Our New Solution To Classification
In 2020, we developed an advanced classifier. This tool adds multiple labels to procurement notices based on the Common Procurement Vocabulary (CPV), with an accuracy of 84%. This classifier gives notices five scored, Level 3 CPV codes based on their text and description, that work for the 90 languages in our database. This classifier can process millions of recommendations within hours. As of writing, we are classifying all our new records each day, with full predictions appearing in our API in early in 2021.
This classifier is a series of algorithms that were trained on millions of rows of procurement data in our database. Our model has an F1 score of at least 84%, a precision of 83% and a recall of 88%.
What Does F1, Precision and Frequency Mean?
These terminologies are industry standard terms within the practise of demonstrating algorithmic accuracy. F1 score is the ‘harmonic’ mean or average of precision and recall. Precision analyses correct and incorrect predictions based on the rows that went into the model. Recall looks at the wider picture: which rows the model got right versus the ones it completely missed.
The Challenges With The Classification
Developing code to categorise spending and tenders has proven technically challenging, due to a number of complicating factors in the underlying published data and the scale of the problem to be addressed.
Multi-Label Data Challenge. All of the documents published in TED and many of the public sources around Europe publish documents with multiple classifications, with no primary classification identified. Many records have more than ten classification references and some buyers deploy a technique of adding diverse requirements to a single tender in different lots. This requires buyers to provide diverse categorisations for the same tender, making it impossible to identify a primary classification for a tender.
So a document with the following codes 72541000, 72510000, 72590000, 30210000 would be classified to 72000000 : IT Services: consulting, software development, Internet and support.
Inaccurate Classification Challenge. We’ve found that a high number of documents have been incorrectly classified. Mostly this is down to confusion around the correct class to apply, but occasionally there is a clear misapplication of codes, usually around services. We often see consulting described as training or research. This problem is exacerbated by the inconsistent nature of CPV codes which are poorly optimised for the description of services. ‘Consultancy’ is mentioned 6 times in the top two layers of the classification and in four different divisions of the vocabulary.
Poor Narrative Data. With some documents having just five words in their combined titles and descriptions, this is barely enough to classify the documents and makes the task of classifying data much harder. This issue is particularly prevalent in Spain where narratives are, on average, half as long as narratives in the UK.
Scale of Classification Challenge. Scale increases the complexity of the analysis we need to do significantly, with thousands of classification options open to the algorithm. The risk of inaccurate classification increases exponentially with the number of classes to be used.
Addressing These Challenges
With our classifier tool, we tackle each of these problems. Our tool is detailed, accurate, and can be deployed at scale. We can use this on private and public sector tenders, contracts and spend. This tool classifies to CPV but we can adapt this tool to other major classifications.
Using The Classification
→ In our api / source data
We can provide data with enhanced classification through our off-the-shelf API or through a custom API. This can be done for either all or a curated part of our data. For instance, if you are a haulage company, we can give you a customised feed of tender, contracts and spend data for logistics, infrastructure, construction.
→ We can classify your data, send and return
We can take your dataset, apply classification labels and send it back to you. Whether you want more details of corporate costs or a B2C analytics firm looking at better segmentation of consumer spend, we can help.
→ Building a custom classification for you
We can work with existing classifications such as CPV or Proclass or even make one for you, tailored to your needs. We can create a new classification that both fully reflects modern buying trends such as consultancy, and improves on existing, outdated taxonomies such as digital and software services.
if you’d like:
- classified procurement data;
- to have your own data classified;
- or to build your own classifier.