ALS召回、LR、GBDT排序的實戰，A/B Test

在生成ALS和LR模型以後，接下來就可以用在代碼中了。

首先ALS，其實在數據已經存在數據庫中了，只要從中取出來，去掉個逗號之類的就好

@Service
public class RecommendService {
    @Autowired
    private RecommendDOMapper recommendDOMapper;
    //找回數據，根據userid召回shopidList
    public List<Integer> recall(Integer userId){
        RecommendDO recommendDO = recommendDOMapper.selectByPrimaryKey(userId);
        if (recommendDO == null){
            recommendDO = recommendDOMapper.selectByPrimaryKey(99999);
        }
        String[] shopIdArr = recommendDO.getRecommend().split(",");
        List<Integer> shopIdList = new ArrayList<>();
        for (int i = 0 ; i < shopIdArr.length ; i ++){
            shopIdList.add(Integer.valueOf(shopIdArr[i]));
        }
        return shopIdList;
    }
}

對於LR：

@Service
public class RecommendSortService {
    private SparkSession spark;
    private LogisticRegressionModel lrModel;
    @PostConstruct
    public void init(){
        //初始化spark運行環境
        spark = SparkSession.builder()
                .master("local")
                .appName("DianpingApp")
                .getOrCreate();
        lrModel = LogisticRegressionModel.load("file:///F:/mouseSpace/project/background/lr/lrmodel");
    }
    public List<Integer> sort(List<Integer> shopIdList , Integer userId){
        //需要根據lrmodel所需要的11維的x生成特徵，然後調用預測方法
        List<ShopSortModel> list = new ArrayList<>();
        for (Integer shopId : shopIdList){
            //造的假數據
            Vector v = Vectors.dense(1,0,0,0,0,1,0.6,0,0,1,0);
            Vector result = lrModel.predictProbability(v);
            double[] arr = result.toArray();
            double score = arr[1];
//            lrModel.predict(v);       如果用這個，就是返回1或者0
            ShopSortModel shopSortModel = new ShopSortModel();
            shopSortModel.setShopId(shopId);
            shopSortModel.setScore(score);
            list.add(shopSortModel);
        }
        list.sort(new Comparator<ShopSortModel>() {
            @Override
            public int compare(ShopSortModel o1, ShopSortModel o2) {
                if (o1.getScore() < o2.getScore()){
                    return -1;
                }else if (o1.getScore() > o2.getScore()){
                    return 1;
                }else {
                    return 0;
                }
            }
        });
        return list.stream().map(shopSortModel -> shopSortModel.getShopId()).collect(Collectors.toList());
    }
}

代碼中自己造了一個數據，所以結果會有些偏差。

對於GBDT

跟lr算法非常像

public class GBDTRecommendSortService {
    private SparkSession spark;
    private GBTClassificationModel gbtClassificationModel;
    @PostConstruct
    public void init(){
        //初始化spark運行環境
        spark = SparkSession.builder()
                .master("local")
                .appName("DianpingApp")
                .getOrCreate();
        gbtClassificationModel = GBTClassificationModel.load("file:///F:/mouseSpace/project/background/lr/gbdtmodel");
    }
    public List<Integer> sort(List<Integer> shopIdList , Integer userId){
        //需要根據lrmodel所需要的11維的x生成特徵，然後調用預測方法
        List<ShopSortModel> list = new ArrayList<>();
        for (Integer shopId : shopIdList){
            //造的假數據
            Vector v = Vectors.dense(1,0,0,0,0,1,0.6,0,0,1,0);
            Vector result = gbtClassificationModel.predictProbability(v);
            double[] arr = result.toArray();
            double score = arr[1];
//            lrModel.predict(v);       如果用這個，就是返回1或者0
            ShopSortModel shopSortModel = new ShopSortModel();
            shopSortModel.setShopId(shopId);
            shopSortModel.setScore(score);
            list.add(shopSortModel);
        }
        list.sort(new Comparator<ShopSortModel>() {
            @Override
            public int compare(ShopSortModel o1, ShopSortModel o2) {
                if (o1.getScore() < o2.getScore()){
                    return -1;
                }else if (o1.getScore() > o2.getScore()){
                    return 1;
                }else {
                    return 0;
                }
            }
        });
        return list.stream().map(shopSortModel -> shopSortModel.getShopId()).collect(Collectors.toList());
    }
}

---

A/B Test

它可以幫助我們決策算法的好壞，提供更多的真實依據的手段。

在真實場景中，假如現有的是LR算法，那麼我現在馬上在線上換成GBDT，當然是有很大風險的，那麼AB TEST就出現了，假如有10條數據，我可以分5條用lr算法，5條用GBDT算法，然後將兩個依次穿插，形成一個結果集發給前端，然後通過記錄點擊率來驗證哪種算法更好。