最近flink较火,尝试使用flink做推荐功能试试,说干就干,话说flink-ml确实比较水,包含的算法较少,且只支持scala版本,以至flink1.9已经将flink-ml移除,看来是准备有大动作,但后期的实时推荐,flink能派上大用场。所幸基于物品的协同过滤算法相对简单,实现起来难度不大。先看目前推荐整体的架构。

先说一下用到的相似算法:
X=(x1, x2, x3, … xn),Y=(y1, y2, y3, … yn)
那么欧式距离为:

很明显,值越大,相似性越差,如果两者完全相同,那么距离为0。

第一步准备数据,数据的格式如下:

actionObject 是房屋的编号,actionType是用户的行为,包括曝光未点击,点击,收藏等。

下面的代码是从hdfs中获取数据,并将view事件的数据清除,其他的行为转化为分数


public static DataSet<Tuple2<Tuple2<String, String>, Float>> getData(ExecutionEnvironment env, String path) {
        DataSet<Tuple2<Tuple2<String, String>, Float>> res= env.readTextFile(path).map(new MapFunction<String, Tuple2<Tuple2<String, String>, Float>> (){

            @Override
            public Tuple2<Tuple2<String, String>, Float> map(String value) throws Exception {
                    JSONObject jj=JSON.parseObject(value);
                    if(RecommendUtils.getValidAction(jj.getString("actionType"))) {                     
                        return new Tuple2<>(new Tuple2<>(jj.getString("userId"),jj.getString("actionObject")),RecommendUtils.getScore(jj.getString("actionType")));                 
                    }else {
                        return null;
                    }

                }   
            }).filter(new FilterFunction<Tuple2<Tuple2<String, String>, Float>>(){
                @Override
                public boolean filter(Tuple2<Tuple2<String, String>, Float> value) throws Exception {           
                    return value!=null;
                }       
            });

           return res;
    }

数据经过简单的清洗后变成如下的格式

按照前两列聚合,

groupBy(0).reduce(new ReduceFunction<Tuple2<Tuple2<String, String>, Float>>() { 

                @Override
                public Tuple2<Tuple2<String, String>, Float> reduce(Tuple2<Tuple2<String, String>, Float> value1,
                        Tuple2<Tuple2<String, String>, Float> value2) throws Exception {
                    // TODO Auto-generated method stub
                    return new Tuple2<>(new Tuple2<>(value1.f0.f0, value1.f0.f1),(value1.f1+value2.f1)); 
                }

            })

结构变成

此时,理论上BJCY56167779_03,BJCY56167779_04 的相似度为 (4-3) ^2+(5-2) ^2, 再开方,继续前进。

去掉第一列,格式如下

因为:
(x1-y1)^2+(x2-y2)^2=x1^2+y1^2-2x1y1+x2^2+y2^2-2x2y2=x1^2+y1^2+x2^2+y2^2-2(x1y1+x2y2), 所以我们先求x1^2+x2^2的值,并注册为item表


.map(new MapFunction<Tuple2<String, Float>, Tuple2<String, Float>>() {
                @Override
                public Tuple2<String, Float> map(Tuple2<String, Float> value) throws Exception {
                    return new Tuple2<>(value.f0, value.f1*value.f1);
                }  
            }).
groupBy(0).reduce(new ReduceFunction<Tuple2<String, Float>>(){

                @Override
                public Tuple2<String, Float> reduce(Tuple2<String, Float> value1, Tuple2<String, Float> value2)
                        throws Exception {
                     Tuple2<String, Float> temp= new Tuple2<>(value1.f0, value1.f1 +  value2.f1);
                     return temp;
                }

}).map(new MapFunction<Tuple2<String, Float>, ItemDTO> (){

                @Override
                public ItemDTO map(Tuple2<String, Float> value) throws Exception {
                    ItemDTO nd=new ItemDTO();
                    nd.setItemId(value.f0);
                    nd.setScore(value.f1);
                    return nd;
                }

}); 

tableEnv.registerDataSet("item", itemdto); // 注册表信息

经过上面的转化,前半部分的值已经求出,下面要求出(x1y1+x2y2)的值

将上面的原始table再次转一下,变成下面的格式

代码如下:

.map(new MapFunction<Tuple2<String,List<Tuple2<String,Float>>>, List<Tuple2<Tuple2<String, String>, Float>>>() {

