Clustering Social Media Influencers by

Pro-environmental Instagram Users

krystal7 — 2020-05-26 21:57:55 UTC

Pro-environmental behavior refers to “behavior that harms the environment as little as possible, or even benefits the environment” (Ertz, Karakas & Sarigöllü, 2016).

Hierarchical Cluster Analysis & Hierarchical Cluster Analysis on Principal Components

krystal7 — 2020-05-31 00:56:13 UTC

Two steps in cluster analysis are:

Measurement of distances between all pairs of objects and
Forming clusters based on distances.

Euclidean distance was used, and clusters were formed based on the "average" method.

The main aim of PCA is:

To replace p metrical correlated variables by a smaller number of uncorrelated variables which contain most of the information in the original data set (the 11 attributes).
Based on the concept of the proportion of total variance that is accounted for by each of the new variables.

Sample Selection

krystal7 — 2020-05-31 00:59:44 UTC

Random sample of N=213 Instagram users selected by searching biographies that included pro-environmental language: "conservation", "sustainability", "sustainable" and "nature".

Data Collection: Apify

krystal7 — 2020-05-31 01:03:51 UTC

Data was collected through online data-scraping tool, Apify. Search parameters included a randomized sample of Instagram users that included the words: "conservation", "sustainability", "sustainable" or "nature" in their biography, suggesting pro-environmental advocacy. The user's follower count ranged from 50 to 5,000,000. (https://apify.com/)

Research Question

krystal7 — 2020-05-31 01:15:41 UTC

Can you identify specific groups of pro-environmental Instagram users based on 11 attributes through a hierarchical cluster analysis?

Research Objectives

krystal7 — 2020-05-31 01:15:46 UTC

Perform a hierarchical cluster analysis on the selected pro-environmental Instagram users.
Perform a hierarchical cluster analysis on principal components: simplify the data structure by performing a Principal Components Analysis and reducing the 11 highly correlated variables into a few independent principal components (without losing as much information as possible).
Interpret the results.

Variables

krystal7 — 2020-05-31 02:14:33 UTC

Attributes measuring social media activity and engagement:
x₁ = Followers count (# of users that follow individual(s) under study)
x₂ = Follows count (# of users the individual under study follows)
x₃ = Post counts (total # of posts)
x₄ = Biography word count (# of words in users' biography)
x₅ = Highlight reel count (# of highlights featured on user's profile)
x₆ = IGTV video count (# of IGTV videos on users' account)
x₇ = Average caption word count (most recent 12 posts*)
x₈ = Average hashtag count (most recent 12 posts)
x₉ = Frequency of posts (average time between posts, most recent 12 posts)
x₁₀ = Average comment count (most recent 12 posts)
x₁₁= Average post likes count (most recent 12 posts)

*Only most recent 12 posts were accessible using Apify.

krystal7 — 2020-05-31 02:25:00 UTC

Bartholomew, D., Steele, F., Moustaki, I. & Galbraith, J., Analysis of Multivariate Social Science Data, Boca Raton: Chapman & Hall, 2008.

Ertz, M., Karakas, F., & Sarigöllü, E. (2016). Exploring pro-environmental behaviors of consumers: An analysis of contextual factors, attitude, and behaviors. Journal of Business Research, 69, 3971-3980.

Hussein, R. & Abeer, M. (2016). Users engagement on Facebook: a cluster analysis. International Journal of Business and Emerging Markets, 8(4), 426-445.

What is an Influencer? - Social Media Influencers Defined. (2020, May 26). Retrieved May 30, 2020, from https://influencermarketinghub.com/what-is-an-influencer/

Types of Influencers

krystal7 — 2020-05-31 02:42:20 UTC

By Follower Numbers

Mega-influencer: > 1 million followers
Macro-influencer: 40,000-1 million followers
Micro-influencer: 1,000-40,000 followers
Nano-influencer: < 1,000 followers

By Level of Influence

Celebrities
Key opinion leaders (i.e. journalists, academics, industry experts, professional advisors)

Instagram Influencers by Business Category

krystal7 — 2020-05-31 02:54:45 UTC

In this sample, top 3 categories are: Non-profits and religious organizations (27%), Other (20%), Personal Goods and General Merchandise (19%). Creators and Celebrities comprised 8% of the sample population.

Hierarchical Cluster Analysis

krystal7 — 2020-05-31 03:00:18 UTC

Visually, it appears 4 clusters may be sufficient to identify specific groups of Instagram users based on the similarities of 11 attributes. Although, I think this is somewhat inconclusive/can be argued against. To simplify data structure, I decide to test if performing a cluster analysis on principal components assists in forming more clear clusters in the results.

Hierarchical Cluster Analysis on 5 Principal Components

krystal7 — 2020-05-31 03:00:26 UTC

Clustering on five principal components that account for almost 70% of total proportion of variance reveal 2 distinct clusters and 1 outlier cluster.

