## Synthetic Experiments with PLSA

[Could not find the bibliography file(s)

Probabilistic Latent Semantic Analysis[?] is a technique for modeling the topic structures of a given set of documents, which is similar to LDA (Latent Dirichlet Allocation) [?], except that the latter has a better (more complete) generative property. In this post, I am conducting some simple experiments for testing the PLSA’s ability to recover the real parameter (here I focused on , the word distribution over topics).

The generative process for the data is described in this post, to make visualization easier, we will not use some random , instead, is set to be visualization friendly as suggested by [?]:

Here, each small box is a for topic , each small areas inside the box indicate words, the white areas mean the probabilities of the word for that given topic is high, while the black areas mean the reverse. These boxes are generated by the following procedure: consider 10 topics and 25 words, for each topic, each is a vector, reshape it into box, and uniformly fill 5 entries according to the above figure.

After the data is generated, we can test it with PLSA (optimized by EM algorithm), and see how the results below. To get more through understandings of the performance, I leverage some parameters used in generative process and learning process of PLSA to create different data and learning scenarios.

### Scenario I – Prototype data

For prototype setting, I used 100 documents, each document with expected 100 words, is set of 0.2 so that documents have more skewed topic distribution. Here shows some of the documents:

You can see they are quite mixed, but still appears some kind of pattern. For the document word matrix, we can re-organize it so that documents with the same (highest) topic will appear together (noted this re-organization require ground truth theta, which is not known to PLSA beforehand):

Here you can see the patterns even more clearly, thus it should not be too difficult for PLSA to recover those parameters like .

The figure below is what PLSA recovered:

Compared with the original boxes, they are quite similar I would say (although there are still some differences from the true , but not far away). So, not surprisingly, PLSA recover the parameters fair enough.

### Scenario II – More uniform topic distributions

Since prototype is like the most basic and simplest setting, it is not surprising PLSA works. Now, consider when documents have much more uniform topic distributions, it can be achieved by setting a grater , say 2. And other settings are the same with Scenario one. Here shows some of the documents and the re-organized document word matrix:

Compared with prototype, it is now much difficult for human to discover some patterns. Now how is PLSA doing on recovering :

Worse I would say, but still can count as fair. So it seems to be good at handling highly mixed topics.

### Scenario III – Few words

Now, I will still use the same settings as in prototype, with the changes on number of expected words per document: now it is 10 instead of 100 in prototype. Some of the documents and the re-organized document word matrix:

This is PLSA recovered:

Well, this is not very good, quite sparse compared with true . So fewer words in the documents are definitely not good for PLSA.

### Scenario IV – Few documents

How about few documents, now change the number of documents from 100 in prototype to 10. Still see some of the documents and the re-organized document word matrix:

Okay, see the recovered :

Wow… This is.. Extremely bad.. So we learn, small number of documents are definitely bad for PLSA.

### Scenario V – Different sized topics (small topics)

All above we assume the size of different topics are uniformly distributed, as we use the same alpha for each topics, now, let’s change it, I will use [1 2 3 4 5 6 7 8 9 10] for the 10 different topics, other than this, all settings are still same as in prototype. See some of the documents and the re-organized document word matrix:

What the.. I don’t think any human can learn anything about  from these figures.. But how much can PLSA recovered

:

Hoho, not much.. This suggest for different sized topics, PLSA perform poorly.

### Scenario VI – Learning with more topics

In all above scenarios, learning for PLSA is conducted with true number of topics. What if we don’t know it and use the wrong number of topics? Here I use the same settings as in prototype, but when learning PLSA, instead of using true number of topics, which is 10, I use 20. Now see what PLSA has recovered:

It seems most topics are broken down into many topics, but it wouldn’t matter too much as long as for each topic word distribution it is not polluted by other ones, and the linear combination of those topics can still well recover the true topics.

### Scenario VII – Learning with fewer topics

How about fewer topics, here using topic number of 5. Recovered :

Huh.. This is like… merged together? Definitely not very good.

