exponential_family_embeddings
Bernoulli Embeddings for Text
view repo
Word embeddings are a powerful approach for capturing semantic similarity among terms in a vocabulary. In this paper, we develop exponential family embeddings, a class of methods that extends the idea of word embeddings to other types of highdimensional data. As examples, we studied neural data with realvalued observations, count data from a market basket analysis, and ratings data from a movie recommendation system. The main idea is to model each observation conditioned on a set of other observations. This set is called the context, and the way the context is defined is a modeling choice that depends on the problem. In language the context is the surrounding words; in neuroscience the context is closeby neurons; in market basket data the context is other items in the shopping cart. Each type of embedding model defines the context, the exponential family of conditional distributions, and how the latent embedding vectors are shared across data. We infer the embeddings with a scalable algorithm based on stochastic gradient descent. On all three applications  neural activity of zebrafish, users' shopping behavior, and movie ratings  we found exponential family embedding models to be more effective than other types of dimension reduction. They better reconstruct heldout data and find interesting qualitative structure.
READ FULL TEXT VIEW PDF
Word embeddings are a powerful approach for analyzing language, and
expo...
read it
The success of machine learning methods heavily relies on having an
appr...
read it
Word embeddings are a powerful approach for unsupervised analysis of
lan...
read it
Growing amounts of online user data motivate the need for automated
proc...
read it
Learning an embedding for a large collection of items is a popular appro...
read it
In data dominated systems and applications, a concept of representing wo...
read it
Archetypal analysis represents a set of observations as convex combinati...
read it
Bernoulli Embeddings for Text
This project collects the different accepted papers and their link to Arxiv or Gitxiv
Word embeddings are a powerful approach for analyzing language (Bengio et al., 2006; Mikolov et al., 2013a, b; Pennington et al., 2014)
. A word embedding method discovers distributed representations of words; these representations capture the semantic similarity between the words and reflect a variety of other linguistic regularities
(Rumelhart et al., 1986; Bengio et al., 2006; Mikolov et al., 2013c). Fitted word embeddings can help us understand the structure of language and are useful for downstream tasks based on text.There are many variants, adaptations, and extensions of word embeddings (Mikolov et al., 2013a, b; Mnih and Kavukcuoglu, 2013; Levy and Goldberg, 2014; Pennington et al., 2014; Vilnis and McCallum, 2015), but each reflects the same main ideas. Each term in a vocabulary is associated with two latent vectors, an embedding and a context vector
. These two types of vectors govern conditional probabilities that relate each word to its surrounding context. Specifically, the conditional probability of a word combines its embedding and the context vectors of its surrounding words. (Different methods combine them differently.) Given a corpus, we fit the embeddings by maximizing the conditional probabilities of the observed text.
In this paper we develop the efemb, a class of models that generalizes the spirit of word embeddings to other types of highdimensional data. Our motivation is that other types of data can benefit from the same assumptions that underlie word embeddings, namely that a data point is governed by the other data in its context. In language, this is the foundational idea that words with similar meanings will appear in similar contexts (Harris, 1954). We use the tools of exponential families (Brown, 1986) and GLM (McCullagh and Nelder, 1989) to adapt this idea beyond language.
As one example beyond language, we will study computational neuroscience. Neuroscientists measure sequential neural activity across many neurons in the brain. Their goal is to discover patterns in these data with the hope of better understanding the dynamics and connections among neurons. In this example, a context can be defined as the neural activities of other nearby neurons, or as neural activity in the past. Thus, it is plausible that the activity of each neuron depends on its context. We will use this idea to fit latent embeddings of neurons, representations of neurons that uncover hidden features which help suggest their roles in the brain.
Another example we study involves shoppers at the grocery store. Economists collect shopping data (called “market basket data”) and are interested in building models of purchase behavior for downstream econometric analysis, e.g., to predict demand and market changes. To build such models, they seek features of items that are predictive of when they are purchased and in what quantity. Similar to language, purchasing an item depends on its context, i.e., the other items in the shopping cart. In market basket data, Poisson embeddings can capture important econometric concepts, such as items that tend not to occur together but occur in the same contexts (substitutes) and items that cooccur, but never one without the other (complements).