Visualization of Correlation Matrix

krystal7 — 2020-05-31 03:13:28 UTC

The larger the dot on the graph, the closer the correlation of the pair of objects is to 1. 1 along diagonal in correlation matrix.

Distance Matrix

krystal7 — 2020-05-31 03:17:43 UTC

Distances were calculated using Euclidean distance formula.

Scree Plot, PCA

krystal7 — 2020-05-31 03:31:23 UTC

Clustering on principal components

krystal7 — 2020-05-31 03:33:38 UTC

PCA: Importance of Components

krystal7 — 2020-05-31 03:43:58 UTC

Component | Cumulative Prop.
1 - 20.1% (λ>1)
2 - 37% (λ>1)
3 - 48.2% (λ>1)
4 - 58.7%* (λ>1)
5 - 68.8%** (λ=1)
6 - 77.7%
7 - 86%
8 - 93%
9 - 97.7%
10 - 100%

Recall: Euclidean Distance

krystal7 — 2020-05-31 03:57:02 UTC

The Euclidean distance between objects i and j is:

PCA Loadings

krystal7 — 2020-05-31 21:28:16 UTC

Component loadings can be interpreted as the correlation coefficient between variable i and component j.

Recall: Average linkage

krystal7 — 2020-05-31 21:41:37 UTC

Suppose P & Q are groups of objects and d(x,y) is the distance measure between objects x and y:

d(P, Q) = ave d(x, y), for x ∈ P, y ∈ Q

{FactomineR} package

krystal7 — 2020-05-31 22:19:16 UTC

To compute the hierarchical clustering I used the FactoMineR package in R. PCA was performed on a correlation matrix and the number of principal components was selected based on eigenvalues greater than 1 and total proportion of variance explained. Then, the function was used to compute clustering on 5 principal components. Hierarchical clustering was performed using the Ward’s criterion on the selected principal components.

3D Representation of clusters on 2 dimensions

krystal7 — 2020-05-31 22:23:56 UTC

Findings

krystal7 — 2020-05-31 22:58:20 UTC

Hierarchical clustering of 213 pro-environmental Instagram users based on 11 attributes measuring activity and engagement revealed 4 possible clusters after a visual inspection of the resulting dendrogram. The results were inconclusive.

After performing a principal component analysis, I found that 5 components explained the total proportion of variance. Then, using the FactoMineR package in R, I was able to cluster based on the principal components. The dendrogram suggested a possible 3 cluster solution.

I found that while it is possible to identify groups of similar social media users using cluster analysis (and a cluster analysis of principal components) based on 11 attributes of activity and engagement, this exploratory analysis is only a starting point for further analysis.

Recall: Ward's Method

krystal7 — 2020-06-01 04:13:33 UTC

Ward's method "seems to yield the clearest picture of any clustering which is present".
Ward's method considers all pairs of clusters and asks how much information would be lost is the pair was combined.
The pair with the least information lost is selected.
Information is measured by a sum of squares.

(Bartholomew et al., 2008)

Cluster 1

krystal7 — 2020-06-01 04:19:23 UTC

These users are most similar in highlight reel count, biography word count and average frequency of posts.

krystal7 — 2020-06-01 04:42:39 UTC

krystal7 — 2020-06-01 04:43:25 UTC

krystal7 — 2020-06-01 04:43:40 UTC

krystal7 — 2020-06-01 04:43:52 UTC

krystal7 — 2020-06-01 04:44:09 UTC

krystal7 — 2020-06-01 04:44:35 UTC

Cluster 2

krystal7 — 2020-06-01 04:45:06 UTC

These users were most similar in IGTV count, biography word count & average caption word count.

Cluster 3

krystal7 — 2020-06-01 04:45:35 UTC

Outlier cluster with one user.

krystal7 — 2020-06-01 04:47:10 UTC

krystal7 — 2020-06-01 04:47:31 UTC

krystal7 — 2020-06-01 04:47:53 UTC

krystal7 — 2020-06-01 04:48:17 UTC

krystal7 — 2020-06-01 04:48:32 UTC

krystal7 — 2020-06-02 00:35:20 UTC

I would say that while it is possible to identify groups of similar social media users using cluster analysis (and a cluster analysis of principal components) based on 11 attributes of activity and engagement, this exploratory analysis is only a starting point for further analysis.

This could be because:

The variables did not provide enough variation between groups (multicollinearity)
It is difficult to cluster non-spherical, overlapping data

To create a more comprehensive cluster analysis you could include variables such as:

Attitudes and motivations towards posting content and sharing information (i.e. for entertainment, for work, etc.)
Internet lifestyles (i.e. “The internet makes an important contribution to almost all parts of my life.” or “I use the internet to make me and those I love achieve their objectives”.

Slide deck

krystal7 — 2020-06-02 03:23:32 UTC

Presentation link

krystal7 — 2020-06-02 19:04:55 UTC