## Conclusions

So here comes a summarization of performances of PLSA in recovering the true parameter :

• PLSA can recover the hidden topics well if the set of documents are more like “prototype”, meaning with a lot of documents, a lot of words per document, each document is not highly mixed with many topics.
• When the mixing of topics are high, so that each document’s topic distribution is more uniform, PLSA performs a little bit worse, but still fair enough, so highly mixing of topics is not a nightmare for PLSA.
• Few words in documents and few documents in the corpus are nightmares, especially when there are only a few documents, PLSA is likely to fail.
• When there are various sized topics, some topics are very popular and appear in many documents, while some topics are cold and only appear on a few documents, PLSA will also perform worse, especially for the cold topics.
• When using not “exact” number of topics when training, it would not be very problematic if the used number is slightly larger than real one, but it would be if the number is smaller.

It seems those conclusions from the synthetic experiments are someway similar to the conclusion of ICML 2014’s best paper ([?], in Section 4), though which is described for LDA and provide much much deeper studies than provided here.

## Some Existing Work on Linked Documents Analysis

[Could not find the bibliography file(s)

Different from Topic Modeling (which analyzes the topic of documents), Linked Documents Analysis (or Topic Modeling with Links) changes the problem setting by considering the links among documents. The goal is still similar, which is to analyze the topic of documents, expect that sometime we may also want to predict the presence of some queried link. The basic idea behind Linked Document Analysis is that, on one hand, link existence can provide some useful information on the topic distribution, on the other hand, topic distribution can also provide useful information on link existence; in other words, they are correlated. In this post, I am going to review some existing work on topic modeling with links. Basically they can be categorized into generative approaches and discriminative approaches, while former tries to explain why link existing or non-existing by topic/membership distributions of nodes, and the latter tries to utilize existing links in a discriminative way to better model topic/membership distributions.

### Generative Approaches

• Citation Influence Model [?]: The basic idea is that if a document has cited others, the way it generates the topics for word occurrences is either (1) draw from the topic distribution of its cited documents, or (2) draw from its own topic distribution. Other similar models that utilize the citation as a pathway of drawing topic assignments: [?].
• PHITS+PLSA [?]: it treats each document as “words” when performing link generation, it is like PLSA twice, once on document-word, once on document-document. This method was later extended into “Link-LDA“, which puts a prior on topic distribution extending PLSA to LDA. Both models suffer from not (explicitly) utilizing word topical information for link analysis, link analysis is purely based on the information of link co-occurrence, and also fails to model the topical dependence between the cited and citing documents explicitly.
• Link-PLSA-LDA[?]: overcoming the difficulties in PHITS+PLSA and “Link-LDA” by connecting word in topic-word and document topic-document by symmetric parametrization of merged cited documents.
• LDA+MMSB[?]: this may be the most natural generative process for both words and links, it doesn’t treat documents as some kind of dictionaries as in PHITS+PLSA and Link-PLSA-LDA,  and thus can be generalized easily to new documents without previously seen links. However, due to the generative process, this model may not be scale-friendly (though some subsampling techniques might help alleviate the computational complexity).
• LDA+ZAlign (RTM[?]): defining link probability according to the alignment of latent variables Z. the alignment is quantified by dot production, and the transformation from alignment to probability is through either a Sigmoid function or an Exponential function. The authors pointed out an common weakness of previous models (PHITS+PLSA, Link-PLSA-LDA, LDA+MMSB) that due to their underlying exchangeability assumptions, they might divide the topics into two independent subsets, and allow for links to be explained by one subset of topics, and the words to be explained by the other; to improve over this, they enforce the word and link are draw from the same set of topic assignments. In terms of scalability, since it also has to consider both existing and non-existing links, it is also not scale-friendly.
• Others [?].

