Load Pre-trained EmbeddingΒΆ

For users who want to use pre-trained user(item) embedding to train their model. We provide a simple way as following.

Firstly, prepare your additional embedding feature file, which contain at least two columns (id & embedding vector) as following format and name it as dataset.suffix (e.g: ml-1m.useremb).




-115.08 13.60 113.69


-130.97 263.05 -129.88

Note that here the header of user id must be different from user id in your .user file or .inter file (e.g: if the header of user id in .user or .inter file is user_id:token, the header of user id in your additional embedding feature file must be different. It can be either uid:token or userid:token).

Secondly, update the args as (suppose that USER_ID_FIELD: user_id):

additional_feat_suffix: [useremb]
    # inter/user/item/...: As usual
    useremb: [uid, user_emb]
fields_in_same_space: [[uid, user_id]]
    uid: user_emb

Then, this additional embedding feature file will be loaded into the Dataset object. These new features can be accessed as following:

dataset = create_dataset(config)

In your model, user embedding matrix can be initialized by your pre-trained embedding vectors as following:

class YourModel(GeneralRecommender):
    def __init__(self, config, dataset):
            pretrained_user_emb = dataset.get_preload_weight('uid')
            self.user_embedding = nn.Embedding.from_pretrained(torch.from_numpy(pretrained_user_emb))