2.1 TLDA Provides Theoretical Foundations, Open Sourced Software, and Scalable Estimates

Our method has three key advantages over existing methods. First, our method has important theoretical properties. Neither the most popular LDA approaches based on Blei, Ng, and Jordan (Reference Blei, Ng and Jordan2003) nor LLMs have yet been shown to have these theoretical foundations (Anandkumar et al. Reference Anandkumar, Foster, Hsu, Kakade and Liu2012, Reference Anandkumar, Foster, Hsu, Kakade and Liu2013, Reference Anandkumar, Ge, Hsu, Kakade and Telgarsky2014). Our approach recovers—in feasible computational time—a provable identification guarantee for the topic–word probabilities and sample complexity bounds, as well as a form of statistical consistency in large samples (Anandkumar et al. Reference Anandkumar, Foster, Hsu, Kakade and Liu2012, Reference Anandkumar, Foster, Hsu, Kakade and Liu2013).Footnote ² These foundations suggest that parameter recovery and large-sample accuracy are achievable when the text data follow a data generation process assumed by LDA, and that the topic–word probability matrix is full rank.

Second, our method is open-source, intentionally designed to be estimated on a wide array of workstations. Although LLMs show tremendous promise as a research method and clearly model human speech patterns more realistically than bag-of-word approaches like we propose, most of the popular LLMs are not open-source, cannot be trained locally without high-end computational resources (Linegar, Kocielnik, and Alvarez Reference Linegar, Kocielnik and Michael Alvarez2023), and demonstrate various social biases with significant implications for research using their model outputs, particularly on prompts concerning gender and race (Bartl, Nissim, and Gatt Reference Bartl, Nissim and Gatt2020; Kocielnik et al. Reference Kocielnik, Prabhumoye, Zhang, Jiang, Alvarez and Anandkumar2023; Nozza, Bianchi, and Hovy Reference Nozza, Bianchi, Hovy and Toutanova2021; Sheng et al. Reference Sheng, Chang, Natarajan, Peng, Inui, Jiang, Ng and Wan2019; Zhao et al. Reference Zhao, Wang, Yatskar, Ordonez and Chang2018). By open-sourcing the training of the model, researchers can better test the sensitivity of their hyperparameter choices, run more robustness checks, and better assess the validity of their model outputs. While LLMs are extremely flexible and powerful and can be fine-tuned to a diverse array of tasks, this flexibility comes at the cost of an unconstrained, high-dimensional parameter space: the prompts needed to define the task that the model will perform. Small changes in these prompts often result in large and unpredictable differences in the model outputs, which make prompt-based methods very challenging to tune and reduce their reliability. The limiting principles of prompt engineering are an area of active research.

At the same time, in terms of interpretation and inference, proprietary LLM model weights, trained parameters, and specifics of training data are hidden from public view. For commercial LLMs, the model underpinnings and mechanics are areas of active research by those developing the LLMs. While it is possible for researchers to fix a particular version of an LLM to use in their research, it is not transparent to users how fixed any particular version of a commercial LLM might be. Unlike some proprietary models, our mathematical underpinnings are clearly and openly communicated and replicated, and the implementation itself is transparent and open source—previous versions are archived and available for reproduction purposes. Relatedly, due to their generative objective, commercial LLMs have been shown to “hallucinate,” to literally shift their responses to prompts to completely different topics and subjects; by construction, our method will not hallucinate. Finally, by implementing a batched, streaming version of our method, researchers are freed from memory constraints that otherwise plague traditional unsupervised methods with large text corpora.

Third, both LLMs and supervised methods impose substantive financial resource hurdles that serve as barriers to their use by many researchers—hand labeling is expensive and proprietary LLM API calls can be extremely costly at scale. Furthermore, in applications where the entire population of documents contains valuable information, down-sampling may not be a viable solution to this cost. For example, 10,000 documents in a 200 million document sample (0.005%) might comprise a cogent and important topic, say congressional speeches opposing war, but due to sampling, such a critical topic may not be identified at all if only a few thousand documents are sampled. For both replication purposes and ready access to applied researchers, we hope our method allows political scientists to more readily answer questions of pressing concern, making more widespread use of large-scale text corpora numbering in the millions and billions of documents.

3.1 Theoretical Guarantees Build on Existing Political Science Methods

Our contribution to the political methodology literature is to introduce topic model techniques from computer science that have statistical theoretical foundations. To cluster and analyze large text datasets, political science researchers make widespread use of unsupervised topic models, which do not always have parameter recovery and accuracy guarantees (Anandkumar et al. Reference Anandkumar, Foster, Hsu, Kakade and Liu2012, Reference Anandkumar, Foster, Hsu, Kakade and Liu2013). Among them, a popular model is LDA (Blei et al. Reference Blei, Ng and Jordan2003; Hoffman, Bach, and Blei Reference Hoffman, Bach, Blei, Lafferty, Williams, Shawe-Taylor, Zemel and Culotta2010). This workhorse model can extract important information without requiring labeling or prior knowledge and has been used to analyze datasets across the social sciences, including studies on coordination among social movements, strategic communication of political elites, news dissemination, and the detection of toxic online behavior (Grimmer, Roberts, and Stewart Reference Grimmer, Roberts and Stewart2022; King and Hopkins Reference King and Hopkins2010; Lauderdale and Clark Reference Lauderdale and Clark2014). Also popular is the structural topic model (STM), which is closely related to LDA (Roberts et al. Reference Roberts2014; Roberts, Stewart, and Tingley Reference Roberts, Stewart, Tingley and Alvarez2016). These methods are popular due to the feasibility and easy implementation, but they often do not scale well to large datasets, so we turn to the computer science literature, where scale is key to unlocking new frontiers in research.

