bayes-smoothing

Click-Through Rate Estimation for Rare Events in Online Advertising

This program implements the algorithm desribed in [http://www.cs.cmu.edu/~xuerui/papers/ctr.pdf]

The fixed-pointed-iteration was accelerated by Steffensen method [https://en.wikipedia.org/wiki/Steffensen%27s_method].

Usage:

   final BayesSmoothingExecutor executor = new BayesSmoothingExecutor(1e-7, 100000);
    System.out.println("start generating bayes smoothing...");

    long t1 = System.currentTimeMillis();
    final List<BayesSmoothingResult> resultList = recordList
            .stream()
            .collect(Collectors.groupingBy(CreativeAdUnitRecord::getAdUnitId, Collectors.toList()))
            .entrySet()
            .stream()
            .parallel()
            .map(entry -> {
                final List<RecordEntry> samples = entry.getValue()
                        .stream()
                        //同一广告位下的样本按照创意ID进行聚合
                        .collect(Collectors.groupingBy(
                                CreativeAdUnitRecord::getCreativeId,
                                Collectors.mapping(CreativeAdUnitRecord::getEntry,
                                        Collectors.reducing(RecordEntry.of(), RecordEntry::merge))
                        ))
                        .entrySet()
                        .stream()
                        .map(Map.Entry::getValue)
                        //对样本进行过滤
                        .filter(re -> {
                            long clickCount = re.getClickCount();
                            long impressionCount = re.getImpressionCount();
                            return clickCount > 5 && impressionCount > 10000 && clickCount < impressionCount;
                        })
                        .collect(Collectors.toList());
                return new AbstractMap.SimpleEntry<>(entry.getKey(), samples);
            })
            .filter(entry -> entry.getValue().size() > 3)
            .map(entry -> executor.executeSmoothing(entry.getKey(), entry.getValue()))
            .collect(Collectors.toList());
    long t2 = System.currentTimeMillis();