Predicting Innovative Firms Using Web Mining and Deep Learning

Abstract

Innovation is considered as a main driver of economic growth. Promoting the development of innovation through STI (science, technology and innovation) policies requires accurate indicators of innovation. Traditional indicators often lack coverage, granularity as well as timeliness and involve high data collection costs, especially when conducted at a large scale. In this paper, we propose a novel approach on how to create firm-level innovation indicators at the scale of millions of firms. We use traditional firm-level innovation indicators from the questionnaire-based Community Innovation Survey (CIS) survey to train an artificial neural network classification model on labelled (innovative/non-innovative) web texts of surveyed firms. Subsequently, we apply this classification model to the web texts of hundreds of thousands of firms in Germany to predict their innovation status. Our results show that this approach produces credible predictions and has the potential to be a valuable and highly cost-efficient addition to the existing set of innovation indicators, especially due to its coverage and regional granularity. The predicted firm-level probabilities can also directly be interpreted as a continuous measure of innovativeness, opening up additional advantages over traditional binary innovation indicators.

Web-Based Innovation Indicators – Which Firm Website Characteristics Relate to Firm-Level Innovation Activity?

“…Our results for product innovators are in line with the results of Kinne and Lenz (2019). Their statistical model has reached a similar accuracy for product innovators only observed in one MIP wave.…”

“…Following the idea of web-based innovation indicators, Kinne & Lenz (2019) attempt to predict innovation at the firm level using textual information on websites and novel machine learning tools. They use traditional firm-level innovation indicators from the MIP 2017 to train an artificial neural network classification model on labelled (innovative/non-innovative) web texts.…”

“…To calculate the termdocument matrix we need to do some preprocessing steps in advance. First, we merge all scraped subpages related to a single firm and we delete irrelevant subpages by using the gold standard approach (e.g., Kinne & Lenz 2019). The model distinguishes between subpages including relevant information like product descriptions and subpages with irrelevant information like the imprint, information about cookies or texts that are prescribed by law.…”

Exploring the Adoption of the International Information Security Management System Standard ISO/IEC 27001: A Web Mining-Based Analysis

“…Web mining describes the application of data mining techniques to uncover relevant data characteristics and relationships (e.g., data patterns, trends, and correlations) from previously web scraped unstructured web data [19]. We do so by using data from the Mannheim Enterprise Panel (MUP) as the firm database, and then categorize web scraped firms using their website texts and conduct multivariate analyses based on firm characteristics and a deep-learning-based product innovator probability indicator [20].…”

Model degradation in web derived text-based models

“…The initial logistic regression model created had an accuracy of 93% and an F1-score of 93% on the test set (Daas & van der Doef, 2020). Kinne and Lenz (2019) demonstrated that the text on the websites of German companies could also be used to detect innovative companies. Their Deep Learning based model had an F1-score of 80%.…”