Tencent Advertising Algorithm Competition 2020: Problem Overview, Data, Model Implementation, and Results

The 2020 Tencent Advertising Algorithm Competition final was held on August 3 at Tencent Binhai Building in Shenzhen, featuring a live broadcast and a sharing session by the internal second‑place team.

The competition task required predicting users' age and gender based on their ad click behavior and associated ad information, essentially a user‑profiling problem.

Provided data fields included time (day granularity), user_id, creative_id, click_times, ad_id, product_id, product_category, advertiser_id, industry, as well as the target labels age (1‑10) and gender (1 or 2).

For model input the team selected five ID sequences—creative_id, ad_id, advertiser_id, product_id, and industry—and applied ID mapping to unify rare and unseen IDs, as shown in the following code:

differ = set(train[col].unique()).symmetric_difference(set(test[col].unique()))
common = set(train[col].unique()) and (set(test[col].unique()))
for v in val_cnt[val_cnt == 1].index:
    id_map[v] = 0
for v in differ:
    id_map[v] = 0
for i, v in enumerate(common):
    id_map[v] = i + 1

Word2Vec embeddings were trained using the skip‑gram model with parameters min_count=1, size=256, window=10, and the resulting vectors were loaded for each selected ID column.

model = models.Word2Vec(list_d, sg=1, min_count=1, size=256, window=10, workers=48, iter=10)
We = []
if '0' in model.wv:
    for i in tqdm(range(len(model.wv.index2word))):
        We.append(model.wv[str(i)].reshape((1,-1)))
else:
    We.append(np.zeros((1,128)))
We = np.vstack(We)

Input construction involved various ordering strategies (forward, reverse, random shuffle) and duplication of records according to click_times, with sequence length capped at 95% of the maximum length.

The team experimented with several model architectures: LSTM, CNN‑Inception, and a transformer‑LSTM hybrid. The final LSTM implementation in Keras and PyTorch, as well as the CNN‑Inception model in PyTorch, are provided below.

## LSTM keras‑version
def LSTM(config, n_cls=10):
    cols = ['creative_id', 'ad_id', 'advertiser_id', 'product_id', 'industry']
    n_in = len(cols)
    inputs, outputs, max_len = [], [], []
    for i in range(n_in):
        We = np.load(f'./w2v_256_10/{cols[i]}_embedding_weight.npy')
        We = np.vstack([We, np.zeros(config.embeddingSize)])
        inp = Input(shape=(config.sequenceLength,), dtype="int32")
        x = Embedding(We.shape[0], We.shape[1], weights=[We], trainable=False)(inp)
        inputs.append(inp)
        outputs.append(x)
        del We
        gc.collect()
    embedding_model = Model(inputs, outputs)
    # ... (rest of the Keras LSTM definition) ...
    return model, lstm_model

class Inception(nn.Module):
    def __init__(self, cin, co, relu=True, norm=True):
        super(Inception, self).__init__()
        assert(co%4==0)
        cos=[co//4]*4
        self.activa=nn.Sequential()
        if norm: self.activa.add_module('norm', nn.BatchNorm1d(co))
        if relu: self.activa.add_module('relu', nn.ReLU(True))
        self.branch1 = nn.Sequential(OrderedDict([('conv1', nn.Conv1d(cin, cos[0], 1, stride=1)]))
        # ... (other branches) ...
    def forward(self, x):
        branch1=self.branch1(x)
        # ... (concatenate branches) ...
        return self.activa(torch.cat((branch1,branch2,branch3,branch4),1))

Performance results showed the best single model achieving approximately 1.475 on the semi‑final A leaderboard, with an ensemble reaching about 1.48, placing second on the internal leaderboard and fourteenth on the external leaderboard.