### Discriminative Approaches

• PLSA+Reg (NetPLSA[?]): adding regularization loss, acts like smoothing, on to original PLSA (log-)likelihood. This model can easily deal weighted links.
• PLSA+Cond (iTopicModel[?]): defining a condition of given its neighbors, which leads to an likelihood of configuration given network of documents, then add this conditioned log-likelihood as a smoothing to original PLSA log-likelihood. Compared with PLSA+Reg, samely this model can also easily deal weighted links, differently this model can fit directed networks more naturally, and also it has a closed form solution of so it is much friendly for optimization compared with PLSA+Reg.
• Others [?].

## Gibbs Sampling for Mixed Membership Stochastic Blockmodels

[Could not find the bibliography file(s)

Mixed Membership Stochastic Blockmodels (MMSB)[?] is a generative model for links in a network. Simply put, give an adjacency matrix , MMSB assume  is governed by the memberships/topics  of each nodes. Since the original paper[?] only introduced the Variational Inference, this post derives a Gibbs Sampling method for the purpose of parameter estimation in MMSB (actually the MMSB presented in this post is slightly different from[?], as I put a prior on interaction matrix ).

The generative process of MMSB is given below:

For each node :

Draw a membership distribution vector

For each entry in :

Draw

For each pair of nodes :

Draw a membership

Draw a membership

Draw

Here is a simple explanation for the notations:

is (number of topics) dimensional hyperparameter for membership/topic distribution.

are two scalar hyperparameters for .

is the membership/topic distribution for node .

is the interaction probability matrix of memberships.

is the membership of node when it interacts with node (the link can be treated as directional, as well as unidirectional). Note that  is abbreviated for .

is the adjacency matrix, each entry is either one or zero.

Later on, I am also going to use the following notations:

is the count of node assigning to membership when it interacts with the rest nodes in the network. For directional network, it can be divided into two sets: and  (), the former is the assignment of node when it links to other nodes, the latter is the assignment of node when other nodes link to it.

is the count of linked node pairs with membership assignments and (sum up  and if don’t consider direction).

is the count of un-linked node pairs with membership assignment and (sum up and if don’t consider direction).

To perform Gibbs Sampling, we need to find the posterior for joint probability of both data and latent variables . Well, here it is:

(1)

Now that we obtained the joint probability of data and latent variables, we can start to derive the conditioned probability of latent variable as following:

(2)

(3)

Given samples of generated from Gibbs Sampling, parameters can be estimated in the following ways:

For , we have

(4)

E.q. 4 is a Dirichlet p.d.f, using Dirichlet variable property, we get the expectation(mean) of :

For , we have

(5)

E.q. 5 is a Beta p.d.f, using Beta variable property, we get the expectation(mean) of :

One may want to sample multiple de-correlated samples of to calculate the above parameters.

#### Likelihood and perplexity

Until now, I didn’t mention the formula for calculating the likelihood of data , which can be important some time for testing of convergence (or using ). In [?], the authors mentioned an approximation technique for computing , it is the harmonic mean of a set of values of from samples, where is sampled from the posterior . So that is:

Where

Using the previous derivation, we can easily compute that:

(6)

Where is a constant:

Now with , we can compute in the way:

(7)

Where , is introduced to make log-sum-exp computed correctly, usually we can set .

The likelihood of hold out data sometime is also an useful quantity, it reflects generalized predictive power of the model, and usually it is measured by perplexity, essentially it is the inverse of the geometric mean per-link likelihood. It is defined by:

This term can be evaluated by first hold out some subset of links and non-links, and evaluated by computing:

(8)

Where and are estimated in the training data.

#### Discussions

The Mixed Membership Stochastic Blockmodels is not scalable as it requires to examining edges (this is a lot even for medium size network with only thousands of nodes!), in sharp contrast to sparse networks as arise more often in real datasets. But in practice, for those non-existing edges, instead of using them all, we can sample a small portion of them, this could effectively reduce the computational complexity, especially for sparse networks where we expect the sample ratio of non-existing edges to be even fewer. However, it is still remained to be examined that the relationship between the loss of performance and sampling ratio of non-existing links.

