Topic modeling in marketing: recent advances and research opportunities

LDA, as introduced by Blei and Jordan (2003a) and Blei et al. (2003) is a method originally developed for soft-clustering large quantities of discrete textual data, in order to find latent structures (Blei 2012, pp. 77; p. 80). The model assumes that textual documents consist of topics, which in turn consist of words from a vocabulary (all words in the documents) (Airoldi et al. 2015, p. 4; Blei 2012, pp. 78; Galyardt 2015, p. 39). The aim is to automatically discover the hidden (not observable) topics in a collection of documents, and how each document exhibits them (Blei and Lafferty 2009, p. 73; Blei 2012, p. 79). The topics are “a recurring pattern of co-occurring words” (Brett 2012), or, more formally speaking, “a distribution over a fixed vocabulary” (Blei 2012, p. 78). To unveil the topic structure in addition to that, the hidden per-document topic distribution needs to be rendered (Blei 2012, pp. 79). For example, “a news article may address multiple topics rather than fitting neatly a single category” (Galyardt 2015, p. 39) and “each document exhibits the topics in different proportion” (Blei 2012, p. 78). Thus, LDA belongs to the family of mixed membership models (Galyardt 2015, p. 43). Contrary to classical clustering methods, where membership is a binary variable, every unit (word) partially belongs to all clusters (topics) with different probabilities, expressed by a (membership) “vector of continuous non-negative latent variables that add up to 1” (Airoldi et al. 2015, p. 4). Similarly, every topic partially belongs to all documents with varying probabilities. Due to that, the basic modeling output in LDA are groups of words (i.e., topics) with their membership probabilities and the proportions of these topics in each document (Blei and Lafferty 2009, p. 74; e.g., Park and Ha 2016, p. 1495). Normally, researchers constrain the output to a low number of the highest probable words and topics (up to 20) and assign a topic label. Figure 1 illustrates this, using an online posting as an example with the highest probable words within four retrieved topics and the assigned topic labels.

To achieve that, LDA computes the posterior distribution (conditional distribution) of the hidden variables performing data analysis on the joint probability distribution over hidden and observed variables. Figure 2 indicates the observable variable (all words) W_d,n within the collection of words in documents N and the collection of documents D. K is the number of all topics, which is usually set by the researcher. There are numerous hidden variables and dependencies. For example, a specific word w_d,n depends on the hidden “topic assignment for the nth word in document d” (Blei 2012, p. 80) (z_d,n), and on all topics β_1:K (where each β_k is a multinomial distribution over all words for a topic k). z_d,n depends on θ_d, which is the (multinomial) per-document topic distribution (Blei 2012, pp. 79). Additionally, LDA uses a Dirichlet prior on the per topic word distribution η and a Dirichlet prior on the per-document topic distribution α, which in practice are mostly set up to be symmetric (Blei and Lafferty 2009, p. 74; Galyardt 2015, p. 40; Wallach et al. 2009, pp. 1).

More formally, Blei (2012, p. 80) defines these dependencies in the generative process for LDA, which depicts the joint distribution of hidden and observable variables in the model:The computational problem is to calculate the conditional (posterior) distribution of the topic structure (Blei 2012, p. 81), where the numerator is the joint distribution of all random variables:It is intractable to directly calculate the posterior due to the denominator, which is the marginal probability of observations. In theory, it would be calculated by summing “the joint distribution over every possible instantiation of the hidden topic structure” (Blei 2012, p. 81), which is exponentially large (Blei 2012, p. 81). Therefore, one needs to infer or approximate it. For inference Blei et al. (2003, pp. 1003; pp. 1019) originally used a convexity-based variational approximation with EM (expectation maximization), where the posterior is approximated by optimizing a distribution to be close to it (Blei and McAuliffe 2010, p. 3). Since then, MCMC (Markov chain Monte Carlo) algorithms (like Gibbs) have become prevalent (e.g., Wallach et al. 2009, p. 2), where samples from the posterior are collected, to approximate it with an empirical distribution (Blei 2012, p. 81). However, both procedures need initialization. A common procedure is to randomly initialize inference (Boyd-Graber et al. 2015, pp. 231), triggering an iterative process. Each iteration provides an update for the algorithm, working slowly towards a local optimum, estimating what the probability of recreating the data is, given the current values of the latent variables (Boyd-Graber et al. 2015, p. 232; Underwood 2012).

The just described basic LDA is clearly related to various clustering and other exploratory data compression techniques. Of particular relevance is model-based clustering, using finite mixture models (FMM) (e.g., McLachlan and Peel 2000, pp. 1–39; Titterington et al. 1985, p. 8), the Products of Experts (PoE) model as introduced by Hinton (2002, p. 1), and exploratory factor analytic (EFA) procedures (e.g., Muthen 1978, p. 407).¹ Below, we briefly focus on similarities and differences between these methods and LDA.

