Amazon SageMaker factorisation machine rating matrix and endpoint - amazon-web-services

I am building a recommender system using sagemaker's built-in factorisation machine model.
My desired result is to have a rating matrix where I can look up a predicted score by a user id and an item id.
I understand that there is a predict API provided by the model:
result = fm_predictor.predict(X_test[1000:1010].toarray())
But I am not sure how I can use it to achieve the desired purpose. If I want to know, say, if user#123 is interested in movie#456, how can I use the above API?
Reference:
https://medium.com/#julsimon/building-a-movie-recommender-with-factorization-machines-on-amazon-sagemaker-cedbfc8c93d8
https://www.slideshare.net/AmazonWebServices/building-a-recommender-system-on-aws-aws-summit-sydney-2018 (p.41,43)
Updated:
I think I understand how to use the API now, you have to build another one-hot encoded dataset as input, for example:
X_new = lil_matrix((1, nbFeatures)).astype('float32')
X_new[0, 935] = 1
X_new[0, 1600] = 1
prediction2 = X_new[0].toarray()
result2 = fm_predictor.predict(prediction2)
print(result2)
But it seems it would be quite inefficient to fill out the recommendation matrix this way. What would be the best practice?

I think one could think about 2 scenarios:
1) if you need very low latency, you can fill up the matrix indeed, i.e. compute all recos for all users, and store it in a key/value backend queried by your app. You can definitely predict multiple users at a time, using the one-hot encoded technique above.
2) predict on-demand by invoking the endpoint directly from the app. This is quite simpler, at the cost of a little latency.
Hope this helps.

Related

Can I make Amazon SageMaker deliver a recommendation based on historic data instead of a probability score?

We have a huge set of data in CSV format, containing a few numeric elements, like this:
Year,BinaryDigit,NumberToPredict,JustANumber, ...other stuff
1954,1,762,16, ...other stuff
1965,0,142,16, ...other stuff
1977,1,172,16, ...other stuff
The thing here is that there is a strong correlation between the third column and the columns before that. So I have pre-processed the data and it's now available in a format I think is perfect:
1954,1,762
1965,0,142
1977,1,172
What I want is a predicition on the value in the third column, using the first two as input. So in the case above, I want the input 1965,0 to return 142. In real life this file is thousands of rows, but since there's a pattern, I'd like to retrieve the most possible value.
So far I've setup a train job on the CSV file using the Linear Learner algorithm, with the following settings:
label_size = 1
feature_dim = 2
predictor_type = regression
I've also created a model from it, and setup an endpoint. When I invoke it, I get a score in return.
response = runtime.invoke_endpoint(EndpointName=ENDPOINT_NAME,
ContentType='text/csv',
Body=payload)
My goal here is to get the third column prediction instead. How can I achieve that? I have read a lot of the documentation regarding this, but since I'm not very familiar with AWS, I might as well have used the wrong algorithms for what I am trying to do.
(Please feel free to edit this question to better suit AWS terminology)
For csv input, the label should be in the first column, as mentioned here: So you should preprocess your data to put the label (the column you want to predict) on the left.
Next, you need to decide whether this is a regression problem or a classification problem.
If you want to predict a number that's as close as possible to the true number, that's regression. For example, the truth might be 4, and the model might predict 4.15. If you need an integer prediction, you could round the model's output.
If you want the prediction to be one of a few categories, then you have a classification problem. For example, we might encode 'North America' = 0, 'Europe' = 1, 'Africa' = 2, and so on. In this case, a fractional prediction wouldn't make sense.
For regression, use 'predictor_type' = 'regressor' and for classification with more than 2 classes, use 'predictor_type' = 'multiclass_classifier' as documented here.
The output of regression will contain only a 'score' field, which is the model's prediction. The output of multiclass classification will contain a 'predicted_label' field, which is the model's prediction, as well as a 'score' field, which is a vector of probabilities representing the model's confidence. The index with the highest probability will be the one that's predicted as the 'predicted_label'. The output formats are documented here.
predictor_type = regression is not able to return the predicted label, according to
the linear-learner documentation:
For inference, the linear learner algorithm supports the application/json, application/x-recordio-protobuf, and text/csv formats. For binary classification models, it returns both the score and the predicted label. For regression, it returns only the score.
For more information on input and output file formats, see Linear
Learner Response Formats for inference, and the Linear Learner Sample
Notebooks.

