneni, Huamin Qu

arXiv:1902.05009v1cs.LGcs.HCstat.ML

TL;DR

AutoML’s black-box results can leave users uncertain about the search and unable to reason about disappointing outcomes. ATMSeer combines multi-granularity visualization with real-time search-space refinement. Its evaluations report utility and usability, while the user study was limited to 13 participants, predominantly graduate students.

  • Problem

    Black-box AutoML leaves users questioning whether the search explored enough, while difficulty analyzing results and modifying a complex search space can hinder reasoning about unsatisfying outcomes.

    Source: §1 · pp. 1–2

  • Method

    ATMSeer visualizes searched models at algorithm, hyperpartition, and hyperparameter levels and supports real-time search-space refinement.

    Source: §1 · p. 2

  • Results

    A series of evaluations demonstrated ATMSeer’s utility and usability, including two case studies, expert interviews, and a user study with 13 end users.

    Source: §1 · p. 2, §9 · p. 10

  • Takeaways & Limitations

    One expert reported that matching prior machine-learning knowledge to ATMSeer’s visualizations creates confidence in the AutoML process and increases the likelihood of adopting AutoML.

    Source: Knowledge Distillation from AutoML · p. 8

    The user studies involved only 13 participants, predominantly graduate students, so further investigation is needed to assess how generally the insights and results apply.

    Source: Limited Evaluation · p. 10

Abstract

from arXiv · show

To relieve the pain of manually selecting machine learning algorithms and tuning hyperparameters, automated machine learning (AutoML) methods have been developed to automatically search for good models. Due to the huge model search space, it is impossible to try all models. Users tend to distrust automatic results and increase the search budget as much as they can, thereby undermining the efficiency of AutoML. To address these issues, we design and implement ATMSeer, an interactive visualization tool that supports users in refining the search space of AutoML and analyzing the results. To guide the design of ATMSeer, we derive a workflow of using AutoML based on interviews with machine learning experts. A multi-granularity visualization is proposed to enable users to monitor the AutoML process, analyze the searched models, and refine the search space in real time. We demonstrate the utility and usability of ATMSeer through two case studies, expert interviews, and a user study with 13 end users.

1 INTRODUCTION

AutoML’s black-box search can leave users unsure whether it explored enough, while limited visibility makes unsatisfying results difficult to improve. ATMSeer provides multi-level visual analysis and real-time search-space refinement, with evaluation through case studies, interviews, and a user study.

1 INTRODUCTION

  • Black-box AutoML can prompt distrust and larger computational budgets, undermining efficiency when users cannot assess what was searched or why results disappoint.Users may be unable to reason about unsatisfying results and can only increase the budget as much as possible.Source: §1 · p. 1
  • Analyzing hundreds of heterogeneous models and modifying a complex hierarchical search space are key challenges for involving users in AutoML.Models differ in algorithms, hyperparameters, and performance; unsatisfying results can also reflect insufficient budget or algorithm limitations.Source: §1 · pp. 1–2
  • ATMSeer summarizes searched models at algorithm, hyperpartition, and hyperparameter levels and lets users modify the search space in real time.Its in-situ configuration connects analysis of model behavior with search-space modification.Source: §1 · p. 2
  • ATMSeer integrates with the ATM framework but is not algorithm-specific and can integrate with other AutoML frameworks.Source: §1 · p. 2
  • The paper contributes an AutoML-use workflow, an interactive tool for monitoring, analysis, and refinement, and evaluations involving two case studies, two expert interviews, and 13 end users.Source: §1 · p. 2

2 RELATED WORK

Model-selection guidance and visualization tools offer useful support, but prior approaches described here do not provide detailed guidance for a particular problem or analyze models across varied algorithms at multiple granularities. ATMSeer addresses these gaps with interactive analysis of searched models.

