Research Fields

This page describes how a research field was assigned to each dataset in the Scholar Data corpus. Field assignments are used to compute normalization factors that make D-index and S-index scores comparable across disciplines.

Approach

Each dataset was assigned a research field using the OpenAlex topics and domains classification, which is a four-level taxonomy comprising 4,516 topics grouped into 252 subfields, 26 fields, and 4 top-level domains. The subfields, fields, and domains follow Scopus's ASJC structure.

The top-level domains and fields are too broad for field-specific normalization since data sharing and reuse practices vary significantly at finer granularity of research communities. After analyzing dataset coverage at both the topic (4,516) and subfield (252) levels, the subfield level was selected as the most suitable classification for normalization. Topic-level classification was found to be too granular, with insufficient dataset coverage per topic.

Assigning Research Fields

OpenAlex Classifications

Where available, the primary topic and associated subfield assigned by OpenAlex were used directly. OpenAlex uses an LLM-based approach to assign topics to indexed works, incorporating both metadata and the reference network of each work, and provides a confidence score for each classification.

Custom Topic Classifier

Not all datasets in the corpus are indexed in OpenAlex. For those without an OpenAlex assignment, a custom high-speed topic classifier was developed to map dataset metadata to the OpenAlex topics taxonomy. Here are key details:

• Base model: a fine-tuned distilled static embedding model (Model2Vec potion-base-32M), chosen for its inference speed of 48,000+ records/second on CPU
• Training data: 10,000 English-language DataCite records with ground truth topic classifications derived from OpenAlex; 8,636 records used for training (80% split), 30 epochs, 1,135 unique topics
• Input: title, description (up to 1,000 characters), and keywords combined into an embedding vector, matched against pre-computed topic embeddings using approximate nearest neighbors
• Confidence: cosine similarity score used as the confidence measure

More details are available in the dedicated GitHub repository.

Classification Strategy

The final topic and subfield assigned to each dataset was determined as follows:

1. OpenAlex confidence ≥ 0.5: the OpenAlex assignment was used, even if the custom model had a higher score. At this confidence level, OpenAlex's use of reference network signals makes it likely more accurate than a metadata-only classifier.
2. OpenAlex confidence < 0.5: the assignment with the higher confidence score between OpenAlex and the custom model was retained. Below 0.5, OpenAlex assignments were observed to be unreliable. For example, the MIMIC-IV v3.1 dataset was assigned "Legal and Regulatory Analysis / Transportation" by OpenAlex with a confidence of 0.16, while the custom model assigned "Family and Patient Care in Intensive Care Units / Radiological and Ultrasound Technology" with a confidence of 0.47, which is a clearly more representative classification.
3. Not indexed in OpenAlex: the custom model's assignment was used regardless of score, to maximize field coverage across the corpus.
4. Non-Latin script metadata, not in OpenAlex: 16,700 datasets with metadata in non-Latin languages could not be classified by the current model and were left unassigned.

All final topic and subfield classifications were saved in an NDJSON file.

Results

Source	Topic in OpenAlex	Topic from model	Topic retained from OpenAlex	Topic retained from model	Total assigned
DataCite	15,324,819	48,983,961	7,394,117	41,598,712	48,992,829
EMDB	0	51,645	0	51,645	51,645
Total	15,324,819	49,035,606	7,394,117	41,650,357	49,044,474