We define an efemb, such as one for neuroscience or shopping data, with three ingredients. (1) We define the context, which specifies which other data points each observation depends on. (2) We define the conditional exponential family. This involves setting the appropriate distribution, such as a Gaussian for realvalued data or a Poisson for count data, and the way to combine embeddings and context vectors to form its natural parameter. (3) We define the embedding structure, how embeddings and context vectors are shared across the conditional distributions of each observation. These three ingredients enable a variety of embedding models.
We describe efemb models and develop efficient algorithms for fitting them. We show how existing methods, such as CBOW (Mikolov et al., 2013a) and negative sampling (Mikolov et al., 2013b), can each be viewed as an efemb. We study our methods on three different types of data—neuroscience data, shopping data, and movie ratings data. Mirroring the success of word embeddings, efemb models outperform traditional dimension reduction, such as exponential family PCA (Collins et al., 2001) and Poisson factorization (Gopalan et al., 2015), and find interpretable features of the data.
Related work. efemb models generalize CBOW (Mikolov et al., 2013a) in the same way that exponential family PCA (Collins et al., 2001) generalizes PCA, GLM (McCullagh and Nelder, 1989) generalize regression, and deep exponential families (Ranganath et al., 2015) generalize sigmoid belief networks (Neal, 1990). A linear efemb (which we define precisely below) relates to contextwindowbased embedding methods such as CBOW or the vLBL (Mikolov et al., 2013a; Mnih and Kavukcuoglu, 2013), which model a word given its context. The more general efemb relates to embeddings with a nonlinear component, such as the skipgram (Mikolov et al., 2013a) or the ivLBL (Mnih and Kavukcuoglu, 2013). (These methods might appear linear but, when viewed as a conditional probabilistic model, the normalizing constant of each word induces a nonlinearity.)
Researchers have developed different approximations of the word embedding objective to scale the procedure. These include noise contrastive estimation
(Gutmann and Hyvärinen, 2010; Mnih and Teh, 2012), hierarchical softmax (Mikolov et al., 2013b), and negative sampling (Mikolov et al., 2013a). We explain in 2.2 and A how negative sampling corresponds to biased stochastic gradients of an efemb objective.We consider a matrix of observations, where each is a vector. As one example, in language is an indicator vector for the word at position and is the size of the vocabulary. As another example, in neural data is the neural activity measured at index pair , where indexes a neuron and indexes a time point; each measurement is a scalar ().
The goal of an efemb is to derive useful features of the data. There are three ingredients: a context function, a conditional exponential family, and an embedding structure. These ingredients work together to form the objective. First, the efemb models each data point conditional on its context; the context function determines which other data points are at play. Second, the conditional distribution is an appropriate exponential family, e.g., a Gaussian for realvalued data. Its parameter is a function of the embeddings of both the data point and its context. Finally, the embedding structure determines which embeddings are used when the th point appears, either as data or in the context of another point. The objective is the sum of the log probabilities of each data point given its context. We describe each ingredient, followed by the efemb objective. Examples are in 2.1.
Context. Each data point has a context , which is a set of indices of other data points. The efemb models the conditional distribution of given the data points in its context.
The context is a modeling choice; different applications will require different types of context. In language, the data point is a word and the context is the set of words in a window around it. In neural data, the data point is the activity of a neuron at a time point and the context is the activity of its surrounding neurons at the same time point. (It can also include neurons at future time or in the past.) In shopping data, the data point is a purchase and the context is the other items in the cart.
Conditional exponential family. An efemb models each data point conditional on its context . The distribution is an appropriate exponential family,
(1) 
where is the natural parameter and is the sufficient statistic. In language modeling, this family is usually a categorical distribution. Below, we will study Gaussian and Poisson.
We parameterize the conditional with two types of vectors, embeddings and context vectors. The embedding of the th data point helps govern its distribution; we denote it . The context vector of the th data point helps govern the distribution of data for which appears in their context; we denote it .
How to define the natural parameter as a function of these vectors is a modeling choice. It captures how the context interacts with an embedding to determine the conditional distribution of a data point. Here we focus on the linear embedding, where the natural parameter is a function of a linear combination of the latent vectors,
(2) 
Following the nomenclature of GLM, we call the link function. We will see several examples of link functions in 2.1.
This is the setting of many existing word embedding models, though not all. Other models, such as the skipgram, determine the probability through a “reverse” distribution of context words given the data point. These nonlinear embeddings are still instances of an efemb.
