Computer Science Journal: Publishing CS and AI Research

Computer science publishing has a distinctive character. The field moves quickly, much of it is conference-driven, and reproducibility has become the dominant review criterion. A CS paper is judged less on novelty in the abstract sense and more on whether a competent reader could rebuild what you built and get the numbers you report.

This page covers what a computer science or AI manuscript needs to contain, the objections reviewers raise most often, and how to report machine learning results honestly. It applies to submissions in software engineering, systems, networks, security, data science and artificial intelligence.

What counts as a contribution

State your contribution explicitly, ideally as a short numbered list at the end of the Introduction. A contribution is something a reader can use: a new method, a new dataset, a new empirical finding, a new analysis of a known method, or a system that demonstrably achieves something previously impractical.

"We built an application using existing libraries" is a project report, not a research contribution — unless the paper also shows something non-obvious: a measured improvement over the standard approach, a deployment study with real users, or an analysis of where the standard approach fails.

Reproducibility: the reporting checklist

• Dataset name, source, version, licence, and exact train/validation/test split with sizes.
• All preprocessing steps in the order applied, including any filtering or class balancing.
• Model architecture in enough detail to reimplement: layers, dimensions, activation functions.
• Every hyperparameter: learning rate, optimiser, batch size, epochs, regularisation, random seed.
• Hardware and software environment: GPU or CPU, framework and version, training wall-clock time.
• Evaluation protocol: metrics, how they are computed, number of runs, and variance across runs.
• Where possible, a link to code and trained weights.

Baselines and ablations

A result without a baseline is not evidence. Compare against the standard method for your problem and, where a published state-of-the-art exists, against that, using the same data split and the same metric. Reporting your model at 94% accuracy without stating what a simple logistic regression or a stock pretrained model achieves on the same split tells a reviewer nothing.

Ablation studies answer the question "which part of your method is doing the work?". Remove each component in turn and report the effect. Papers that propose a five-part architecture without ablations routinely receive major-revision decisions, because the reviewer cannot tell whether the improvement came from the novel component or from the extra parameters.

Reporting metrics honestly

Choose metrics appropriate to the problem. Accuracy on an imbalanced dataset is misleading; report precision, recall, F1 and, where relevant, per-class results and a confusion matrix. For ranking and retrieval report the metric your field uses, and state the cut-off. For regression report the error metric and the scale of the target.

Report variance. A single run of a stochastic training process is an anecdote. Give mean and standard deviation over multiple seeds, and state how many runs. Do not report a test-set number that was used to select hyperparameters — that is the most common methodological error in student CS papers and reviewers look for it specifically.

Objections reviewers raise most often

• No baseline, or a baseline that is deliberately weak.
• Test set used for model selection, contaminating the reported result.
• Dataset undisclosed or not reproducible.
• Results from a single run with no variance reported.
• Related work section that lists papers without contrasting them with the present approach.
• Claims of real-time or scalable performance without latency, throughput or memory measurements.
• Figures that are screenshots of a framework's default plots, unreadable at print size.
• No discussion of limitations or failure cases.

Special considerations for AI and machine learning papers

Where a model is trained on data about people, address provenance and consent, and discuss bias in the dataset and in the outputs. Where a model could plausibly be misused, a short limitations and ethics paragraph is expected practice, not an optional flourish.

If you fine-tuned or prompted a large pretrained model, name the model and version explicitly. Results obtained from a hosted model that has since changed are not reproducible unless the version and date are recorded.

Publishing computer science research with IJVAST

IJVAST publishes computer science, artificial intelligence, data science, networking, security and electronics research within its multidisciplinary scope, routing each manuscript to a reviewer from the relevant area. Papers appear in IEEE two-column format, open access under CC BY 4.0, with an editorial decision targeted within 48 hours.

The fee is a flat Rs.1,200 for Indian authors including 18% GST, or $25 USD for international authors. IJVAST is published by SkyInnovate Technologies. Its ISSN application and Crossref DOI membership application are both in progress; until those are granted the journal states this openly rather than displaying a number it does not yet hold.

Frequently asked questions