IFST PFSG Discussion Forum Summary: Cluster Analysis in Sensory and Consumer Research

Guest Speaker:  Carol Raithatha, Carol Raithatha Ltd

The PFSG held its 4th Discussion Forum on October 13th 2010 at the University of Nottingham, where Carol Raithatha (secretary of the PFSG) kindly agreed to lead a discussion on the use of cluster analysis (CA) in sensory and consumer research.  The session was developed by Carol with input from a few PFSG members with experience in cluster analysis.

The forum was very well attended, with around 20 PFSG members and invitees coming from a range of academic and industrial organisations.  The common theme was that everyone present wanted to learn more about cluster analysis!

The session began with an introductory presentation on the use of cluster analysis in consumer and sensory research from a non statistician point of view.  The presentation lasted around 30 minutes and in brief covered:

Why this discussion forum?

Carol explained how she had recently been involved in a sensory profiling project where cluster analysis was used to group products, and how she felt the use of the technique had added value in terms of crystallising themes and communicating results.  This led her to consider what else is possible with cluster analysis and to want to learn more about its variations and limitations.

What is cluster analysis?

Cluster analysis is a multivariate statistical technique used in a range of disciplines.  In essence it takes a group of objects and creates subsets of objects that are similar.

The most common type of cluster analysis used in sensory research appears to be agglomerative hierarchical cluster analysis, although the use of ‘k-means’ or non-hierarchical cluster analysis is also found in sensory literature. 

Applications in sensory and consumer testing

Carol explained that there are applications for cluster analysis in both consumer research and objective sensory assessment. 

The most common use of cluster analysis in sensory consumer research is consumer segmentation.  Consumers can be clustered by taste preferences and these clusters can be used in preference mapping.

Cluster analysis can be used in sensory profiling to look at relationships between attributes, product groupings, and also potentially panellist groupings.

Examples and opportunities

Sensory Dimensions provided some data generated when consumers used a 9 point hedonic scale to assess a group of products.  Performing hierarchical cluster analysis revealed several groups of consumers whose preference patterns were different than that suggested by overall product means.  The consumer clusters also highlighted the fact that one product was polarising – it was the most liked by one cluster, while being the least liked by another.  This example illustrated the insight that can be gain by applying cluster analysis to consumer sensory data.

A new approach to cluster analysis for consumer/sensory applications was also discussed – latent and fuzzy techniques which are based on probabilistic models and therefore are claimed to result in more robust clusters.  In addition, a recent paper has been published in which fuzzy clustering has been used when different consumer groups test different sets of products (Johansen, S.B, Hersleth, M., & Næs,T. (2010). A new approach to product set selection and segmentation in preference mapping. Food Quality and Preference, 21,188-196).

Key Questions

Some key unresolved questions (posed in preparing the presentation by Hal MacFie) were briefly discussed.  These are:

  • How do we know if the clusters formed are meaningful?
  • Should we use non-discriminating attributes in forming assessor or attribute clusters?

Group Discussion

Following the presentation there was a quick chance for initial questions and then the group broke up into four subsets to discuss issues in detail and share experiences of using cluster analysis in practice.  After around 45 minutes, the whole group reconvened to feedback on the main/interesting points covered in the sub-groups.  Summary comments for some of the discussion points included:

