Building an Advertising Recommendation Model with Python and PyTorch

Being a Python developer can feel contradictory: you are immersed in big‑data environments where everyone is "naked swimming" while you also have to analyze and block intrusive ads.

The example starts by gathering a large set of ad‑placement data, including app information, ad slot IDs, media IDs, material details, titles, descriptions, and other vector features.

To handle categorical fields such as pkgname , ver , slotid , mediaid , and material , label encoding is applied to both training and test sets:

for col in ["pkgname", "ver", "slotid", "mediaid", "material"]: lbl = LabelEncoder() lbl.fit(train_df[col].tolist() + test_df[col].tolist()) train_df[col] = lbl.transform(train_df[col]) test_df[col] = lbl.transform(test_df[col])

After encoding, textual features are transformed into vectors using Torch's Embedding layer, converting each categorical value into a dense representation based on the logarithm of its cardinality.

The core model is a multilayer perceptron (MLP) defined with PyTorch. It creates embedding dictionaries for each categorical field, concatenates them with other feature vectors, passes the result through configurable fully‑connected layers, applies ReLU and dropout, and finally outputs a logit:

class MLP(nn.Module): def __init__(self, category_dict, layers=[45 + 240, 32], dropout=False): super().__init__() self.category_dict = category_dict self.embedding_dict = { key: torch.nn.Embedding(self.category_dict[key] + 1, int(np.log2(self.category_dict[key]))) for key in category_dict.keys() } self.fc_layers = torch.nn.ModuleList() for _, (in_size, out_size) in enumerate(zip(layers[:-1], layers[1:])): self.fc_layers.append(torch.nn.Linear(in_size, out_size).to(device)) self.output_layer = torch.nn.Linear(layers[-1], 1).to(device) def forward(self, feed_dict, embed_dict): embedding_feet = {key: self.embedding_dict[key](feed_dict[key]) for key in self.category_dict.keys()} x = torch.cat(list(embedding_feet.values()), 1) x = torch.cat([x, embed_dict], 1) for idx, _ in enumerate(range(len(self.fc_layers))): x = self.fc_layers[idx](x) x = F.relu(x) x = F.dropout(x) logit = self.output_layer(x) return logit

Training is then launched ("Training starts~"), demonstrating a typical workflow for an ad recommendation pipeline, while acknowledging that more sophisticated techniques may be needed for production‑grade systems.