LightFM 如何为新用户进行预测（冷启动）——用户 ID 8 不在用户 ID 映射中

Question

我正在构建一个推荐系统，以便根据用户特征和项目特征向员工推荐培训，根据文档，LightFM 是一个很棒的算法。

我的用户数据框：

    User-Id name    age los ou  gender  skills
0   1   Luis    21  IFS architecture    M   python
1   2   Peter   22  ADV pmo M   pm
2   3   Jurgen  23  IFS architecture    M   sql
3   4   Bart    24  IFS architecture    M   python
4   5   Cristina    25  ADV pmo F   pm
5   6   Lambert 33  IFS development M   sql
6   7   Rahul   44  IFS development M   python

我的trainingds数据框

Training-Id training name   main skill
0   1   basic python    python
1   2   advanced python python
2   3   basic scrum pm
3   4   advanced scrum  pm
4   5   basic sql   sql
5   6   advanced sql    sql

我接受过的培训数据框（10 表示用户接受过该培训）所以我的权重只有 10s

User-Id Training-Id TrainingTaken
0   1   1   10
1   1   2   10
2   2   3   10
3   2   4   10
4   3   5   10
5   3   6   10
6   4   1   10
7   4   2   10

我找到了这个创建矩阵的好帮手： https://github.com/Med-ELOMARI/LightFM-Dataset-Helper

所以：

items_column = "Training-Id"
user_column = "User-Id"
ratings_column = "TrainingTaken"

items_feature_columns = [
    "training name",
    "main skill"
]

user_features_columns = ["name","age","los","ou", "gender", "skills"]

dataset_helper_instance = DatasetHelper(
    users_dataframe=usersdf,
    items_dataframe=trainingsdf,
    interactions_dataframe=trainingstakendf,
    item_id_column=items_column,
    items_feature_columns=items_feature_columns,
    user_id_column=user_column,
    user_features_columns=user_features_columns,
    interaction_column=ratings_column,
    clean_unknown_interactions=True,
)

# run the routine
# you can alslo run the steps separately one by one | routine function is simplifying the flow
dataset_helper_instance.routine()

上面的helper returns交互矩阵，权重矩阵等

dataset_helper_instance.weights.todense()
Output menu
matrix([[10., 10.,  0.,  0.,  0.,  0.],
        [ 0.,  0., 10., 10.,  0.,  0.],
        [ 0.,  0.,  0.,  0., 10., 10.],
        [10., 10.,  0.,  0.,  0.,  0.],
        [ 0.,  0.,  0.,  0.,  0.,  0.],
        [ 0.,  0.,  0.,  0.,  0.,  0.],
        [ 0.,  0.,  0.,  0.,  0.,  0.]], dtype=float32)


dataset_helper_instance.interactions.todense()

matrix([[1., 1., 0., 0., 0., 0.],
        [0., 0., 1., 1., 0., 0.],
        [0., 0., 0., 0., 1., 1.],
        [1., 1., 0., 0., 0., 0.],
        [0., 0., 0., 0., 0., 0.],
        [0., 0., 0., 0., 0., 0.],
        [0., 0., 0., 0., 0., 0.]], dtype=float32)

然后我训练测试拆分并拟合模型

from lightfm import LightFM
from lightfm.cross_validation import random_train_test_split
(train, test) = random_train_test_split(interactions=dataset_helper_instance.interactions, test_percentage=0.2)

model = LightFM(loss='warp')

model.fit(
    interactions=dataset_helper_instance.interactions,
    sample_weight=dataset_helper_instance.weights,
    item_features=dataset_helper_instance.item_features_list,
    user_features=dataset_helper_instance.user_features_list,
    verbose=True,
    epochs=50,
    num_threads=20,
)

然后我检查AUC和精度：

from lightfm.evaluation import precision_at_k
from lightfm.evaluation import auc_score

train_precision = precision_at_k(model, train,item_features=dataset_helper_instance.item_features_list, user_features=dataset_helper_instance.user_features_list , k=10).mean()
test_precision = precision_at_k(model, test, item_features=dataset_helper_instance.item_features_list, user_features=dataset_helper_instance.user_features_list,k=10).mean()
train_auc = auc_score(model, train,item_features=dataset_helper_instance.item_features_list, user_features=dataset_helper_instance.user_features_list).mean()
test_auc = auc_score(model, test,item_features=dataset_helper_instance.item_features_list, user_features=dataset_helper_instance.user_features_list).mean()
print('Precision: train %.2f, test %.2f. '% (train_precision, test_precision))
print('AUC: train %.2f, test %.2f.' % (train_auc, test_auc))

Precision: train 0.15, test 0.10. 
AUC: train 0.90, test 1.00.

然后我为现有用户做预测

scores = model.predict(user_ids=6, item_ids=[1,2,3,5,6])
print(scores)

[ 0.01860116 -0.20987387  0.06134995  0.08332028  0.13678455]

太好了，我可以得到一些针对用户 ID 6 的训练预测。

现在我要为新用户预测，（冷启动）

我尝试了以下方法：

dataset = Dataset()
new_user_feature = [8,{'name:John', 'Age:33', 'los:IFS','ou:development', 'skills:sql'} ]    
new_user_feature = [8,new_user_feature]
new_user_feature = dataset.build_user_features([new_user_feature])
#predict new users User-Id  name    age los ou  gender  skills
model.predict(0, item_ids=[1,2,3,5,6], user_features=new_user_feature)

但是我得到这个错误：

ValueError: user id 8 not in user id mappings.

我在这里错过了什么？

Answer 1

我无法测试，但我认为问题出在你写的时候：

new_user_feature = [8,{'name:John', 'Age:33', 'los:IFS','ou:development', 'skills:sql'} ]    
new_user_feature = [8,new_user_feature]

根据文档，dataset.build_user_features(..) 想要一个 (user id, [list of feature names]) 或 (user id, {feature name: feature weight}).

形式的可迭代对象

在你的情况下，我认为你应该将上面的两行替换为：

new_user_feature = [8,{'name':'John', 'Age':33, 'los':'IFS','ou':'development', 'skills':'sql'} ]
# Is the gender missing?

如果不行，输入格式可能是这样的：

new_user_feature = [8,['John', 33, 'IFS', 'development', 'sql'] ]

如果问题解决了请告诉我

LightFM 如何为新用户进行预测（冷启动）——用户 ID 8 不在用户 ID 映射中

LightFM how to make predictions for new users (cold start) - user id 8 not in user id mappings

python

recommendation-engine

pandas

lightfm