  • Have you used cluster analysis –if so when and how?
    • A range of experience in those present, most of it in terms of clustering consumer data, some in terms of clustering products for PCA representations
    • Performing the cluster analysis was the role of statisticians within some of the participants’ organisations – a comment was that this could lead to a lack of control/losing touch with the data and end results
  • If you haven’t used cluster analysis–why not? What are the issues/problems? 
    • Some scepticism over method – too vague and dependent on method used
    • Opportunities to use the method do not arise for everyone
  • How can cluster analysis be used with objective sensory profiling? Experiences or ideas?
    • Very little experience of using cluster analysis in this way but some enthusiasm
    • General interest in using the technique for attribute and product segmentation
    • One group discussed clustering data for each sample separately so that assessor x sample interactions could be seen more easily
    • Doubt over its use in terms of panel performance – other established methods are better
  • How can cluster analysis help within consumer segmentation? Experiences or ideas?
    • Good exploratory method
    • A good application is highlighting important segments that may be lost in an average
    • Providing insight into preference patterns
    • Helping to add focus to advertising and promotional projects when used along with demographic and usage and attitude data
  • Ease of access to this method? Stats packages you have available? Experience with these? Tips?
    • XLstat good because of flexibility with outputs
    • Software in general can provide a lot of choices, so it is necessary to have some understanding of the statistics
  • How can cluster analysis be made to be “user friendly? for clients/customers? How is it best to communicate results of cluster analysis?
    • Depends on intended audience and their experience of multivariate methods,etc.
    • Use of colour and simplified images suggested
    • Dendograms too complicated for inexperienced
    • Interpretation should be done jointly with statisticians and individuals who know the products/market
    • Face to face debriefs essential
  • Statistical issues: similarity vs. dissimilarity matrices –why? How –which methods are used? Clustering algorithm –hierarchical vs. non-hierarchical? Why? Which aggregation method used in hierarchical? Why? Where to truncate? How many clusters?
    • Standardisation of data first if measured on different scales
    • Most experience was with hierarchical approach
    • Most commonly used methods as per the literature and which give ‘sensible’ results were Euclidean or Pearson methods for establishing dissimilarity/similarity matrices followed by Average Linkage (unweighted mean paired group average)
    • Also usage of Ward’s clustering method once outliers were removed
    • General agreement on difficulty/confusion knowing which methods to use – largely using what has been advised by statisticians
    • A general ‘rule’ discussed in one group was using a minimum of n=100 consumers before applying cluster analysis and ideally n=150
    • Point raised that need at least 30 consumers in a cluster and this has implications for the number of people to be recruited in consumer studies
  • How does cluster analysis compare to other methods for partitioning sample sets –e.g. discriminant analysis?
    • The perceived vagueness of Cluster Analysis seems to be a key deterrent to its use for some
  • What novel ways have you seen of using cluster analysis within the sensory and consumer research field?
    • University of Nottingham have used cluster analysis to group people according to their eating behaviour for chocolate and confectionary (successfully)
  • Where do you think cluster analysis can add the most value? Why?
    • Consumer segmentation: You cannot please all the people all of the time so it’s a great tool to understand differences in behaviour
    • Product segmentation: Helping to identify gaps in market
    • Providing a more structured way to identify groups on PCA plots
    • Identifying different preference clusters within a target market and then using along with Preference Mapping to understand the key drivers of liking for each cluster
    • A valuable aid for product development and marketing

To round this report off, here are some thoughts on the topic from Carol based on preparation for the above presentation and reflecting on points raised within the discussion. . . .

  • Cluster analysis is a good exploratory technique to look for patterns in data.  It can be used to establish hypotheses which can be further tested by probabilistic methods.
  • Linking consumer taste preference clusters with relevant consumer demographics could be an effective mode to establish unique target segments.
  • Communicating output of cluster analysis to end clients needs careful consideration in terms of explaining methodology, visualising results, and discussing implications of findings.
  • Cluster analysis can be a useful tool for understanding/developing robust sensory profiling vocabularies and understanding patterns/groupings of samples within large sensory profiling projects.
  • The challenge in using cluster analysis is wading through the multitude of statistical options.  Establishment of clear guidelines/guidance from sensometricians would be valuable to the sensory/consumer research community.

The next discussion forum will consider the use of Diagnostic Scales and will be held at Sensory Dimensions Nottingham 11.00 -1.00pm February 16th 2011, prior to the PFSG Committee meeting (Contact tracey@sensorydimensions.com).  PFSG members are welcome to stay for the committee meeting.  PFSG is always looking for new committee members so why not come along, meet the committee and see what we do. Suggestions from members on potential topics for future discussion forums would be welcomed by the committee.  Although these events are free to PFSG members, numbers are limited so please book your place soon.

Back to top