LDA is defined as a generative mixed membership model (Galyardt 2015, pp. 40), and thus represents a generalization (Sammut and Webb 2011, p. 689), or a special application (Sun et al. 2012, p. 260) of (generative) finite mixture models. FMMs assume that L observed subjects (with observations contained in a vector y) of an overall (empirical) distribution stem from several latent groups (sub-distributions) (C₁,…,C_S) in proportions π₁,…,π_S, each encompassing the same distributional family, but with varying values (Frühwirth-Schnatter 2006, p. 5; Wedel and Kamakura 1999, p. 76; Wedel 2002, p. 364). Let v be a latent indicator vector for the multinomial group memberships of subject l and ф denote the parameters of the probability density function f_S(y|ф_S). When Ф = {π, ф} denotes the parameters in the model, the unconditional (overall) probability distribution of y is:Once the parameters are obtained via maximizing the corresponding likelihood function (e.g., using an expectation maximization procedure; Wedel and Kamakura 1999, p. 80), the posterior probability of a subject l being member of group C_S can be derived using the Bayes rule (Wedel 2002, p. 365):Thus, due to unobserved heterogeneity in the data, FMMs assign a probability to the cluster assignment and thus belong to the family of (probabilistic) fuzzy clustering methods (Clement and Boßow-This 2007, pp. 169). The above described LDA is related to FMMs, such that technically speaking, in LDA, every document is modeled with an FMM (Teh and Jordan 2010, p. 170), with the mixture proportions being drawn uniquely for each document, but the mixture components (i.e., the topics) are shared across the text corpus (Blei et al. 2010, pp. 2). That is, the mixture proportions in FMMs, π, are equivalent to the topic proportions in a specific document, θ_d, in LDA, and the mixture components in FMMs are equivalent to the topics, β_1:K, which are shared across the corpus, in LDA. Specifically, the generative process of words in a specific document in LDA requires multiple draws (of topics and of words) from this finite mixture. When considering a different document, there is again a finite mixture, “with the same mixture components (the topics), but with a different set of mixing proportions (the document-specific vector θ_d)” (Teh and Jordan 2010, p. 170). Broadly speaking, there are further similarities between both methods. For example, the determination of a suitable number of topics / mixture components to obtain reasonable results (Melnykov and Maitra 2010, p. 88; Teh and Jordan 2010, p. 260). Also, similar to LDA, FMMs are quite flexible in certain aspects, like in choosing the family of distributions under consideration (Costa Filho 2010, p. 9).

Another important related technique to LDA is the PoE model to learn sharp distributions in high dimensional spaces by multiplying and renormalizing sub-distributions (called experts). Because of its capability to detect specific subsets of dimensions in a high dimensional space and to better conjointly describe them, PoE can improve the performance of FMMs for certain applications (Hinton 2002, pp. 1; Salakhutdinov and Hinton 2009). The overall distribution, P(x), represented by a PoE is as follows:with f_i(x|σ_i) denoting un-normalized experts and Q the overall normalization constant (Welling et al. 2002, p. 2). Translated into the conceptual framework of LDA, a “product of topics” could give a higher probability to a cross-cutting specific word, where the word is not as strongly predicted by each individual topic, thus better accounting for interrelations between topics (Salakhutdinov and Hinton 2009, p. 1). On the other hand, due to multiplying every expert, there is no explicit mixture proportion (i.e., π in FMM, and θ_d in LDA), but the experts themselves are the weights, which does not make sense for the LDA framework when directly transferring this into it, due to the need of a per-document topic distribution, θ_d. However, there are attempts to integrate PoEs into topic modeling. For example, Gormley et al. (2012, pp. 783), conceptualize each single topic as P(x), consisting of experts, thus adding an additional layer to standard LDA, to e.g., avoid overparameterization, and to account for topic relations.

Finally, LDA is also related to a third family of methods, namely, exploratory or model-based factor analysis for binary variables used for dimensional reduction. Similar to LDA, factor analytic models aim at compressing high-dimensional data sets into a smaller set of (latent) common factors (equivalent to the topics in LDA), while conserving as much of the original information as possible (Bartholomew et al. 2011, p. 209; Fabrigar and Wegener 2012, p. 10). There are many methodological variants for finding latent groups for binary data (e.g., Hruschka 2016, p. 3; Muthen 1978, pp. 551; Muthen and Christoffersson 1981, pp. 407). While similar by idea, there are also some notable differences to LDA, regarding the conceptualization, structure, and output. For example, in LDA, the co-occurences of words in probability distributions that are part of the method’s structure, form the topics (i.e., the factors), not correlations between variables.

