CRAN R build status Build status codecov DOI

Construct Consistent Time Series from Textual Data

RollingLDA is a rolling version of the Latent Dirichlet Allocation. By a sequential approach, it enables the construction of LDA-based time series of topics that are consistent with previous states of LDA models. After an initial modeling, updates can be computed efficiently, allowing for real-time monitoring and detection of events or structural breaks.


Please cite the package using the BibTeX entry, which is obtained by the call citation("rollinglda").

Please also have a look at this short overview on topic modeling in R: * Wiedemann, G. (2022). The World of Topic Modeling in R. M&K Medien & Kommunikationswissenschaft, 70(3), pp. 286-291.


This R package is licensed under the GPLv3. For bug reports (lack of documentation, misleading or wrong documentation, unexpected behaviour, …) and feature requests please use the issue tracker. Pull requests are welcome and will be included at the discretion of the author.



For the development version use devtools:


(Quick Start) Example

Load the package and the example dataset rom Wikinews consisting of 576 articles - tosca or quanteda can be used to manipulate text data to the format requested by rollinglda: The texts should be provided as a uniquely named list of tokenized texts, and the associated dates should be provided either as a named vector of dates or (at least) in the same order as the passed texts.


Then, the modeling is similar to the modeling of a standard latent Dirichlet allocation (LDA) by specifying the data texts and dates, the parameters K, alpha (default: 1/K), eta (default: 1/K) and num.iterations (default: 200), as well as the parameters chunks, memory, init and type relevant for the RollingLDA. By means of chunks the user determines at which interval steps the texts are to be modeled, starting from one day after init, the date specifying the end of the initialization period for which a standard LDA (type = "lda") or LDAPrototype (type = "ldaprototype") is modeled. In addition, memory specifies how much knowledge about the past model should be used for each interval (chunk).

In the case below, the 576 Wikinews texts are initially modeled up to July 3rd, 2008. Starting from that, the modeling is executed quarterly, namely with the start dates July 4th, 2008 and October 4th, 2008 (see getChunks). The texts published in the corresponding periods are modeled together, each with the last three quarters as memory, thus corresponding to October 4th, 2007 and January 4th, 2008, respectively. Note that the modeling is stochastic for both scenarios, using type = "lda" and using the default type = "ldaprototype" (see ldaPrototype package) as initial modeling step, i.e. the results will be fully reproducible only when using the same seeds.

roll_lda = RollingLDA(texts = economy_texts,
                      dates = economy_dates,
                      chunks = "quarter",
                      memory = "3 quarter",
                      init = "2008-07-03",
                      K = 10,
                      type = "lda",
                      seeds = 42)
# Fitting LDA as initial model.
# Exporting objects to package env on master for mode: local
# Fitting Chunk 1/2.
# Fitting Chunk 2/2.
# Compute topic matrix.

Using the function getChunks a lot of information about the modeling can be displayed. For some of these values further parameters of the method (see ?RollingLDA) are also relevant.

#     memory   n n.discarded n.memory n.vocab
# 1:        0 2007-01-01 2008-07-03       <NA> 470           2       NA    2691
# 2:        1 2008-07-05 2008-09-30 2007-10-04  50           0      204    2720
# 3:        2 2008-10-04 2008-12-29 2008-01-04  54           0      186    2814

It is noticeable that the of the first chunk is not 4th July, 2008. This is due to the fact that there are no texts for this day. The table shows the actual minimum and maximum dates per chunk. From n.vocab one can see how the vocabulary of the model increases due to the (frequent enough, see parameters vocab.abs, vocab.rel and vocab.fallback) use of new words within the observation intervals.

You can use getLDA to convert a RollingLDA object into a standard LDA object, which can be further processed using several functions from the ldaPrototype and tosca packages. You can also use getVocab to get the entire vocabulary of the model.

# RollingLDA Object named "rolling-lda" with elements
# "id", "lda", "docs", "dates", "vocab", "chunks", "param"
#  3 Chunks with Texts from 2007-01-01 to 2008-12-29
#  vocab.abs: 5, vocab.rel: 0, vocab.fallback: 100, doc.abs: 0
# LDA Object with element(s)
# "param", "assignments", "topics", "document_sums"
#  574 Texts with mean length of 120.68 Tokens
#  2814 different Words
#  K: 10, alpha: 0.1, eta: 0.1, num.iterations: 200

# LDA Object with element(s)
# "param", "assignments", "topics", "document_sums"
#  574 Texts with mean length of 120.68 Tokens
#  2814 different Words
#  K: 10, alpha: 0.1, eta: 0.1, num.iterations: 200

Finally, such an existing model roll_lda can be updated using the updateRollingLDA function. Note that the RollingLDA function can also be used for updating if the first argument in the function call is the RollingLDA object to be updated. Have a look at the help page ?updateRollingLDA for a minimal example of updating an existing model.