3.2 Feasible Improvements for Existing Scalable LDA

There have been numerous efforts to make LDA more scalable. Specifically, Yu et al. (Reference Yu, Hsieh, Yun, Vishwanathan and Dhillon2015) develop a method for faster sampling and distributing computation across multiple CPU cores. More recently, efficient GPU LDA implementations have been proposed: some have developed improved GPU workload partitioning (Wang et al. Reference Wang, Liu, Gaihre and Yu2023; Xie et al. Reference Xie, Liang, Li and Tan2019), while Wang et al. (Reference Wang, Liu, Gaihre and Yu2023) developed a new branched sampling method. However, all of these implementations rely on traditional LDA methods based on Gibbs sampling, variational Bayes, or expectation maximization. The parallelization and scalability of these methods are inherently algorithmically challenging, as they are limited significantly by the sampling required to estimate the topics. Furthermore, unlike our method, these implementations are not available open-source, so they cannot be used easily by applied researchers.

Table 1

Runtime of our TLDA method on GPU for 260 million and 1.04 billion documents using the COVID dataset.

Note: None of the previous LDA methods scale to billions of documents.

Therefore, instead of relying on traditional LDA methods, we leverage tensor methods (Kolda and Bader Reference Kolda and Bader2009), which are embarrassingly parallel and have been proposed in order to scale to larger datasets (Papalexakis, Faloutsos, and Sidiropoulos Reference Papalexakis, Faloutsos and Sidiropoulos2016; Sidiropoulos et al. Reference Sidiropoulos, De Lathauwer, Fu, Huang, Papalexakis and Faloutsos2017). Using tensor methods, it is possible to learn latent variable models with accuracy guarantees under mild regularity conditions (Anandkumar et al. Reference Anandkumar, Ge, Hsu, Kakade and Telgarsky2014; Janzamin et al. Reference Janzamin, Ge, Kossaifi and Anandkumar2019). In particular, Tensor LDA relies on the computation of third-order moments, i.e., the three-way co-occurrence of words, and decomposes them to recover the topics. This approach has been shown to have similar performance as traditional LDA (Huang et al. Reference Huang, Niranjan, Hakeem and Anandkumar2015), which we also demonstrate empirically in the results of this article.

However, previous implementations of these tensor methods are limited by (1) the explicit construction of the second- or third-order moments and (2) a lack of hardware acceleration due to being developed either entirely for CPU or with a costly exchange between CPU and GPU. In particular, researchers have developed tensor LDA methods that face memory constraints due to the computation of high-dimensional low-order tensors (Anandkumar et al. Reference Anandkumar, Foster, Hsu, Kakade and Liu2012, Reference Anandkumar, Foster, Hsu, Kakade and Liu2013, Reference Anandkumar, Ge, Hsu, Kakade and Telgarsky2014). Further, these methods only run on CPU and do not benefit from hardware acceleration on GPU. But Huang et al. (Reference Huang, Niranjan, Hakeem and Anandkumar2015) developed a stochastic tensor gradient descent (STGD) approach to estimate the third-order decomposition, allowing for further scaling of the method. However, this method has CPU–GPU exchange and relies on the explicit construction of the second-order moment, so it cannot be run fully online. Similarly, Swierczewski et al. (Reference Swierczewski, Bodapati, Beniwal, Leen and Anandkumar2019) proposed a method for learning the third-order decomposition using alternating least squares but is limited by a CPU-based implementation.

By contrast, we derive an efficient centered, online version of the TLDA that scales linearly to any dataset size (with constant memory). We provide evidence supporting this claim in Table 1. We provide an end-to-end GPU-accelerated implementation that will be open-sourced along with this article to enable its application to any dataset by other researchers.

4.1 Computing the Centered Cumulants

In this section, we provide a technical overview of the method. For a summary of our notation and the mathematical background for the method, please refer to the Supplementary Material (Table A1). We begin by introducing the model, the data generation process, and the estimation routine.

The data are a corpus of documents which we note in the following way. Let the data be a document term matrix $\boldsymbol {F}$ with rows ${\boldsymbol {f}}_t := (f_{t, 1}, f_{t, 2},. .. , f_{t, V}) \in \mathbb {R}^{V}$ denoting the vector of word counts for the t-th document, where V is the number of words in the vocabulary, and let N be the number of documents. Finally, we will let K denote the total number of topics and h be the topic labels. Then, the first-order moment is

(1) $$ \begin{align} {\boldsymbol{M}}_1 := \frac{1}{N}\sum_{i=1}^N {\boldsymbol{f}}_i. \end{align} $$

Our first innovation is to center the data by forming $\tilde{{\boldsymbol{x}}}_i := {\boldsymbol{f}}_i - {\boldsymbol{M}}_1$ . Given this set-up, we can simplify the usual moments for spectral LDA (see Anandkumar et al. Reference Anandkumar, Foster, Hsu, Kakade and Liu2012, Reference Anandkumar, Foster, Hsu, Kakade and Liu2013, Reference Anandkumar, Ge, Hsu, Kakade and Telgarsky2014, as well as Section 1.2 of the Supplementary Material). This is because the diagonalization terms in $\boldsymbol{M}_2$ and $\boldsymbol{M}_3$ become $\mathbf {0}$ in expectation, as the first moment is now ${\boldsymbol {0}}$ for the centered data matrix $\tilde {\boldsymbol {X}}$ . This vastly reduces the number of off-diagonal calculations required to estimate the higher-order moments.

By removing the corresponding terms that are now $\mathbf {0}$ in expectation from the expression of the moments in Anandkumar et al. (Reference Anandkumar, Foster, Hsu, Kakade and Liu2013), we now have the following simplified empirical moments for the centered data $\tilde {\mathbf {X}}$ , where $\alpha _{0}$ is the topic mixing parameter

(2) $$ \begin{align} \tilde{\mathbf{M}}_2 &:= \frac{(\alpha_{0} +1)}{N}\sum_{t=i}^N \tilde{{\boldsymbol{x}}}_i \otimes \tilde{{\boldsymbol{x}}}_i \end{align} $$

(3) $$ \begin{align} \tilde{\mathbf{M}}_3 &:= \frac{(\alpha_{0} + 1)(\alpha_{0} + 2)}{2N}\sum_{i=1}^N \tilde{{\boldsymbol{x}}}_i \otimes \tilde{{\boldsymbol{x}}}_i \otimes \tilde{{\boldsymbol{x}}}_i,\end{align} $$

where $\otimes $ is the Kronecker product.