Due to its open framework, LDA is highly extendable and can easily be applied to various kinds of data (Airoldi et al. 2015, p. 4; Blei 2012, p. 83). The prerequisite is a large set of documents, each consisting of discrete units, which are distributed unevenly. Whatever the documents and units specifically might be, plays a minor role from that perspective. For example, mixed membership models in general have been applied to texts (e.g., Wang and Blei 2011, pp. 450), surveys (Gross and Manrique-Vallier 2015, pp. 119), political voting data (Gormley and Murphy 2015, p. 441), population genetics (Shringarpure and Xing 2015, p. 397), image analyses (Cao et al. 2014, pp. 8959), image and text analyses (Blei and Jordan 2003a, p. 128), and more (e.g., Blei and Lafferty 2009, p. 71). Data in marketing relevant research exemplarily encompass purchase histories and consumer data (Hruschka 2016, p. 7; Ishingaki et al. 2015, p. 17; Jacobs et al. 2016, p. 397), the internet browser’s cookies (Trusov et al. 2016, pp. 409), and mobile apps usage data (Do and Gatica-Perez 2010, p. 3).

Additionally, in LDA, even basic assumptions like the so-called “bag-of-words” property can be relaxed. The latter assumption characterizes the fact that the basic LDA ignores the order of words in documents and the order of documents in a text corpus (Blei 2012, pp. 82). Also, basic statistical assumptions like the assumed distributions were altered.² This flexibility leads to a substantial amount of extensions of LDA (e.g., Airoldi et al. 2015, p. 4; Balasubramanyan and Cohen 2015, p. 256; Blei and Jordan 2003a, pp. 2). In fact, the model has served as a basis for the advent of numerous other types of topic models (Blei and Lafferty 2009, p. 72). These include the Correlated Topic Model (CTM) for discovering correlations between topics (Blei and Lafferty 2007, p. 17; Blei and Lafferty 2009, pp. 82), Dynamic LDA for modeling topics as changing over time (Blei and Lafferty 2009, p. 82), the Supervised Topic Model (sLDA), where an additional response variable (e.g., the rating of a text) is integrated for better fitting on the data (Blei and McAuliffe 2010, pp. 2), Online LDA, which reduces the computational time needed in LDA for massive document streams (Hoffman et al. 2010, p. 2), the Author Topic Model (ATM), which associates each author with a topic probability (Rosen-Zvi et al. 2004, p. 487), and the Author-Recipient Topic Model (ART), which further extends this idea by “building a topic distribution for every author recipient pair” (Balasubramanyan and Cohen 2015, p. 260). For instance, other models also include hierarchies between documents like the Hierarchically Supervised Latent Dirichlet Allocation (HSLDA) (Wood and Perotte 2015, pp. 311). Various network topic models mainly differ in the view of what a network is, what a link consists of, if the links are conceptualized as within or between documents, and if additional factors (e.g., time dependency) are included (e.g., Airoldi et al. 2015, p. 7). These form a substantial amount of the latest attempts in the field to analyze complex, multi layered real world data. Examples are so-called Relational Topic Models (RTMs) (Chang and Blei 2009, p. 81), and Block-LDA (Balasubramanyan and Cohen 2015, pp. 255). However, only a small fraction of these more recently introduced models are currently in use (see for example: Table 6).

Stemming from the variability of the approach, in conjunction with a diverse range of data and research interests, there are numerous procedures and metrics used to evaluate the models. These are applied at various stages of the modeling process and include manual, semi-automated, and fully automated methods (Roberts et al. 2015, pp. 12). Likewise, evaluating topic models could include the computational performance (e.g., Jacobs et al. 2016, pp. 394), indicators for optimal parameter settings, model fit (in sample and predicitive out of sample), and the assessment of the clustering output (topics). In general, a widely used practice is to run the same model with a different number of topics, and by varying other model parameters (e.g., the prior distributions), or to render different topic models on the same data for comparison (e.g., Hruschka 2014, p. 270; Roberts et al. 2015, p. 18; Wang et al. 2015, p. 4). This may enable the researcher to find a good model, feasible parameter settings, reasonable topics, and to assess the stability of the output and of the covariate effects (Roberts et al. 2015, pp. 14; pp. 22; p. 24). When doing that, a quite common procedure is to split the data into a training set and a hold-out validation set. This enables scholars to examine how predictive models behave on unseen data, e.g., utilizing hold-out-likelihood (e.g., Blei and McAuliffe 2010, p. 8). However, a substantial amount of research has gone into the assessment of the clustering output (topics). One way to examine topics is to evaluate semantic coherence, which is a summary measure to capture “the tendency of a topic’s high probability words to co-occur in the same document” (Mimno et al. 2011 quoted from Roberts et al. 2015, pp. 13; see also Newman et al. 2010, p. 100). Newman et al. (2009, pp. 3) use a slightly different approach. They propose a model to measure the likelihood of co-occurrence of words, by using external text data sources to provide regularization instead of the internal text data in the documents. Another important aspect of topics is exclusivity, which “captures whether those high probability words are specific to a single topic” (Roberts et al. 2015, p. 14). A manual approach to condense both procedures is to ask human raters if topics are interpretable and can be associated with a single concept (Newman et al. 2009, pp. 2). Also, one can validate that topics capture a single concept by “reading several example documents” (Roberts et al. 2015, pp. 12), or by comparing already present categories to the automatically generated clusters (Roberts et al. 2015, pp. 12).

