houchenyu/l2r

A Python version of RankNet, LambdaRank and LambdaMart

ReadMe

This repository is used to share some L2R algorithms implemted by Python.

So far, this repository contains RankNet , LambdaRank and LambdaMART

RankNet

I utilize Pytorch to implement the network structure.

In order to use the interface, you should input following parameters:

• n_feaure: int, features numble
• h1_units: int, the unit numbers of hidden layer1
• h2_units: int, the unit numbers of hidden layer2
• epoch: int, iteration times
• learning_rate: float, learning rate
• plot: boolean, whether plot the loss.

LambdaRank

The usage is similar with RankNet.

LambdaMart

This is a Python version of LambdaMART.

I implement it based on the code of lezzago

‼️I have made some modification because I think there is a mistake on calculating $\lambda$ in lezzago's code.

Dataset

The dataset is the same as that of lezzago. I have preprocessed it and store in train.npy and test.npy.

You can directly used np.load() to import dataset.

The first column is label, the second column is qid, and the following columns are features (total 46 features).

l2r's Issues

iDCG = 0？

iDCG是否可以是0？这种情形应该怎么计算NDCG和lambda呢？

LambdaRank算法的模型里的predict()

请问关于predict()里面的 for tree in self.trees ，整个代码里没有看到trees呀？代码运行起来预测的结果全是0

lambdaRank: lambda calculate wrong

    delta = abs(single_dcgs[(i,j)] + single_dcgs[(j,i)] - single_dcgs[(i,i)] -single_dcgs[(j,j)])/IDCG

for this line，you mean the ith document in true_scores should be position i ? after checking the data, it seems wrong, maybe we should sort true socres and save the position of documents in true score which can be used here.

Questions about predicted_scores

Hi, sorry to bother you. I am an undergraduate student of sun yat-sen university and I am learning LambdaMART recently. Thank you for sharing the code. I have some doubts about the variable "predicted_scores"
Could you please tell me what it stands for? Does it mean a predicted label or a grade of document location? What does the sign mean?
I would appreciate it very much if you could response my issue. You can also contact me by email:[email protected]
Regards

lambdaMART

w[i] += rho * rho_complement * delta
w[i] -= rho * rho_complement * delta
你好，我想问下这里w表示什么意思，后面好像也没有用到这个值

请问如何保存模型呢？并且输出y_pred

如上

how do we select value of K for calculating NDCG score?

Hello,

I am not sure If I fully understand how to select the value of k for getting accurate results. Can you please guide me on which factors should be considered by deciding the k value.

Thank you!

RankNet code

Maybe there is a mistake in [split_pair(order_pairs)] Function, relevant_doc and irrelecant_doc do not move while qid increasing, it seems d1 and d2 always swing in the first part of the training_data. I think it might be add #code like below.

def split_pairs(order_pairs):
    """
    split the pairs into two list, named relevant_doc and irrelevant_doc.
    relevant_doc[i] is prior to irrelevant_doc[i]
    :param order_pairs: ordered pairs of all queries
    :return: relevant_doc and irrelevant_doc
    """
    relevant_doc = []
    irrelevant_doc = []
    # addindex = 0
    query_num = len(order_pairs)
    for i in range(query_num):
        pair_num = len(order_pairs[i])
        for j in range(pair_num):
            d1, d2 = order_pairs[i][j]
            # d1 = d1 + addindex
            # d2 = d2 + addindex
            relevant_doc.append(d1)
            irrelevant_doc.append(d2)
        # addindex = addindex + pair_num
    return relevant_doc, irrelevant_doc

Format of input data

Hello, you used a .npy format for the input data. What is the content and format of the input data in the raw text?