Following Anandkumar et al. (Reference Anandkumar, Foster, Hsu, Kakade and Liu2013), we know that the moments for the centered data can be factorized as

(4) $$ \begin{align} \mathbb{E}_{ } \boldsymbol{ \big[ }\, { \tilde{{\boldsymbol{M}}}_1 } \,\boldsymbol{ \big] } &:= {\boldsymbol{0}} \end{align} $$

(5) $$ \begin{align}\mathbb{E}_{ } \boldsymbol{ \big[ }\, { \tilde{\boldsymbol{M}}_2 } \,\boldsymbol{ \big] } &:= \sum_{i=1}^K \frac{\alpha_{i}}{\alpha_{0}} {\boldsymbol{{\boldsymbol{\nu}}}}_i \otimes {\boldsymbol{\nu}}_i\end{align} $$

(6) $$ \begin{align}\mathbb{E}_{ } \boldsymbol{ \big[ }\, { \tilde{\boldsymbol{M}}_3 } \,\boldsymbol{ \big] } &:= \sum_{i=1}^K \frac{\alpha_{i}}{\alpha_{0}} {\boldsymbol{\nu}}_i \otimes {\boldsymbol{\nu}}_i \otimes {\boldsymbol{\nu}}_i, \end{align} $$

where $\tilde {{\boldsymbol {M}}}_1$ , $\boldsymbol {\tilde {M}}_2$ , and $\tilde {\mathbf {M}_3}$ are the first, second, and third moments of the centered data, and $\boldsymbol {\nu } = [{\boldsymbol {\nu }}_1,. .. , {\boldsymbol {\nu }}_K]$ , the learned decomposition of $\tilde {\mathbf {M}}_3$ . Note that we use the centered analog from Anandkumar et al. (Reference Anandkumar, Foster, Hsu, Kakade and Liu2013), which showed that the singular value decomposition of the third-order moment tensor yields estimates for the LDA model parameters.

Finally, to recover the actual topic–word probability matrix as derived in Anandkumar et al. (Reference Anandkumar, Foster, Hsu, Kakade and Liu2013) using the topics computed from the centered data, we prove Theorem 1, demonstrating that the ground-truth factors can be recovered by de-centering the factors from the centered ones.

Theorem 1 Given the factors ${\boldsymbol {\nu }}_i$ learned from the centered data, we show that

$$ \begin{align*}\mathbb{E}[\mathbf{M}_3] = \sum_{i = 1}^K \frac{\alpha_i}{\alpha_0} ({\boldsymbol{\nu}}_i + \boldsymbol{M}_1) \otimes ({\boldsymbol{\nu}}_i + \boldsymbol{M}_1) \otimes ({\boldsymbol{\nu}}_i + \boldsymbol{M}_1).\end{align*} $$

That is, the true third-order cumulant can be recovered directly by re-centering the factors of the decomposition of the centered cumulant, indicating that the vectors ${\boldsymbol {\nu }}_i + \boldsymbol {M}_1$ are equivalent to the ground-truth factors from Anandkumar et al. (Reference Anandkumar, Foster, Hsu, Kakade and Liu2013).

For the proof, see Section 1.3 of the Supplementary Material.

Following from Theorem 1, we have

$$ \begin{align*} \mathbb{E}_{ } \boldsymbol{ \big[ }\, { \mathbf{M}_3 } \,\boldsymbol{ \big] } &= \sum_{i = 1}^K \frac{\alpha_i}{\alpha_0} ({\boldsymbol{\nu}}_i + {\boldsymbol{M}}_1) \otimes ({\boldsymbol{\nu}}_i + {\boldsymbol{M}}_1) \otimes ({\boldsymbol{\nu}}_i + {\boldsymbol{M}}_1) \\ &= \sum_{i = 1}^K \frac{\alpha_i}{\alpha_0} {\boldsymbol{\mu}}_i \otimes {\boldsymbol{\mu}}_i \otimes {\boldsymbol{\mu}}_i, \end{align*} $$

where ${\boldsymbol {\mu }} = [\mu _1,. .. , \mu _K]$ and ${\boldsymbol {\mu }}_i = Pr (f_j | h = i),$ where ${\boldsymbol {h}}$ are the set topic labels, i is the i’th topic, and j is the j’th word in the vocabulary. In other words, ${\boldsymbol {\mu }}$ is the topic word matrix of the uncentered data $\boldsymbol {F}$ .

4.2 Batched Implementation Enables Arbitrary Scaling

In the sections below, we present a batched implementation for TLDA. Each moment is individually calculated incrementally as we feed in data to the method, and then we estimate the topic–word probabilities using stochastic gradient descent.

In the section that follows the online decomposition for the second and third moments, we present a fully online implementation, which we recommend for very large datasets on the scale of billions of documents. For such datasets, the individual contribution of one data point is extremely small, so we average over many documents with minimal loss to accuracy given the enormous gains to scale.Footnote ⁵ In practice, this means we can iterate through the documents just once to still achieve our convergence criteria and to achieve accurate inference. As a byproduct of this implementation, we can then update the model by streaming new data points into the training API, giving a means to offer a fully online version of the model as part of this library.

