Node2VecSkipGramEnsmallen lacks num_walks parameter
abbynewbury opened this issue · comments
Hi,
I am using the Node2VecSkipGramEnsmallen function and noticed that it does not take a num_walks parameter to represent the number of walks per node. The parameters that I see that are available are:
mappingproxy({'embedding_size': <Parameter "embedding_size: int = 100">,
'epochs': <Parameter "epochs: int = 30">,
'clipping_value': <Parameter "clipping_value: float = 6.0">,
'number_of_negative_samples': <Parameter "number_of_negative_samples: int = 10">,
'walk_length': <Parameter "walk_length: int = 128">,
'iterations': <Parameter "iterations: int = 10">,
'window_size': <Parameter "window_size: int = 5">,
'return_weight': <Parameter "return_weight: float = 0.25">,
'explore_weight': <Parameter "explore_weight: float = 4.0">,
'max_neighbours': <Parameter "max_neighbours: Optional[int] = 100">,
'learning_rate': <Parameter "learning_rate: float = 0.01">,
'learning_rate_decay': <Parameter "learning_rate_decay: float = 0.9">,
'central_nodes_embedding_path': <Parameter "central_nodes_embedding_path: Optional[str] = None">,
'contextual_nodes_embedding_path': <Parameter "contextual_nodes_embedding_path: Optional[str] = None">,
'normalize_by_degree': <Parameter "normalize_by_degree: bool = False">,
'stochastic_downsample_by_degree': <Parameter "stochastic_downsample_by_degree: Optional[bool] = False">,
'normalize_learning_rate_by_degree': <Parameter "normalize_learning_rate_by_degree: Optional[bool] = False">,
'use_scale_free_distribution': <Parameter "use_scale_free_distribution: Optional[bool] = True">,
'random_state': <Parameter "random_state: int = 42">,
'dtype': <Parameter "dtype: str = 'f32'">,
'ring_bell': <Parameter "ring_bell: bool = False">,
'enable_cache': <Parameter "enable_cache: bool = False">,
'verbose': <Parameter "verbose: bool = True">})
I was wondering if iterations is the equivalent to num_walks? Thanks!
Hi! If I recall correctly, there is no equivalent of that parameter because we compute new random walks on the fly for instead of pre-computing them.
In our experiments this lead to better performance
No, @zommiommy, the parameter @abbynewbury is looking for is the one called iterations.
Okay, thank you!