Embedding structure. The goal of an efemb is to find embeddings and context vectors that describe features of the data. The embedding structure determines how an efemb shares these vectors across the data. It is through sharing the vectors that we learn an embedding for the object of primary interest, such as a vocabulary term, a neuron, or a supermarket product. In language the same parameters and are shared across all positions . In neural data, observations share parameters when they describe the same neuron. Recall that the index connects to both a neuron and time point . We share parameters with and to find embeddings and context vectors that describe the neurons. Other variants might tie the embedding and context vectors to find a single set of latent variables, .
The objective function. The efemb objective sums the log conditional probabilities of each data point, adding regularizers for the embeddings and context vectors.^{1}^{1}1
One might be tempted to see this as a probabilistic model that is conditionally specified. However, in general it does not have a consistent joint distribution
(Arnold et al., 2001). We use log probability functions as regularizers, e.g., a Gaussian probability leads to regularization. We also use regularizers to constrain the embeddings,e.g., to be nonnegative. Thus, the objective is(3) 
We maximize this objective with respect to the embeddings and context vectors. In 2.2 we explain how to fit it with stochastic gradients.
3 can be seen as a likelihood function for a bank of GLM (McCullagh and Nelder, 1989). Each data point is modeled as a response conditional on its “covariates,” which combine the context vectors and context, e.g., as in 2; the coefficient for each response is the embedding itself. We use properties of exponential families and results around GLM to derive efficient algorithms for efemb models.
We highlight the versatility of efemb models with three example models and their variations. We develop the gemb for analyzing real observations from a neuroscience application; we also introduce a nonnegative version, the ngemb. We develop two Poisson embedding models, pemb and apemb, for analyzing count data; these have different link functions. We present a categorical embedding model that corresponds to the CBOW word embedding (Mikolov et al., 2013a). Finally, we present a bemb for binary data. In 2.2 we explain how negative sampling (Mikolov et al., 2013b) corresponds to biased stochastic gradients of the bemb objective. For convenience, these acronyms are in 1.
efemb  exponential family embedding 

gemb  Gaussian embedding 
ngemb  nonnegative Gaussian embedding 
pemb  Poisson embedding 
apemb  additive Poisson embedding 
bemb  Bernoulli embedding 
Example 1: Neural data and Gaussian observations. Consider the (calcium) expression of a large population of zebrafish neurons (Ahrens et al., 2013). The data are processed to extract the locations of the neurons and the neural activity across location and time . The goal is to model the similarity between neurons in terms of their behavior, to embed each neuron in a latent space such that neurons with similar behavior are close to each other.
We consider two neurons similar if they behave similarly in the context of the activity pattern of their surrounding neurons. Thus we define the context for data index
to be the indices of the activity of nearby neurons at the same time. We find the KNN of each neuron (using a Balltree algorithm) according to their spatial distance in the brain. We use this set to construct the context
. This context varies with each neuron, but is constant over time.With the context defined, each data point is modeled with a conditional Gaussian. The conditional mean is the inner product from 2, where the context is the simultaneous activity of the nearest neurons and the link function is the identity. The conditionals of two observations share parameters if they correspond to the same neuron. The embedding structure is thus and for all . Similar to word embeddings, each neuron has two distinct latent vectors: the neuron embedding and the context vector .
These ingredients, along with a regularizer, combine to form a neural embedding objective. gemb uses regularization (i.e., a Gaussian prior); ngemb constrains the vectors to be nonnegative ( regularization on the logarithm. i.e., a lognormal prior).
Example 2: Shopping data and Poisson observations. We also study data about people shopping. The data contains the individual purchases of anonymous users in chain grocery and drug stores. There are different items and trips to the stores among all households. The data is a sparse matrix of purchase counts. The entry indicates the number of units of item that was purchased on trip . Our goal is to learn a latent representation for each product that captures the similarity between them.
We consider items to be similar if they tend to be purchased in with similar groups of other items. The context for observation is thus the other items in the shopping basket on the same trip. For the purchase count at index , the context is .
