Source-linked AI summary
A Survey of Visual Analytics Techniques for Machine Learning
Jun Yuan, Changjian Chen, Weikai Yang, Mengchen Liu, Jiazhi Xia, Shixia Liu
TL;DR
Visual analytics for machine learning lacks a comprehensive account spanning the entire learning pipeline. This paper systematically reviews 259 papers, builds a three-stage taxonomy, and identifies representative tasks and future research directions. The survey organizes techniques before, during, and after model building while highlighting open problems including weakly supervised data quality and online training diagnosis.
Problem
The rapid growth of visual analytics for machine learning created a need for a comprehensive review covering how visualization techniques are designed and applied throughout machine learning pipelines.
Method
The authors manually reviewed relevant top-tier visualization venues from 2010–2020 and organized the literature according to three machine learning pipeline stages.
Results
The survey classified 259 papers into techniques before, during, and after model building, detailing typical analysis tasks and representative works.
Takeaways & Limitations
The taxonomy and six proposed research directions provide a starting point for practitioners and researchers developing visual analytics techniques for machine learning.
Takeaways & Limitations
Existing visual analytics work devotes fewer efforts to exploiting unlabeled data and addressing inexact data in weakly supervised learning.
Abstract
from arXiv · showhide
Visual analytics for machine learning has recently evolved as one of the most exciting areas in the field of visualization. To better identify which research topics are promising and to learn how to apply relevant techniques in visual analytics, we systematically review 259 papers published in the last ten years together with representative works before 2010. We build a taxonomy, which includes three first-level categories: techniques before model building, techniques during model building, and techniques after model building. Each category is further characterized by representative analysis tasks, and each task is exemplified by a set of recent influential works. We also discuss and highlight research challenges and promising potential future research opportunities useful for visual analytics researchers.
1 Introduction
Visual analytics for machine learning has grown rapidly as a way to make learning processes more explainable, trustworthy, and reliable. This survey responds by systematically reviewing the field across the full machine learning pipeline.
- The recent success of artificial intelligence applications depends on the performance and capabilities of machine learning models.
- Visual analytics combines interactive visualization and machine learning to facilitate analysis of major components in the learning process and improve performance.
- The field needs a comprehensive review explaining how visualization techniques are designed and applied across machine learning pipelines.
- The survey reviews 259 papers and organizes them into techniques before, during, and after model building.
- It characterizes representative tasks, including data and feature quality improvement, model understanding, diagnosis, steering, and post-deployment data understanding.
2 Survey Landscape
The survey constructs a pipeline-based landscape from a systematic review of visualization-community research. It groups techniques by when they operate and distinguishes their principal analysis goals.
- Paper Selection: Three Ph.D. candidates manually reviewed top-tier visualization venues from 2010–2020, selecting 259 papers after title and abstract screening.
- Taxonomy: The taxonomy maps visual analytics techniques to preprocessing before model building, model building itself, and deployment after the model is built.
- Before Model Building: Before model building, techniques prepare data through data quality improvement and feature engineering.
- During Model Building: During model building, methods address model understanding, diagnosis, and steering, including interactive topic refinement.
- After Model Building: After deployment, techniques target model users by intuitively presenting and exploring model outputs, categorized as static or temporal data analysis.
3 Techniques before Model Building
Before model building, visual analytics focuses on preparing data and features by identifying quality problems and supporting interactive correction, labeling, and feature analysis.
- Low-quality data and features degrade machine learning performance, making data processing and feature engineering critical before model building.
- Instance-level Improvement: Data-quality methods address instance-level problems such as missing values, duplication, and out-of-distribution samples.
- Instance-level Improvement: OoDAnalyzer combines ensemble detection with grid visualization to detect out-of-distribution samples and explore their presence in context.
- Instance-level Improvement: Interactive data-quality tools expose preprocessing effects, uncertainty, privacy risk, and utility changes so users can compare operations and choose among them.
- Label-level Improvement: Label-level methods improve noisy labels or support interactive labeling through trusted-item propagation, misclassification exploration, grouping, filtering, and sorting.
4 Techniques during Model Building
During model building, visual analytics addresses the opacity and practical difficulty of machine learning models through model understanding, diagnosis, and steering. These methods help developers inspect mechanisms, identify training problems, and improve performance interactively.
- Machine learning models are often treated as black boxes, limiting their practical use in risky scenarios such as self-driving cars and financial investment.
- Visual analytics methods during model building reveal model mechanisms and support developers in building well-performing models.