Despite the above-mentioned recent advances and extensions of the basic LDA, the method is not immune against critizism and limitations. First, there are obvious problems like intervening variables (such as author and environment specific covariates).³ Despite recent advances allowing to infer the number of topics from the data, such as the Hierarchical Dirichlet Process as an extension of LDA (Teh et al. 2006, p. 1575), it is still prevalent to choose the number of topics beforehand (Blei and Lafferty 2009, p. 81) and to employ post hoc procedures to validate the suitablility of choice. However, if the number of topics is chosen wrong, this can result in a poor performance (Tang et al. 2014, p. 7). Another eligible critique is the need for extensive parameter optimization before running (Asuncion et al. 2009, p. 30), possibly more arranging topic models to fit the needs of the researcher, than capturing what is really there (Schmidt 2013). Also, the underlying bag-of-words assumption (where information on word order is lost) has been criticized for oversimplifying documents (Shafiei and Milios 2006, p. 1).

Putting these deficiencies aside, Schmidt (2013) questions assumptions that are at the heart of the method. More specifically, analysts assume that topics are coherent (i.e., they share some common aspect) and stable (i.e., they apply to several documents the same way), leading them to the opinion that the co-occurrence patterns of words are “more meaningful than the words that constitute them” (Schmidt 2013) and appropriately capture a concept by being semantically coherent. However, as Schmidt (2013) nicely illustrates, in some instances, the top few words characterizing a topic are not necessarily a decent summary of the large amount of words constituting the whole probability distribution. Transferred to language processing, the most frequent words don’t necessarily create the meaning. To a certain extent, these problems can be at least partly solved by techniques like word removal, or changing the bag-of-words assumption to incorporate more information. However, these are also related to the practice of looking just on the top words in the output of the model. Newman et al. (2009, p. 2) express a similar concern on this issue by saying that some topics learned by a model “(while statistically reasonable) are not particularly useful for human use”. Intertwined with that, Crain et al. (2012, pp. 144–148) note that LDA tends to learn broad (more diffuse) topics, where adding concepts to the same topic are favored if these share the same aspects. Thus, the suggestion of Schmidt (2013) is to put extensive effort into visualizing and validating the model before interpreting the results. This, of course, is considerably easier when analyzing position data on a map than with words and their respective semantic implications. Tang et al. (2014, pp. 4–8) point to limiting factors of LDA using a posterior contraction analysis. For example, they depict that a small number of documents (no matter how much words these contain) makes it impossible to guarantee a valid topic identification. The underlying topics need to be well separated for good LDA performance in the sense of Eucledian metric, which is the case if, for example, these are concentrated at a small number of words.

Finally, there are numerous papers which compare the performance of LDA to other related methods in certain problem domains, pointing towards possible deficiencies of LDA for these. Even if LDA is optimized (e.g., in terms of the number of topics), other methods are capable of outperforming it for specific tasks, data and setups. The methods under consideration comprise a multitude of quantitative methods. For example, Hruschka (2016, pp. 8) compared the relative performance of LDA, CTM, a Binary Factor Analytic Model, Restricted Boltzmann Machines (RBMs) and Deep Belief Nets (DBNs) for predicting purchase incidences in a market basket. The author shows that Binary Factor Analysis vastly outperformed both topic models, and was itself outperformed by RBMs and DBNs. Schröder (2017, pp. 31) used a Multidimensional Item Response Theory Model (MIRT) to analyze market baskets for identifying latent traits of households and predicting purchase behaviour. Based on AIC (Akaike Information Criterion), and AICc (Corrected Akaike Information Criterion), MIRT outperforms LDA for both, the binary and the polytomous purchase data scenarios. Salakhutdinov and Hinton (2009, pp. 6) introduce the Replicated Softmax Model, to automatically model low dimensional latent semantic representations in academic and in newspaper articles. Compared to LDA, their model makes better predictions and has a higher retrieval accuracy.

