Branding Models With Textual Data and Deep Learning
By John V. Colias, Ph.D.
How would one develop a branding model using text data and deep learning methods?
Let’s define a branding model as a mathematical equation or set of equations that:
- Ingests not only customers’ structured data but also textual feedback.
- Explains one or more key performance indicators (KPIs) based on the inputs.
- Delivers insight into how experience with and perceptions of the brand lead to better performance on KPIs.
What steps would one take to build such a branding model based on both structured data and unstructured textual feedback from customers? What kind of modeling techniques could be used?
To answer these questions, let’s suppose that we have customer-experience survey data that includes both brand ratings and text responses to open-end questions that evince emotions and perceptions, likes and dislikes related to the brand and its products, and reactions to recent experiences with the brand, customer service, and/or products. Although other sources of textual feedback from customers are available (such as social media commentary and product reviews), let’s leave the discussion of these other data sources aside for now and focus on our two questions.
Several essential data engineering steps would be:
- Clean the text data, converting text to lower case, stripping out punctuation and stop-words, and stemming. Many software tools are available to easily do this.
- Label individual customers’ text. For example, “I liked that the technician had a lot of experience doing this kind of repair and could explain the work so that I could understand” could be labelled with the codes, “experienced technician” and “easy to understand.” Experienced coders should be employed to label all, or a random sample, of the text data.
- Define and select the KPIs based on discussion with those in the organization who are likely to use the analysis.
- This is particularly important with survey data .For example, calculate the frequency of each coded textual response. Answer codes with fewer than 20 or 30 observations may be used for back-end tables and charts with appropriate caveats. However, low-frequency responses should be excluded from statistical modeling due to the small sample size.
Once the data engineering steps have been completed, then several types of modeling should be explored. A useful open-source modeling software package is Keras (Chollet 2015), and the following modeling methods are recommended:
- Multi-label text classification producing a model to predict coded labels based on textual input. Such a model would be used much like a “scoring model” in segmentation research, except that the model assigns labels (rather than a segment) to each customer.
- Word embedding to convert textual data into a vector of continuous variables, summarizing the key points of meaning in the text. Word embedding with unstructured text data is the analog of factor analysis with structured data. If the word embedding layer is developed as part of the aforementioned multi-label text classification model, then the embedding model layer is predictive of the coded labels. Being predictive of the coded labels is extremely valuable, since the labels would have been developed by experienced human coders who understood the nuances within the textual data.
- Latent Class Discrete Factor modeling to explain multiple KPIs as a function of customer feedback variables (Vermunt and Magidson 2013). A latent discrete factor is an ordered variable with discrete categories, similar to the levels of survey “scaled-response questions” (https://www.decisionanalyst.com/library/sglossary/). The customer feedback variables would include (a) outputs of the word embedding layer of the multi-label text classification model and (b) brand attribute ratings from the survey. The final model combines the word embedding vectors and the brand ratings into discrete latent factors which explain the KPIs. A practical advantage of Latent Class Discrete Factor models is that they also output the discrete factors that can be profiled (based on coded labels and brand ratings) to draw actionable insights out of the modeling process.
While many branding models have been developed over the years, this model outlined offers several unique benefits:
- Incorporation of textual feedback from customers, which may contain unexpected or developing themes.
- Linkage to KPIs deemed important by the organization.
- Easy-to-interpret profiles of discrete factors that are automatically output by the model.
- Chollet, Francois and others 'Keras', accessed.
- Vermunt, Jeroen K. and Magidson, Jay (2013), 'Technical guide for Latent GOLD 5.0: Basic, advanced, and syntax', Belmont, MA: Statistical Innovations Inc.
About the Author
John Colias, Ph.D. (firstname.lastname@example.org) is a Senior VP Research & Development at Decision Analyst. He may be reached at 1-800-262-5974 or 1-817-640-6166.
Copyright © 2023 by Decision Analyst, Inc.
This posting may not be copied, published, or used in any way without written permission of Decision Analyst.
Scaled-Response Questions: A set of answer choices designed to measure the intensity or degree of a respondent’s answer. The answer choices of “excellent,” “good,” “fair” or “poor” would be a scaled-response example.
Scoring Model: One or more mathematical equations used to calculate a score based on answers to a survey. A simple scoring model is the well-known “top 2 box” score.