haowei01/pytorch-examples

train models in pytorch, Learn to Rank, Collaborative Filter, Heterogeneous Treatment Effect, Uplift Modeling, etc

pytorch-examples's Issues

Relevance sort direction

Hi! Thanks for this excellent repo, it's very informative.

I've noticed that in your implementation of LambdaRank (specifically, on line 178) you sort the rank data frame by relevance in ascending order (which is the default for pandas' sort_values function).
Every other implementation of LambdaRank I've looked at (allRank, tensorflow-LTR) seems to order relevance levels in a descending order - and I'd have to agree with them (since you want higher relevance items towards the front of the array).

Sorry if I've misunderstood. Thanks.

Data Preprocessing might add more flavour!

Thank you for publishing the great repo and I definitely am learning a lot from this repo!!!
One thing I just noticed when I was working on investigating the dataset(Personalize Expedia Hotel Searches - ICDM 2013) was that some columns contain the large amount of Null so that removing those columns or imputing the missing values might improve the result!

Anyway, I will try myself as well!

Code

pytorch-examples/ranking/data_loaders/load_expedia.py

Lines 17 to 38 in 6c217bb

	def __init__(self):
	cur_file = os.path.abspath(__file__)
	self.data_dir = os.path.join(os.path.dirname(cur_file), DATA_DIR)
	print('{} loading data from DATA dir {}'.format(get_time(), self.data_dir))
	pkl_file = os.path.join(self.data_dir, 'train.pkl')
	if os.path.isfile(pkl_file):
	train_df = pd.read_pickle(pkl_file)
	else:
	train_df = pd.read_csv(os.path.join(self.data_dir, 'train.zip'))
	train_df.to_pickle(pkl_file)
	self.random_df = train_df[train_df[RANDOM_COL] == 1]
	self.biased_click_df = train_df[train_df[RANDOM_COL] == 0]
	print('{} unbiased training data rows {}'.format(get_time(), self.random_df.shape[0]))
	print('{} biased training data rows {}'.format(get_time(), self.biased_click_df.shape[0]))
	test_pkl_file = os.path.join(self.data_dir, 'test.pkl')
	if os.path.isfile(test_pkl_file):
	self.test_df = pd.read_pickle(test_pkl_file)
	else:
	self.test_df = pd.read_csv(os.path.join(self.data_dir, 'test.zip'))
	self.test_df.to_pickle(test_pkl_file)
	print('{} test file size {}'.format(get_time(), self.test_df.shape[0]))

Data Analysis on the dataset

• https://www.kaggle.com/rishitdagli/unsupervised-learning-workflow-working-with-huge
• https://www.kaggle.com/bimimi/time-series-anomaly-detection#Note

Something Wrong with decay_diff ?

Hi~Thank you for providing such an excellent work! But I have some questions about the calculation of the "decay_diff" term. ( LambdaRank.py, line 197)

I noticed that you are using "sort_order" as the discounted factor.
Sort_order is the the indices that would sort Y, which means sort_order[0] is the index of the document which is the most relevant to the query. But this is not cosistent with the gain_diff calculation!
Gain_diff (i, j) is the gain difference between douc i and douc j，but decay_diff is not the decay difference between douc i and douc j.