Abstract
Exploratory data analytics (EDA) is a sequential decision making process where analysts choose subsequent queries that might lead to some interesting insights based on the previous queries and corresponding results. Data processing systems often execute the queries on samples to produce results with low latency. Different downsampling strategy preserves different statistics of the data and have different magnitude of latency reductions. The optimum choice of sampling strategy often depends on the particular context of the analysis flow and the hidden intent of the analyst. In this paper, we are the first to consider the impact of sampling in interactive data exploration settings as they introduce approximation errors. We propose a Deep Reinforcement Learning (DRL) based framework which can optimize the sample selection in order to keep the analysis and insight generation flow intact. Evaluations with 3 real datasets show that our technique can preserve the original insight generation flow while improving the interaction latency, compared to baseline methods.
Introduction
Exploratory Data Analysis (EDA) is an iterative, "human-in-the-loop" process where analysts run queries and refine their search based on visual feedback. To maintain interactivity on massive datasets, systems often employ Approximate Query Processing (AQP), running queries on small data samples. However, a significant challenge arises: "intent divergence." If a sample produces a distorted visualization, the analyst may be misled into chasing false trends or missing critical insights.
This paper introduces APPROXEDA, a framework that treats sample selection as a sequential decision-making problem. By using Deep Reinforcement Learning, the system automatically selects the optimal sampling strategy for each step to maximize speed while ensuring the analyst reaches the same conclusions they would have found on the full dataset.
The APPROXEDA Framework
APPROXEDA utilizes an Advantage Actor-Critic (A2C) agent to navigate the trade-off between query latency and accuracy. The framework is built on several core components:
- Intent Identification: Since goals are rarely explicit, the system uses Biterm Topic Modeling (BTM) to cluster query sequences into latent "intents," providing necessary context for the RL agent.
- State and Action Space: The agent monitors query history, visualization results, and cumulative latency. It can choose from 29 different sampling actions, including uniform, stratified, and cluster-based sampling at various rates.
- Reward Function: The agent is rewarded for reducing latency, minimizing "intent divergence" (keeping the sampled path close to the ground-truth path), and achieving high insight recall at the end of a session.
Experimental Evaluation
Due to the lack of real-world EDA logs with ground-truth intent, the authors utilized the ATENA simulator to generate thousands of realistic analysis sequences. The system was tested on three public datasets: Flights (5.2M records), Housing (324k records), and Income (426k records).
Key Results
- Significant Speedups: APPROXEDA achieved a median latency reduction of nearly 90% compared to full-dataset processing.
- Superior Accuracy: The framework outperformed intelligent baselines like BlinkDB and CIGREEDY in preserving the original intent distribution and final insights.
- Adaptive Strategy: Qualitative analysis revealed that the agent learned to use very small samples for simple exploratory queries while switching to more robust, larger samples at critical decision points in the analysis.
Conclusion and Future Work
APPROXEDA demonstrates that Reinforcement Learning can effectively manage the risks of approximation in interactive data exploration. While the current model assumes consistent intent throughout a session, future work will address "intent-shift"—where users abruptly change goals—and explore methods to reduce the "cold start" training requirements for new datasets.
Summary updated