4.3 Online Decomposition of the Second Moment

As a function of centering the data, our method streamlines the pipeline that Anandkumar et al. (Reference Anandkumar, Foster, Hsu, Kakade and Liu2012) proposes for calculating the second moment and the whitening matrix. Specifically, instead of constructing $\boldsymbol {\tilde {M}}_2$ , which is very memory-intensive for large data, we implicitly form $\boldsymbol {\tilde {M}}_2$ by computing a singular value decomposition of the centered data matrix.

Using the singular values and singular vectors from the centered data, we construct a whitening matrix $\boldsymbol {W}$ such that

(7) $$ \begin{align} \boldsymbol{W}^{\top} \boldsymbol{\tilde{M}}_2 \boldsymbol{W}=I. \end{align} $$

We let D be the whitening dimension size. We note that from Anandkumar et al. (Reference Anandkumar, Foster, Hsu, Kakade and Liu2012, Reference Anandkumar, Foster, Hsu, Kakade and Liu2013, Reference Anandkumar, Ge, Hsu, Kakade and Telgarsky2014), letting $D = K$ is sufficient to compute the third-order decomposition, although a slightly larger D can be chosen to improve the dimensionality reduction. Then, from the centered data, we have

$$ \begin{align*}\boldsymbol{W} = \frac{\sqrt{\alpha_0+1}}{N} \boldsymbol{U}\, \boldsymbol{\Sigma}^{-\frac{1}{2}},\end{align*} $$

where $\boldsymbol {U}$ and $\boldsymbol {\Sigma }$ (the variance matrix of the centered data) are the top D singular vectors and singular values of the centered data, obtained through computing the PCA of $\boldsymbol {\tilde {X}}$ , which is equivalent to its SVD since the data are centered.

Then, the $\boldsymbol {\tilde {M}}_3$ tensor is implicitly formed using the whitened counts of the centered data. Whitening renders the tensor symmetric and orthogonal (in expectation). Most importantly, it reduces the dimensionality of the third moment from size $N^3$ to $D^3 \approx K^3$ , where K is the number of topics. Given the nature of speech in social environments, the number of topics will almost always be at least an order of magnitude smaller than the number of words.

To estimate the implicit third moment, the method calculates the whitened counts $\boldsymbol {x} = \boldsymbol {W} \boldsymbol {\tilde {X}}.$ We will use these whitened counts to construct the implicit third-order tensor. Using that implicit tensor, the method utilizes stochastic gradient descent to find the spectral decomposition of the third-order moments.

4.4 Online Learning of the Third Moment

Here, we formulate the TLDA framework in a vectorized form, solving for a batch of data, resulting in a much more efficient implementation. Let ${\boldsymbol {\Phi }} = [\Phi _1|\Phi _2|. .. |\Phi _K]$ be the eigenvectors of the third-order moment for the whitened, centered data. We note that each eigenvector $\Phi _i$ is of length D and denote the full sample size as N.

Now, note that the decomposition of the third-order moment for the whitened, centered data $\boldsymbol {X}$ (of size $D \times D \times D \approx K \times K \times K$ ) is

$$ \begin{align*}\boldsymbol{T} = \sum_{i \in K} {\boldsymbol{\Phi}}_i \otimes {\boldsymbol{\Phi}}_i \otimes {\boldsymbol{\Phi}}_i.\end{align*} $$

With the whitened tensor in hand, the method follows Huang et al. (Reference Huang, Niranjan, Hakeem and Anandkumar2015) in implementing a batched STGD algorithm for tensor CP decomposition.

Specifically, we consider a mini-batch of $n_B$ centered and whitened samples ${\boldsymbol {x}}_1, \ldots {\boldsymbol {x}}_{n_B}$ , which we collect in a matrix $\boldsymbol {X} \in \mathbb {R}^{n_B \times D}$ . We want to learn a tensor factorization of the third-order whitened and centered cumulant:

(8) $$ \begin{align} \tilde{\boldsymbol{M}}_3 = & \,\, \frac{(\alpha_0 + 1)(\alpha_0 + 2)}{2N} \sum_{n=1}^N {\boldsymbol{x}}_n \otimes {\boldsymbol{x}}_n \otimes {\boldsymbol{x}}_n\nonumber \\ = & \,\, \frac{(\alpha_0 + 1)(\alpha_0 + 2)}{2N} \sum_{n=1}^N {\boldsymbol{x}}_n \otimes^3. \end{align} $$

We are trying to learn a rank-K CP factorization with factors ${\boldsymbol {\Phi }}$ of $ \tilde {\boldsymbol {M}}_3$ such that $\tilde {\boldsymbol {M}}_3 = \mathbf {T} = \sum _{i=1}^K {\boldsymbol {\Phi }}_i \otimes {\boldsymbol {\Phi }}_i \otimes {\boldsymbol {\Phi }}_i $ . In other words, we solve the following optimization problem:

(9) $$ \begin{align} \operatorname*{\mbox{arg min}}_{{\boldsymbol{\Phi}};\, ||{\boldsymbol{\Phi}}_i||_F^2=1} \underbrace{|| \tilde{\boldsymbol{M}}_3 - \mathbf{T} ||_F^2}_{\text{reconstruction loss}} + \underbrace{ \frac{(1 + \theta)}{2} || \mathbf{T} ||_F^2}_{\text{orthogonality loss}}. \end{align} $$

In plain words, we minimize the reconstruction loss while inducing orthogonality on the decomposition factors. This can be seen by noting that the factors (and therefore the rank-1 components) are normalized, meaning the Frobenius norm of the second term simplifies to only the inner product between the components.

