Summary: 78 instances, 44 unique

Text	Count
# TODO consider other formulations metapath2vec	2
# TODO check that the masking step works	1
# TODO possibly filter out again the frequent nonzero columns	1
# TODO Check if this title is important or not.. does it exist in other dictionaries	4
# TODO find all zero indices rows and remove.	1
# TODO negative samples	2
# TODO sum for the link prediction to not consider the zero messages	1
TODO add metapaths	2
TODO add motifs	1
#TODO fix this is wrong had to add all edges in the testign graph	1
# TODO implement this layer	1
# TODO maybe represent triplets as three arrays to make it faster.	2
# TODO 2 different type of embeddings context and embedding to improve performance	1
# TODO use sum instead of mean	1
# TODO THIS IS WRONG!!! I have to add the train valid and test	1
#values_o = np.random.randint(num_entity_o, size=num_to_generate) TODO add later	1
# TODO check the masked 1 with without mask that returns the same	1
# TODO is it an entry per person and movie or list of persons	2
# TODO change to torch	1
# TODO (lingfan): implement filtered metrics	1
# TODO filter edges	2
# TODO conver to torch th.array()	1
# TODO possibly high complexity here??	2
# TODO consider other formulations metapath2vec	1
# TODO do not pass the whole value key only the subsets...	2
# TODO how to treat missing attributes	1
# TODO filter positive edges	1
# TODO check datatype	1
# TODO 1. Make sure the graph is connected (how) ? rgcn does not	1
# TODO summary per node type or across all node types? for infomax	1
# TODO move to dataloader so that it is faster	5
# TODO should we include job title and character played ? job category, more meaningfull	2
# TODO is it an entry per person and movie or list of persons	2
# TODO load first to dictionary and then process in parallel for the nlp model ...	2
# TODO Possible Bug. How about when only one or 2 directors or writers exists?	4
# TODO genre to multihot encoding	2
# TODO test data dumping and loading	4
# TODO Parameterize per node type pairs!!	1
# TODO have to map the etype and ntype to their integer ids.	2
# TODO decide if these attributes are needed.	2
# TODO filter positive edges	1
# TODO implement regularization	7
TODO add metapaths	1
# TODO different layers may have different number of hidden units current implementation prevents	2