We use conditional Poisson distributions to model the count data. The sufficient statistic of the Poisson is
, and its natural parameter is the logarithm of the rate (i.e., the mean). We set the natural parameter as in 2, with the link function defined below. The embedding structure is the same as in gemb, producing embeddings for the items.We explore two choices for the link function. pemb uses an identity link function. Since the conditional mean is the exponentiated natural parameter, this implies that the context items contribute multiplicatively to the mean. (We use regularization on the embeddings.) Alternatively, we can constrain the parameters to be nonnegative and set the link function . This is apemb, a model with an additive mean parameterization. (We use regularization in logspace.) apemb only captures positive correlations between items.
Example 3: Text modeling and categorical observations. efemb are inspired by word embeddings, such as CBOW (Mikolov et al., 2013a). CBOW is a special case of an efemb; it is equivalent to a multivariate efemb with categorical conditionals. In the notation here, each is an indicator vector of the th word. Its dimension is the vocabulary size. The context of the th word are the other words in a window around it (of size ), .
The distribution of is categorical, conditioned on the surrounding words ; this is a softmax regression. It has natural parameter as in 2 with an identity link function. The embedding structure imposes that parameters are shared across all observed words. The embeddings are shared globally (, ). The word and context embedding of the word is the row of and respectively. CBOW does not use any regularizer.
Example 4: Text modeling and binary observations. One way to simplify the CBOW objective is with a model of each entry of the indicator vectors. The data are binary and indexed by , where is the position in the text and indexes the vocabulary; the variable is the indicator that word is equal to term . (This model relaxes the constraint that for any only one will be on.) With this notation, the context is ; the embedding structure is and .
We can consider different conditional distributions in this setting. As one example, set the conditional distribution to be a Bernoulli with an identity link; we call this the bemb model for text. In 2.2 we show that biased stochastic gradients of the bemb objective recovers negative sampling (Mikolov et al., 2013b). As another example, set the conditional distribution to Poisson with link . The corresponding embedding model relates closely to Poisson approximations of distributed multinomial regression (Taddy et al., 2015).
We fit the embeddings and context vectors by maximizing the objective function in 3. We use SGD with Adagrad (Duchi et al., 2011). We can derive the analytic gradient of the objective function using properties of the exponential family (see the Supplement for details). The gradients linearly combine the data in summations we can approximate using subsampled minibatches of data. This reduces the computational cost.
When the data is sparse, we can split the gradient into the summation of two terms: one term corresponding to all data entries for which , and one term corresponding to those data entries
. We compute the first term of the gradient exactly—when the data is sparse there are not many summations to make—and we estimate the second term by subsampling the zero entries. Compared to computing the full gradient, this reduces the complexity when most of the entries
are zero. But it retains the strong information about the gradient that comes from the nonzero entries.This relates to negative sampling, which is used to approximate the skipgram objective (Mikolov et al., 2013b)
. Negative sampling redefines the skipgram objective to distinguish target (observed) words from randomly drawn words, using logistic regression. The gradient of the stochastic objective is identical to a noisy but biased estimate of the gradient for a bemb model. To obtain the equivalence, preserve the terms for the nonzero data and subsample terms for the zero data. While an unbiased stochastic gradient would rescale the subsampled terms, negative sampling does not. Thus, negative sampling corresponds to a biased estimate, which downweights the contribution of the zeros. See the Supplement for the mathematical details.
We study efemb models on realvalued and countvalued data, and in different application domains—computational neuroscience, shopping behavior, and movie ratings. We present quantitative comparisons to other dimension reduction methods and illustrate how we can glean qualitative insights from the fitted embeddings.
Data. We analyze the neural activity of a larval zebrafish, recorded at single cell resolution for time frames (Ahrens et al., 2013). Through genetic modification, individual neurons express a calcium indicator when they fire. The resulting calcium imaging data is preprocessed by a nonnegative matrix factorization to identify neurons, their locations, and the fluorescence activity of the individual neurons over time (Friedrich et al., 2015). Using this method, our data contains 10,000 neurons (out of a total of 200,000).
We fit all models on the lagged data to filter out correlations based on calcium decay and preprocessing.^{2}^{2}2We also analyzed unlagged data but all methods resulted in better reconstruction on the lagged data. The calcium levels can be measured with great spatial resolution but the temporal resolution is poor; the neuronal firing rate is much higher than the sampling rate. Hence we ignore all “temporal structure” in the data and model the simultaneous activity of the neurons. We use the gemb and ngemb from 2.1 to model the lagged activity of the neurons conditional on the lags of surrounding neurons. We study context sizes and latent dimension .
