写在前面话
UCF通常是User-base Collaborative Filter的简写;大体的算法思路是根据用户行为计算相似群体(邻居),为用户推荐其邻居喜好的内容;感觉是不是很简单、那废话不多说先撸个SQL。
SQL
- select uid1,uid2,sim
- from (
- select uid1
- ,uid2
- ,cnt12 / sqrt(cnt1*cnt2) sim
- ,row_number() over(partition by uid1 order by cnt12 / sqrt(cnt1*cnt2) desc) sim_rn
- from (
- select a.uid uid1
- ,b.uid uid2
- ,count(a.iid) cnt12
- from tb_behavior a
- join tb_behavior b
- on a.iid = b.iid
- where a.uid <> b.uid
- group by a.uid,b.uid
- ) a12
- join (select uid,count(iid) cnt1 from tb_behavior group by uid) a1
- on a12.uid1 = a1.uid
- join (select uid,count(iid) cnt2 from tb_behavior group by uid) a2
- on a12.uid1 = a2.uid
- ) tb_neighbour
- where sim > 0.1 and sim_rn <= 30
读者实现的话只需要把上面的tb_behavior表替换成自己业务的用户行为即可;iid,uid分别对应物品id和用户id;
根据共现相似度,即共同喜好的物品个数比上各自喜好物品总数乘积取平方;最后截断用户最相似的前30个邻居作为推荐的依据。
上面构造了邻居表,下面就是根据邻居的喜好为用户推荐了,具体sql如下:
- select uid1,iid
- from (
- select uid1
- ,iid
- ,max(sim) score
- ,row_number() over(partition by uid1 order by max(sim) desc) user_rn
- from tb_neighbour a12
- join (select uid,iid from tb_behavior) a2
- on a12.uid2 = a2.uid
- join (select uid,collect_set(iid) iids1 from tb_behavior group by uid) a1
- on a12.uid1 = a1.uid
- where not array_contaions(iids1,a2.iid)
- group by uid1,iid
- ) tb_rec
- where user_rn <= 500
这里说明下包括上面的top30邻居和用户top500的最大推荐列表都是工程优化,截断节约些存储;具体读者可以根据自己业务需要进行设置;
然后大概说下各个表的含义:a1表是用户已消费过的物品,a2表是用户每个邻居喜好的物品;那么也就是说从邻居喜好的物品中过滤掉已经消费的
物品整体根据共现相似度进行排序。
思考
但思路很简单、实际作者开发中总会遇到各种各样的问题,下面就捡几个主要的和大家一起讨论下:
- 1.join引起的数据倾斜问题:tb_neighbour表很大,往往热点物品会占据80%的曝光和消费记录,如何解决?
- 2.增量更新问题:上面的框架,tb_behavior表每次都是全量计算,是否能改造成增量更新邻居表和推荐结果,并减少计算时间呢?
join引起的数据倾斜问题
先思考问题1,既然我们目的是求相似邻居,物品join只是为了关联上一组用户对,那自然的想法是可以根据feed做近似采样、相似度精度也几乎无损失。
下面我试着实现下这种思路:
- with tb_behavior_sample as (
- select uid,iid
- from (
- select uid
- ,iid
- ,row_number() over(partition by iid order by rand()) feed_rn
- from tb_behavior
- ) bh
- where feed_rn <= 50000
- )
- select uid1,uid2,sim
- from (
- select uid1
- ,uid2
- ,cnt12 / sqrt(cnt1*cnt2) sim
- ,row_number() over(partition by uid1 order by cnt12 / sqrt(cnt1*cnt2) desc) sim_rn
- from (
- select a.uid uid1
- ,b.uid uid2
- ,count(a.iid) cnt12
- from tb_behavior_sample a
- join tb_behavior_sample b
- on a.iid = b.iid
- where a.uid <> b.uid
- group by a.uid,b.uid
- ) a12
- join (select uid,count(iid) cnt1 from tb_behavior group by uid) a1
- on a12.uid1 = a1.uid
- join (select uid,count(iid) cnt2 from tb_behavior group by uid) a2
- on a12.uid1 = a2.uid
- ) tb_neighbour
- where sim > 0.1 and sim_rn <= 30
这里用了hive的with as语法,读者可自行查阅,篇幅有限,就不展开了;feed_rn就是随机采样了50000条,实际操作时读者可以先统计下item的分布、大概找到一个阈值;