Also due to the scalability issue, one may consider the MMSB without prior for each document topic distribution, which can be also by EM rather than Gibbs Sampling. Also, one can derive Gibbs Sampling for combined LDA and MMSB [?] in similar ways to Gibbs Sampling for LDA (see here) and MMSB individually (basically key is to derive , where ).

## Notes on Latent Dirichlet Allocation: The Generative Model and Gibbs Sampling Learning

[Could not find the bibliography file(s)

Latent Dirichlet Allocation, originally proposed in [?], is a generative graphical model for mining latent topics of texts or any data with similar underlying statistical structures.

## The Generative Model

The generative process of words in a document are as following:

For each of the topic:

Choose

For each document in the corpus:

Choose topic distribution .

For each of the word occurrences in the document:

Choose a topic .

Choose a word

Here presents a simple explanation of the notations:

is (number of topics) dimensional hyperparameter for topic distribution.

is (number of words) dimensional hyperparameter for word distribution over a given topic.

is the probability of choosing word given its drawn from topic .

is the topic distribution for document .

is the corresponding topic for the -th word in document .

Later on, I am also going to use the following notations:

, , are the observed words.

, , are the topic assignments for word occurrence.

is a dimensional vector, in which is the count of word in the corpus assigned to topic .

is without counting information at word occurrence at position .

is a dimensional vector, in which is the count of words in document been assigned for topic .

is without counting information at word occurrence at position .

The graphical representation of above generative process is as following (each node represents a random variable, filled with gray when observed, links denote dependences, and rectangle plates mean repetition, i.e. the big plate indicates there are documents like that, the small plate indicates there are occurrences words in each document):

To keep (writing) this note simple, I am not going to explain the generative process in details here.

## The Likelihood

What we are interested in this model are random variables , , and maybe . But before going to inferences of these latent random variables, we may first follow the routine and write down the likelihood of the data, i.e.:

(1)

Where

## Gibbs Sampling Learning

Gibbs sampling learning for LDA was first proposed in [?]. The learning problem or inference problem targeted by [?] is to estimate and . I didn’t find any good justification yet for using Gibbs sampling for LDA, one justification is that given the valid sample , we estimate , by either maximizing it or getting its expectation/mean (since this expression is a Dirichlet p.d.f., getting its expectation/mean is easy, and in literature this is often the common practice rather than using mode obtained by maximum likelihood) can give us an estimation of , similar procedures hold for as well. However, this justification is not satisfying in the sense that it didn’t consider the distribution of rather than just a sample of it, and it shows nothing about why a sample of will be good enough. By considering the distribution of (which is intractable thus requires sampling methods), one can derive that

(2)

Where each sample is drawn from distribution . Here the expectation/mean is easier to obtain than the mode by maximum likelihood because we have a summation over p.d.f. and the p.d.f. is of Dirichlet distribution. The expectation of a Dirichlet variable from one sample of  can be found in book or Wiki, and the expectation of the same variable over a set of samples is just an average of expectations of each sample. Technically, we need to obtain multiple de-correlated samples of as according to E.q. 2, however, one sample may be good enough depends on the data (for example, it can be shown that for , even one sample of , it could contain many times of samples given the same word ).

I think E.q. 2 gives us enough motivation and justification for doing Gibbs sampling on LDA. Basically we need to be able to generate samples from , which is an intractable distribution, one popular way for this purpose is Gibbs sampling, which generates samples from each variable once at a time while others fixed.