The problem in equation (9) thus simplifies to

(10) $$ \begin{align} \operatorname*{\mbox{arg min}}_{{\boldsymbol{\Phi}};\, ||{\boldsymbol{\Phi}}_i||_F^2=1} & \frac{(1 + \theta)}{2} || \sum_{i=1}^K {\boldsymbol{\Phi}}_i \otimes^3 ||_F^2 \\ \nonumber - & \langle {\sum_{i=1}^K {\boldsymbol{\Phi}}_i \otimes^3}, { \frac{(\alpha_0 + 1)(\alpha_0 + 2)}{2N} \sum_{n=1}^N {\boldsymbol{x}}_n \otimes^3 } \rangle.\end{align} $$

This can be equivalently written in matrix form using the Khatri–Rao product:

(11) $$ \begin{align} \operatorname*{\mbox{arg min}}_{{\boldsymbol{\Phi}};\, ||{\boldsymbol{\Phi}}_i||_F^2=1} & \,\, \frac{(1 + \theta)}{2} || {\boldsymbol{\Phi}} \left({\boldsymbol{\Phi}} \odot {\boldsymbol{\Phi}}\right)^{\top} ||_F^2 \\ \nonumber & - \frac{(\alpha_0 + 1)(\alpha_0 + 2)}{2n} \langle {{\boldsymbol{\Phi}} \left({\boldsymbol{\Phi}} \odot {\boldsymbol{\Phi}}\right)^{\top}}, { \boldsymbol{X}^{\top} \left( \boldsymbol{X} \odot \boldsymbol{X} \right) } \rangle. \end{align} $$

By taking the derivative with respect to the decomposition factor ${\boldsymbol {\Phi }}$ , we get

(12) $$ \begin{align} \frac{\partial{\mathcal{L}}}{\partial{{\boldsymbol{\Phi}}}} = 3(1 + \theta) {\boldsymbol{\Phi}} ({\boldsymbol{\Phi}}^{\top}{\boldsymbol{\Phi}} * {\boldsymbol{\Phi}}^{\top}{\boldsymbol{\Phi}}) - \frac{3(\alpha_0 + 1)(\alpha_0 + 2)}{2n} \boldsymbol{X}^T (\boldsymbol{X}{\boldsymbol{\Phi}}*\boldsymbol{X}{\boldsymbol{\Phi}}). \end{align} $$

We then update the factor via batched stochastic gradient update:

(13) $$ \begin{align} {\boldsymbol{\Phi}}_{t+1} = {\boldsymbol{\Phi}}_{t} - \beta \frac{\partial{\mathcal{L}}}{\partial{{\boldsymbol{\Phi}}}}, \end{align} $$

with $\beta $ as the learning rate.

4.5 Fully Online Implementation

In a fully batched implementation above, the method relies on computing each higher-order moment sequentially, even though each moment is individually learned online. By contrast, in the fully online version of our TLDA method presented here, we learn both moments by jointly learning both the second- and third-order moments online.Footnote ⁶ We first find initial values for the factors by running the batched TLDA to convergence on a small portion of the data. Then, when given a new batch of data, we first update $\boldsymbol {M}_1$ , then update the incremental PCA (and use the new version of the PCA to whiten the data), and finally perform a gradient update of the third-order moment using the new batch of whitened data. As a result, we can update the third-order moment decomposition only using the new batch without looping through any prior data. This is in contrast to the batched version of our method, where we loop through the entire dataset three times to compute the first-order moment, second-order moment decomposition, and third-order moment decomposition, respectively.

Although we only perform one gradient descent step per batch of data in this version of the method, we expect that for large datasets, there are sufficiently many documents so that this does not significantly impact the quality of the factors produced.Footnote ⁷ We use online LDA to obtain topic coherence values that are similar to or better than existing methods.

4.6 Recovering the Topic Model Parameters

Once we have learned the factorized form $\Phi $ of the third-order moment, we describe how we recover the uncentered, unwhitened moment to recover the topics. First, we obtain $\boldsymbol {\nu }$ , the estimate of the decomposition of $\tilde {\boldsymbol {M}}_3$ , by unwhitening the components $\Phi $ of the decomposition:

$$ \begin{align*}\boldsymbol{\nu} = \boldsymbol{W}^{T^{\dagger}} \boldsymbol{\Phi},\end{align*} $$

where $\dagger $ denotes the pseudo-inverse. Using this, we can find

$$ \begin{align*}\alpha_i = \gamma^2{\boldsymbol{\nu}}_i^{-2}.\end{align*} $$

Here, $\gamma $ is a scaling factor such that $\sum _{i=1}^k \alpha _i = 1$ .

As derived in Theorem 1, we can then re-center $\boldsymbol {\nu }$ to compute the estimate for the uncentered topic–word probabilities

$$ \begin{align*}{\boldsymbol{\mu}}_i = {\boldsymbol{\nu}}_i + \boldsymbol{M}_1.\end{align*} $$

5.1 Input and Output

The method takes a data matrix where each entry is the centered word frequencies for each of V words for N documents. Each column represents a word in the dataset and each row is a document, for a matrix of size $N\times V$ . The method produces two key outputs: first, the topic–word matrix and second, the learned weights, $\alpha _i$ . These outputs can be fed as inputs into a standard VI method that calculates the topic–document matrix, as well. We have included a standard VI method in our API so users can calculate the document–topic matrix for their applications.

5.2 Package Availability

The package is fully open-source as part of the TensorLy (Kossaifi et al. Reference Kossaifi, Panagakis, Anandkumar and Pantic2019) project, under the BSD-3 license, which makes it suitable for any use, academic or industrial. It is well tested and has extensive documentation.

6.1 Parameter Recovery and Comparison to Previous TLDA Methods

In this section, we demonstrate that our method results in gains in accuracy, topic correlation, and speed in comparison to existing TLDA methods in a simulated setting. We use the traditional LDA Data Generation Process for generating the simulated data (see Section 1.4 of the Supplementary Material and Blei et al. Reference Blei, Ng and Jordan2003). We present a comparison to two key existing versions of the TLDA method: (1) the spectral decomposition algorithm in Anandkumar et al. (Reference Anandkumar, Foster, Hsu, Kakade and Liu2012, Reference Anandkumar, Foster, Hsu, Kakade and Liu2013, Reference Anandkumar, Ge, Hsu, Kakade and Telgarsky2014) and (2) the SGD-based TLDA derived in Huang et al. (Reference Huang, Niranjan, Hakeem and Anandkumar2015), in which the third-order moment is computed online. To do so, we show comparisons of all three TLDA methods on synthetic data. Due to the small scale of the synthetic experiment (20,000 documents), we run the batched version of our method.

