Recommendation Systems for Exploratory Data Tasks

1 INTRODUCTION

• Data-dependent action utility and combinatorial choices make it difficult to select the next step, while poor choices can waste effort, reinforce confirmation bias, and leave insights undiscovered.Even a restricted example of aggregations over attributes yields about one million possible choices.Source: §1 · p. 1
• Recommendations must account for subjective user utility: the same data and intent can make different findings valuable to different users, while unvalidated suggestions risk consequential decisions.The paper contrasts an investigative journalist's interest in demographic disparities with a manager's interest in resource utilization.Source: §1 · pp. 1–2
• Multi-task bundles and workflows require coordination across dependent tasks, because later recommendations may need to account for earlier choices and their effects.The vision example uses discovery-inferred time-series semantics during wrangling and avoids unreliable trends after imputation.Source: §1 · p. 2
• The vision makes EDT recommendation a core data-system capability, targeting relevant, personalized, context-aware recommendations, practical efficiency, and support across single- and multi-task settings.The efficiency challenge includes enumerating combinatorial candidates and materializing them over data.Source: §1 · p. 2
• Existing data-system recommenders are often bespoke to individual tasks, while content recommenders assume pre-enumerated candidates and readily scored items rather than data-dependent actions.EDT action relevance requires materializing candidates on the specific data instance before measuring their results.Source: §1 · p. 2
• LLMs and AI agents may support semantic pruning and cross-task coordination, but the paper frames them as enablers rather than complete solutions to search and validation.The authors characterize EDT recommendation as a search-and-optimization problem, not merely generation.Source: §1 · p. 2

2 EDT RECOMMENDATION SYSTEM VARIANTS

• EDT action recommendation identifies the best action(s) for a user, data instance, constraints, and context to maximize progress toward implicit exploratory goals.Source: §2 · p. 2
• The framework categorizes recommendations by single-task or multi-task setting and by single-action, bundle, or workflow modality, yielding four variants.Source: §2 · p. 3, Figure 2 · p. 3

2.1 Top-𝑘recommendation for single task

• Single-task top-k systems rank candidate actions by user utility, while a generalizable framework could reuse shared components and require only a task-specific action language and utility notion.Source: §2.1 · p. 3

2.2 Bundle recommendation for single task

• Single-task bundle recommendation selects actions jointly for global objectives or constraints, but existing task-specific solutions do not generalize across EDTs.Because action distance depends on results materialized over a specific data instance, objectives and constraints must be pushed into recommendation.Source: §2.2 · p. 3

2.3 Workflow recommendation for single task

• Workflow recommendation is sequence optimization: early actions constrain later ones, and data transformations can make subsequent actions incompatible.The paper calls for targeted techniques that use action semantics to ensure workflow compatibility.Source: §2.3 · p. 3

2.4 Unified recommendations across tasks

• In multi-task EDTs, iterative and interdependent discovery, wrangling, and summarization require coordination across recommenders rather than isolated recommendations.Source: §2.4 · p. 3

3 EDT RECOMMENDATION COMPONENTS

• The paper favors search and optimization over generation-based recommendations without utility guarantees or ranking that requires pre-materialized candidates and often-unavailable training data.Search and optimization explores promising regions and can approximate utility and use efficient approximate search algorithms.Source: §3 · pp. 3–4
• The system combines context-aware personalization with search and optimization, compatibility checking, and outputs of top-k actions, bundles, or workflows.Source: §3 · p. 4

3.1 Contexts

• Four context sources inform recommendations: user, task, data, and ecosystem; existing and emerging EDTs provide limited support for these contexts.Source: §3.1 · p. 4, Table 4 · p. 4
• User responses support personalization and adaptivity, task context guides utility and pruning, data context supports semantic understanding and utility estimation, and ecosystem context enables cross-task coordination.Source: §3.1 · p. 4

