Question

我正在做新闻推荐系统，我需要为用户和他们阅读的新闻建立一个表格。我的原始数据就像这样：

001436800277225 ["9161492","9161787","9378531"]
009092130698762 ["9394697"]
010003000431538 ["9394697","9426473","9428530"]
010156461231357 ["9350394","9414181"]
010216216021063 ["9173862","9247870"]
010720006581483 ["9018786"]
011199797794333 ["9017977","9091134","9142852","9325464","9331913"]
011337201765123 ["9161294","9198693"]
011414545455156 ["9168185","9178348","9182782","9359776"]
011425002581540 ["9083446","9161294","9309432"]

我使用spark-SQL做爆炸和一个热门编码，

df = getdf()
df1 = df.select('uuid',explode('news').alias('news'))
stringIndexer = StringIndexer(inputCol="news", outputCol="newsIndex")
model = stringIndexer.fit(df1)
indexed = model.transform(df1)
encoder = OneHotEncoder(inputCol="newsIndex", outputCol="newsVec")
encoded = encoder.transform(indexed)
encoded.show(20,False)

之后，我的数据变为：

+---------------+-------+---------+----------------------+
|uuid           |news   |newsIndex|newsVec               |
+---------------+-------+---------+----------------------+
|014324000386050|9398253|10415.0  |(105721,[10415],[1.0])|
|014324000386050|9428530|70.0     |(105721,[70],[1.0])   |
|014324000631752|654112 |1717.0   |(105721,[1717],[1.0]) |
|014324000674240|730531 |2282.0   |(105721,[2282],[1.0]) |
|014324000674240|694306 |1268.0   |(105721,[1268],[1.0]) |
|014324000674240|712016 |4766.0   |(105721,[4766],[1.0]) |
|014324000674240|672307 |7318.0   |(105721,[7318],[1.0]) |
|014324000674240|698073 |1241.0   |(105721,[1241],[1.0]) |
|014324000674240|728044 |5302.0   |(105721,[5302],[1.0]) |
|014324000674240|672256 |1619.0   |(105721,[1619],[1.0]) |
|014324000674240|730236 |2376.0   |(105721,[2376],[1.0]) |
|014324000674240|730235 |14274.0  |(105721,[14274],[1.0])|
|014324000674240|728509 |1743.0   |(105721,[1743],[1.0]) |
|014324000674240|704528 |10310.0  |(105721,[10310],[1.0])|
|014324000715399|774134 |8876.0   |(105721,[8876],[1.0]) |
|014324000725836|9357431|3479.0   |(105721,[3479],[1.0]) |
|014324000725836|9358028|15621.0  |(105721,[15621],[1.0])|
|014324000730349|812106 |4599.0   |(105721,[4599],[1.0]) |
|014324000730349|699237 |754.0    |(105721,[754],[1.0])  |
|014324000730349|748109 |4854.0   |(105721,[4854],[1.0]) |
+---------------+-------+---------+----------------------+

但是一个ID有多行，所以我想groupBy('uuid')然后add这些向量。但只是使用groupBy然后添加将有错误。我怎么能这样做？

Answer 1

从indexed开始，我们可以将列newsIndex收集为列表，并使用SparseVector将其转换为udf。

要声明稀疏向量，我们需要功能的数量和包含位置和值的元组列表。因为我们正在处理分类变量，所以我们将使用1.0作为值。索引将是newsIndex列：

from pyspark.sql.functions import collect_list, max, lit
from pyspark.ml.linalg import Vectors, VectorUDT

def encode(arr, length):

  vec_args =  length, [(x,1.0) for x in arr]

  return Vectors.sparse(*vec_args)   

encode_udf = udf(encode, VectorUDT())

功能的数量为max(newsIndex) + 1（因为StrinIndexer从0.0开始）：

feats = indexed.agg(max(indexed["newsIndex"])).take(1)[0][0] + 1

将所有这些结合在一起：

indexed.groupBy("uuid") \
       .agg(collect_list("newsIndex")
       .alias("newsArr")) \
       .select("uuid", 
               encode_udf("newsArr", lit(feats))
               .alias("OHE")) \
       .show(truncate = False)
+---------------+-----------------------------------------+
|uuid           |OHE                                      |
+---------------+-----------------------------------------+
|009092130698762|(24,[0],[1.0])                           |
|010003000431538|(24,[0,3,15],[1.0,1.0,1.0])              |
|010720006581483|(24,[11],[1.0])                          |
|010216216021063|(24,[10,22],[1.0,1.0])                   |
|001436800277225|(24,[2,12,23],[1.0,1.0,1.0])             |
|011425002581540|(24,[1,5,9],[1.0,1.0,1.0])               |
|010156461231357|(24,[13,18],[1.0,1.0])                   |
|011199797794333|(24,[7,8,17,19,20],[1.0,1.0,1.0,1.0,1.0])|
|011414545455156|(24,[4,6,14,21],[1.0,1.0,1.0,1.0])       |
|011337201765123|(24,[1,16],[1.0,1.0])                    |
+---------------+-----------------------------------------+

如何使用Spark SQL在group by之后添加稀疏向量？

1 个答案: