TFRS之上下文特征

在推荐模型中&＃xff0c;有很多因素会影响id之外的特性是否有用:

• 上下文的重要性&＃xff1a;如果用户的首选项跨上下文和时间相对稳定&＃xff0c;那么上下文特性可能不会提供太多好处。然而&＃xff0c;如果用户偏好是高度上下文相关的&＃xff0c;那么添加上下文将显著改善模型。例如&＃xff0c;在决定是推荐一个短片还是一部电影时&＃xff0c;一周的哪一天可能是一个重要的特征:用户在一周中可能只有时间看短内容&＃xff0c;但可以在周末放松并享受一部完整长度的电影。类似地&＃xff0c;查询时间戳可能在建模流行动态中扮演重要角色:一部电影可能在周围非常流行
• 数据稀疏&＃xff1a;如果数据稀疏&＃xff0c;使用非id特征可能是关键。由于对给定用户或项目的观察很少&＃xff0c;模型可能难以估计每个用户或每个项目的良好表示。为了建立一个准确的模型&＃xff0c;其他的特征&＃xff0c;如项目类别、描述和图像&＃xff0c;必须被用来帮助模型泛化训练数据以外的数据。这在冷启动的情况下尤其重要&＃xff0c;因为在冷启动的情况下&＃xff0c;一些项目或用户的数据相对较少。

import os
import tempfileimport numpy as np
import tensorflow as tf
import tensorflow_datasets as tfdsimport tensorflow_recommenders as tfrs


特征数据

ratings &＃61; tfds.load("movielens/100k-ratings", split&＃61;"train")
movies &＃61; tfds.load("movielens/100k-movies", split&＃61;"train")ratings &＃61; ratings.map(lambda x: {"movie_title": x["movie_title"],"user_id": x["user_id"],"timestamp": x["timestamp"],
})
movies &＃61; movies.map(lambda x: x["movie_title"])


特征词汇表

timestamps &＃61; np.concatenate(list(ratings.map(lambda x: x["timestamp"]).batch(100)))max_timestamp &＃61; timestamps.max()
min_timestamp &＃61; timestamps.min()timestamp_buckets &＃61; np.linspace(min_timestamp, max_timestamp, num&＃61;1000,
)unique_movie_titles &＃61; np.unique(np.concatenate(list(movies.batch(1000))))
unique_user_ids &＃61; np.unique(np.concatenate(list(ratings.batch(1_000).map(lambda x: x["user_id"]))))


定义模型

user model
代码中改动&＃xff1a;增加是否使用时间戳这个特征的选择。

class UserModel(tf.keras.Model):def __init__(self, use_timestamps):super().__init__()self._use_timestamps &＃61; use_timestampsself.user_embedding &＃61; tf.keras.Sequential([tf.keras.layers.StringLookup(vocabulary&＃61;unique_user_ids, mask_token&＃61;None),tf.keras.layers.Embedding(len(unique_user_ids) &＃43; 1, 32),])if use_timestamps:self.timestamp_embedding &＃61; tf.keras.Sequential([tf.keras.layers.Discretization(timestamp_buckets.tolist()),tf.keras.layers.Embedding(len(timestamp_buckets) &＃43; 1, 32),])self.normalized_timestamp &＃61; tf.keras.layers.Normalization(axis&＃61;None)self.normalized_timestamp.adapt(timestamps)def call(self, inputs):if not self._use_timestamps:return self.user_embedding(inputs["user_id"])return tf.concat([self.user_embedding(inputs["user_id"]),self.timestamp_embedding(inputs["timestamp"]),tf.reshape(self.normalized_timestamp(inputs["timestamp"]), (-1, 1)),], axis&＃61;1)


movie model

class MovieModel(tf.keras.Model):def __init__(self):super().__init__()max_tokens &＃61; 10_000self.title_embedding &＃61; tf.keras.Sequential([tf.keras.layers.StringLookup(vocabulary&＃61;unique_movie_titles, mask_token&＃61;None),tf.keras.layers.Embedding(len(unique_movie_titles) &＃43; 1, 32)])self.title_vectorizer &＃61; tf.keras.layers.TextVectorization(max_tokens&＃61;max_tokens)self.title_text_embedding &＃61; tf.keras.Sequential([self.title_vectorizer,tf.keras.layers.Embedding(max_tokens, 32, mask_zero&＃61;True),tf.keras.layers.GlobalAveragePooling1D(),])self.title_vectorizer.adapt(movies)def call(self, titles):return tf.concat([self.title_embedding(titles),self.title_text_embedding(titles),], axis&＃61;1)


组合模型

class MovielensModel(tfrs.models.Model):def __init__(self, use_timestamps):super().__init__()self.query_model &＃61; tf.keras.Sequential([UserModel(use_timestamps),tf.keras.layers.Dense(32)])self.candidate_model &＃61; tf.keras.Sequential([MovieModel(),tf.keras.layers.Dense(32)])self.task &＃61; tfrs.tasks.Retrieval(metrics&＃61;tfrs.metrics.FactorizedTopK(candidates&＃61;movies.batch(128).map(self.candidate_model),),)def compute_loss(self, features, training&＃61;False):# We only pass the user id and timestamp features into the query model. This# is to ensure that the training inputs would have the same keys as the# query inputs. Otherwise the discrepancy in input structure would cause an# error when loading the query model after saving it.query_embeddings &＃61; self.query_model({"user_id": features["user_id"],"timestamp": features["timestamp"],})movie_embeddings &＃61; self.candidate_model(features["movie_title"])return self.task(query_embeddings, movie_embeddings)


实验

准备数据集

tf.random.set_seed(42)
shuffled &＃61; ratings.shuffle(100_000, seed&＃61;42, reshuffle_each_iteration&＃61;False)train &＃61; shuffled.take(80_000)
test &＃61; shuffled.skip(80_000).take(20_000)cached_train &＃61; train.shuffle(100_000).batch(2048)
cached_test &＃61; test.batch(4096).cache()


Baseline :没有时间戳特性

model &＃61; MovielensModel(use_timestamps&＃61;False)
model.compile(optimizer&＃61;tf.keras.optimizers.Adagrad(0.1))model.fit(cached_train, epochs&＃61;3)train_accuracy &＃61; model.evaluate(cached_train, return_dict&＃61;True)["factorized_top_k/top_100_categorical_accuracy"]
test_accuracy &＃61; model.evaluate(cached_test, return_dict&＃61;True)["factorized_top_k/top_100_categorical_accuracy"]print(f"Top-100 accuracy (train): {train_accuracy:.2f}.")
print(f"Top-100 accuracy (test): {test_accuracy:.2f}.")


利用时间特征捕捉时间动态

model &＃61; MovielensModel(use_timestamps&＃61;True)
model.compile(optimizer&＃61;tf.keras.optimizers.Adagrad(0.1))model.fit(cached_train, epochs&＃61;3)train_accuracy &＃61; model.evaluate(cached_train, return_dict&＃61;True)["factorized_top_k/top_100_categorical_accuracy"]
test_accuracy &＃61; model.evaluate(cached_test, return_dict&＃61;True)["factorized_top_k/top_100_categorical_accuracy"]print(f"Top-100 accuracy (train): {train_accuracy:.2f}.")
print(f"Top-100 accuracy (test): {test_accuracy:.2f}.")