Most of the data used in topic modeling is uncstructured text in various forms like websites and online advertisements (e.g., Le et al. 2008, pp. 889), social media postings (e.g., Schieber et al. 2011, pp. 3), online product reviews (e.g., Tirullinai and Tellis 2014, pp. 465), and more (for a more complete overview see Table A2 in the appendix). Of course, there are also other forms of data like images (e.g., Cao et al. 2014, pp. 8959), purchase records (e.g., Jacobs et al. 2016, pp. 397), mobile apps usage records (Do and Gatica-Perez 2010, pp. 3) and traces of internet browsing behavior (Schröder et al. 2017, pp. 40; Trusov et al. 2016, pp. 409). Some of these comprise discrete units to be crawled and to be pre-processed beforehand (e.g., words, products, etc.) (e.g., Blanchard et al. 2017, pp. 403; pp. 407; Büschken and Allenby 2016, p. 958; Ishingaki et al. 2015, pp. 5; Knights et al. 2009, pp. 244; Schieber et al. 2011, p. 3; Wang et al. 2015, p. 3). For others, scholars need to extract these units automatically or manually before the actual application of a topic model. An example of an automatized extraction is the application of algorithms for the recognition of shapes and patterns prior to any attempts of clustering (e.g., Cao et al. 2014, pp. 8958; Wang et al. 2015, p. 3). A rather manual approach is proposed by Do and Gatica-Perez (2010, pp. 4), who generate artificial words for mobile apps usage, which describe categorical groups, manually predefined by the authors, like the level (high, medium, low), or the time of the day (night, morning, afternoon, evening) of mobile apps usage. For the sake of completeness, however, it should be noted that a third distinct approach is to use synthetic data in a simulation study. In the papers examined by us, this kind of approach was solely used to evaluate topic models before the actual main analyses by knowing the structure of the artificially generated data. There are two ways of doing that. Firstly, one takes an algorithm (often a topic model) to create data and distributions (e.g., Blanchard et al. 2017, p. 402; Büschken and Allenby 2016, p. 971; Knights et al. 2009, p. 243). For example, Knights et al. (2009, p. 243) used LDA to generate documents from synthetic topics with each topic following a random distribution over a set of symbolic words (Knights et al. 2009, p. 243), to subsequently evaluate their Compound Topic Model. Büschken and Allenby (2016, p. 971) evaluate the efficiency of their Sentence-constrained LDA (SC-LDA) model, where a sentence is constrained to have one topic against basic LDA. Secondly, Ishingaki et al. (2015, pp. 13) create purchase records based on marketing variables from a real customer base (Ishingaki et al. 2015, pp. 13), which makes sense for their research aim, in which they try to connect these to latent purchase patterns. By doing so, they intend to assess model performance in terms of the estimators precision and computational time needed for predicting household purchase patterns. We briefly summarized these strategies in Table 1 in conjunction with a few exemplary publications.

In terms of LDA, the critical definition of documents, topics and words vastly differs between papers (Table A2 in the appendix—column: conceptualization of data). Obviously, it varies with the examined data. For example, in the paper of Sun et al. (2013, p. 2–4), documents are a user’s purchase history with each purchased product being a word in this document and topics (i.e., the clusters to be retrieved) are the customers’ purchasing preferences. Schieber et al. (2011, pp. 4) try to model topics and individuals’ opinions about products in Twitter. Accordingly, they define a document as a single Tweet, which consists of words and (possibly) contains a few topics. However, the research interest is another important determinant of how the data is conceptualized. Sticking to the example provided, as already mentioned, a document can be set up as a single posting of a user (Schieber et al. 2011, p. 5). In contrast, Weng et al. (2010, p. 264), who try to identify influential users on Twitter by the following structure and topic similarities, classify all postings of a user as a document (Weng et al. 2010, p. 264). For Paul and Girju (2009, pp. 1412), who intend to find differences in services related to traveling, an online discussion or thread (involving postings of several users) is a document. There are further assumptions that reflect themselves in the methodological conceptualization of the data. For example, scholars often restrict a certain unit in a document (e.g., a sentence) to have a single topic (e.g., Büschken and Allenby 2016, p. 954).

In most publications either basic LDA (e.g., Chen et al. 2013, p. 1), common adaptations like Labeled LDA (e.g., Ramage et al. 2010, p. 132), sLDA (e.g., Blei and McAuliffe 2010, pp. 2), temporal LDA (e.g., Wang et al. 2012, p. 124), the Author Topic Model (e.g., Do and Gatica-Perez 2010, pp. 4), CTM (e.g., Trusov et al. 2016, pp. 413), or other custom adjustments of LDA (e.g., Büschken and Allenby 2016, pp. 957; Paul and Girju 2009, p. 1410; Tirullinai and Tellis 2014, p. 468) are used. However, noticeably less commonly, scholars utilize more exotic models like the User Aware Sentiment Topic Model (USTM) (Yang et al. 2015, pp. 415–417), or the Visual Sentiment Topic Model (VSTM) (Cao et al. 2014, pp. 8959). Apart from that, when taking a broader perspective, in the papers under consideration, there is a continuum of scientific strategies in how to use topic models. One easily to implement approach consists of utilizing the method for the clustering output to represent the actual research results in an exploratory manner (e.g., Cao et al. 2014, p. 8964; Christidis and Mentzas 2013, pp. 4375; Karpienko and Reutterer 2017, p. 17; Luo et al. 2015, pp. 1185; Schröder et al. 2017, pp. 42; Sun et al. 2013, p. 7; Wang et al. 2015, p. 3; Yang et al. 2015, pp. 419).

