DeepSoft: a vision for a deep model of software

Hoa Khanh Dam, Truyen Tran, John C. Grundy, Aditya K. Ghose

Abstract

DeepSoft is a proposed end-to-end deep learning framework designed to model the complex evolution of software projects over time. By utilizing hierarchical Long Short-Term Memory (LSTM) networks, DeepSoft processes irregular, episodic sequences of software issues and releases. This approach seeks to eliminate the reliance on manual feature engineering, enabling advanced predictive analytics such as project risk forecasting and automated code patch recommendation.

Introduction

Software analytics, including bug prediction and project delay forecasting, currently relies heavily on manual feature engineering. Metrics such as "code complexity" or "code churn" are handcrafted by domain experts, a process that is expensive and often fails to generalize across diverse projects. Furthermore, traditional models often treat software as static snapshots, failing to account for the dynamic nature of software evolution.

The evolution of software is characterized by long-term dependencies, where historical design choices impact current bugs, and irregular timing, as commits and releases do not follow a uniform schedule. To address these challenges, the authors propose DeepSoft, a model capable of "memorizing" a project's entire history to automatically learn underlying patterns.

The DeepSoft Framework

DeepSoft is a compositional, multi-layered Recurrent Neural Network (RNN) built on LSTM units. The architecture is designed to model software at three distinct hierarchical levels:

  1. Issue Level: An LSTM processes sequences of resolved issues. Inputs include issue diagnoses (text embedded via NLP) and resolutions (e.g., code patches).
  2. Release Level: Hidden states from the issue level are aggregated via pooling into vectors representing specific releases, which are then processed by a second LSTM layer.
  3. Project Level: Release-level states are pooled to create a holistic project "health" vector, suitable for downstream predictive tasks.
The DeepSoft Hierarchical Architecture

Time-Aware Memory Modeling

A key innovation in DeepSoft is its handling of irregular timing. Standard LSTMs assume uniform time steps, which does not reflect the reality of software development. DeepSoft modifies the LSTM's forget gate to be a function of the elapsed time ($\Delta_t$) between events. This allows the network to decay memory of older, less relevant changes while retaining high-impact historical information.

Time-Aware LSTM Unit

Key Insights

DeepSoft conceptualizes software development as an evolving organism, similar to healthcare, where issues are "diagnosed" and patches are "treatments." By replacing handcrafted metrics with end-to-end learning, the model bypasses human bias. The resulting holistic vector representations can be utilized for various machine learning tasks, providing a unified mathematical representation of a project's history.

Future Research Agenda

As a vision paper, DeepSoft sets a roadmap for future work, including:

  • Source Code Modeling: Embedding Abstract Syntax Trees (ASTs) to capture structural interactions.
  • Resolution Recommendation: Using sequence-to-sequence models to generate code patches automatically.
  • Release Planning: Forecasting defect rates and feature implementation timelines.
  • Developer Modeling: Estimating team capability and automating bug triaging based on developer history.

While conceptual, DeepSoft provides a robust framework for the next generation of software analytics.

Summary updated

BibTeX
@inproceedings{dam2016deepsoft,
  author = {Hoa Khanh Dam and Truyen Tran and John C. Grundy and Aditya K. Ghose},
  title = {{DeepSoft: a vision for a deep model of software}},
  booktitle = {SIGSOFT FSE},
  year = {2016},
  doi = {10.1145/2950290.2983985},
}
Figure