Preparing for derivation of Gibbs sampling for LDA, let me first lightlight some important functions (as for the mathematics behind MCMC and Gibbs Sampling, readers can refer to some good tutorials on the Internet):

Dirichlet Distribution:

Multinomial Beta function:

Gamma function:

Multinomial distribution:

Noted that in LDA or many of other bayesian statistical models, people use conjugate distribution for likelihood and prior, e.g., Multinomial and Dirichlet. The trick behind this is that posterior will be the same distribution as the prior, so that we can easily find a closed form for the integration of the posterior. In the case of Multinomial and Dirichlet, we have

Since

i.e.:

This piece of formula is extremely useful for deriving the following learning algorithm. (One more note would be: although people call it multinomial distribution in the generative process above, but actually it is one time multinomial, which has already degenerated into discrete distribution, however, this doesn’t affect much for the learning, as we see the formula here will also hold for a sequence of discrete distribution.)

Then let’s derive the joint distribution of and which will be useful just one step later (notations are explained in the beginning of the post):

(3)

Now we can start to derive key formula for Gibbs sampling, which is , here means the topic assignments for all word occurrences other than the one at position , without loss of generality, we assume here that the position is in document and associated with the observed word , since

We have:

(4)

The above E.q. 4 gives us a very clear picture for implementing Gibbs sampling, basically, one just needs to initialize , setting counter for three terms in final of E.q. 4, and each time when sample a word, decrease the corresponding counters, sample according to normalized E.q. 4, than increase corresponding counters according to the sampling result. After “burn-in”, one can read out (de-correlated) samples and start estimation of parameters. By E.q. 2, bayes’ rule and expectation of Dirichlet Distribution, we obtain the expectation of and given a sample of as following:

(5)

(6)

As mentioned above, one might also want to average parameters over multiple de-correlated samples according to E.q. 2.

Likelihood and perplexity

Until now, I didn’t mention the formula for calculating the likelihood of data , which can be important some time for testing of convergence (or using ). In [?], the authors mentioned an approximation technique for computing , it is the harmonic mean of a set of values of from samples, where is sampled from the posterior . So that is:

Where

Using the previous derivation, we can easily compute that:

(7)

Where we have assumed a symmetric . However, the E.q. 7 can easily lead to numerical underflow/overflow if directly computed (due to the large factorial terms). Here we can take a log of it and see what we got:

(8)

Where is constant:

The log gamma function should be built in the programming language so you can use it directly. But now since we only have , and it can be quite “negatively large” (maybe depending on the corpus), this again create a computational difficulty when we compute the Harmonic average of . Again we are going to log , and using some trick below enable us to compute it accurately:

(9)

Here , is introduced to make log-sum-exp computed correctly, usually we can set . Now, even the corpus contains millions of documents, the above likelihood calculation should still work as the scale is under the range of floating point number.

The likelihood of hold out data sometime is also an useful quantity, it reflects generalized predictive power of the model, and usually it is measured by perplexity, essentially it is the inverse of the geometric mean per-link likelihood. It is defined by:

This term can be evaluated by: first hold out some documents, and then sample topic assignment for words in those documents given a point estimate , then the likelihood (there go the perplexity) of hold out documents can be calculated as stated above.

Some resources on Gibbs Sampling for LDA

GibbsLDA++: A C/C++ Implementation of Latent Dirichlet Allocation using Gibbs sampling.

WinBUGS: Gibbs sampling tools for graphical models.

## Notes on EM and pLSA

[Could not find the bibliography file(s)

(Appendix added on August 4th, 2014, see at the end)

## Introduction to EM

EM (Expectation Maximization) is a general technique for solving model learning problem including latent variables. Given a set of observed data and a set of latent variables , we want to learn the parameters by maximizing the log-likelihood of data, i.e. , including latent variable, we have the objective function for learning:

(1)

However, this makes it difficult for us to utilize some standard optimization techniques like Gradient Descent, etc.. Facing this difficulty, we can include two ideas to help: 1, we can do the the whole learning step by step, i.e. iteratively, so that at each step we will have a old set of parameters, from the -th iteration, that we can use; 2, we can find an lower bound for objective function 1, and we optimize towards it, since it is an lower bound that we can guarantee to improve (or not decrease), we can (somehow) guarantee the objective function to be increasing (or not decreasing). Keep these two ideas in mind, we can do the derivation as below:

