ML Service Interactions

ML Backend	Type	Data Input Format	Model Creation	Query Format
Azure ML	service	http://channel9.msdn.com/Blogs/Windows-Azure/Getting-and-Saving-Data-in-Azure-ML-Studio local upload or from any source with a URL CSV, TSV ARFF OData Hive SQL Azure DB Azure storage table data It's looking for a single table / set of columns for training appears to require user interaction	http://channel9.msdn.com/Blogs/Windows-Azure/Predictive-Modelling-with-Azure-ML-studio Requires user interaction Requests exist to automate both model creation/training and retraining: http://feedback.azure.com/forums/257792-machine-learning/suggestions/6179666-enable-re-training-of-the-model-via-webservice-a http://feedback.azure.com/forums/257792-machine-learning/suggestions/6179714-expose-api-for-model-building-publishing-automat	http://channel9.msdn.com/Blogs/Windows-Azure/Data-Science-Series-IPython-and-Azure-ML Send feature set as JSON two REST services: synchronous single query, async batch query Needs JSON feature set and api key simple result comes back (as JSON?)
Big ML	service	https://bigml.com/developers/sources CSV ARFF <= 64GB, can be gzipped (4TB if the file is stored on Amazon S3) can create sources from remote locations: supported protocols include azure, https, odata, s3, and dropbox Creating a source is a service, as is updating a source, so this can be easily automated.	https://bigml.com/developers/models Service call: send dataset's id creation is async: you must query to determine if it's finished Can also use a service call to update the model, but that only updates the model's metadata, it does not actually retrain it you can easily create ensemble models, which is pretty cool	https://bigml.com/developers/predictions async batch predictions are also available send model id and input data as JSON, plus username and api key get back prediction plus a bunch of metadata, including a confidence score, as JSON can statically refer to this prediction later by a unique ID
Google Predictions	service	Upload a CSV to Google Cloud Storage Can call a service endpoint to update the data set of a model (I don't think this auto-retrains the model, though)	https://developers.google.com/prediction/docs/reference/v1.6/trainedmodels/insert send project id, id for the model being created, and Google Cloud Storage location of the training data training is async, you must query to find out what the status is with the get service endpoint Can retrain an existing model: https://developers.google.com/prediction/docs/reference/v1.6/trainedmodels/update	https://developers.google.com/prediction/docs/reference/v1.6/trainedmodels/predict you can batch requests, but that does not seem to be an async operation send model and project ids, along with a list of input features (JSON structure mostly consisting of a JSON list) get back JSON with output value. If task was classification, also get back a list of all class labels with their associated probabilities. also gets back a url to re-request the same prediction again