As a consequence, scholars often perform further analyses and visualizations of certain aspects of the data. For example, it is quite common to use a topic model as an in-between-step for an overall model or research aim (e.g., Cao et al. 2014, pp. 8959; Christidis and Mentzas 2013, pp. 4373; Luo et al. 2015, pp. 1180; Sun et al. 2013, pp. 4; Yang et al. 2015, pp. 420). For instance, Luo et al. (2015, pp. 1180) try to find marketing topics in social media postings and their influence, which is measured by the reaction of users. They utilize LDA to get a topic probability vector for each micro-blog-post and subsequently calculate the topic influence, the topic response, and the topic trends by implementing various custom formulas for further processing. The topic influence is a measure of the proportion of microblog-posts to contain a certain topic, topic response indicates how much customers actively engage in reposting, and the latter indicates the development of the former two over time and across topics (Luo et al. 2015, p. 1182). Different in context but similar by idea, Karpienko and Reutterer (2017, pp. 11) apply LDA to the abstracts of a large compilation of academic marketing articles to derive latent topics of scholarly interest as a preprocessing step before inferring communities of topic combinations, using a version of social network analysis. Subsequently, the authors study the evolution of marketing topics over time and academic journals. Yet another example for using LDA in conjunction with other data compression techniques is the study by Schröder et al. (2017, pp. 42), who use LDA to derive latent shopping interests based on users’ website browsing behavior. Based on the derived latent shopping interests, the authors examine the existence of different online shopper segments using k-means clustering and study implications on shopping behavior. Sun et al. (2013, p. 2), who try to predict customers’ propensities to join group-purchasing events on an online social platform, use LDA to capture the purchasing preferences of customers beforehand. They propose two models, which use this output for further calculation. The first one, PCIM (product-centric inference model), tries to apprehend if the specific product determines that a customer would join a group purchasing event, given the user’s purchasing preferences. Accordingly, the second one, GICIM (group-initiator-centric model), assumes that the group initiator in the social network plays the decisive part in that process, when accounting for the user’s and initiator’s topic mixture (Sun et al. 2013, p. 5). By doing so, they extend the application of topic models from an exploratory approach to one which supports hypotheses testing. Dan et al. (2017, pp. 42) also build a predictive model based on the output derived by LDA for the case of online hotel reviews. Using the latent topics derived from a sentence-constrained LDA version, they develop a latent rating regression approach for making inferences on the relative contribution of each latent topic on guests’ overall hotel evaluations. Another example of that kind of utilizing the method represents a practical approach. Christidis and Mentzas (2013, pp. 4373) try to build a topic based recommender system for buyers and sellers on an e-auction platform, consisting of buyer item recommendations and seller text and item recommendations. They use basic LDA to extract the probabilities of words in topics and topics in documents (i.e., the item descriptions / items in the online marketplace). Subsequently, they calculate the cosine similarities between items (using the topic distributions of each item) and similarities between topics and terms to establish the recommendation functionality.

Obviously, a vital step in current research is to incorporate the output of a complex method or model into topic models (e.g., Cao et al. 2014, pp. 8959; Wang et al. 2015, pp. 2; Yang et al. 2015, pp. 415–418). Self-evidently, this is highly associated with more sophisticated forms of data extraction and preparation. An example for that kind of approach is the USTM framework, which aims at modeling user meta-data, topical aspects, and sentiments in online consumer reviews (Yang et al. 2015, p. 414), to aggregate the opinions of various market segments. To capture the sentiments, the authors utilize seed words and two sentiment lexica and incorporate the sentiment information by using asymmetric Dirichlet priors to assign e.g., positive words with a higher probability for positive topics (Yang et al. 2015, p. 414–419). Another example is the Image-regulated Graph Topic Model (IGTM), where the authors utilize the SIFT feature algorithm and k-means clustering, to extract 500 discrete visual words from images, and build a bag of visual words model for detecting weighted relationships between images via an image relation graph, which consists of nodes (images) and edges (similarities between images). By using further variables, IGTM aims at jointly modeling text and images to enrich topic detection with sentiments (Wang et al. 2015, pp. 2). They allocate a topic assignment and an image assignment for each word in a document, where each topic is a multinomial distribution over words and an image is a multinomial distribution over topics (Wang et al. 2015, pp. 2). There are further instances like Cao et al. (2014, pp. 8958) using an Adjective Noun Pairs (ANPs) based detector for image annotations and Visual Sentiment Ontology (VSO) detectors to construct SentiBank. Subsequently, they incorporate this information in their topic models, trying to enhance the prediction of sentiments in images.

However, at the same time, these authors mark a fourth noticeable implementation strategy, which is to combine a set of topic models in a framework, to improve prediction. More specifically, Cao et al. (2014) intend to build a topic model, which analyzes the distributions of visual sentiment features in topics. Since some non-discriminative (i.e., topic irrelevant) sentiment features have high probabilities in topics, they introduce a background topic model and additional estimators to distinguish these from discriminative ones.

However, it is noteworthy that scholars often tend to use a few strategies (Table 2) in conjunction, meaning that these are not mutually exclusive (e.g., Cao et al. 2014; Sun et al. 2013; Yang et al. 2015).

In addition to these various implementation strategies, authors extend topic models (or use common extensions) (e.g., Blei and Jordan 2003a, p. 3; Blei 2012, pp. 82) (Table 3). These are intertwined with each other and vary with the research aim, modeling purpose and data.

