ACM Recsys Challange hosted by Xing.

A month or so ago Xing hosted the ACM RecSys Challenge Workshop in Boston, concluding a great challenge on predicting Job Postings for our jobs market. A selected group of participants including the top 5 winners presented their solution in the workshop. In the following I will present some of my impressions on the challenge and the workshop.

The winners and organisers of the challenge

The Challenge

The details of the challenge can be found in our paper [1], however I will still describe some of the task for the convenience of the reader.  

As some of you know I currently work as a data scientist Xing is a social network in Hamburg for professionals. The network also includes a job market where users can browse one million job offers
from other companies. For the challenge we created a semi- synthetic dataset, handing out anonymised user profiles, job postings as well as interaction data.

The task of the challenge was to predict a users future interactions using the provided dataset.
In order to anonymise the dataset, we created synthetic (random) users and their interactions.

From Xing's perspective the challenge was a huge success with the top solutions providing new insight and solutions into our data. In my mind the "coolest" approach to the challenge was that multiple teams used gradient boosting of trees (xgboost [2]) to predict the users clicks. In both cases
the algorithm was fed user item pairs and their features from the interaction data and then the tree
learned to predict users clicks.  Another team with a high rank used the temporal data to train a recurrent neural network (LSTM).

Happily we are also confirmed to host the next years challenge. For 2017 we plan to release the dataset periodically and evaluate high scoring solutions in our live recommendation system.

All information about the 2016 challenge and the upcoming 2017 challenge can be found at [3] 

References

[1] Abel, Benczur, Kohlsdorf, Larson, Palovics: "RecSys Challenge 2016: Job Recommendations" http://dl.acm.org/citation.cfm?id=2959207

[2] https://xgboost.readthedocs.io/en/latest/tutorials/index.html
[3] http://2016.recsyschallenge.com/

POS Tagging with Word Embeddings

I found an interesting paper on Part-Of-Speech (POS) tagging with word embedding. Through my current work at Xing and my previous work on hidden Markov Models, I find this new model very interesting for several reasons. As mentioned in previous posts, word embeddings such as word2vec map all words in a dictionary into a d-dimensional, real valued vector space. In this new space, the similarity of two words decreases with distance of their respective word vectors. The idea is now to
see a document as a d- dimensional time series and use a hidden Markov model with Gaussian observations to model this real valued sequence. Now each state of the hidden Markov model
can be regarded as a POS category. Now this model can be trained using the Baum Welch algorithm and decoding can be performed using the Viterbi path. In my opinion this model is superior to classic hidden Markov models with multinomial observations and word embeddings alone in several ways.
First, the model's observation space has way fewer dimensions. While the multinomial observation for text can have thousands of parameters to estimate (one dimension / word), the word embeddings
only need several hundred dimensions and the semantic representation of each state can be captured more efficiently. Second, a recent article on kaggle suggested summing up all word vectors in a document as a representation. However, the average word vector might not be a sufficient representation, since a lot of the fine differences in a document are lost. In the hidden Markov model,
we still have to average vectors but the average is taken per state. In this way, the model introduces structure. In the POS paper, the temporal model with word embeddings outperformed the other multinomial approach by a large margin. It would be interesting to see such a model used in different NLP tasks.

Speed Bake-Off Tensorflow / (Numpy | SciPy | Matplotlib) / OpenCV

I recently got interested into scaling some of the algorithms from my PhD thesis. Especially the feature learning part. I used a K-Means to learn a convolutional codebook over audio spectrogram.
Why you can use K-Means to learn something like a convolutional neural net is described in: Coates, Lee and NG as well as my thesis. Anyways in my thesis experiments the bottle neck was the feature learning. So I need a fast clustering implementation as well as a fast convolution in order to resolve the speed issues. The three main implementations I look at are:

• Google Tensor Flow
• SciKit Learn
• OpenCV

All three provide a python interface but use a low level, speedy C or Fortran implementation under the hood. My goal is to evaluate these three libraries for their performance (speed) under the following conditions:

The library is fastest when testing on a single Macbook. It can use the available parallel devices such as a GPU and multi threading. Furthermore, the library should be fastest on medium sized problems.

