import sparsesvd
import scipy
import csv
from itertools import islice
import numpy as np
import math


def recommendation(test_user_id_list, row, col, data, m, n, k, goods_num):
    lr = 0.05
    l = 0.01
    iter = 6
    coo_mat = scipy.sparse.coo_matrix((data, (row, col)), shape=(m, n))
    csc_mat = coo_mat.tocsc()
    ut, s, vt = sparsesvd.sparsesvd(csc_mat, k)

    u = ut.T

    rui = np.zeros([m,n])
    p = [] #存储pu,pu为用户的行向量
    q = [] #存储qi,qi为物品的列向量

    for i in range(m):
        pu = np.reshape(u[i], [1, k])
        p.append(pu)

    for j in range(n):
        qi = np.reshape(vt[:, j], [k, 1])
        q.append(qi)
    print("数据准备完成..")

    count = 0
    for i in range(m):
        pu = p[i]
        for j in range(n):
            qi = q[j]
            rui[i][j] = np.matmul(pu, qi)[0][0]
        count += 1
        if count % 100 == 0:
            print("计算rui进度：{:.2f}%".format(count / m * 100))

    for a in range(iter):
        loss = 0
        for i in range(m):
            pu = p[i]
            for j in range(n):
                qi = q[j]
                err = rui[i][j] - np.matmul(pu, qi)[0][0]
                loss += err
                qi = qi + lr * (err * pu - l * qi)
                pu = pu + lr * (err * qi - l * pu)
                p[i] = pu
                q[j] = qi
        lr *= 0.9
        print("梯度下降进度：{:.2f}%，损失：{}".format(a / iter * 100, loss))

    user_recommendation_list = {}
    count = 0
    for user_id in test_user_id_list:
        recommendation_score = {}
        pu = p[user_id]
        # 遍历所有物品，依次计算评分
        for i in range(n):
            # i是物品的id
            score = np.matmul(q[i], pu)[0][0]
            recommendation_score[i] = score
        recommendation_score = sorted(recommendation_score.items(), key=lambda x: x[1], reverse=True)
        topk_recommendation = []
        for id, score in recommendation_score[:goods_num]:
            topk_recommendation.append(id)
        user_recommendation_list[user_id] = topk_recommendation
        if count % 100 == 0:
            print("当前推荐进度: {:.2f}%".format(count / len(test_user_id_list) * 100))
        count += 1
    return user_recommendation_list

def load_test_data(data_path):
    file = csv.reader(open(data_path, "r", encoding="utf-8"))
    user_id_list = []
    for line in islice(file, 1, None):
        user_id_list.append(int(line[0]))
    return user_id_list

def load_train_data(data_path):
    row = []
    col = []
    data = []
    file = csv.reader(open(data_path, "r", encoding="utf-8"))
    for line in islice(file, 1, None):
        row.append(int(line[0]))
        col.append(int(line[1]))
        data.append(1)

    return row, col, data



def write_res_to_csv(user_recommendation_res):
    file = open("submission_lfm.csv", "w", encoding="utf-8", newline="")
    csv_writer = csv.writer(file)
    csv_writer.writerow(["user_id", "item_id"])

    for user_id, recommendation_list in user_recommendation_res.items():
        for item_id in recommendation_list:
            csv_writer.writerow([str(user_id), str(item_id)])
    file.close()

if __name__ == "__main__":
    train_data_path = "./dataset/book_train_dataset.csv"
    test_data_path = "./dataset/book_test_dataset.csv"
    # n = 10000物品的数量为10000
    # goods_num = 10即用户喜欢物品A，则寻找与物品最相似的 K种物品
    # k为svd分解系数
    # m为用户的数量
    n = 10000
    k = 10
    goods_num = 10

    row, col, data = load_train_data(train_data_path)
    print("训练数据加载完成...")
    m = max(row) + 1

    test_user_id_list = load_test_data(test_data_path)
    print("测试数据加载完成...")
    user_recommendation_list = recommendation(test_user_id_list, row, col, data, m, n, k, goods_num)
    print("用户物品推荐完成, 正在写入文件...")
    write_res_to_csv(user_recommendation_list)
    print("测试结果写入文件完成！")