3.2 Recommendation engine

• The search module prioritizes candidates aligned with user preferences and task requirements, pruning irrelevant, semantically invalid, or low-potential actions to reduce materialization.Source: §3.2.1 · p. 4

3.2.1 Search & optimization.

• For bundles and workflows, search enforces global constraints during candidate generation and materialization, fully materializing only promising candidates to balance quality and latency.Source: §3.2.1 · p. 4
• Compatibility checking tracks action order and removes incompatible sequences; cross-task recommenders must coordinate, track intertask dependencies, and obtain relevant external information.Source: §3.2.2 · p. 4

• Learn user preferences over action primitives and compose them to estimate action relevance; use richer data semantics to detect invalid actions and distinguish expected from surprising insights.The paper identifies LLMs as a possible means of inferring semantic context.Source: §4 · p. 5
• Reduce LLM cost and latency with caching, precomputation, or lightweight proxies, and reduce search through context-relevant candidates, semantic pruning, global constraints, and clustering.Source: §4 · p. 5
• Estimate utility from semantics, statistics, and context, or use anytime materialization and reinforcement learning to predict utility without materializing every action.Workflow materialization could stop when incompatibility is detected.Source: §4 · p. 5
• Develop decomposable global objectives and characterize when properties such as submodularity and monotonicity permit efficient greedy algorithms with approximation guarantees.Source: §4 · p. 5
• Action embeddings could decompose constraint computations using structural, semantic, and data-impact representations; compatibility graphs over action clusters could model task-specific ordering constraints.Source: §4 · p. 5
• Learn temporal preference embeddings from edits, refinements, and rejections to balance recent feedback with historical behavior.Source: §4 · p. 5
• Use multi-agent collaboration as an orchestration layer for cross-task coordination, with per-task recommenders providing context-request and conflict-report interfaces.Source: §4 · p. 5

• Evaluate generalizability across tasks and domains, scalability across data and candidate-space sizes, and interactive efficiency through latency and reductions in search and materialization costs.Source: §5 · p. 6
• Measure syntactic and semantic validity, bundle utility and constraint satisfaction, workflow utility and coherence, and multi-task success in achieving the end goal.Source: §5 · p. 6
• Validate user alignment through user studies and measure satisfaction, adaptivity, personalization, and efficiency across iterations and sessions.Suggested efficiency measures include iteration counts and time to task completion.Source: §5 · p. 6
• Assess robustness under noise, data skew, intent ambiguity, and task complexity, and analyze sensitivity to system parameters through ablations.Source: §5 · p. 6
• Human-validated utility models could support scalable evaluation by selecting a model through small-scale user studies, then using it to evaluate recommendations at scale.The proposed models cover criteria such as relevance, novelty, diversity, and coverage.Source: §5 · p. 6
• Build benchmarks from extracted or user-study workflows and brute-force ground truths; personalization and adaptivity benchmarks could pair intents with expert-grounded workflows.Source: §5 · p. 6

• Existing suggestion systems target specific tasks, including visualization, exploration, wrangling, analysis, explanation, and summarization.Source: §6 · p. 6
• Many machine-learning approaches assume small candidate spaces, require materializing all candidates, or provide only top-k suggestions.Auto-Formula and Auto-Suggest predict top-k suggestions for spreadsheet formulas and data-preparation steps.Source: §6 · p. 6
• Prior systems are often bespoke, ignore workflow and ecosystem context, and rely heavily on history-driven machine learning that is ineffective for EDTs.Source: §6 · p. 6

• The vision makes recommendation a first-class data-system capability rather than an afterthought or superficial chatbot feature.Source: §7 · p. 6
• The paper characterizes EDT recommendation across single-task and multi-task settings, and single-action, bundle, and workflow modalities.Source: §7 · p. 6
• Its research agenda spans context modeling, reducing search-space and materialization costs, cross-task coordination, and an evaluation framework.The paper calls for integrating data-management, recommendation-system, and constrained-optimization techniques with large language models and AI agents.Source: §7 · p. 6