比如取top10的item的出现次数作为阈值;那计算相似度时分子最多减小10,分母不变。这对大多数情况精度应该足够了,而且因为避免了数据倾斜,大大降低了计算时间。
增量更新问题
问题2是一个工程问题,lambda架构能使初始结果效果不错,可直接上线灰度了;在此基础上再加小时或者天增量;kappa架构相对就比较繁琐、需要一开始就设计增量流程。
精度方面也需要一定的累积;不过如何选择,读者可以根据自己的数据量和熟悉程度自行选择;作者这里仅以kappa架构说明。
重新review上面sql,我们发现我们仅需要记录下cnt12,cnt1,cnt2,iids1这些计算关键即可,其中iids2是用户邻居喜好的物品数组;数值类型可累加更新、
数组类型合并起来比较麻烦,一种解决方案是注册UDF;这里采取另一种这种的方案:把iids1合并成字符串,过滤的时候再分割为字符串数组。
- with tb_behavior_sample_incr as (
- select uid,iid
- from (
- select uid
- ,iid
- ,row_number() over(partition by iid order by rand()) feed_rn
- from tb_behavior_incr
- ) bh
- where feed_rn <= 50000
- )
- insert overwrite table tb_neighbour
- select uid1,uid2,sim
- from (
- select uid1
- ,uid2
- ,sum(cnt12) / sqrt(sum(cnt1)*sum(cnt2)) sim
- ,row_number() over(partition by uid1 order by sum(cnt12) / sqrt(sum(cnt1)*sum(cnt2)) desc) sim_rn
- from (
- select uid1,uid2,cnt12,cnt1,cnt2
- from tb_neighbour
- union all
- select a.uid uid1
- ,b.uid uid2
- ,count(a.iid) cnt12
- ,cnt1
- ,cnt2
- from tb_behavior_sample_incr a
- join tb_behavior_sample_incr b
- on a.iid = b.iid
- where a.uid <> b.uid
- group by a.uid,b.uid
- ) a12
- join (select uid,count(iid) cnt1 from tb_behavior_incr group by uid) a1
- on a12.uid1 = a1.uid
- join (select uid,count(iid) cnt2 from tb_behavior_incr group by uid) a2
- on a12.uid1 = a2.uid
- group by uid1,uid2
- ) tb_neighbour
- where sim > 0.1 and sim_rn <= 30
其中tb_behavior_sample_incr,tb_behavior_incr是相应tb_behavior_sample,tb_behavior的增量表;使用union all和group by聚合相同用户对的结果
kappa架构初次计算即是增量,不断累积每次增量的结果更新tb_neighbour;相当于lambda初始全量计算的一种回放,直至追到最新的时间分区。
- insert overwrite table tb_user_consume
- select uid,substring_index(concat_ws(",",collect_list(iids1)),",",10000) iids1
- from (
- select uid,concat_ws(",",collect_set(cast(iid as string))) iids1
- from tb_behavior_incr
- union all
- select uid,iids1
- from tb_user_consume
- ) a
- group by uid
- select uid1,iid
- from (
- select uid1
- ,iid
- ,max(sim) score
- ,row_number() over(partition by uid1 order by max(sim) desc) user_rn
- from tb_neighbour a12
- join (select uid,cast(iid as string) iid from tb_behavior_incr) a2
- on a12.uid2 = a2.uid
- join (select uid,split(iids1,",") iids1 from tb_user_consume) a1
- on a12.uid1 = a1.uid
- where not array_contaions(iids1,a2.iid)
- group by uid1,iid
- ) tb_rec
- where user_rn <= 500
使用tb_user_consume缓存用户最近消费的前10000条记录,将用户邻居最新喜好物品推荐给用户。
写在后面的话
呼!终于写完了;虽然说有了上面这一套操作,UCF推荐基本完成;但有没有更好的方式呢?我想应该就是embedding大法了吧;比如item2vec对用户聚类,根据聚类
推荐;再或者根据好友关系,推荐好友喜好的物品。前者表征更细致,值得一说的是其也有负采样策略和checkpoint增量更新;后者好友信任度更高,解释性更强。