Beyond data structure and implementation, the examined papers employing topic models in marketing can be distinguished along another dimension, namely, the way they evaluate results. All papers examined (except for 11, which do not evaluate the topic model at all⁴) utilize quantitative techniques. The range of these includes H-Scores, Cosine Similarities (Wang et al. 2015, pp. 3), the Kullback–Leibler Divergence (Cao et al. 2014, pp. 8962), Mean Average Precision Scores (Phuong and Phuong 2012, p. 67), the Log Marginal Density (Büschken and Allenby 2016, pp. 962), the Log-Likelihood (Park and Ha 2016, p. 1493; Paul and Girju 2009, p. 1415), Perplexity (Yang et al. 2015, pp. 417), RMSE (Root Mean Square Error) (Iqbal et al. 2015, p. 4), descriptive statistics and correlations (Do and Gatica-Perez 2010, p. 7), Fleiss’k (Ramage et al. 2010, p. 134), and more. Additionally, a considerable portion of publications explicitly or implicitly use qualitative techniques (e.g., Tirullinai and Tellis 2014, p. 470) since it is a common, but in case the model validation relies to a high degree on it, inappropriate practice to use the interpretation of generated topics for model validation (Chang and Blei 2009, p. 3). In general, scholars either provide indicators that encompass the whole model or analyze components of it (e.g., Liu et al. 2010, pp. 204; Yang et al. 2015, pp. 418) (or both). To put this into perspective on an aggregate level, we employ the methodological categories discussed in Sect. 2 (e.g., Roberts et al. 2015, pp. 12), and summarize these in Table 4.

Within the evaluation techniques depicted in Table 4, scholars perform comparisons with various goals in mind. More precisely, in about 80% of publications, the authors conduct some form of comparison for methodological evaluation. We categorized them into 5 types and include them in Table 5 along with a brief description. These five comparison types rely on human ratings / scores (e.g., Tirullinai and Tellis 2014, pp. 470), external reports and categories (e.g., Tirullinai and Tellis 2014, pp. 470), traditional clustering techniques (e.g., Trusov et al. 2016, p. 417), specific metrics in a field (e.g., Weng et al. 2010, pp. 267), and topic models, either in the form of different types of topic models (e.g., Jacobs et al. 2016, pp. 398), or as variations of the same model (e.g., Christidis and Mentzas 2013, p. 4377; Trusov et al. 2016, p. 417).

Taking a broader perspective, one approach to assess the clustering output is to evaluate the outcome in retrospect (e.g., Tirullinai and Tellis 2014, pp. 472), another is to compare predictions of the model from limited data to hold-out data (e.g., Jacobs et al. 2016, p. 397). Another evaluation strategy is to use synthetically generated data (with known distributions beforehand) to compare expected with actually retrieved results (e.g., Blanchard et al. 2017, p. 402; Büschken and Allenby 2016, p. 971; Knights et al. 2009, p. 243).

By simulatenously looking at Table 6 and table A1 in the appendix, we can see that topic modeling extensively depends on somewhat arbitrary factors, like setting the hyperparameters, the number of topics (etc.), and is experimental to a certain degree. Consistently, such uncertainties are visible in an emphasis on model validation. 41% of all papers aim at both (model validation and explorative research), but 34% mainly focus on evaluating their model and the remaining ones (23%), with the exception of one publication which entails no aim, implement the latter. Closely related is the extensive use of evaluation procedures with a focus on (1) assessing model fit of the proposed models, (2) optimal parameter settings, and (3) a detailed analysis of the clustering output. Scholars also introduce comparisons to examine their models (80% of papers). Concisely, they tend to compare (1) variations of the same topic model and (2) their model to other topic models as the dominant strategies. However, an outstanding fact is the frequent inclusion of human ratings and scores (20%). Despite these efforts, there are 11 publications in which there is no applied evaluation technique at all. Scholars in sales / retailing emphasize more on model fit and optimal parameter settings than on an analysis of the clustering output. Intertwined with that is the relatively prominent use of comparing variations of the same topic model. For social media, scholars evaluate topic models less often. One reason for that is that a few authors focus on utilizing topic models as part of a more complex model or research aim, evaluating the overarching model (not topic models directly) (e.g., Herzig et al. 2014, pp. 52). This also reflects itself in the “implementing TM” columns in Table 6. Apart from that, in the online textual consumer reviews section, optimal parameter settings, an analysis of the clustering output, and model fit are predominant, which seems to be a reasonable pattern when looking at the aims of research and the data.

In terms of employed data, scholars heavily rely on already present units in the data (89%). However, a considerable amount (mostly in addition) performs an automatized recognition of units from high dimensional data (20%), and 11% extract manually predefined categories beforehand. It is noteworthy that seven papers don’t rely on already present units but solely use another form of the depicted data retrieval strategies (e.g., Cao et al. 2014, pp. 8958; Do and Gatica-Perez 2010, pp. 4). Scholars also utilize artificially generated data for evaluation purposes in 6 publications. The importance of already present units in the data repeats itself in all the specific fields of research, except for sales / retailing, where scholars often deal with manual categorizations of products and brands (e.g., Do and Gatica-Perez 2010, p. 4).