Models. We compare efemb to probabilistic FA, fitting dimensional factors for each neuron and
dimensional factor loadings for each time frame. In FA, each entry of the data matrix is Gaussian distributed, with mean equal to the inner product of the corresponding factor and factor loading.
Evaluation. We train each model on a random sample of of the lagged time frames and hold out each for validation and testing. With the test set, we use two types of evaluation. (1) Leave one out: For each neuron in the test set, we use the measurements of the other neurons to form predictions. For FA this means the other neurons are used to recover the factor loadings; for efemb this means the other neurons are used to construct the context. (2) Leave out: We randomly split the neurons into 4 folds. Each neuron is predicted using the three sets of neurons that are out of its fold. (This is a more difficult task.) Note in efemb, the missing data might change the size of the context of some neurons. See 5 in C
for the choice of hyperparameters.
single neuron held out  of neurons held out  

Model  
FA  
gemb (c=10)  
gemb (c=50)  
ngemb (c=10) 
Analysis of neural data: mean squared error and standard errors of neural activity (on the test set) for different models. Both efemb models significantly outperform FA; gemb is more accurate than ngemb.
Results. 2 reports both types of evaluation. The efemb models significantly outperform FA in terms of mean squared error on the test set. gemb obtains the best results with components and a context size of . 1 illustrates how to use the learned embeddings to hypothesize connections between nearby neurons.
We study the Poisson models pemb and apemb on two applications: shopping and movies.


Market basket data. We analyze the IRI dataset^{3}^{3}3We thank IRI for making the data available. All estimates and analysis in this paper, based on data provided by IRI, are by the authors and not by IRI. (Bronnenberg et al., 2008), which contains the purchases of anonymous households in chain grocery and drug stores. It contains trips in 2012. We remove items that appear fewer than 10 times, leaving a dataset with items. The context for each purchase is the other purchases from the same trip.
MovieLens data.
We also analyze the MovieLens
dataset (Harper and Konstan, 2015), which contains movie ratings on a scale from to . We keep only positive ratings, defined to be ratings of 3 or more (we subtract 2 from all ratings and set the negative ones to 0). The context of each rating is the other movies rated by the same user. After removing users who rated fewer than movies and movies that were rated fewer than times, the dataset contains users and movies; the sparsity is about .Models. We fit the pemb and the apemb models using number of components . For each we select the Adagrad constant based on best predictive performance on the validation set. (The parameters we used are in 5.) In these datasets, the distribution of the context size is heavy tailed. To handle larger context sizes we pick a link function for the efemb model which rescales the sum over the context in 2 by the context size (the number of terms in the sum). We also fit a pemb model that artificially downweights the contribution of the zeros in the objective function by a factor of , as done by Hu et al. (2008) for matrix factorization. We denote it as “pemb (dw).”
We compare the predictive performance with HPF (Gopalan et al., 2015) and Poisson PCA (Collins et al., 2001). Both HPF and Poisson PCA factorize the data into dimensional positive vectors of user preferences, and dimensional positive vectors of item attributes. apemb and HPF parameterize the mean additively; pemb and Poisson PCA parameterize it multiplicatively. For the efemb models and Poisson PCA, we use stochastic optimization with regularization. For HPF, we use variational inference. See 5 in C for details.
Evaluation. For the market basket data we hold out of the trips to form the test set, also removing trips with fewer than two purchased different items. In the MovieLens data we hold out of the ratings and set aside an additional of the nonzero entries from the test for validation. We report prediction performance based on the normalized loglikelihood on the test set. For pemb and apemb, we compute the likelihood as the Poisson mean of each nonnegative count (be it a purchase quantity or a movie rating) divided by the sum of the Poisson means for all items, given the context. To evaluate HPF and Poisson PCA at a given test observation we recover the factor loadings using the other test entries we condition on, and we use the factor loading to form the prediction.
Predictive performance. 3 summarizes the test loglikelihood of the four models, together with the standard errors across entries in the test set. In both applications the pemb model outperforms HPF and Poisson PCA. On shopping data pemb with provides the best predictions; on MovieLens pemb with is best. For pemb on shopping data, downweighting the contribution of the zeros gives more accurate estimates.
