系列随笔:
(二)基于商品属性的相似商品推荐算法——Flink SQL实时计算实现商品的隐式评分
(三)基于商品属性的相似商品推荐算法——批量处理商品属性,得到属性前缀及完整属性字符串
(四)基于商品属性的相似商品推荐算法——推荐与评分高的商品属性相似的商品
2020.04.15 补充:协同过滤推荐算法.pptx
提取码:4tds
注:如果你没有使用日志埋点和实时计算(接口直接累计也是可行的),你可以直接跳到这一节~
Flink SQL实时计算实现商品的隐匿评分
一、导入log service日志源表
二、导入评分配置维度表(用户行为的评分配置)
三、导入用户商品评分维表
四、用户评分结果表
四、预处理日志数据
-- 处理日志数据 CREATE VIEW probe_log0_view AS SELECT t1.cid, CAST(memberCode as INT) as memberCode, t1.event, t1.eventApp, TO_TIMESTAMP(CAST(CAST(__timestamp__ as DOUBLE) as BIGINT)*1000) as eventTime, CAST(IF (SUBSTRING(t1.eventProps,0,1)='%', REGEXP_EXTRACT(t1.eventProps, concat(t2.code_name,'\%22:(\d+),'), 1), JSON_VALUE (t1.eventProps, concat('$.',t2.code_name))) as INT) as goodsCode, t2.score FROM probe_log0 t1 LEFT JOIN rc_config_dimension FOR SYSTEM_TIME AS OF PROCTIME() AS t2 ON t1.event=t2.event AND t2.status=1 WHERE t1.event IN ('viewGoods','shareGoods','collectGoods','addToCart');
注:eventProps为埋点的扩展json数据,因为小程序的埋点不太规范,所以加了额外的判断;正常来说,直接使用 JSON_VALUE 函数即可
五、写入结果表
-- 入库 INSERT INTO rc_member_goods (member_code, cid, goods_code, score, update_time) SELECT t1.memberCode, t1.cid, t1.goodsCode, CAST(IF(t2.score IS NOT NULL, t2.score, 0) + SUM(t1.score) as INT) AS score, MAX(t1.eventTime) as update_time FROM probe_log0_view t1 LEFT JOIN rc_member_goods_dimension FOR SYSTEM_TIME AS OF PROCTIME() AS t2 ON t1.memberCode=t2.member_code AND t1.cid=t2.cid AND t1.goodsCode=t2.goods_code WHERE t1.goodsCode IS NOT NULL AND (t1.eventTime > t2.update_time OR t2.update_time IS NULL) GROUP BY t1.memberCode, t1.cid, t1.goodsCode, t2.score;
注:这里的难点在于 CAST(IF(t2.score IS NOT NULL, t2.score, 0) + SUM(t1.score) as INT) AS score 和 AND (t1.eventTime > t2.update_time OR t2.update_time IS NULL)
意思是:如果rc_member_goods表中没有记录的,就直接加入;如果 rc_member_goods 中有记录的,则判断 eventTime 是否大于 上前的更新时间(防止重复更新),最后累计上当前的日志分
PS:如果没有 t2.update_time IS NULL 则左连接会变成 left outer join