As a result, through this process, we obtain ground-truth factors that adhere to the assumptions of LDA. By running TLDA on the synthetic document vectors, we can then compare each factor in the learned topic–word matrix to the corresponding factor in the generated ground-truth topic–word matrix by computing their correlation.

However, the topics in the learned topic–word matrix can be in any order, so we limit the number of topics to $K = 2$ and use the permutation that maximizes the average correlation to the ground truth. We use parameters $\alpha _0 = 0.01$ and whitening size $D = 2$ for all TLDA methods, as well as learning rate $1 \times 10^{-4}$ for the SGD-based TLDA and our method. Table 2 shows the correlation between the learned and ground-truth factors in corpora with $20,000$ documents. The results are averaged over ten random seeds for each combination of parameters. This table illustrates that under a variety of vocabulary sizes, our method is more accurate than existing tensor methods, as evidenced by the higher mean and lower standard deviation of correlations among all runs.

Table 2

Comparison of topic recovery on synthetic data for various TLDA methods.

Note: (1) Anandkumar et al. (Reference Anandkumar, Ge, Hsu, Kakade and Telgarsky2014); (2) Huang et al. (Reference Huang, Niranjan, Hakeem and Anandkumar2015); and (3) Our method. The method achieving the highest average correlation and lowest standard deviation among the correlations is bolded at each vocabulary size.

In Table 3, we compare the average runtime of the three TLDA methods for the synthetic data experiments in Table 2. In Table A4 in the Supplementary Material, we also compare against other scalable LDA methods—we note our TLDA method compares favorably and that none of the methods against which we compare have theoretical accuracy guarantees. We run this analysis on the CPU due to the relatively small scale of the experiment. Our version of the TLDA method results in a runtime that is between 6 and 20 times faster than the existing TLDA methods. Furthermore, as shown in Table 3, as the vocabulary size increases, the runtime of our method increases its relative speed advantage over the others. This demonstrates the value of the simplifications we introduce in the method section; we significantly outperform the existing TLDA methods in terms of runtime while also making non-trivial gains in accuracy.

Table 3

Comparison of CPU runtime on synthetic data for various TLDA methods.

Note: (1) Anandkumar et al. (Reference Anandkumar, Ge, Hsu, Kakade and Telgarsky2014); (2) Huang et al. (Reference Huang, Niranjan, Hakeem and Anandkumar2015); and (3) Our method.

7.1 The #MeToo Movement: Scaling, Ablation, and Substantive Studies

7.1.1 Studying Mass Movements and Collective Action with Large-Scale Datasets

The #MeToo movement is a prolific women’s rights movement that gained traction extremely quickly on Twitter in October 2017, with over 7.9 million tweets containing the #MeToo hashtag from October 2017 to October 2018 alone (Brown Reference Brown2018). This movement is an important example of what Clark-Parsons has termed “Network Feminism,” where social media platforms have become a crucial organizational tool for mobilization of social and political movements (Clark-Parsons Reference Clark-Parsons2022).

Going back to the early theoretical work of Mancur Olson, studying social movements and protest politics as a lens for collective action has been an important literature in political science (Olson Jr. Reference Olson1971). In particular, researchers have long tried to understand the political origins of protest politics and mass movements, because as Olson noted, participation can be costly and the results of participation can be difficult for individuals to assess.

Studying how mass movements and protest politics arise, how they are organized, and how they are sustained in the long run, is also complicated by a lack of available data. Movements and protests arise quickly, and authorities often act to stop and prevent protesting and organizing, which means in many cases that political scientists cannot often collect data about protests and movements. Surveys of protest participants can be done after the fact, but they can be difficult to find, difficult to persuade to participate in a survey, and their survey responses may be inaccurate with the passage of time. Thus, much of the literature has resorted to case studies of historical examples (Chong Reference Chong1991).

In recent years, the use of social media by protest and movement participants has sparked new research about collective action, specifically regarding protests and mass movements. By collecting data from participants in the protests, while they are protesting or acting collectively, has proven to be an important way to generate datasets to test existing theories, for example, about the Arab Spring or Black Lives Matter movements (Kann et al. Reference Kann, Hashash, Steinert-Threlkeld and Michael Alvarez2023; Steinert-Threlkeld Reference Steinert-Threlkeld2017).

In a similar way, the tweets related to #MeToo thus provide rich data for investigating the evolution of a modern social movement initiated by online discussions on social media. What topics engaged participants in the #MeToo movement? Did the topics of conversation change over the course of the movement? Can the language of social media conversations help us determine the motivations of participants in the movement, where they motivated by self-interest or collective concerns? These data can be crucial for understanding this important social and political movement, and for testing theories about how movements like these arise and are sustained.

We analyze the topics present in a corpus of #MeToo tweets collected from January 2017 to September 2019, which contains 7.97 million tweets after pre-processing. Figure 4 shows the initial proliferation of tweets in January 2017 related to the Women’s March and Movement, as well as the viral growth of the #MeToo movement in September and October 2017.

Figure 4

Tweets per month in the #MeToo data, in millions.

7.1.3 #MeToo Scaling and Convergence Speed Comparison

In this set of results, we first run the batched and online versions of our TLDA method on the entire #MeToo Twitter dataset on one GPU core to analyze how quickly both versions of our approach converge with varying numbers of topics. We then compare the scaling of our TLDA method to that of Gensim by computing convergence time on subsets of the #MeToo dataset containing 1M, 2M, 5M, and 7.97M documents.