- Model understanding reduces reliance on time-consuming trial and error by explaining how models work and produce outputs.
- Model diagnosis targets issues in failed or unsatisfactory training processes, while model steering supports interactive performance improvement.
4.1 Model Understanding
Model understanding methods explain either how parameter changes affect outputs or how models produce outputs. Behaviour explanations include network-centric, instance-centric, hybrid, and surrogate-model approaches.
- Model understanding examines parameter effects and model behaviours.
- Network-centric methods visualize model structure and how components cooperate to produce outputs.Directed graph layouts can suffer from visual clutter as model structures become more complex.
- Instance-centric methods analyze representations, individual instances, and relationships between instances.Examples use projections, parallel coordinates, and bipartite graphs to study learned spaces, hidden states, and sentence-level explanations.
- Hybrid methods combine instance-level analysis with network architecture context to support understanding of model mechanisms.Summit combines an embedding view of activations with an attribute graph of influential neuron connections.
- Surrogate explainable models treat classifiers as black boxes and visualize rules extracted from their inputs and outputs.These methods are useful for users with limited machine learning knowledge.
4.2 Model Diagnosis
Model diagnosis methods analyze either training results or training dynamics. They address performance, fairness, vulnerabilities, and failures in model training through visual analytics systems.
- Model diagnosis analyzes training results or training dynamics.
- Analyzing Training Results: Performance-oriented tools organize samples by prediction classes and encode true and false positives and negatives to estimate metrics at multiple detail levels.Squares uses boxes and textures for this purpose.
- Analyzing Training Results: AEVis analyzes adversarial examples by extracting prediction datapaths from normal and adversarial inputs, then visualizing path patterns and critical feature maps.Its river-based visualization shows diverging and merging paths, while a layer-level view shows critical feature maps.
- Analyzing Training Dynamics: Training-dynamics techniques support debugging by preserving outliers and major distributions and tracing neuron interactions during failed generative-model training.DGMTracker uses bluenoise polyline sampling and credit assignment to investigate failure causes and propagation.
- Analyzing Training Dynamics: ReVision combines constrained hierarchical clustering with an uncertainty-aware tree visualization to help users refine hierarchical topic-modeling results.
4.3 Model Steering
Model steering places users in the loop through interactive refinement or selection among model ensembles. These approaches incorporate human knowledge, correct results, or compare models across configurations.
- Model steering uses two strategies: refining models with human knowledge or selecting the best model from an ensemble.
- Refining Models with Human Knowledge: Visual analytics techniques place users in the model-refinement loop through flexible interaction.
- Refining Models with Human Knowledge: ProtoSteer lets users add, delete, and revise learned prototypes before retraining the prototype sequence network.
- Refining Models with Human Knowledge: ReVision incorporates user constraints into hierarchical clustering and lets users refine the constraint hierarchy by dragging.An uncertainty-aware tree visualization displays constraints and clustering results.
- Model Selection from an Ensemble: Ensemble-selection systems guide choices among clustering and regression models using recommendations, varied algorithms, hyperparameters, instance weights, and feature selection.Clustrophile 2 supports clustering analysis, while BEAMES supports multimodel regression steering and selection.
- Model Selection from an Ensemble: Static text-visualization examples reveal relationships through topic-graph layouts or user-similarity projections.TopicPanorama compares topic graphs from multiple sources, while DemographicVis uses t-SNE after analyzing posting contents.
5 Techniques after Model Building
After model building, visual analytics helps users understand model outputs and data-analysis results. The survey distinguishes static from dynamic results, with especially extensive work in text analytics and temporal analysis.
- Post-model-building techniques help users understand model outputs and high-dimensional data-analysis results.Methods are categorized by data type and by whether the analyzed results are static or dynamic.
- Static Data Analysis: Research on static analysis results focuses mainly on textual data, with fewer studies addressing other data types.
- Static Data Analysis: Visual text analytics combines interactive visualization with text mining methods such as document clustering, topic models, and word embeddings.
- Static Data Analysis: TopicPanorama extracts and matches topic graphs from multiple sources, clusters them hierarchically, and displays them with hybrid visualizations.
- Dynamic Data Analysis: Dynamic-result analysis includes topic, event, and trajectory analysis, and can be conducted offline or online.
- Dynamic Data Analysis: Topic-evolution systems commonly use river metaphors to show changing themes, including topic birth, death, merging, and splitting.ThemeRiver shows theme-volume changes, while TextFlow represents topic transitions with a river-based metaphor.