Choosing Machine Learning Models

  • No single machine-learning model works best for every problem, so practitioners rely on algorithm knowledge, data understanding, and time-consuming trial and error.Source: Choosing Machine Learning Models · p. 2
  • Existing theoretical and empirical model-selection guidance offers general pros and cons or dataset-wide comparisons, while ATMSeer aims to help users analyze models for a particular dataset.The cited empirical study evaluated 179 classifiers on 121 datasets and found random forests most likely to be a good classifier, followed by support vector machines and neural networks.Source: Choosing Machine Learning Models · p. 2

Visualizing Automated Machine Learning

  • Prior visualization tools include search-space definition, parallel coordinates, and user-in-the-loop optimization, but the described systems support one model type at a time.ATMSeer supports 14 machine-learning algorithms and multi-granularity visualization of searched models.Source: Visualizing Automated Machine Learning · p. 2

D2. Computational

The supplied material depicts model-choice and search-space decisions alongside a broader account of visual analytics for machine learning. It distinguishes model-behavior analysis tools from tools that must handle many searched model configurations in AutoML.

D2. Computational

  • The workflow fragment frames model choice as a decision to analyze or reason about, with prior knowledge, preferences, efficiency, and model familiarity among the depicted considerations.Source: D2. Computational · p. 3
  • The workflow fragment identifies algorithms, hyperpartitions, and hyperparameters as ranges users may modify.Source: D2. Computational · p. 3

Visualizing Machine Learning Models

  • Visual analytics tools address understanding, diagnosis, and refinement of models including generative models, reinforcement learning, and convolutional and recurrent neural networks.Source: Visualizing Machine Learning Models · p. 3
  • Instance-level tools connect model performance summaries to individual examples, but do not directly address analysis of the configurations of many models searched by AutoML.The configurations include algorithm, hyperpartition, and hyperparameter.Source: Visualizing Machine Learning Models · p. 3

3 SYSTEM REQUIREMENTS AND DESIGN

The design targets users with machine-learning expertise who face costly manual search, and treats each AutoML model as a structured data point. Interviews inform a workflow of three user decisions, which in turn motivates requirements for overview, model–search-space context, modification guidance, in-situ configuration, and multi-granularity analysis.

Goals & Target Users

  • ATMSeer targets users with machine-learning expertise who have experienced time-consuming and error-prone manual model search.Source: Goals & Target Users · p. 3

Data Abstraction

  • Each AutoML model is represented as a multivariate data point with algorithm, hyperpartition, hyperparameter, and performance attributes.Algorithms are categorical, hyperparameters are numerical, and performance is numerical.Source: Data Abstraction · p. 3

User Interview

  • Six interviewees—three domain practitioners and three machine-learning experts—discussed their experience, expectations and concerns about AutoML, and use of a pilot system.Interviews lasted approximately 45 minutes; half used their own data and half used example data.Source: User Interview · p. 3

The Workflow

  • Interviews identified decisions about search-space modification, computational budget, and model choice, with choices informed by prior knowledge, ongoing results, potential improvement, expenses, and domain requirements.Source: The Workflow · pp. 3–4

Design Requirements

  • The design requirements call for process overview, linking models to the search space, modification guidance, in-situ configuration, and multi-granularity analysis.Source: Design Requirements · p. 4

4 ATMSEER

ATMSeer combines controls for managing AutoML with visual summaries that let users inspect searched models at multiple levels and refine the search space. Its profiler supports comparisons from algorithms down to hyperparameter values, while the interface updates as a process runs.

System Overview

  • The client-server system coordinates AutoML, data and model storage, and a visual client that controls processes and presents summaries.The client can create, configure, start, and pause processes; it can also request summaries of recorded data.Source: System Overview · p. 5

Interface

  • The overview panel reports process statistics, algorithm and hyperpartition coverage, performance distributions, and top models for comparison.Focus mode highlights the algorithms and hyperpartitions associated with selected top models.Source: AutoML Overview · p. 5
  • The profiler organizes analysis at algorithm, hyperpartition, and hyperparameter levels, with finer views available on demand for advanced users.These views help users assess whether to resume a process, refine its search space, or choose among models.Source: AutoML Profiler · p. 5
  • Hyperpartition bars group models by fixed nontunable settings, while box height encodes each model’s performance.A hyperpartition is an algorithm configuration with fixed nontunable hyperparameters; only numeric hyperparameters can be tuned.Source: AutoML Profiler · pp. 5–6
  • Hyperparameter scatter plots relate each tunable value to performance, and distribution plots show which values have been tried extensively.ATMSeer uses scatter plots rather than parallel coordinates because users preferred their simplicity and found parallel-coordinate analysis interaction-intensive.Source: AutoML Profiler · p. 6