Item similarity in the shopping data. Embedding models can capture qualitative aspects of the data as well. 4 shows four example products and their three most similar items, where similarity is calculated as the cosine distance between embedding vectors. (These vectors are from pemb with downweighted zeros and .) For example, the most similar items to a soda are other sodas; the most similar items to a yogurt are (mostly) other yogurts.
Maruchan chicken ramen  Yoplait strawberry yogurt  Mountain Dew soda  Dean Foods 1 % milk 

M. creamy chicken ramen  Yoplait apricot mango yogurt  Mtn. Dew orange soda  Dean Foods 2 % milk 
M. oriental flavor ramen  Yoplait strawberry orange smoothie  Mtn. Dew lemon lime soda  Dean Foods whole milk 
M. roast chicken ramen  Yoplait strawberry banana yogurt  Pepsi classic soda  Dean Foods chocolate milk 
The pemb model can also identify complementary and substitutable products. To see this, we compute the inner products of the embedding and the context vectors for all item pairs. A high value of the inner product indicates that the probability of purchasing one item is increased if the second item is in the shopping basket (i.e., they are complements). A low value indicates the opposite effect and the items might be substitutes for each other.
We find that items that tend to be purchased together have high value of the inner product (e.g., potato chips and beer, potato chips and frozen pizza, or two different types of soda), while items that are substitutes have negative value (e.g., two different brands of pasta sauce, similar snacks, or soups from different brands). Other items with negative value of the inner product are not substitutes, but they are rarely purchased together (e.g., toast crunch and laundry detergent, milk and a toothbrush). D gives examples of substitutes and complements.
Topics in the movie embeddings. The embeddings from MovieLens data identify thematically similar movies. For each latent dimension , we sort the context vectors by the magnitude of the th component. This yields a ranking of movies for each component. In E we show two example rankings. (These are from a pemb model with .) The first one contains children’s movies; the second contains sciencefiction/action movies.
We described efemb, conditionally specified latent variable models to extract distributed representations from high dimensional data. We showed that CBOW (Mikolov et al., 2013b) is a special case of efemb and we provided examples beyond text: the brain activity of zebrafish, shopping data, and movie ratings. We fit the efemb objective usng stochastic gradients. Our empirical study demonstrates that an efemb can better reconstruct data than existing dimensionalityreduction techniques based on matrix factorization. Further, the learned embeddings capture interesting semantic structure.
This work is supported by the EU H2020 programme (Marie SkłodowskaCurie grant agreement 706760), NFS IIS1247664, ONR N000141110651, DARPA FA87501420009, DARPA N6600115C4032, Adobe, the John Templeton Foundation, and the Sloan Foundation.
Innovations in Machine Learning
, pages 137–186. Springer.A generalization of principal components analysis to the exponential family.
In Neural Information Processing Systems, pages 617–624.Uncertainty in Artificial Intelligence
.Conference on Empirical Methods on Natural Language Processing
, volume 14, pages 1532–1543.We fit the embeddings and context vectors by maximizing the objective function in 3. We use SGD.
We first calculate the gradient, using the identity for exponential family distributions that the derivative of the lognormalizer is equal to the expectation of the sufficient statistics, i.e., . With this result, the gradient with respect to the embedding is
(4) 
The gradient with respect to has the same form. In B, we detail this expression for the particular models that we study empirically (3).
The gradient in 4 can involve a sum of many terms and be computationally expensive to compute. To alleviate this, we follow noisy gradients using SGD. We form a subsample of the terms in the summation, i.e.,
(5) 
where
denotes the size of the subsample and where we scaled the summation to ensure an unbiased estimator of the gradient.
5 reduces computational complexity when is much smaller than the total number of terms. At each iteration of SGD we compute noisy gradients with respect to and (for each ) and take gradient steps according to a stepsize schedule. We use Adagrad (Duchi et al., 2011) to set the stepsize.Relation to negative sampling.
In language, particularly when seen as a collection of binary variables, the data are sparse: each word is one of a large vocabulary. When modeling sparse data, we split the sum in
4 into two contributions: those where and those where . The gradient is(6)  
We compute the first term of the gradient exactly—when the data is sparse there are not many summations to make—and we estimate the second term with subsampling. Compared to computing the full gradient, this reduces the complexity when most of the entries are zero. But, it retains the strong information about the gradient that comes from the nonzero entries.