Preserve Order for Cross Validation in Weka

I am using the Weka GUI for classifying sensor data.
I have measures of 10 people, the data is sorted. So the first 10% correspond to participant 1, the second 10% to participant 2 etc.
I would like to use 10 fold cross validation to build a model on 9 participants and test it on the remaining participant. In my case I believe I could accomplish this by simply not randomizing the data splits.
How would I best go about doing this?
I don't know how to do this in the Explorer.
In the KnowledgeFlow GUI, there is a CrossValidationFoldMaker used to create cross-validation folds. This has an option to Preserve instances order, which says it preserves the order of instances rather than randomly shuffling.
There's a video describing the KnowledgeFlow interface here:
https://www.youtube.com/watch?v=sHSgoVX9z-8&t=7s

Overfitting with random forest though very successful cross validation results

I have moderate experience with data science. I have a data set with 9500 observations and more than 4500 features most of which are highly correlated. Here is briefly what I have tried: I have dropped columns where there are less than 6000 non-NAs and have imputed NAs with their corresponding columns' median values when there are at least 6000 non-NAs. As for correlation, I have kept only features having at most 0.7 correlation with others. By doing so, I have reduced the number of features to about 750. Then I have used those features in my binary classification task in random forest.
My data set is highly unbalanced where ratio of (0:1) is (10:1). So when I apply RF with 10-fold cv, I observe too good results in each cv (AUC of 99%) which is to good to be true and in my test set I got way worse results such as 0.7. Here is my code:
import h2o
from h2o.estimators import H2ORandomForestEstimator
h2o.init(port=23, nthreads=4)
train = fs_rf[fs_rf['Year'] <= '201705']
test = fs_rf[fs_rf['Year'] > '201705']
train = train.drop('Year',axis=1)
test = test.drop('Year',axis=1)
test.head()
train = h2o.H2OFrame(train)
train['BestWorst2'] = train['BestWorst2'].asfactor()
test = h2o.H2OFrame(test)
test['BestWorst2'] = test['BestWorst2'].asfactor()
training_columns = train.drop('BestWorst2',axis=1).col_names
response_column = 'BestWorst2'
model = H2ORandomForestEstimator(ntrees=100, max_depth=20, nfolds=10, balance_classes=True)
model.train(x=training_columns, y=response_column, training_frame=train)
performance = model.model_performance(test_data=test)
print(performance)
How could I avoid this over-fitting? I have tried many different parameters in grid search but none of them improved the results.
This is not what I would call "overfitting". The reason you are seeing really good cross-validation metrics compared to your test metrics is that you have time-series data and so you can't use k-fold cross-validation to give you an accurate estimate of performance.
Performing k-fold cross-validation on a time-series dataset will give you overly-optimistic performance metrics because you are not respecting the time-series component in your data. Regular k-fold cross-validation will randomly sample from your whole dataset to create a train & validation set. Essentially, your validation strategy is "cheating" because you have "future" data included in your CV training sets (if that makes any sense).
I can see by your code that you understand that you need to train with "past" data and predict on "future" data, but if you want to read more about this topic, I'd recommend this article or this article.
One solution is to simply look at test set performance as way to evaluate your model. Another option is to use what's called "rolling" or "time-series" cross-validation, but H2O does not currently support that (though it seems like it might be added soon). Here's a ticket for this if you want to keep track of the progress.

When training a single batch, is iteration of examples necessary (optimal) in python code?

Say I have one batch that I want to train my model on. Do I simply run tf.Session()'s sess.run(batch) once, or do I have to iterate through all of the batch's examples with a loop in the session? I'm looking for the optimal way to iterate/update the training ops, such as loss. I thought tensorflow would handle it itself, especially in the cases where tf.nn.dynamic_rnn() takes in a batch dimension for listing the examples. I thought, perhaps naively, that a for loop in the python code would be the inefficient method of updating the loss. I am using tf.losses.mean_squared_error(batch) for a regression problem.
My regression problem is given two lists of word vectors (300d each), and determines the similarity between the two lists on a continuous scale from [0, 5]. My supervised model is Deepmind's Differential Neural Computer (DNC). The problem is I do not believe it is learning anything. this is due to the fact that the all of the output from the model is centered around 0 and even negative. I do not know how it could possibly be negative given no negative labels provided. I only call sess.run(loss) for the single batch, I do not create a python loop to iterate through it.
So, what is the most efficient way to iterate the training of a model and how do people go about it? Do they really use python loops to do multiple calls to sess.run(loss) (this was done in the training file example for DNC, and I have seen it in other examples as well). I am certain I get the final loss from the below process, but I am uncertain if the model has actually been trained entirely just because the loss was processed in one go. I also do not understand the point of update_ops returned by some functions, and am uncertain if they are necessary to ensure the model has been trained.
Example of what I mean by processing a batch's loss once:
# assume the model has been defined prior through batch_output_logits
train_loss = tf.losses.mean_squared_error(labels=target,
predictions=batch_output_logits)
with tf.Session() as sess:
sess.run(init_op) # pseudo code, unnecessary for question
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord)
# is this the entire batch's loss && model has been trained for that batch?
loss_np = sess.run(train_step, train_loss)
coord.request_stop()
coord.join(threads)
Any input on why I am receiving negative values when the labels are in the range [0, 5] is welcomed as well(general abstract answers for this are fine, because its not the main focus). I am thinking of attempting to create a piece-wise function, if possible, for my loss, so that for any values out of bounds face a rapidly growing exponential loss function. Uncertain how to implement, or if it would even work.
Code is currently private. Once allowed, I will make the repo public.
To run DNC model, go to the project/ directory and run python -m src.main. If there are errors you encounter feel free to let me know.
This model depends upon Tensorflow r1.2, most recent Sonnet, and NLTK's punkt for Tokenizing sentences in sts_handler.py and tests/*.
In a regression model, the network calculates the model output based on the randomly initialized values for your model parameters. That's why you're seeing negative values here; you haven't trained your model enough for it to learn that your values are only between 0 and 5.
Unless I'm missing something, you are only calculating the loss, but you aren't actually training the model. You should probably be calling sess.run(optimizer) on an optimizer, not on your loss function.
You probably need to train your model for multiple epochs (training your model for one epoch = training your model once on the entire dataset).
Batches are used because it is more computationally efficient to train your model on a batch than it is to train it on a single example. However, your data seems to be small enough that you won't have that problem. As such, I would recommend reducing your batch size to as low as possible. As a general rule, you get better training from a smaller batch size, at the cost of added computation.
If you post all of your code, I can take a look.

