Document for Arguments

Common Argument Options For EXP1~4

--dataset

This project HiRM provides Three Dataset: "amazon-book", "gowalla", "yelp2018". You can choose one of this datasets. For example,

--dataset="amazon-book"

--topks

This is for calculating metrics of multiple topKs. Please type the list of topK. For example,

--topks="[10,20]"

--simple_model

This argument is for selecting the simple model. We provide 6 simple models:

• "gf-cf"
• "exp1"
• "exp2"
• "exp3"
• "exp4"
• Our model, "HiRM"

--simple_model="gf-cf"

--expdevice

This argument is for selecting the device for inference. You could use "cpu", "cuda:0", "cuda:1", etc. For example,

--expdevice="cpu"

--svdtype

This argument sets the sparsesvd package. We provide these packages:

• "sparsesvd"
• "scipy"
• "sklearn-rand"
• "fbpca"
• "torch"
• "torch_cuda"

Optionally, we provide "cupy" also. For example,

--svdtype="sparsesvd"

--svdvalue

This argument defines the low rank dimension for SVD. The default is 256, and we suggest to type positive integer. For example,

--svdvalue=256

--alpha_start

This argument sets the start real value of proportion(=alpha). We define any real value, which means, negative value is also okay. For example,

--alpha_start=0.0

--alpha_end

This argument sets the end real value of proportion(=alpha). We define any real value, which means, negative value is also okay. Also, if you want to check for only one alpha value, then set the alpha_end with same value of alpha_start. For example,

--alpha_end=1.0

--alpha_step

This argument sets the difference between each step of different alpha values. If you set alpha_step as 0.05, then alpha will diverge with adding 0.05, until this is greater or equal then alpha_end. For example,

--alpha_step=0.05

--filter

This argument sets the filter type for the experience 2 and 3. We provied 11 Filters:

• "['linear']"
• "['ideal-low-pass']"
• "['gaussian', 0.1]"
• "['heat-kernel', 0.1]"
• "['butterworth',1]"
• "['butterworth',2]"
• "['butterworth',3]"
• "['gfcf-linear-autoencoder', 0.1]"
• "['gfcf-Neighborhood-based]"
• "['inverse']"
• "['sigmoid-low-pass']".

For filter gaussian, heat-kernel, gfcf-linear-autoencoder, you can add hyperparameter as a second element of the list. For example,

--filter="['ideal-low-pass']"

--testbatch

For inference step, you can increase the size of the batch. We recommend to use 2048 size batch. But for our model HiRM, 256 is recommended only for "amazon-book" dataset, for VRAM size constraint. For example,

--testbatch=2048