百度电影推荐比赛参赛：评分预测问题.zip

package baidu.nb; import java.util.ArrayList; import java.util.HashMap; import java.util.List; import java.util.Map.Entry; import org.apache.log4j.Logger; import baidu.entity.Record; /** * FNM(Factorized Neighborhood Model)<b/> * * 使用 mu+bu+bi+qi*(xu+yu)进行预测 * * @author WangFengwei */ public class FNM extends AbstractMethod { protected int feature; protected HashMap<Integer, HashMap<Integer, Double>> userRates = new HashMap<Integer, HashMap<Integer, Double>>(); protected List<List<Double>> P = new ArrayList<List<Double>>(); // userCount // × feature protected List<List<Double>> Q = new ArrayList<List<Double>>(); // itemCount // × feature protected List<List<Double>> X = new ArrayList<List<Double>>(); // itemCount // × feature protected List<List<Double>> Y = new ArrayList<List<Double>>(); // itemCount // × feature protected static final Logger LOG = Logger.getLogger(FNM.class); public FNM(int feature, int iterCount, double alpha, double lambda) { this.feature = feature; this.iterCount = iterCount; this.alpha = alpha; this.lambda = lambda; } public static void main(String[] args) { int feature = 50, iterCount = 100; double alpha = 0.01, lamda = 0.002; FNM fnm = new FNM(feature, iterCount, alpha, lamda); // fnm.crossValidate(); String trainf = "resource/processedData/trainingSet"; String predictf = "resource/processedData/predict"; fnm.train(trainf); fnm.initPredict(predictf, false); fnm.predict(); fnm.outputPredict("resource/zjl/BaseLine-iter" + iterCount + "-alpha" + alpha); } @Override public void train(String trainf) { try { // 读取record readData(trainf); // 初始化参数 initParas(); // 初始化Bui trainBui(); // FNM训练 trainFNM(); } catch (Exception e) { e.printStackTrace(); } } // TODO 梯度下降有问题，会发散 private void trainFNM() { double localAlpha = alpha; for (int iter = 0; iter < iterCount; iter++) { for (Integer userid : userRates.keySet()) { // Calculate pu List<Double> pu = P.get(userid); List<Double> pu1 = newRandList(feature, 0, 0); List<Double> pu2 = newRandList(feature, 0, 0); double rateCoef = 0d, binCoef = 0d; for (Entry<Integer, Double> itemrate : userRates.get(userid) .entrySet()) { int itemid = itemrate.getKey(); double rate = itemrate.getValue(); double eui = rate - bPredict(userid, itemid); for (int f = 0; f < feature; f++) { pu1.set(f, pu1.get(f) + eui * X.get(itemid).get(f)); rateCoef += eui * eui; // TODO problem? pu2.set(f, pu2.get(f) + Y.get(itemid).get(f)); binCoef += 1;// 1*1; } } rateCoef = Math.pow(rateCoef / feature, -0.5d); binCoef = Math.pow(binCoef, -0.5d); for (int f = 0; f < feature; f++) { pu.set(f, pu1.get(f) * rateCoef + pu2.get(f) * binCoef); } // gradient descent RateInfo bu = users.get(userid); // List<Double> sum = newRandList(feature, 0, 0); for (Entry<Integer, Double> itemrate : userRates.get(userid) .entrySet()) { int itemid = itemrate.getKey(); double rate = itemrate.getValue(); double pui = predict(userid, itemid); double Eui = rate - pui; List<Double> Qi = Q.get(itemid); RateInfo bi = items.get(itemid); // LOG.info(String.format("FNM\titer:%d\trate:%s\tpredict:%f", // iter, df.format(rate), pui)); System.out.format("FNM\titer:%d\trate:%s\tpredict:%f\n", iter, df.format(rate), pui); for (int f = 0; f < feature; f++) { double gradientQif = -Eui * pu.get(f) + lambda * Qi.get(f); Qi.set(f, Qi.get(f) - localAlpha * gradientQif); double gradientBu = -Eui + lambda * bu.getAvg(); bu.setAvg(bu.getAvg() - localAlpha * gradientBu); double gradientBi = -Eui + lambda * bi.getAvg(); bi.setAvg(bi.getAvg() - localAlpha * gradientBi); } for (Entry<Integer, Double> jrate : userRates.get(userid) .entrySet()) { int j = jrate.getKey(); double r = jrate.getValue(); List<Double> Xj = X.get(j); List<Double> Yj = Y.get(j); for (int f = 0; f < feature; f++) { double gradientXjf = -rateCoef * (r - bPredict(userid, j)) * Eui * Qi.get(f) + lambda * Xj.get(f); Xj.set(f, Xj.get(f) - localAlpha * gradientXjf); double gradientYjf = -binCoef * Eui * Qi.get(f) + lambda * Yj.get(f); Yj.set(f, Yj.get(f) - localAlpha * gradientYjf); } } } } localAlpha *= 0.95; // calculate RMSE double rmse = calcRMSE(); LOG.info("FNM\titer:" + iter + "\trmse:" + rmse); } } /** * 使用梯度下降求解Bui */ private void trainBui() { double lastRMSE = 100d; double localAlpha = alpha; for (int iter = 0; iter < iterCount; iter++) { for (Record rd : records) { int u = rd.getUserId() - 1; int i = rd.getMovieId() - 1; double rui = rd.getScore(); double pui = predict(u, i); double bu = users.get(u).getAvg(); double bi = items.get(i).getAvg(); double gradientBu = lambda * bu - (rui - pui); double gradientBi = lambda * bi - (rui - pui); users.get(u).setAvg(bu - localAlpha * gradientBu); items.get(i).setAvg(bi - localAlpha * gradientBi); } double rmse = calcRMSE(); if (rmse > 1) localAlpha *= 1.05; else localAlpha *= 0.95; LOG.info("trainBui\titer:" + iter + "\trmse:" + rmse); if (lastRMSE - rmse < 0.00001) break; else lastRMSE = rmse; } } private double bPredict(int u, int i) { return mu + users.get(u).getAvg() + items.get(i).getAvg(); } @Override protected double predict(int userid, int itemid) { return bPredict(userid, itemid) + getProduct(Q.get(itemid), P.get(userid)); } @Override protected void initParas() { double tot = 0d; for (Record record : records) { int userid = record.getUserId() - 1; int itemid = record.getMovieId() - 1; double rate = record.getScore(); if (!users.containsKey(userid)) { users.put(userid, new RateInfo(rate, 1)); userRates.put(userid, new HashMap<Integer, Double>()); } else { users.get(userid).addRate(rate); userRates.get(userid).put(itemid, rate); } if (!items.containsKey(itemid)) items.put(itemid, new RateInfo(rate, 1)); else items.get(itemid).addRate(rate); tot += rate; } mu = tot / records.size(); for (RateInfo bu : users.values()) { bu.calcAvg(); P.add(newRandList(feature, 0, 0)); } for (RateInfo bi : items.values()) { bi.calcAvg(); Q.add(newRandList(feature, 0d, Math.random() / feature)); X.add(newRandList(feature, 0d, Math.random() / feature)); Y.add(newRandList(feature, 0d, Math.random() / feature)); } LOG.info("Init Over: users=" + users.size() + " items=" + items.size()); } @Override public String toString() { return "FNM-iter" + iterCount + "-alpha" + df.format(alpha) + "-lamda" + df.format(lambda); } @Override protected double calcCost() { throw new RuntimeException("Not implemeted yet..."); } }