AWS Machine Learning Data

I'm using the AWS Machine Learning regression to predict the waiting time in a line of a restaurant, in a specific weekday/time.
Today I have around 800k data.
Example Data:
restaurantID (rowID)weekDay (categorical)time (categorical)tablePeople (numeric)waitingTime (numeric - target)1 sun 21:29 2 23
2 fri 20:13 4 43
...
I have two questions:
1)
Should I use time as Categorical or Numeric?
It's better to split into two fields: minutes and seconds?
2)
I would like in the same model to get the predictions for all my restaurants.
Example:
I expected to send the rowID identifier and it returns different predictions, based on each restaurant data (ignoring others data).
I tried, but it's returning the same prediction for any rowID. Why?
Should I have a model for each restaurant?
There are several problems with the way you set-up your model
1) Time in the form you have it should never be categorical. Your model treats times 12:29 and 12:30 as two completely independent attributes. So it will never use facts it learn about 12:29 to predict what's going to happen at 12:30. In your case you either should set time to be numeric. Not sure if amazon ML can convert it for you automatically. If not just multiply hour by 60 and add minutes to it. Another interesting thing to do is to bucketize your time, by selecting which half hour or wider interval. You do it by dividing (h*60+m) by some number depending how many buckets you want. So to try 120 to get 2 hr intervals. Generally the more data you have the smaller intervals you can have. The key is to have a lot of samples in each bucket.
2) You should really think about removing restaurantID from your input data. Having it there will cause the model to over-fit on it. So it will not be able to make predictions about restaurant with id:5 based on the facts it learn from restaurants with id:3 or id:9. Having restaurant id there might be okay if you have a lot of data about each restaurant and you don't care about extrapolating your predictions to the restaurants that are not in the training set.
3) You never send restaurantID to predict data about it. The way it usually works you need to pick what are you trying to predict. In your case probably 'waitingTime' is most useful attribute. So you need to send weekDay, time and number of people and the model will output waiting time.
You should think what is relevant for the prediction to be accurate, and you should use your domain expertise to define the features/attributes you need to have in your data.
For example, time of the day, is not just a number. From my limited understanding in restaurant, I would drop the minutes, and only focus on the hours.
I would certainly create a model for each restaurant, as the popularity of the restaurant or the type of food it is serving is having an impact on the wait time. With Amazon ML it is easy to create many models as you can build the model using the SDK, and even schedule retraining of the models using AWS Lambda (that mean automatically).
I'm not sure what the feature called tablePeople means, but a general recommendation is to have as many as possible relevant features, to get better prediction. For example, month or season is probably important as well.
In contrast with some answers to this post, I think resturantID helps and it actually gives valuable information. If you have a significant amount of data per each restaurant then you can train a model per each restaurant and get a good accuracy, but if you don't have enough data then resturantID is very informative.
1) Just imagine what if you had only two columns in your dataset: restaurantID and waitingTime. Then wouldn't you think the restaurantID from the testing data helps you to find a rough waiting time? In the simplest implementation, your waiting time per each restaurantID would be the average of waitingTime. So definitely restaurantID is a valuable information. Now that you have more features in your dataset, you need to check if restaurantID is as effective as the other features or not.
2) If you decide to keep restaurantID then you must use it as a categorical string. It should be a non-parametric feature in your dataset and maybe that's why you did not get a proper result.
On the issue with day and time I agree with other answers and considering that you are building your model for the restaurant, hourly time may give a more accurate result.