I choose these conditions since I plan to design programs that animal communication researchers can use in the field. Meaning, if you are to analyse dolphin communication on a boat in the Bahamas or  

you observe Birds in the rainforest, the only thing you might have is your laptop with no internet connection.

Convolution

The convolution test is based on the three functions:

• Tensorflow: tensorflow.nn.conv2d
• SciPy: scipy.signal.convolve2d
• OpenCV: cv2.filter2d

My test is to compute the edge image using a sobel filter in x-direction and on in y-direction:

I am using the following image and expect the following edge image:

The results in seconds:

• Tensorflow: 0.125 [s]
• SKLearn:     0.049 [s]
• OpenCV:     0.019 [s] 

Here it looks like the OpenCV implementation is the fastest. So openCV it is. For the clustering there are a lot of libraries and even the sklearn implementation is very fast.

Update:

After Himanshu's comment I chose to check Tensorflow not including the variable assignment
and then it took 0.021 seconds. However, the image copying into the variable and the session setup
matter in my use case, too. And it is still lower than Open CV. It is also interesting that for these problems, the speed between the libraries is not that different. However, I belief that for larger problems, the tensor flow version that can run on a cluster will show way better performance. Also I don't know right now if the current tensor flow version works with opencl.

Word Embeddings

Since my grad school interest was focused on machine learning for perception,  I did not notice a class of methods called word embeddings. However, recently I got interested more into text mining so I started to read up on these method and implement some.
A word embedding maps words into a multi dimensional euclidean space in which semantically similar words are close. In other words, each word in your dictionary is represented by a multi dimensional vector.

A word embedding can capture many semantics. For example, on the word2vec webpage,
an embedding from google, it is noticed that:

1. "vector('king') - vector('man') + vector('woman') is close to vector('queen')"
2. "vector('Paris') - vector('France') + vector('Italy') [...] is very close to vector('Rome')"

In other words, the representation is capturing concepts such as gender and captial city.
The two most prominent methods so far seem to be word2vec (by google), glove (by stanford's nlp group). Furthermore, there is a very recent combination of the two called swivel (again by google).

All the methods are based on the idea that the usage of a word gives insight into the words meaning or that similar words are used in a similar context. Here context can be defined as a small neighbourhood. For example, a context definition could be defined as the words to the left of the target word and the three words to the right.

Google's Word2Vec obtains the vectors by using a simple neural net that predicts the target word from it's context (Continuous Bag Of Words) or vice versa (Skip Gram). As usual the neural net can be trained using stochastic gradient descent (back propagation). The neural net can be seen as learning a representation for words (word vectors) and a representation for contexts (context vectors).
Glove sloves a similar problem. However, instead of predicting the actual context around a word, glove learns vectors predictive of the a global coocurance matrix, extracted from the complete corups. Glove is trained using adagrad. In general we solve an optimisation problem of the form:

Basically the two methods differ in the way f(.) and C are defined. C is the target function and we aim to minimise the difference between the prediction and the target, scaled by a function of the co-occurrence count f(.). For word2vec the target function is the pointwise mutual information between two words and for glove it is the log  co-occurrence count. 

Both method come with several implementations already. Semantic similarity can be used for several 

NLP tasks. For example, sentiment analysis or query expansion. 

Thesis Published: DATA MINING IN LARGE AUDIO COLLECTIONS OF DOLPHIN SIGNALS

The study of dolphin cognition involves intensive research of animal vocalisations recorded in the field. In this dissertation I address the automated analysis of audible dolphin communication. I propose a system called the signal imager that automatically discovers patterns in dolphin signals. These patterns are invariant to frequency shifts and time warping transformations. The discovery algorithm is based on feature learning and unsupervised time series segmentation using hidden Markov models. Researchers can inspect the patterns visually and interactively run comparative statistics between the distribution of dolphin signals in different behavioral contexts. The required statistics for the comparison describe dolphin communication as a combination of the following models: a bag-of-words model, an n-gram model and an algorithm to learn a set of regular expressions. Furthermore, the system can use the patterns to automatically tag dolphin signals with behavior annotations. My results indicate that the signal imager provides meaningful patterns to the marine biologist and that the comparative statistics are aligned with the biologists’ domain knowledge.

