Recommendation – Building Babylon

Steffen Rendle and Christoph Freudenthaler (University of Konstanz), WSDM 2014.

The authors present a modification of the Bayesian Pairwise Ranking (BPR) for implicit feedback (i.e. one class) recommendation datasets in which the negative samples are drawn according both to the models current belief and the user/context in question (“adaptive oversampling“). They show that the prediction accuracy of BPR models trained with adaptive oversampling matches that of BPR models trained with uniform sampling but that convergence is 10x-20x faster.

Consider the problem of recommending items to users (or more generally: contexts, e.g. user on a particular page). The observed data consists then of context-item pairs

(c, i)

, where item

i

was the choice made in context

c

. The authors work in the context of pairwise learning, which amounts to a binary classification task where context-item-item triples

(c, i, j)

are labelled as true i.f.f. item

i

was chosen in context

c

but item

j

is not:

where

\hat{y}

is a scoring function (e.g. the dot product of the context- and the item- latent vectors, for matrix factorisation). It is infeasible to consider all the negative examples

j

, so how should we choose which to consider?

Negative sampling from static distributions

We could draw negative examples from the uniform distribution over all items, or instead from the observed distribution over all items (i.e. by popularity). Both are inexpensive to perform and easy to implement. However:

Uniform sampling tends to yield uninformative samples, i.e. those for which the probability of being incorrectly labelled is very likely already low: popular items are precisely those that appear often as positive examples (and hence tend to be highly ranked by the model), while a uniformly-drawn item is likely to be from the tail of the popularity distribution (so likely lowly ranked by the model).
Sampling according to popularity is demonstrated by the authors to converge to inferior solutions.

The authors point out that these sampling schemes depend neither upon the current context (user) nor the current belief of the model. This contrasts with their own method, adaptive over-sampling.

Adaptive over-sampling

The authors propose a scheme in which the negative samples chosen are those that the model would be likely to recommend to the user in question, according to its current state. In this sense it is reminiscent of the Gibbs sampling used by restricted Boltzmann machines.

Choosing negative samples dependent on the current model and user is computationally expensive if performed in the naive manner. The authors speed this up by working with the latent factors individually, and only periodically re-computing the ranking of the items according to each latent factor. Specifically, in the case of matrix factorisation, when looking for negative samples for a context

c

, a negative sample is chosen by:

sampling a latent factor $l$ according to the absolute values of the latent vector associated to $c$ (normalised, so it looks like a probability distribution);
sampling an item $j$ that ranks highly for $l$ . More precisely, sample a rank $r$ from a geometric distribution over possible ranks, then find the item that has rank $r$ when the $l$ th coordinates of the item latent vectors are compared.

(We have ignored the sign of the latent factor. If the sign is negative, one choses the rth-to-last ranked item). The ranking of the items according to each latent factor is precomputed periodically with a period such that the extra overhead is independent of the number of items (i.e. is a constant).

Problems with the approach

The samples yielded by the adaptive oversampling approach depend heavily upon the choice of basis for the latent space. It is easy to construct examples of where things go wrong:

Non-negativity constraints would solve this. Regularisation would also deal with this particular example – however regularisation would complicate the expression of the scoring function $\hat{y}$ as a mixture (since you need to divide though by $Z_{c}$ .

Despite these problems, the authors demonstrate impressive performance, as we’ll see below.

Performance

The authors demonstrate that their method does converge to solutions slightly better than those given by uniform sampling, but twenty times faster. It is also interesting to note that uniform sampling is vastly superior to popularity based sampling, as shown in the diagrams below.

Note that a single epoch of the adaptive oversampling takes approximately 33% longer than a single epoch of uniform sampling BPR.

Implementation

According to the paper, the method is implemented in libFM, a C++ software package that Rendle has published. However, while I haven’t looked exhaustively, I can’t see anything in that package about the adaptive oversampling (nor in Rendle’s other package, MyMediaLite).

What about adaptive oversampling in word2vec?

Word2vec with negative sampling learns a word embedding via binary classification task, specifically, “does the word appear in the same window as the context, or not?”. As in the case of implicit feedback recommendation, the observed data for word2vec is one-class (just a sequence of words). Interestingly, word2vec samples from a modification of the observed word frequency distribution (i.e. of the “distribution according to popularity”) in which the frequencies are raised to the $0.75$ th power and renormalised. The exponent was chosen empirically. This raises two questions:

Would word2vec perform better with adaptive oversampling?
How does BPR perform when sampling from a similarly-modified item popularity distribution (i.e. raising to some exponent)?

Corrections

Corrections and comments are most welcome. Please get in touch either via the comments below or via email.

We consider the “Weighted Approximate-Rank Pairwise-” (WARP-) loss, as introduced in the WSABIE paper of Weston et. al (2011, see references), in the context of making recommendations using implicit feedback data, where it has been shown several times to perform excellently. For the sake of discussion, consider the problem of recommending items $i$ to users $u$ , where a scoring function $f_{u} (i)$ gives the score of item $i$ for user $u$ , and the item with the highest score is recommended.