                @Override
                public List<Tuple2<Tuple2<String, String>, Float>> map(Tuple2<String,List<Tuple2<String,Float>>> value) throws Exception {
                    List<Tuple2<String, Float>> ll= value.f1;                   
                    List<Tuple2<Tuple2<String, String>, Float>> list = new ArrayList<>();
                    for (int i = 0; i < ll.size(); i++) {
                        for (int j = 0; j < ll.size(); j++) {
                            list.add(new Tuple2<>(new Tuple2<>(ll.get(i).f0, ll.get(j).f0),
                                    ll.get(i).f1 * ll.get(j).f1));
                        }
                    }
                    return list;        
                }

            })

tableEnv.registerDataSet("item_relation", itemRelation); // 注册表信息

下面就是将整个公式连起来,完成最后的计算。

Table similarity=tableEnv.sqlQuery("select ta.firstItem,ta.secondItem,"
        + "(sqrt(tb.score + tc.score - 2 * ta.relationScore)) as similarScore from item tb " +
        "inner join item_relation ta  on tb.itemId = ta.firstItem and ta.firstItem <> ta.secondItem "+
        "inner join item tc on tc.itemId = ta.secondItem "          
        );

DataSet<ItemSimilarDTO> ds=tableEnv.toDataSet(similarity, ItemSimilarDTO.class);

现在结构变成

感觉离终点不远了,上述结构依然不是我们想要的,我们希望结构更加清晰,如下格式

代码如下:

DataSet<RedisDataModel> redisResult= ds.map(new MapFunction<ItemSimilarDTO, Tuple2<String, Tuple2<String, Float>>> (){

            @Override
            public Tuple2<String, Tuple2<String, Float>> map(ItemSimilarDTO value) throws Exception {               
                return new Tuple2<String, Tuple2<String, Float>>(value.getFirstItem(), new Tuple2<>(value.getSecondItem(), value.getSimilarScore().floatValue()));
            }
        }).groupBy(0).reduceGroup(new GroupReduceFunction<Tuple2<String, Tuple2<String, Float>> , Tuple2<String, List<RoomModel>>>() { 

            @Override
            public void reduce(Iterable<Tuple2<String, Tuple2<String, Float>>> values,
                    Collector<Tuple2<String, List<RoomModel>>> out) throws Exception {

                List<RoomModel> list=new ArrayList<>();
                String key=null;
                for (Tuple2<String, Tuple2<String, Float>> t : values) {
                    key=t.f0;
                    RoomModel rm=new RoomModel();
                    rm.setRoomCode(t.f1.f0);
                    rm.setScore(t.f1.f1);
                    list.add(rm);
                }       

                //升序排序
                Collections.sort(list,new Comparator<RoomModel>(){
                    @Override
                    public int compare(RoomModel o1, RoomModel o2) {                                            
                        return o1.getScore().compareTo(o2.getScore());                      
                    }
                 });

                out.collect(new Tuple2<>(key,list));            
            }

        }).map(new MapFunction<Tuple2<String, List<RoomModel>>, RedisDataModel>(){

            @Override
            public RedisDataModel map(Tuple2<String, List<RoomModel>> value) throws Exception {
                RedisDataModel m=new RedisDataModel();
                m.setExpire(-1); 
                m.setKey(JobConstants.REDIS_FLINK_ITEMCF_KEY_PREFIX+value.f0);      
                m.setGlobal(true);
                m.setValue(JSON.toJSONString(value.f1));
                return m;
            }

        });

最终将这些数据存入redis中,方便查询

RedisOutputFormat redisOutput = RedisOutputFormat.buildRedisOutputFormat()
                    .setHostMaster(AppConfig.getProperty(JobConstants.REDIS_HOST_MASTER))
                    .setHostSentinel(AppConfig.getProperty(JobConstants.REDIS_HOST_SENTINELS))
                    .setMaxIdle(Integer.parseInt(AppConfig.getProperty(JobConstants.REDIS_MAXIDLE)))
                    .setMaxTotal(Integer.parseInt(AppConfig.getProperty(JobConstants.REDIS_MAXTOTAL))) 
                    .setMaxWaitMillis(Integer.parseInt(AppConfig.getProperty(JobConstants.REDIS_MAXWAITMILLIS)))
                    .setTestOnBorrow(Boolean.parseBoolean(AppConfig.getProperty(JobConstants.REDIS_TESTONBORROW)))
                    .finish();   
            redisResult.output(redisOutput);

            env.execute("itemcf");

大功告成,其实没有想象中的那么难。当然这里只是一个demo,实际情况还要进行数据过滤,多表join优化等。