Hello,

Question 1 :
I have a binary classification problem (click/no click).
I would like to retrieve the prediction values from test set (0 or 1) after training, testing and evaluation the NeuralFM (or FM) model in order to calculate some metrics like precision-recall.
Is that possible ?

Question 2:
What is the "keep_prob" argument in NeuralFM.py ? I don't know how to use it.

Anyway, thanks for this python library, that's a great too to play with.
Thanks in advance for the answers.

跑了一下。NFM代码效果并不比FM效果好，不知为什么

Hi there,

Could you please provide the command for evaluating NeuralFM model? -i.e. I was expecting something similar presented at https://github.com/hexiangnan/attentional_factorization_machine.

Best,

all_weights['feature_bias'] = tf.Variable(tf.random_uniform([self.features_M, 1], 0.0, 0.0), name='feature_bias') # features_M * 1
elements in all_weights['feature_bias'] are all zeros.
so in
# _________out _________ Bilinear = tf.reduce_sum(self.FM, 1, keep_dims=True) # None * 1 self.Feature_bias = tf.reduce_sum(tf.nn.embedding_lookup(self.weights['feature_bias'], self.train_features) , 1) # None * 1 Bias = self.weights['bias'] * tf.ones_like(self.train_labels) # None * 1 self.out = tf.add_n([Bilinear, self.Feature_bias, Bias]) # None * 1

all of elements in self.Feature_bias are zeros.
So, all_weights['feature_bias'] should be initialized from 0 and 1.

ok?

Q1:
It seems there's a deprecated API tf.sub in your implementation, which will throw exception like
AttributeError: 'module' object has no attribute 'sub'
in tensorflow 1.3.0+.

Change that to tf.subtract will fix it.

Q2:
Furthermore, have you ever consider using the tf.estimator.Estimator to replace sklearn.base.BaseEstimator ?
A tf.estimator.Estimator model with model.export_model will enable you to deploy a trained model with tensorflow Serving for a product env, while it can also make your training parallelized with high level api tf.contrib.learn.Experiment.

do you know code example to run criteo-1tb-benchmark fully locally , without spark?
kind of online learning?
for
https://labs.criteo.com/2013/12/download-terabyte-click-logs-2/

As the code it seems do not load the deep layer params if use the pretrain model.

https://github.com/hexiangnan/neural_factorization_machine/blob/master/NeuralFM.py#L202-L217

Is that on purpose which means daily model has a brand new deep layer params?

FM.py中
elif self.loss_type == 'log_loss':
self.out = tf.sigmoid(self.out)
if self.lambda_bilinear > 0:
self.loss = tf.contrib.losses.log_loss(self.out, self.train_labels, weight=1.0, epsilon=1e-07, scope=None) + tf.contrib.layers.l2_regularizer(self.lamda_bilinear)(self.weights['feature_embeddings']) # regulizer
else:
self.loss = tf.contrib.losses.log_loss(self.out, self.train_labels, weight=1.0, epsilon=1e-07, scope=None)

self.lambda_bilinear 变量有误，应该是 self.lamda_bilinear

I change the parameter layers from [64] to [64,128,256,128,64],and I meet the error "InvalidArgumentError (see above for traceback): slice index 4 of dimension 0 out of bounds."
How could I fix it

https://github.com/hexiangnan/neural_factorization_machine/blob/master/LoadData.py#L47

In the read_features() function, you just init a dict, and record the feature names and the first user which get this feature!! Nonsense!! Lost of Data.

This is a preview of ml-tag.test.libfm file:

-1.0 51798:1 2473:1 37583:1
-1.0 66335:1 61344:1 29842:1
-1.0 89085:1 60033:1 47050:1
1.0 61293:1 8073:1 3903:1
-1.0 81335:1 56575:1 50067:1
-1.0 65166:1 48181:1 12510:1
-1.0 75300:1 26027:1 38510:1
1.0 10219:1 2122:1 383:1
1.0 80855:1 80856:1 24728:1
1.0 67033:1 721:1 19495:1

I rewrite the code and test on upper file, and clearly you lost many data! Wrong Badly.

def read_features(file): # read a feature file
    features = {}
    i = len(features)
    with open(file) as f:
        for line in f:
            items = line.strip().split(' ')
            for item in items[1:]: # ['51798:1', '2473:1', '37583:1']
                if item not in features:
                    features[item] = i
                    i = i + 1
                else:
                    print('nfm load code error', i, item)
    return features

In deepfm, continuous features is multiplied by its responding embedding vector, while is your codes no multiplication was not seen. Can nfm not deal with continuous features?

Model.

_________ sum_square part _____________

get the summed up embeddings of features.

nonzero_embeddings = tf.nn.embedding_lookup(self.weights['feature_embeddings'], self.train_features)
self.summed_features_emb = tf.reduce_sum(nonzero_embeddings, 1) # None * K

get the element-multiplication

self.summed_features_emb_square = tf.square(self.summed_features_emb) # None * K

The above code is in the NerualFM.py

when you computed the self.summed_features_emb, the axis you wrote is '1', I think it should be 0. Did I understand wrong?