Every now and then someone (you know who you are) asks if the feature vectors one passes into LDA should be vectors of word counts (i.e., vectors of non-negative integers) or vectors of word presence/absence (i.e., vectors of binary values). Now the former gives strictly more information so the short answer is that you should always use word counts when available. But in my experience the difference is less than you might think.
To that end, I decided to put together a little side-by-side. Today’s corpus is Cora abstracts (with stopword and infrequent word filtering). I’m running everything through stock LDA-C with alpha set to 1.0 and K set to 10. Let me just preface what follows with the warning that your mileage may vary (and if it does, let me know!).
First up, let’s look at the topics (or specifically, the top 10 words in each topic) produced for the two feature representations. No earth-shattering differences here and absent an objective way to measure the quality of a topic, I’d be hard-pressed to say that one produced results any better than the other.
Topic 1 Topic 2 Topic 3 Topic 4 Topic 5 learning problem performance model research control problems results models report state search classification data part reinforcement method paper bayesian technical paper selection methods analysis paper dynamic algorithms data probability grant system solution method markov university systems space classifier time supported simulation optimization parallel distribution science robot test application methods artificial Topic 6 Topic 7 Topic 8 Topic 9 Topic 10 network learning algorithm genetic knowledge neural decision error model system networks paper number evolutionary design input examples show fitness reasoning training features algorithms visual case weights algorithms class results theory recurrent algorithm learning population paper hidden rules results evolution systems trained approach function crossover cases output tree model strategies approach
Topic 1 Topic 2 Topic 3 Topic 4 Topic 5 system research performance learning algorithms behavior report paper data genetic systems abstract approach present search complex technical parallel classification algorithm design part proposed algorithm problem model paper implementation show results development university results training optimal paper science level accuracy show computational supported memory results problems environment computer multiple decision find Topic 6 Topic 7 Topic 8 Topic 9 Topic 10 method bayesian neural function learning methods reasoning network distribution paper problem models networks algorithm learn applied model input general problem problems cases learning class system time case model model reinforcement demonstrate paper hidden linear approach technique framework training functions knowledge large markov trained show learned learning knowledge weights results programming
Next, let’s look at the entropy of the topic multinomials to get an idea of what the general shape of these topic distributions look like. Here I’ve computed the topic entropies for both sides of the comparison; I then sort them by entropy (silly exchangeability!). Finally, I’m showing the scatter plot of the entropy of the word-counts topics vs. the entropy of the word-presence topics. The blue line runs along the diagonal. The differences are not too phenomenal; as expected binary features mean higher entropy since binary features amount to a smaller number of observations with which to overcome smoothing. But in the end these are really just 3-5% differences.
Next up, let’s look at convergence of the likelihood bound (these aren’t strictly comparable in any meaningful way since of course the likelihood is computed over two different data representations). Here I’m showing the variational bound on the log likelihood as a function of iteration. In black, word counts are used as features; in red, strictly binary features are used. Because the likelihood scales are different, I’ve rescaled the two curves so that they’re approximately equal (the two axes reflect the two original scales). As with the other evaluations, the shape of the two curves is very similar.
Finally, let’s look at the entropy of the per-document topic-proportions to get an idea of how topics are assigned in each case. As with the previous scatter plot, this plot shows the entropy of each document’s topic proportions under word counts vs. word presence. As before, less data in the binary feature case means generally higher entropy. But the differences are more notable in this case. While for many documents the differences are within 10% for some the difference is as large as 100%. This is most likely due to the fact that in some cases (e.g., when the number of distinct words in a document is small), feature binarization changes the character and amount of data in the document by quite a lot.
I think it is in these corner cases that using full word counts data is likely to be most useful. But overall I think the differences are not that great and not worth expending too many grey cells over.