Interaction Design

5 CASE STUDY

The case studies show experts using ATMSeer to inspect model performance, select models, and refine an AutoML search space. In one comparison, an expert reached 0.922 within 210 models, while none of 20 unattended runs reached 0.92.

5 CASE STUDY

Select and analyze models (D2, D3)

  • For the arsenic-female-bladder dataset, E1 selected KNN after observing that its performance distribution was concentrated toward the top.The top 10 models scored 0.936–0.939 across four algorithms, and the dataset search covered all algorithms after 250 models.Source: Select and analyze models (D2, D3) · pp. 6–7
  • Among KNN models, small n_neighbors values were associated with poor performance, while individual hyperpartitions included both strong and poor models.The figure caption reports no obvious relationship between hyperpartition choice and performance.Source: Select and analyze models (D2, D3) · p. 7, Figure 6 · p. 7

Refine Search Space (D1, D2)

  • E2 focused the search on ET, raising its best performance from 0.887 to 0.906 after 30 more models, then reached 0.922 after setting max_features to [0.7, 1.0] and searching 50 more models.E2 stopped after the last 20 models produced no further improvement.Source: Figure 7 · p. 7, Refine Search Space (D1, D2) · p. 7, Figure 6 · p. 7
  • 0.922 was reached by the expert within 210 models, while none of 20 unattended runs reached 0.92.The authors say human involvement has the potential to improve AutoML performance and efficiency, possibly through more aggressive search-space changes.Source: Refine Search Space (D1, D2) · pp. 7–8, Figure 7 · p. 7

6 EXPERT INTERVIEW

Expert interviews suggest ATMSeer helps users interpret model behavior and interact with AutoML, while also exposing possible algorithm and reward-function issues. The experts saw these visual insights as useful for understanding, adoption, and diagnosing AutoML systems.

6 EXPERT INTERVIEW

  • Two closely collaborating AutoML experts, including a co-author of the ATM framework, provided feedback in weekly meetings over more than two months.Their interviews informed three usage scenarios for ATMSeer.Source: §6 · p. 8

Knowledge Distillation from AutoML

  • Some algorithms showed more stable performance distributions, while the strongest hyperpartitions and hyperparameters for the same algorithm varied across datasets.The experts said these observations could inform users about the importance of tuning certain algorithms.Source: Knowledge Distillation from AutoML · p. 8, Figure 8 · p. 8

Human-Machine Interaction in AutoML

  • The experts said matching prior machine-learning knowledge to ATMSeer’s visualizations built confidence in AutoML and could help users learn unfamiliar algorithms.They also believed that human observation and prior knowledge can sometimes be more efficient than AutoML algorithms when search spaces are large and budgets limited.Source: Knowledge Distillation from AutoML · p. 8, Human-Machine Interaction in AutoML · p. 8

Diagnosis of AutoML

  • ATMSeer revealed that MLP was searched much more often than ET despite similar best performance and a lower MLP performance-distribution mode.The experts traced this pattern to the bandit setup, where each hyperpartition is a separate arm, disadvantaging algorithms with fewer hyperpartitions.Source: Diagnosis of AutoML · p. 8
  • For a given dataset, SVM-RBF’s higher best score and more stable distribution contrasted with nearly equal search counts for SVM-RBF and SVM-Sigmoid (27 vs. 26).The experts attributed the similar counts to a reward based on the average performance of each hyperpartition’s best five models and reconsidered the reward-function design.Source: Diagnosis of AutoML · pp. 8–9

7 USER STUDY

A 13-person user study examined participants’ use of ATMSeer to search for a good model and analyze an existing AutoML run. Participants generally answered model-analysis questions correctly and reported confidence in their selections, while their search-space and budget decisions reflected a coarse-to-fine process and multiple signals.