This relates to negative sampling, which is used to approximate the skipgram objective (Mikolov et al., 2013b). Negative sampling redefines the skipgram objective to distinguish target (observed) words from randomly drawn words, using logistic regression. The gradient of the stochastic objective is identical to a noisy but biased estimate of the gradient in 6 for a bemb model. To obtain the equivalence, preserve the terms for the nonzero data and subsample terms for the zero data. While an unbiased stochastic gradient would rescale the subsampled terms, negative sampling does not. It is thus a biased estimate, which downweights the contribution of the zeros.
To specify the gradients in Equation 4 for the SGD procedure we need the sufficient statistic , the expected sufficient statistic , the gradient of the natural parameter with respect to the embedding vectors and the gradient of the regularizer on the embedding vectors. In this appendix we specify these quantities for the models we study empirically in Section 3.
Using the notation and reflecting the embedding structure , , the gradients with respect to each embedding and each context vector becomes
(7)  
(8) 
By restricting the parameters to be nonnegative we can learn nonnegative synaptic weights between neurons. For notational simplicity we write the parameters as and and update them in logspace. The operator stands for element wise multiplication. With this notation, the gradient for the ngemb can be easily obtained from Equations 7 and 8
by applying the chain rule.
(9)  
(10)  
We proceed similarly as for the gemb model.
(11)  
(12) 
Here, we proceed in a similar manner as for the ngemb model.
(13)  
(14) 
Model 






neuro  gemb  n/a  
neuro  ngemb  n/a  
shopping  all models  n/a  
movies  all models  n/a 
Table 6 shows some pairs of items with high inner product of embedding vectors and context vector. The items in the first column have higher probability of being purchased if the item in the second column is in the shopping basket. We can observe that they correspond to items that are frequently purchased together (potato chips and beer, potato chips and frozen pizza, two different sodas).
Similarly, Table 7 shows some pairs of items with low inner product. The items in the first column have lower probability of being purchased if the item in the second column is in the shopping basket. We can observe that they correspond to items that are rarely purchased together (detergent and toast crunch, milk and toothbrush), or that are substitutes of each other (two different brands of snacks, soup, or pasta sauce).
Inner product  Item 1  Item 2 

Diet 7 Up lemon lime soda  Diet Squirt citrus soda  
Old Dutch original potato chips  Budweiser Select 55 Lager beer  
Lays potato chips  DiGiorno frozen pizza  
Coca Cola zero soda 
Coca Cola soda  
Soyfield vanilla organic yogurt  La Yogurt low fat mango 
Inner product  Item 1  Item 2 

General Mills cinnamon toast crunch  Tide Plus liquid laundry detergent  
Doritos chilli pepper  Utz cheese balls  
Land O Lakes 2% milk  Toothbrush soft adult (private brand)  
Beef Swanson Broth soup 48oz  Campbell Soup cans 10.75oz  
Ragu Robusto sautéed onion & garlic pasta sauce  Prego tomato Italian pasta sauce 
LABEL:tab:kids_movies,tab:scifi_movies show clusters of ranked movies that are learned by our pemb model. These rankings were generated as follows. For each latent dimension we sorted the context vectors according their value in this dimension. This gives us a ranking of context vectors for every . LABEL:tab:kids_movies,tab:scifi_movies show the 10 top items of the ranking for two different values of . Similar as in topic modeling, the latent dimensions have the interpretation of topics. We see that sorting the context vectors this way reveals thematic structure in the collection of movies. While Table 8 gives a table of movies for children, Table 9
shows a cluster of sciencefiction and action movies (with a few outliers).
#  Movie Name  Year  Rank 

1  Winnie the Pooh and the Blustery Day  1968  
2  Cinderella  1950  
3  Toy Story  1995  
4  Fantasia  1940  
5  Dumbo  1941  
6  The Nightmare Before Christmas  1993  
7  Snow White and the Seven Dwarfs  1937  
8  Alice in Wonderland  1951  
9  James and the Giant Peach  1996 
#  Movie Name  Year  Rank 

1  Die Hard: With a Vengeance  1995  
2  Stargate  1994  
3  Star Trek IV: The Voyage Home  1986  
4  Manon of the Spring (Manon des sources)  1986  
5  Fifth Element, The  1997  
6  Star Trek VI: The Undiscovered Country  1991  
7  Under Siege  1992  
8  GoldenEye  1995  
9  Supercop  1992 
Comments
There are no comments yet.