(2)

In the above derivation, we use last iteration estimated parameters to find a lower bound, which we can just maximize to indirectly improve the original likelihood function. The equality holds when in Jensen’s inequality is a constant variable, in above case that is , is a constant. This equality gives us a estimate of latent variable under the model:

(3)

Many textbook about EM utilizes the so-called Q-function:

(4)

Where we can view as the complete data log-likelihood ( is incomplete data, are latent variables), as predictions of latent variables by using current estimated parameters, and Q-function as an expectation of the log-likelihood of the complete data. Then we can formalize the procedure for applying the EM algorithm:

Pre-Initialization:

Prepare the Q-function 4

Initialization:

Randomly initialize the parameters

E step:

Estimate latent variables according current estimated parameters:

M step:

Maximize the Q-function 4 w.r.t. the parameters given last estimated parameters

Some properties about EM: it is proved that under EM procedure, if the lower bound Q-function is increasing (or not decreasing), then the log-likelihood function will be increasing (or not decreasing) and the iteration will finally go to a converged point (one way to see it is subtract with ). It is worth noting that EM can only find local maxima (a steady high point, but not the highest one), and so the initialization of EM matters.

## Applying EM in pLSA

pLSA (Probabilistic Latent Semantic Analysis [?]) is a technique for mining topics of texts. In pLSA, the log-likelihood function is:

(5)

Where indicates documents, denotes words, means topic assignments. is the number of documents, is the size of vocabulary, is number of topics.

In the log-likelihood function 5, we find a structure with latent variable that is very similar to the log-likelihood function 1 of a general EM problem, so we can utilize the similar derivation that utilized by EM to further derive the log-likelihood function 5, however, we can also do it by applying EM algorithm. To do so, we need to derive Q-function for EM. First, as the equation 5 is not the complete data log-likelihood, which is:

(6)

To apply Q-function, we have to be very careful now, as it will get very tricky as we include the random variable , which is a multivariate random variable that has slots with each slot of states. In Q-function, we have to sum over all possible , which in this case seems lead to an exponential of terms. However that is not the case, as is independent to each other. We can write the Q-function as:

(7)

(Still, the above derivation is not at the most detailed level compared to how tricky it is).

Now we have prepared the Q-function 7 and established the E step, which is to calculate (using Bayes’ Rule), we can go ahead and optimize the Q-function. Before doing so, we should notice the constraints: and , we can include Lagrange multipliers and into the Q-function to form a complete objective function:

(8)

Setting equation 8 to zero w.r.t. to parameters and yields:

(9)

By substituting equation 9 into normalization constraints and , we find a closed form to our maximized parameters estimation:

(10)

To summarize, the EM algorithm for solving pLSA is:

Initialization:

Initialize the parameters and (note that implementation-wise, for , random is better than uniform, which will get stuck, for , uniform is better than random)

E step:

M step

## Appendix

Two more general ways of understanding EM:

Lower bound optimization: given we are optimizing an objective function includes latent variables , since this is directly difficult, we apply Jensen’s inequality in a specific way (see E.q. 2) to get a lower bound of original objective, now the trick of EM is that we improve the objective function by improving the tight lower bound: in E step, we estimate the distribution of latent variable such that (after E step) lower bound is equal to the objective (to do so, the equality condition of Jensen’s equality is used); in M step, we optimize the lower bound w.r.t. parameters (getting rid of constant, M step is optimizing Q-function), since from E step, the objective is tightly lower bounded, the improvement of the objective in M step is guaranteed to be greater or equal to the improvement of the lower bound. Repeat E and M steps, you can see 1. the objective will not decrease (as the lower bound will not decrease) 2. the objective will converge, since it will not decrease, and there must be a upper limit.

Similar to lower bound optimization point of view, what is shown on Bishop’s book “Pattern Recognition and Machine Learning”, chapter 9.4 is even more general. I very much recommend readers to take a look at it.

The end.