Full #MeToo Timing Comparison

To choose the optimal parameters for our method, we run an extensive grid search over the number of topics K, whitening sizes D, topic mixing parameters $\alpha _0$ , and learning rates $\beta $ . For each number of topics, we determine the optimal parameters by finding the parameter combination with the highest mean coherence score across all metrics. For $K = 10$ , we report the 20 top words for the batched model with the best parameters in Table A2 in the Supplementary Material. We include the parameters used for each number of topics in Table A6 in the Supplementary Material.

Using these optimal parameters, we run the batched and online versions of our model to convergence and provide a benchmark comparison of speed and coherence on GPU to the most popular off-the-shelf CPU-based LDA method, Gensim LDAMulticore, which is fully parallelized (Rehurek and Sojka Reference Rehurek and Sojka2011). For the Gensim LDA model, we keep all default parameters, except we increase the number of passes and iterations until the coherence of the model converges for each specified number of topics. We include the final parameters used in Table A7 in the Supplementary Material. We compute the coherence measures using the Gensim CoherenceModel library by providing the $20$ top words in each topic and using the default parameters for each coherence measure.

In the Supplementary Material, we also compare how our methods scale against other state-of-the-art LDA methods (see Table A4 in the Supplementary Material). In particular, we analyze LightLDA (Yuan et al. Reference Yuan2015) and WarpLDA (Chen et al. Reference Chen, Li, Zhu and Chen2016), both of which purport tremendous increases in scale. We note neither method offers the accuracy guarantees of TLDA, nor are they actively maintained as part of a larger suite of packages (such as Gensim’s LDAMulticore or Tensorly’s TLDA, presented here). We note in any case that our method outperforms both in terms of time, even without the theoretical guarantees of our method (Anandkumar et al. Reference Anandkumar, Foster, Hsu, Kakade and Liu2013).Footnote ¹⁰ We also note that there are many other methods purporting to perform topic modeling at scale. As far as our review of them can determine, none purport to have the accuracy guarantees of our method, nor have they been widely adopted. For this reason, we focus our analysis on Gensim’s implementation, which is, by far, the most popular and widely adopted.

As shown in Table 4, we find that the batched version of our method, running on a single GPU, achieves 10 $\times $ –95 $\times $ speedup over Gensim running in parallel on 79 CPU cores. The online version running on a single GPU achieves a 15 $\times $ –143 $\times $ speedup over Gensim run in parallel on 79 CPU cores. Notably, our method has a relatively constant convergence speed and consistent coherence for all numbers of topics. On the other hand, as the number of topics increases, the convergence time for the Gensim model increases nonlinearly and the final coherence score to which it converges decreases significantly. The runtime comparison as the number of topics increases is visualized in Figure 5. As these results indicate, our method is especially useful for analyzing larger numbers of topics, thanks to the near-constant scaling of convergence time as the number of topics increases: for more than 60 topics, our method is over 100 $\times $ faster on GPU than the fully parallelized Gensim CPU LDA method.

Table 4

TLDA convergence timing comparison on the full #MeToo dataset.

Note: See Rehurek and Sojka (Reference Rehurek and Sojka2011) for details about Gensim LDA Multicore and Röder, Both, and Hinneburg (Reference Röder, Both and Hinneburg2015) for the coherence metrics.

Figure 5

Runtime comparison for TLDA on GPU vs. Gensim for the full #MeToo corpus and varying numbers of topics.

Note: This shows that the runtime of our method scales nearly constantly with respect to the number of topics, while Gensim scales more than linearly.

Scaling Study

We fit our TLDA method for 10, 20, 40, 60, 80, and 100 topics on four subsets of the #MeToo movement containing 1, 2, 5, and 7.97 million tweets. We keep the same vocabulary and optimal parameters from the full #MeToo timing comparison while running the model on each of these subsets. As seen in Figure 6, which displays the TLDA fit time (excluding preprocessing) for 100 topics, the convergence time for our method scales linearly with the number of documents on the GPU, while the convergence time for the Gensim LDAMulticore method increases significantly faster. Combined with our empirical finding from the full-scale #MeToo study that our method scales nearly constantly with the number of topics, this indicates that it is feasible to run both the batched and online versions of our method on even larger data. We include the plots for the remaining number of topics in Figure A2 in the Supplementary Material.

Figure 6

TLDA vs. Gensim fitting time.

Note: We compare the time to fit Gensim’s LDAMulticore and our online TLDA, not including pre-processing, for 100 topics. We plot the runtime in seconds as a function of the size of the subset from the #MeToo dataset, from 1 million to 7.97 million tweets.

7.1.4 CPU Ablation Study

To demonstrate the benefits of our GPU implementation, we perform an ablation study comparing CPU and GPU TLDA runtime.

As shown in Table 5 (which does not include pre-processing time), we find that the CPU version of our method has relatively constant runtime with respect to the number of topics and takes under 1 hour to converge for each number of topics on the full #MeToo dataset. Furthermore, we perform an empirical scaling study (see Figure A3 in the Supplementary Material), in which we confirm that the convergence time of our method scales linearly with the number of topics on CPU as well as on GPU. These results indicate that our method is feasible to use on CPU for datasets on the scale of tens of millions of documents for researchers who do not have access to GPU machines. However, the GPU implementation results in significant gains in TLDA convergence speed and allows for scaling to hundreds of millions and billions of documents.

Table 5

Comparison of CPU and GPU runtime on the #MeToo dataset (7.97 million tweets).

7.2 Qualitative Analysis of the #MeToo Movement

Finally, we present a full-scale analysis of the evolution of two years of #MeToo Twitter discussion over time. Previous studies have shown the importance of accounting for the dynamic nature of conversation in the #MeToo movement (Liu et al. Reference Liu, Srikanth, Adams-Cohen, Alvarez and Anandkumar2019), and here we analyze the topical development of tweets concerning #MeToo from just before the start of the movement in August 2017 through September 2019. Understanding the topical development of the #MeToo movement would be important for scholars who wish to test theories about how social movements organize online, for studies of feminism or gender politics, for two examples.