7 USER STUDY

  • The study compared user behavior with a characterized workflow rather than a baseline, which lacked detailed AutoML-result information for informative decisions and search modification.Participants received a tutorial, could think aloud and ask questions, and completed a roughly 40-minute session.Source: §7 · p. 9, Procedure · p. 9

Datasets & Tasks

  • The study used two tasks: searching for a good model on German Credit data and analyzing a pre-run 200-model search on the Friedman dataset.The analysis task included objective questions about the three-level profiler and subjective questions about decision-making.Source: Datasets & Tasks · p. 9

Usability

  • Participants answered 99 of 104 objective model-analysis questions correctly or fluently, with most errors or omissions on hyperparameter-level questions linked to limited familiarity with machine-learning models.Two participants skipped a question about SGD classifiers because they were unfamiliar with them and unsure of their observations.Source: Usability · p. 9, Table 1 · p. 10

Revisiting the Workflow

  • Among 13 participants, 10 refined the search space coarse-to-fine, first narrowing algorithms and later adjusting hyperparameter ranges.Participants made these changes after observing models from ongoing searches.Source: Revisiting the Workflow · p. 9

Usability

  • Most participants agreed ATMSeer was easy to learn and use; 84.6% were confident in their selected model and 92.3% were willing to use ATMSeer again.Two participants were not confident in their selections and wanted additional validation with familiar tools.Source: Figure 9 · p. 9, Usability · p. 9

Revisiting the Workflow

  • Participants considered multiple signals when deciding whether to increase the computational budget: performance histograms and algorithm coverage were each cited by nine participants.Five participants each considered the number of searched models and the best performance.Source: Revisiting the Workflow · p. 10
  • Six advanced participants searched an average of 117 models, compared with 97.5 models for seven novice participants; nine participants also preferred familiar models alongside good performance.Three participants said ATMSeer helped them understand models, increasing their familiarity and willingness to use them.Source: Revisiting the Workflow · p. 10

8 DISCUSSION & FUTURE WORK

The discussion identifies learning and AutoML debugging as potential uses for ATMSeer, while noting that evaluation involved a small, predominantly graduate-student sample. The visualization supported analysis of processes with over 1,000 models, and future work includes broader evaluation and intelligent human-in-the-loop control.

General Applicability

  • ATMSeer may help beginners see how algorithms, hyperpartitions, and hyperparameters influence performance, and help AutoML designers identify possible bugs and improvement opportunities.The authors identify these as potential scenarios beyond the tool’s initial design for machine-learning experts.Source: General Applicability · p. 10

Limited Evaluation

  • The user studies involved only 13 participants, predominantly graduate students, so broader applicability of the findings remains to be validated.The authors propose further investigation with a larger and more diverse user group.Source: Limited Evaluation · p. 10

Scalability of the Visualization

  • ATMSeer’s visualization supported fluent analysis of AutoML processes with over 1,000 models.The paper contrasts this capacity with typical AutoML searches of 100 to 400 models and systems supporting 8 to 15 algorithms.Source: Scalability of the Visualization · p. 10

Future Work

  • Future work includes evaluating ATMSeer in real-world applications and combining it with human-in-the-loop reinforcement learning to detect when human involvement may be needed.A possible trigger is an AutoML process stuck in a local optimum.Source: Future Work · p. 10

9 CONCLUSION

ATMSeer combines a workflow derived from expert interviews with multi-granularity visualizations and in-situ configuration to support real-time analysis and refinement of AutoML searches. Evaluations indicated utility and usability, and the user study found coarse-to-fine refinement and greater willingness to interact among participants with higher machine-learning expertise.

9 CONCLUSION

  • ATMSeer’s multi-granularity visualization and in-situ configuration support real-time examination at algorithm, hyperpartition, and hyperparameter levels.The design was guided by a workflow identifying search-space updates, computational-budget changes, and model-choice reasoning as key decisions.Source: §9 · p. 10

Related papers

Loading 1902.05009v1…