Data Preprocessing
This page provides extended details of the metadata normalization, consolidation, and deduplication process used before study screening in the survey Fuzzing AI Systems: Foundations, Techniques, and Open Challenges.
The initial literature search retrieved 3,905 records from ACM Digital Library, ScienceDirect, Scopus, Semantic Scholar, and Google Scholar. Because these sources used different metadata formats, the exported records were standardized before filtering and eligibility assessment.
Metadata Consolidation
Records collected from the five scholarly sources were combined into a single dataset.
The consolidated dataset used a common structure containing the following fields:
| Metadata field | Description |
|---|---|
| Title | Title of the study |
| Authors | Listed study authors |
| Publication year | Year associated with the publication |
| Venue | Conference, journal, workshop, or preprint venue |
| Source | Scholarly database from which the record was retrieved |
| Abstract | Available abstract text |
| DOI | Digital Object Identifier, when available |
| URL | Publisher, indexing, or publication link |
| Citation information | Citation metadata available from the source |
Standardizing these fields enabled records from different databases to be compared and processed consistently.
Metadata Normalization
Metadata formats differed across sources in field names, ordering, missing values, and formatting conventions.
The normalization process included:
- mapping source-specific fields to the common metadata schema;
- standardizing title information for duplicate comparison;
- consolidating author and publication information;
- retaining DOI and URL information when available;
- standardizing publication-year fields; and
- preserving source information to support traceability.
The normalized records were then combined into a unified tabular dataset using a Python-based processing workflow.
Duplicate Detection
Because the same publication could appear in multiple scholarly databases, duplicate removal was necessary before screening.
We used two primary matching mechanisms.
Title-Based Duplicate Detection
Study titles were compared across the combined dataset to identify records that referred to the same publication.
Title-based matching was useful when:
- the same study appeared in multiple databases;
- DOI information was unavailable;
- DOI values were missing from one or more records; or
- metadata fields differed across indexing sources.
Potential duplicate records identified through title matching were consolidated before screening.
DOI-Based Matching
DOI values were used as a second identifier for duplicate detection.
Records sharing the same DOI were treated as references to the same study, even when differences existed in:
- title capitalization;
- author formatting;
- venue information;
- source-specific metadata; or
- URL structure.
Using both title-based and DOI-based matching reduced the risk of retaining duplicate database records.
Preprocessing Outcome
The preprocessing pipeline reduced the initial dataset as follows:
| Processing stage | Number of records |
|---|---|
| Records retrieved from all sources | 3,905 |
| Unique records after normalization and deduplication | 1,996 |
| Duplicate or overlapping records removed | 1,909 |
The resulting 1,996 unique records formed the input to the multi-stage screening process.
Relationship to Study Selection
After preprocessing, the 1,996 unique records were evaluated through:
- automated keyword filtering;
- precision-oriented filtering;
- manual title and abstract screening;
- backward and forward snowballing;
- quality and eligibility assessment;
- full-text eligibility screening; and
- duplicate-version removal.
Detailed screening and eligibility procedures are available on the Study Selection and Multi-Stage Filtering page.
The preprocessing stage removed duplicate database records before screening. A separate final consistency check was later used to remove duplicate study versions, such as cases in which both a preprint and a published version of the same work remained in the candidate corpus.