WARP considers each observed user-item interaction $(u, i)$ in turn, choses another “negative” item $i^{'}$ that the model believed was more appropriate to the user, and performs gradient updates to the model parameters associated to $u$ , $i$ and $i^{'}$ such that the models beliefs are corrected. WARP weights the gradient updates using (a function of) the estimated rank of item $i$ for user $u$ . Thus the updates are amplified if the model did not believe that the interaction $(u, i)$ could ever occur, and are dampened if, on the other hand, if the interaction is not surprising to the model. Conveniently, the rank of $i$ for $u$ can be estimated by counting the number of sample items $i^{'}$ that had to be considered before one was found that the model (erroneously) believed more appropriate for user $u$ .

Minimising the rank?

Ideally we would like to make updates to the model parameters that minimised the rank of item $i$ for user $u$ . Previous work of Usunier (one of the authors) showed that the precision at k was best optimised when the logarithm of the rank was minimised. (to read!)

The problem with the rank is that, while it does depend on the model parameters, this dependence is not continuous (the rank being integer valued!). So it is not possible to speak of gradients. So what is to be done instead? The approach of the authors is to derive a differentiable approximation to the logarithm of the rank, and to minimise this instead.

Derivation: approximating the (log of the) rank

WARP has been shown several times to perform very well for implicit feedback recommendation. However, the derivation of the approximation of the log of the rank used in WARP, as outlined in the WSABIE paper, is nonsense. I can only think that the authors arrived at WARP in another way. Let’s look at it more closely. In the following:

$f_{u} (i)$ is the score assigned by the model to item $i$ for user $u$ .
$L$ is some function that defines the error as a function of the rank. In the WSABIE paper, $L (k) = \sum_{j = 1}^{k} \frac{1}{j}$ is approximately the natural logarithm (for the derivation below, however, it doesn’t matter what $L$ is)

warp derivation The most obvious problem with the derivation is the approximation marked with an asterix (*). At this step, the authors approximate the indication function $I [x > y]$ by $I [x > y] \cdot (x - y + 1)$ . While the latter is familiar as the hinge loss from SVMs, it is (begin unbounded!) a dreadful approximation for the indicator $I [x > y]$ . It seems to me that the sigmoid of the difference of the scores would be a much better differentiable approximation to the indicator function.

To appreciate why the derivation is nonsense, however, you have to notice that the it has nothing to do with $L$ . That is, the derivation above would yield an approximation for $L$ , whatever $L$ happened to be, even a constant function.

Optimisation

WARP considers each observed interaction $(u, i)$ in turn, repeatedly sampling items $i^{'}$ from the uniform distribution over all items until it finds one in $V_{u, i}^{1}$ , i.e. until it finds an item $i^{'}$ whose score for the user $u$ is at worst 1 less than the score of the observed interaction. For this triple $(u, i, i^{'})$ , it performs gradient updates to minimise:

$L ({rank}_{u}^{1} (i)) \cdot (f_{u} (i^{'}) + 1 - f_{u} (i))$

The naive approach to computing ${rank}_{u}^{1} (i)$ is to calculate all the scores for the given user in order to determine the rank of the item $i$ . WARP performs a nice trick to do much better: it estimates ${rank}_{u}^{1} (i)$ by counting how many candidate negative items $i^{'}$ it had to consider before finding one in $V_{u, i}^{1}$ . This yields

${rank}_{u}^{1} (i) \approx \frac{total number of items - 1}{number of i’ we had to draw}$

However it is still the case that $L ({rank}_{u}^{1} (i))$ is not differentiable. So when we compute the gradients, this quantity has to be treated as a constant. Thus it simply becomes a weighting applied to the gradient of the difference of the scores (hence the name WARP, I guess).

WARP optimises for item to user recommendations

With its negative sampling technique, WARP optimises for recommending items to a user. For instance, the problem of recommending users to items (so, transposing the interaction matrix) is not trained for. I wonder if some extra uplift could be obtained by training for both problems simultaneously.

Normalising for the total number of items

With the optimisation stated as above, the learning rate will need to be re-tuned for datasets that have different numbers of items, since the gradient weighting $L ({rank}_{u}^{1} (i))$ is ranges from $L (0)$ to $L (total number items)$ . It would make more sense to weigh the gradient updates by:

$\frac{L ({rank}_{u}^{1} (i))}{L (total number items)}$

which ranges between 0 and 1.

Implementations

There are two implementations of WARP for recommendation that I know of, both in Python:

LightFM, written by Maciej Kula. Works well. Also implements BPR with uniform sampling and WARP k-OS (which I’ve not investigated yet).
MREC, written by Levy and Jack at Mendeley, has a matrix factorisation recommender trained using WARP. I’ve not tried this one out yet.

References

Jason Weston, Samy Bengio and Nicolas Usunier, Wsabie: Scaling Up To Large Vocabulary Image Annotation, 2011, (PDF).

Category: Recommendation

Improving Pairwise Learning for Item Recommendation from Implicit Feedback 2014