To develop this dynamic analysis, we iteratively grow the corpus of tweets and estimate the TLDA model. That is, first, we fit the model on August and September 2017 data to capture the discussion immediately preceding the time when #MeToo discussion went viral on Twitter. Then, we add the next month and estimate the entire model again. We repeat this process for each subsequent month until we reach the end of the dataset. In Figure 1, we report the topic for selected months with the largest weight $\alpha _i$ for both pro- and counter-#MeToo topics. We call these topics the most prominent. The relative size of the labels indicates the relative prominence of the respective topics. In Figure 2, we report the political topic with the largest $\alpha _i$ , which we call the most politically salient.

This approach leads to three key qualitative findings: first, topics related to politically salient news events are generally ephemeral, changing often as we grow the dataset over time. In contrast, topical prominence related to personal testimonies, coordinating protests, and supporting other participants in the #MeToo movement is persistent over time. Third, discussion around counter-#MeToo topics declines in prominence over time and is subsumed into one topic by September 2019. These findings suggest avenues for the study of the temporal evolution of political and social coordination in mass movements, especially on social media, as most research using this type of data to study social movements has not examined the longer-term dynamics of these forms of political engagement (Jost et al. Reference Jost2018; Steinert-Threlkeld Reference Steinert-Threlkeld2017). Moreover, these results indicate that dynamic changes in topic evolution on social media may be in response to changing news events, another area ripe for new research using large social media and news media datasets using methods like TLDA.

7.3 Election 2020: Elite Political Communication and the Losers’ Effect

In this second application, we look at another source of real-time social media data relevant to how political elites communicate with and coordinate their supporters, especially in the face of electoral defeat. A peaceful transfer of power is a necessary indicator of a healthy democracy. Studies into how political elites respond to electoral defeat offer critical insights into how democracies persist, even in polarized or politically fraught periods of history. In light of the 2020 election, longstanding questions around how politicians organize their supporters around election were especially salient. Previous studies into political anger and the losers’ effect have relied on survey-based evidence, and produced valuable insights into voter psychology and political behavior (Craig et al. Reference Craig, Martinez, Gainous and Kane2006; Sinclair, Smith, and Tucker Reference Sinclair, Smith and Tucker2018); however, these methods are static and rely on recalled emotions after an election. Building on this important work, our methodology allows researchers to study how a politician coordinates with his or her politically engaged supporters online. With this type of data, we can study online political behaviors in real-time—President Trump was active online and especially on Twitter (Li et al. Reference Li, Cao, Adams-Cohen, Michael Alvarez, Ceolin, Caselli and Tulin2023), so these conditions allow us to better understand how electoral losers respond to electoral defeat. For policy makers, such methods might allow for real-time detection of the spread of effective attempts to inform the public, while highlighting messaging that angers, rather than informs, the voting public.

As with #MeToo, this data was collected based on a keyword search employing the social media data collection techniques in Cao, Adams-Cohen, and Alvarez (Reference Cao, Adams-Cohen and Michael Alvarez2021) and Li et al. (Reference Li, Cao, Adams-Cohen, Michael Alvarez, Ceolin, Caselli and Tulin2023). The data are lightly structured, so our method can offer insights into the latent structure of this politically engaged demographic: online supporters. In this case, a comprehensive collection of tweets related to keywords related to the administration of the 2020 Presidential election served as the basis of the data collection. Overall, the data are composed of 29,711,862 tweets collected from September 1, 2020 through the Inauguration on January 20, 2021. These data were collected in real time as the tweets were being posted. In large part thanks to this real-time collection of data, we were able to capture many tweets prior to their being deleted or moderated, giving us an unfiltered look at social media activity during a critical period in American electoral politics.

In Figure 7, we report the daily number of tweets collected by our keyword search. The data are most voluminous on November 4th, the day of the election, with nearly 3 million election-related tweets. There is a second peak of just under 1.5 million tweets on November 7th, the day the media outlets declared Joe Biden the winner of the election. Activity remains high following the election with nearly 250 thousand to 500 thousand tweets following the election.

Figure 7

Number of tweets per day.

The TLDA framework offers potential new insights as a data discovery method for a critically important area of political science—how the public and politicians organize messaging surrounding the legitimacy of elections on social media. Large-scale social media data concerning the legitimacy of the election are particularly suited for the TLDA framework because discerning meaningful structure from these data would be impractical without automated methods. More than merely large, the data are unstructured because they are directly collected in real-time from people’s authentic online thoughts. In terms of generating new theories about the role of online behavior in disseminating information about the election amplifying the loser’s effect, actively organizing political supporters, and its direct role in helping to coordinate the rally and riots on January 6th, a data discovery method like TLDA is the most computationally feasible way to study these data real-time and at scale.

7.4 Qualitative Analysis of Trump-Related Social Media Activity in the 2020 Election

We now report an analysis based on the outputs of our TLDA analysis—we hope this serves as an example of how political scientists might employ our framework to engage in data discovery for unstructured data at large scale. We report descriptive findings on a full population of tweets covering a topic of critical importance to students of political behavior and electoral politics. To that end, we recover 30 topics overall. For this analysis, we classify the topics into three main categories based on the following criteria: of the discovered topics, we find three broadly related categories: discrediting the election, discussing legal challenges to the election, and tweets verifying the election result.

Notably, in Figure 8, we observe relative stability in the daily share of tweets belonging to each category both across time and between the relative share of each of the categories. We see that tweets discrediting the election account for 25% of all tweets on average, with a relative peak on January 6th. We see that in the days after January 6th, the share of tweets related to verifying the election results actually drops.

Figure 8

Topical composition over time: In this figure, we report the average share of tweets belonging to one of three main categories of topics with a greater than 90% probability.