[Paper Review] Table2Charts: Learning Shared Representations for Recommending Charts on Multi-dimensional Data

Mengyu Zhou, Qingtao Li+3 more|arXiv (Cornell University)|Aug 24, 2020

Time Series Analysis and ForecastingComputer Science3 references6 citations

TL;DR

Table2Charts proposes a deep reinforcement learning framework that learns shared representations from 196k tables and 306k charts to recommend appropriate visualizations. Using sequence generation with a copying mechanism and heuristic search, it achieves >0.61 recall at top-3 and >0.49 at top-1 for both single- and multi-type chart recommendations.

ABSTRACT

It is common for people to create different types of charts to explore a multi-dimensional dataset (table). However, to build an intelligent assistant that recommends commonly composed charts, the fundamental problems of multi-dialect unification, imbalanced data and open vocabulary exist. In this paper, we propose Table2Charts framework which learns common patterns from a large corpus of (table, charts) pairs. Based on deep Q-learning with copying mechanism and heuristic searching, Table2Charts does table-to-sequence generation, where each sequence follows a chart template. On a large spreadsheet corpus with 196k tables and 306k charts, we show that Table2Charts could learn a shared representation of table fields so that tasks on different chart types could mutually enhance each other. Table2Charts has >0.61 recall at top-3 and >0.49 recall at top-1 for both single-type and multi-type chart recommendation tasks.

Motivation & Objective

• Address the challenge of recommending appropriate charts for multi-dimensional tables in data exploration.
• Overcome fundamental issues in chart recommendation, including multi-dialect unification, class imbalance, and open-vocabulary generation.
• Enable mutual enhancement across different chart types by learning a shared representation of table fields.
• Develop a scalable, end-to-end framework for table-to-chart generation that generalizes across diverse chart types.

Proposed method

• Leverage deep Q-learning with a copying mechanism to generate chart sequences from table inputs.
• Model chart recommendation as a table-to-sequence generation task, where each sequence corresponds to a chart template.
• Use heuristic search to improve generation quality and explore diverse chart configurations.
• Train on a large-scale corpus of 196k tables and 306k (table, chart) pairs to learn shared representations of table fields.
• Apply attention mechanisms and sequence modeling to align table attributes with chart structure and semantics.
• Optimize for both single-type and multi-type chart recommendation by learning a unified representation space.

Experimental results

Research questions

• RQ1 Can a unified representation of table fields improve chart recommendation performance across diverse chart types?
• RQ2 How effective is deep Q-learning with copying mechanisms in generating accurate and diverse chart sequences from tabular data?
• RQ3 To what extent do multi-type chart recommendation tasks benefit from shared representations learned during training?
• RQ4 How does the framework handle data imbalance and open-vocabulary challenges in real-world chart recommendation?
• RQ5 What is the performance of the model in top-k chart recommendation, particularly in terms of recall at top-1 and top-3?

Key findings

• Table2Charts achieves a recall of >0.61 at top-3 for both single-type and multi-type chart recommendation tasks.
• The model attains a recall of >0.49 at top-1 across all recommendation scenarios, indicating strong precision in top-ranked suggestions.
• Shared representation learning enables mutual enhancement between different chart types, improving overall generalization.
• The copying mechanism effectively handles rare or out-of-vocabulary chart elements, supporting open-vocabulary generation.
• Heuristic search combined with deep Q-learning improves sequence generation quality and diversity.
• The framework demonstrates scalability and robustness on a large-scale corpus of 196k tables and 306k charts.