Entire Space Multi‑Task Model (ESMM) for Post‑Click Conversion Rate Estimation

The article, authored by Gao Yue from Alibaba and sourced from Zhihu, reviews the SIGIR 2018 paper "Entire Space Multi‑Task Model: An Effective Approach for Estimating Post‑Click Conversion Rate". It presents the ESMM model, which addresses two key challenges in CVR estimation: Sample Selection Bias (SSB) and Data Sparsity (DS).

In industrial recommendation systems, tasks extend beyond single‑goal CTR prediction to downstream actions such as comments, favorites, add‑to‑cart, purchases, and watch time. The paper formulates CVR as the probability of conversion given a click, and distinguishes it from CTCVR (conversion given exposure). Traditional CVR models suffer from bias because training data (clicked samples) differ from the full sample space used at inference.

ESMM solves these problems by employing Multi‑Task Learning (MTL) with two tasks: CTR and CTCVR. It shares the same feature embeddings for both tasks and learns an implicit pCVR variable without direct supervision. The model jointly optimizes the two tasks on the entire sample space, allowing CVR to be inferred implicitly from the learned CTR and CTCVR outputs.

The objective function combines the losses of the CTR and CTCVR tasks, effectively learning CVR without requiring explicit labels for unclicked samples.

Implementation details are provided using Alibaba's EasyRec recommendation framework. The following code snippets illustrate the core components:

def build_predict_graph(self):
    """Forward function.
    Returns:
      self._prediction_dict: Prediction result of two tasks.
    """
    # ... (omitted tensor generation logic)
    cvr_tower_name = self._cvr_tower_cfg.tower_name
    dnn_model = dnn.DNN(self._cvr_tower_cfg.dnn, self._l2_reg, name=cvr_tower_name, is_training=self._is_training)
    cvr_tower_output = dnn_model(all_fea)
    cvr_tower_output = tf.layers.dense(inputs=cvr_tower_output, units=1, kernel_regularizer=self._l2_reg, name='%s/dnn_output' % cvr_tower_name)
    ctr_tower_name = self._ctr_tower_cfg.tower_name
    dnn_model = dnn.DNN(self._ctr_tower_cfg.dnn, self._l2_reg, name=ctr_tower_name, is_training=self._is_training)
    ctr_tower_output = dnn_model(all_fea)
    ctr_tower_output = tf.layers.dense(inputs=ctr_tower_output, units=1, kernel_regularizer=self._l2_reg, name='%s/dnn_output' % ctr_tower_name)
    tower_outputs = {cvr_tower_name: cvr_tower_output, ctr_tower_name: ctr_tower_output}
    self._add_to_prediction_dict(tower_outputs)
    return self._prediction_dict

def build_loss_graph(self):
    """Build loss graph.
    Returns:
      self._loss_dict: Weighted loss of ctr and cvr.
    """
    cvr_tower_name = self._cvr_tower_cfg.tower_name
    ctr_tower_name = self._ctr_tower_cfg.tower_name
    cvr_label_name = self._label_name_dict[cvr_tower_name]
    ctr_label_name = self._label_name_dict[ctr_tower_name]
    ctcvr_label = tf.cast(self._labels[cvr_label_name] * self._labels[ctr_label_name], tf.float32)
    cvr_loss = tf.keras.backend.binary_crossentropy(ctcvr_label, self._prediction_dict['probs_ctcvr'])
    cvr_loss = tf.reduce_sum(cvr_loss, name="ctcvr_loss")
    self._loss_dict['weighted_cross_entropy_loss_%s' % cvr_tower_name] = self._cvr_tower_cfg.weight * cvr_loss
    ctr_loss = tf.reduce_sum(tf.nn.sigmoid_cross_entropy_with_logits(labels=tf.cast(self._labels[ctr_label_name], tf.float32), logits=self._prediction_dict['logits_%s' % ctr_tower_name]), name="ctr_loss")
    self._loss_dict['weighted_cross_entropy_loss_%s' % ctr_tower_name] = self._ctr_tower_cfg.weight * ctr_loss
    return self._loss_dict

def build_metric_graph(self, eval_config):
    """Build metric graph.
    Args:
      eval_config: Evaluation configuration.
    Returns:
      metric_dict: Calculate AUC of ctr, cvr and ctrvr.
    """
    metric_dict = {}
    cvr_tower_name = self._cvr_tower_cfg.tower_name
    ctr_tower_name = self._ctr_tower_cfg.tower_name
    cvr_label_name = self._label_name_dict[cvr_tower_name]
    ctr_label_name = self._label_name_dict[ctr_tower_name]
    for metric in self._cvr_tower_cfg.metrics_set:
        ctcvr_label_name = cvr_label_name + '_ctcvr'
        cvr_dtype = self._labels[cvr_label_name].dtype
        self._labels[ctcvr_label_name] = self._labels[cvr_label_name] * tf.cast(self._labels[ctr_label_name], cvr_dtype)
        metric_dict.update(self._build_metric_impl(metric, loss_type=self._cvr_tower_cfg.loss_type, label_name=ctcvr_label_name, num_class=self._cvr_tower_cfg.num_class, suffix='_ctcvr'))
        cvr_label_masked_name = cvr_label_name + '_masked'
        ctr_mask = self._labels[ctr_label_name] > 0
        self._labels[cvr_label_masked_name] = tf.boolean_mask(self._labels[cvr_label_name], ctr_mask)
        pred_prefix = 'probs' if self._cvr_tower_cfg.loss_type == LossType.CLASSIFICATION else 'y'
        pred_name = '%s_%s' % (pred_prefix, cvr_tower_name)
        self._prediction_dict[pred_name + '_masked'] = tf.boolean_mask(self._prediction_dict[pred_name], ctr_mask)
        metric_dict.update(self._build_metric_impl(metric, loss_type=self._cvr_tower_cfg.loss_type, label_name=cvr_label_masked_name, num_class=self._cvr_tower_cfg.num_class, suffix='_%s_masked' % cvr_tower_name))
    for metric in self._ctr_tower_cfg.metrics_set:
        metric_dict.update(self._build_metric_impl(metric, loss_type=self._ctr_tower_cfg.loss_type, label_name=ctr_label_name, num_class=self._ctr_tower_cfg.num_class, suffix='_%s' % ctr_tower_name))
    return metric_dict

The article also notes practical observations: experiments on the public AliCCP dataset show a pronounced "seesaw" effect where improving CTR often harms CVR and vice‑versa. Links to the EasyRec repository and the ESMM implementation are provided.

References include the original SIGIR 2018 paper, the EasyRec documentation, and additional resources on multi‑task learning models such as MMoE, PLE, and DBMTL.