Ask Minmax!

Ask Minmax is an expert system for optimization problems targeted towards the non-expert user.

How it works

• Ask-minmax associates a prior to every problem and question in the database.
• The prior of a problem is proportional to the number of times there is a query intended for that particular problem.
• At the beginning of every iteration, it asks the user for a human readable summary.
• Document vectors have been trained on abstracts downloaded from arxiv and google-scholar.
• A softmax classifier is built on top of these document vectors to output a probability distribution on the problems.
• The prior thus reflects this probability distribution together with the frequency.
• The prior of a question reflects the gain in information by asking the question - in other words it is inversely proportional to the expected conditional entropy of the distribution of problem posteriors conditioned on the response to this question.
• At every step a question is sampled proportional to it's posterior.
• The posteriors of a problem are updated according to the confidence level in your answer.
• The posteriors of a question are updated reflecting the information gain provided for this new distribution.
• The algorithm outputs the most popular problems by doing a 1 dimensional k-means (a.k.a Jenks natural breaks ).
• A similar (clustering) idea is applied to questions, at every step only the "most useful questions" are sampled from.
• The algorithm is greedy in the sense that it only cares about questions that reduce the entropy on the posterior distribution supported on the "most relevant" (coming from Jenks/k-means) problems.
• One can visualize the changing distribution in a simple matplotlib plot.

Prerequisites:

• Python 3
• A running Mongodb server
• pymongo
• jenks
• scipy
• numpy
• matplotlib
• nltk
• scikit-learn 0.17
• gensim

Quickstart:

Clone the repository and use

sudo python3 setup.py install

To create an instance of the Expert system first import the Expert class and then use the basic constructor

from askminmax.expert import Expert
expert = Expert()

It will ask you to either import stuff from an existing database or it will give you the option to create one of your own. Finally to run the expert system use

expert.run(data_set, doc2vec_model_path, classifier_path)

where data_set is the path to the data set (for e.g. dataset), doc2vec_model_path is the path to the Doc2Vec model (for e.g. models/model_2015-11-25_08-44-27.d2v), and classifier_path is the path to the softmax classifier (for e.g. models/model_2015-11-28_21-50-40.log).

Doc2Vec model examples

You can train your own doc2vec model on your domain specific dataset. In this case we train it on abstracts downloaded from arxiv and google-scholar. Here are some interesting examples in IPython:

In [1]: from gensim.models import Doc2Vec
In [2]: model = Doc2Vec.load('model/[Insert model name]')
In [3]: model.doesnt_match(['bipartite', 'non-bipartite', 'stable', 'matching', 'scheduling'])
Out[3]: 'scheduling'
In [4]: model.doesnt_match(['facility location', 'minimum cut', 'maximum cut', 'sparsest cut'])
Out[4]: 'facility location'
In [5]: model.doesnt_match(['scheduling', 'routing', 'maximum', 'TSP', 'facility location'])
Out[5]: 'maximum'