Methodologically, scholars most often present the output of topic models (82%), but mostly also use this output for further processing in a more complex model (66%). However, in a noticeably smaller amount of publications, just one of the two strategies is exclusively utilized (41%), and 11 papers do not present the output of topic models at all, just using it for a research aim that goes beyond the scope of just deriving topics from the data. A popular strategy is to use the output of another method or model as input for topic models in addition (25%). As previously mentioned, for social media, scholars tend to use the output of topic models for further processing noticeably more often than in other fields of research, almost inverting established methodological patterns. There is an emphasis on utilizing the output of topic models as research results in the Online Textual Consumer Reviews and Services Research branch. These patterns are intertwined with field-specific research interests. In general, scholars tend to use LDA (e.g., Park and Ha 2016, pp. 1491), common extensions and / or perform customizations (e.g., Trusov et al. 2016, pp. 413). Therefore, a total of 87% of papers contain a customized topic model (in comparison to original LDA). Speaking of extending topic models, scholars mostly use MCMC (64%) (and within that Gibbs) for inference, in contrast to Blei et al. (2003). Strikingly, extended topic models seem to be prevalent with 59% of publications introducing additional variables, 51% changing basic assumptions, and 41% introducing constraints. This overall pattern is more or less consistent in the specific fields of research.

The already covered research areas seem to be connected to a certain extent to relatively easy to get and to process (online and digital) data. An additional inclusion of offline data and of high dimensional data (e.g., Do and Gatica-Perez 2010, pp. 4) by using sophisticated strategies of data retrieval could further extend the field. Also, as shown by Blei and Lafferty (2006, pp. 5), who propose a Dynamic Topic Model to analyze changes of topics over time, it is worth it to consider the inclusion of time-varying information. Adopting topic modeling frameworks that allow to integrate additional information, in particular time-dependent marketing covariates, should have great potential in the field of marketing research. Furthermore, integrating some form of guidance in the process of topic generation can be helpful for interpreting the derived solutions (see e.g., Andrzejewski and Zhu 2009, pp. 43; Blei and McAuliffe 2010, pp. 2).

Modeling marketing phenomena entails complexity, reflecting itself in methodologically elaborated conceptualizations and procedures (see the “implementing TM” and “extending TM” columns in Table 6). We expect this aspect to proceed and intensify in the future—both on the implementation level of topic models, as well as on the level of extending the method itself. We highlighted numerous examples of implementing the method. For example, an intriguing approach is presented by Sun et al. (2013), shifting topic modeling from an explorative method to one that supports hypotheses testing, by integrating the output of the topic model into two subsequent models. As we noted above (Table 6), extending the method itself is quite an important strategy in dealing with marketing problems, entailing the introduction of additional variables, constraints and changing basic assumptions & the inference method respectively. This reflects itself in the literature to a certain extent. Several authors also see the need for an improvement of inference algorithms, the invention of tools to more easily develop and implement topic models, and more automatization (e.g., in choosing the number of topics) (Blei et al. 2010, p. 1; p. 10; Blei 2012, pp. 82; Blei 2014, p. 25). Some of these efforts are already in the implementation phase. For example, Bart (2009, pp. 1) technically changes Gibbs Sampling to get a faster performing inference method.

Despite the large variety of techniques and tools available to suitably visualize the outcome of topic models, such as topic proportions and topic distributions across corporae (e.g., Chaney and Blei 2012, pp. 420; Grün and Hornik 2011, p. 16), many scholars in the field continue to call for developing new algorithms to visualize topics and present data and corpora (e.g., Chaney and Blei 2012, p. 419). Additionally, they stress the need for simplified, yet more sophisticated and interactive frameworks for scholars & practicioners (e.g., Blei 2012, p. 84; Blei 2014, p. 25; Kjellin and Liu 2016, pp. 485–460; Zinman and Fritz n.d., p. 2).

Researchers utilize numerous methods and measures to validate modeling outcomes. However, there are some shortcomings as well. Firstly, some authors do not use the full spectrum of evaluation techniques, making their research results somewhat opaque. Secondly, despite indicators that point into the right direction, most of the authors fail to meet the critical problems mentioned by Schmidt (2013), both, in assessing the clustering output, and in model validation (Sect. 2). There are numerous other problems needing to be solved. As already mentioned above, Tang et al. (2014, pp. 4-8) point to some limiting factors of LDA in the form of conditions when applying the method and Crain et al. (2012, pp. 144–148) and Newman et al. (2009, p. 2) elaborate on problems of learning topics. Blei (2012) also sees the need for the development of further methods for evaluating and selecting topic models, when confronted with a particular problem domain or dataset. As we have also mentioned in Sect. 2.5., some related methods have been demonstrated to outperform LDA-type models in specific tasks, using specific data, or setups, which calls for future research to gain a more thorough understanding on the relative advantages of competing methods.