Identifying Thematic Clusters using Latent Dirichlet Allocation

Explorative Survey in Energy Domain considering 5 premier journals over a period of 25 years

By Arun Abraham Thomas


Introduction


Identifying thematic clusters helps in reviewing literature in a given area of research. Typically in traditional manual analysis, a researcher starts by defining topics of interest through identifying a list of keywords. These keywords are used in online search engines for potential relevant literature for the topic of interest. This traditional method of manual review has a few shortcomings like choosing non-precise keywords, missing journals, and errors due to the large volume of literature that a reseacher has to skim through (Delen & Crossland, 2008) .


Thus, if such thematic clusters are developed and made available for multiple journals, it will aid and support in the traditional manual analysis during the exploratory phase in research. It helps in providing an overview of the research field and is particularly useful for novice academics and reseachers.
Clustering of large number of documents is a non-trivial problem. Since any textual documents can contain words from a vocabulary, representation of textual documents is high- dimensional. High dimensionality of NLP problems leads to difficulty in clustering and failure of “distance-metric” based algorithms.


The very basic step of NLP clustering algorithms is to reduce the dimensionality of the problem, using bag of words, removing stop words etc. Even after such methods, the document space remains very high dimensional. David M. Blei, Andrew Y. Ng and Michael I. Jordan, 2003 in their seminal paper, introduced the Latent Dirichlet Algorithm (LDA), which thereafter has been used to provide interpretable lower dimensional representation of documents.


Topic Modelling


Topic Modelling has been used for exploratory analysis of large number of papers (Jelodar, et al., 2019). It has gained much prominence in public policy, political science and rhetoric analysis (J & BM, 2013) , (Debnath & Bardhan, 2020) , finance (Feuerriegel, et al., 2016) , (Shirota, et al., 2014) , biomedical research (HJ, et al., 2019) . Topic based clustering model was used to group Indian legal documents into various clusters (Kumar & Raghuveer, 2012) . In the Energy domain, Topic Modelling has been used for tracking the evolution of the policy of New Energy Vehicles (NEVs) in China (Jia & Wu, 2018) , sustainability reporting (Székely & Brocke, 2017)


Latent Dirichlet Algorithm (LDA)


Latent Dirichlet Algorithm (LDA) is a probabilistic generative algorithm that extracts the thematic structure in a large corpus. The model considers that a topic is a distribution of words in a vocabulary space and every document (described over the same vocabulary) is a distribution of a small subset of those topics. These latent topics that the models learn are highly interpretable and provide insights and qualitative understanding of the text corpora. For example, in our case, we would expect the model to learn “renewable” as one of the topics and cluster all papers related to renewable energy under the topic “renewable”. Once LDA generates a sparse representation in much lower dimensions (1000 or lower), it is easily amenable for standard clustering algorithms. We have used the pyLDAvis package for interpreting the topics/clusters and make interactive web-based visualizations of the journals (Sievert & Shirley, 2014) .


LDA calculates the joint probability distribution between the observed words in a paper and the unobserved (the hidden structure). This method evaluates the frequency of words and the semantics are ignored (Asmussen & Møller , 2019) . LDA is an unsupervised algorithm and a key feature of such unsupervised machine learning methods is hyperparameters like the number of topics/clusters. Generally, cross validation using perplexity score is used to find the optimal number of topics/clusters. A larger number of topics/clusters provide a more detailed clustering while a low number of topics/clusters provide a general overview. Depending on the research questions trying to be answered, the number of topics can differ.


The Results


It is interesting to note that the methodology and algorithms described above automatically identifies and clusters the broad areas of research like “Electricity Market”. “Climate Change”, “Renewables” and groups various papers into the identified thematic groups . The broad themes remain more or less the same for the three niche, focused journals, i.e. Energy Economics, Energy Policy and Resource & Energy Economics. For the journals with broader scope, i.e. Applied Energy and Energy, technical thematic clusters like “Power Systems”, “Thermal Storage”, “Fuel Cells” are automatically identified in addition to other themes like “Renewables”, “Climate Change”. Thus, the algorithm and methodology

# An interactive demo is explored here!!!


Share: Facebook LinkedIn