[PDF]

Random Intersection Trees

Found an interesting paper with a new idea on how to build interpretable decision rules. The paper
describes an algorithm to find a set of binary features that are together indicative of a class. For example, a set of binary features might be:

• Humidity = high: true / false
• Temperature > 10 C: true / false
• Sunny = true / false
• Rain = true false

A feature is said to be active whenever the result evaluates to true. For two classes C1 and C2, the goal is to find a subset of active features S  such that P(S|C1) > O2 and P(S|C2) < O1 for two thresholds 0 < O1 < O2 < 1. 

For example, let the class be good weather and bad weather. Furthermore, I specified the thresholds

as O1 = 0.5 and O2 = 0.25. A good result might then be 

• P(S = {temperature > 10, sunny} | good weather)  > .5
• P(S = {temperature > 10, sunny} | bad weather)    < .25

Such a result would be indeed interpretable and we could agree with the results that a higher temperature and sun would result in higher probabilities for good weather.

The algorithm works in two steps: 

1. Candidate generation for subsets S.
2. Check subsets for constraint: P(S|C1) > O2 and P(S|C2) < O1 

The main algorithm in the paper is an efficient candidate generation algorithm.

The main idea is to organise the subset search into a random tree of fixed depth and breadth.

At every node in the tree we pick a random example from the dataset with a positive label and then

compute the intersection of that example with the active feature set of the parent node.

By intersecting with positive examples only, the set of features gets more and more refined at each level. All remaining sets at the specified depth are candidates. Afterwards, we only have to check gif the subset holds to the constraints. 

A very interesting method to find rules in a dataset which is an interesting alternative to decision tree learning and apriori rule discovery.

Social Network Sampling

I will start working for a social network in August. My new job is data scientist at the business network Xing. I decided to play with their open API a little. The Xing API gives access to users in the network as well as their contact list. So an external app can analyze the social graph. Well actually there are restrictions on the number of queries so it is possible to
analyze the graph partly. My goal is to extract an unbiased sample from the complete graph by performing a random walk from a start node. That means at every node the algorithm expands all
the node's successors and then picks a random one. The chosen successor is the next node. With
a probability of 0.15 the algorithm returns to the start node. This will ensure that we explore the graph around the start node. The going back is ensuring a trade off between breadth of the exploration and the depth of the exploration. I use the ruby version of the API.
A code snippet for a random walk step in the Xing API
is shown below:

##
# Random walk step: 
#   1) With probability 0.15 go back to start
#   2) Randomly choose a node from the successors
def random_walk_step(user_id, start)
  reset = false
  if Random.rand < 0.15
    user_id = start    
    reset = true
  end
  num_contacts = 5 # control branching
  contacts     = XingApi::Contact.list(user_id, limit: num_contacts)[:contacts][:users]
  next_user    = Random.rand(num_contacts)    
  contacts.map! {|x| x[:id]}
  if reset
    return random_walk_step(contacts[next_user], start)
  end
  return [contacts[next_user], contacts, user_id]
end

The random step function takes a user id as the current node and a node we started the random walk at. First we decide if we restart the search at the start node, then it samples a neighbor. Repeating the process and updating the user id with the sampled neighbor results in a graph. The function below, explores the graph for 25 nodes and returns the graph as a dot file:

##
# Print graph as graphviz file
def dot(path)
  expended = []
  File.open(path, 'w') do |file|
    file.write "digraph my_neighbors {\n"
    start_id = XingApi::User.me[:users].first[:id]
    ids = random_walk_step(start_id, start_id)  
    45.times do
      id = ids[2]
      username = XingApi::User.find(id)[:users].first[:display_name]
      username.gsub! /\W/, ""      
      if not expended.include? username
        expended = expended + [username]
        ids[1].each do |neighbor| 
          neighborname = XingApi::User.find(neighbor)[:users].first[:display_name]
          neighborname.gsub! /\W/, ""       
          file.write "#{username} -> #{neighborname};\n"        
        end
      end
      ids = random_walk_step(ids[0], start_id)      
    end
    file.write "}"
  end
end

A dot file is a graph description that can be converted into graph images. A sample graph of my network is shown below.