- Dynamic Data Analysis: Event-analysis methods discover and display sequential patterns at multiple granularities, from patterns and segments to sequences and events.
- Dynamic Data Analysis: Trajectory-analysis methods combine spatiotemporal clustering, semantic enrichment, and frequent-pattern visualization for movement data.Geographic and temporal views can jointly show frequent routes, destinations, and temporal patterns.
6 Research Opportunities
The survey identifies research opportunities spanning data quality, explainable feature engineering, and the broader challenges of weakly supervised learning. Key needs include better use of unlabeled or inexact data and more interpretable handling of deep features.
- 6.1 Opportunities before Model Building: Weakly supervised learning requires improved handling of inaccurate, incomplete, and inexact labels, but unlabeled and coarse-grained data remain underexplored.The survey highlights these gaps as potential future research directions.
- 6.1.1 Improving Data Quality for Weakly Supervised Learning: Visual analytics could improve semi-supervised learning by helping experts exploit unlabeled data alongside limited labeled data.Graph-based methods depend on relationships between labeled and unlabeled instances.
- 6.1.1 Improving Data Quality for Weakly Supervised Learning: Inexact labels, such as scan-level tumor diagnoses without precise slice or location information, require fine-grained validation to improve data quality.Interactive visualization could expose data distributions and wrong predictions while supporting more finely grained labeling with limited expert effort.
- 6.1.2 Explainable Feature Engineering: Deep neural-network features can outperform handcrafted features yet remain difficult to interpret, complicating feature engineering.Data bias may cause extracted features to emphasize incidental properties, such as color, instead of discriminating concepts such as shape.
- 6.1.2 Explainable Feature Engineering: Visual analytics is needed to convey redundancy among deep features so experts can remove redundant representations.Removing redundancy can reduce storage requirements and improve generalization, but judging redundancy is difficult without understanding feature meanings.
6.2 Opportunities during Model Building
During model building, the survey emphasizes moving from offline diagnosis toward online monitoring and making iterative model refinement more intelligent. These directions aim to support timely intervention and faster refinement.
- 6.2.1 Online Training Diagnosis: Online diagnosis is needed because offline tools collect data only after training, limiting assistance when training lasts days or weeks.Continuous analysis could help developers identify anomalies and adjust training promptly.
- 6.2.1 Online Training Diagnosis: The central challenge for online diagnosis is detecting training anomalies accurately and quickly as process data continuously arrives.Interactive visualization is presented as a way to locate potential errors even while automatic real-time detection remains difficult.
- 6.2.2 Intelligent Model Refinement: Interactive model refinement can use uncertainty scores and visual hints to direct users toward outputs requiring examination.Further guidance may incorporate information beyond uncertainty scores.
- 6.2.2 Intelligent Model Refinement: Because refinement is iterative, later guidance could learn from users’ earlier interactions, such as must-link and cannot-link clustering constraints.These constraints can instruct a model during subsequent refinement.
6.3 Opportunities after Model Building
After model building, the survey highlights multimodal data understanding and concept-drift analysis as important opportunities. These areas address richer real-world inputs and changing relationships between inputs and outputs.
- 6.3.1 Multimodal Data Understanding: Real-world applications increasingly combine modalities such as text, audio, and images, creating opportunities for multimodal visual analytics.The survey notes that existing content-analysis work has focused largely on single-modal data.
- 6.3.1 Multimodal Data Understanding: Vision-and-language visual grounding can identify image regions corresponding to descriptions and support natural-language image retrieval.This connects a classic multimodal task with interaction in visual environments.
- 6.3.2 Analyzing Concept Drifts: Concept drift occurs when the mapping from input data to output values changes over time, causing models trained on historical data to degrade on new data.Machine learning research studies drift detection, understanding, and adaptation.
- 6.3.2 Analyzing Concept Drifts: Visual analytics can complement numerical drift measures by helping users analyze changing model behavior and understand the source of drift.The cited discussion states that numerical values alone make drift difficult to understand.
7 Conclusions
The survey organizes visual analytics for machine learning by three analysis stages and identifies six future research directions across the pipeline. It aims to summarize the state of the art and inform future research.
- 7 Conclusions: The survey classifies visual analytics techniques into before, during, and after model building, with representative tasks and works for each stage.It systematically reviews recent progress and developments in the area.
- 7 Conclusions: Six future directions cover weakly supervised data quality, explainable feature engineering, online diagnosis, intelligent refinement, multimodal understanding, and concept-drift analysis.These directions span all three stages of the machine learning pipeline.