推荐策略之als基于tensorflow2

今天分享一个在用户规模很大时的模型训练更新代码，还有很多需要优化的地方

'''=================================================
@Function -> 用TensorFlow2实现协同过滤矩阵的分解
@Author ：郭艳丹
@Date   ：2023-01-11
=================================================='''

import numpy as np
import tensorflow as tf
from keras.callbacks import ModelCheckpoint

def get_all_user_ids():
    #TODO change your real data
    return ["user1","user2","user3","user4","user5","user6"]

def get_all_item_ids():
    # TODO change your real data
    return ["item1","item2","item3","item4","item11","item22","item33","item44","item111","item222","item333","item444"]

def add_negtive_samples(item_ids:set,user_click_item_ids:set,total_len=20):
    # 此处采集和正样本等数的负样本
    candidate_set = list(item_ids - user_click_item_ids)  # 热度采样
    neg_list = np.random.choice(candidate_set, size=total_len-len(user_click_item_ids), replace=True)  # 对于每个正样本，选择n个负样本
    return

import pandas as pd
def get_pretrained_data():
    #TODO 后续调整为tf.data
    user_click_matrix = [("user1","item1"),
                         ("user2","item3"),
                         ("user1","item3"),
                         ("user2","item2"),
                         ("user4","item3"),
                         ("user4","item33"),
                         ("user3","item3"),
                         ("user5","item3"),
                         ("user5","item11"),
                         ("user1","item3"),
                         ("user1","item2"),
                         ]
    original_click_data = pd.DataFrame(user_click_matrix,columns=["user","item"])
    original_click_data["target"] = 1
    item_ids = get_all_item_ids()
    user_ids = get_all_user_ids()
    negtive_samples = np.random.choice(item_ids, size=original_click_data.shape[0],replace=True)
    #负样本构造
    negtive_click_datas = pd.DataFrame({"user":original_click_data["user"],"item": negtive_samples})
    negtive_click_datas["target"] = 0
    result = pd.concat([original_click_data,negtive_click_datas],ignore_index=True,sort=False)
    return item_ids,user_ids, {"user":result["user"],"item":result["item"]},result["target"]


def als_by_batch_train():
    item_ids,user_ids,train_data,target = get_pretrained_data()
    #输入为(None,)表示输入的为一个一维的向量
    user_input = tf.keras.layers.Input(shape=(None,),name="user",dtype=tf.string)
    item_input = tf.keras.layers.Input(shape=(None,),name="item",dtype=tf.string)

    user_string_lookup = tf.keras.layers.StringLookup(vocabulary=user_ids)(user_input)
    item_string_lookup = tf.keras.layers.StringLookup(vocabulary=item_ids)(item_input)

    user_embedding = tf.keras.layers.Embedding(len(user_ids)+5,64)(user_string_lookup)
    item_embedding = tf.keras.layers.Embedding(len(item_ids)+5,64)(item_string_lookup)
    cons_sim_result = tf.keras.layers.Dot(-1,normalize=True)([user_embedding,item_embedding])

    model = tf.keras.Model(inputs=[user_input,item_input], outputs=cons_sim_result)
    model.compile(optimizer="adam", loss=tf.keras.losses.MSE)
    callbacks = [
        ModelCheckpoint(filepath='base_path/als_models/'+'{epoch: 02d}.h5')
    ]

    model.fit(train_data, target, batch_size=4, epochs=2, verbose=1,
              validation_split=0.01,callbacks=callbacks)
    return tf.keras.Model(inputs=[user_input], outputs=user_embedding),tf.keras.Model(inputs=[item_input], outputs=item_embedding)

def get_embedding(model,inputs):
    inputs = tf.constant(inputs)
    return model(inputs)


if __name__ == '__main__':
    user_embedding_model,item_embedding_model = als_by_batch_train()

    user_embedding_result = get_embedding(user_embedding_model,["user1","user2","user100"])
    item_embeddings_result = get_embedding(item_embedding_model,["item2","item1","itemk"])
    print(user_embedding_result)
    print(item_embeddings_result)

其中，tf.keras.layers.Dot是对应的矩阵中的每一行对应相乘并求和，而且可以通过设置normalize可以在进行对应行相乘之前，对其进行行正则化，从而实现了余弦相似性的计算，相关代码已经开源到：sparkle_code_guy/rec_sys - 码云 - 开源中国 (gitee.com)