Add FIXED_BIAS_WEIGHTS positional bias type to ranking_agent

`FIXED_BIAS_WEIGHTS` is a new type where fixed positional bias weights for all slots are given as an array. These weights can be from user's pre-knowledge, learned from offline analysis (e.g. TopN randomization) and so on.

PiperOrigin-RevId: 617544784
Change-Id: I3cd891bffdc65626d3e0ba513f050827f74ce62a

Author: TF-Agents Team

tf_agents/bandits/agents/examples/v2/train_eval_ranking.py

@@ -64,12 +64,19 @@
     'bias_type',
     '',
     'Whether the agent models the positional '
-    'bias with the basis or the exponent changes. If unset, the'
+    'bias with the basis, the exponent or fixed bias weights. If unset, the'
     ' agent applies no positional bias.',
 )
 flags.DEFINE_float(
     'bias_severity', 1.0, 'The severity of the bias adjustment by the agent.'
 )
+flags.DEFINE_list(
+    'bias_weights',
+    [],
+    'The positional bias weights. For FIXED_BIAS_WEIGHTS type, the agent will'
+    ' use these weights to adjust the rewards. The length of the list must be'
+    ' equal to the number of slots.',
+)
 flags.DEFINE_bool(
     'bias_positive_only',
     False,
@@ -174,12 +181,15 @@ def _relevance_fn(global_obs, item_obs):
     positional_bias_type = ranking_agent.PositionalBiasType.BASE
   elif FLAGS.positional_bias_type == 'exponent':
     positional_bias_type = ranking_agent.PositionalBiasType.EXPONENT
+  elif FLAGS.positional_bias_type == 'fixed_bias_weights':
+    positional_bias_type = ranking_agent.PositionalBiasType.FIXED_BIAS_WEIGHTS
   else:
     raise NotImplementedError(
         'Positional bias type {} is not implemented'.format(
             FLAGS.positional_bias_type
         )
     )
+  positional_bias_weights = [float(w) for w in FLAGS.positional_bias_weights]
 
   agent = ranking_agent.RankingAgent(
       time_step_spec=environment.time_step_spec(),
@@ -190,6 +200,7 @@ def _relevance_fn(global_obs, item_obs):
       feedback_model=feedback_model,
       positional_bias_type=positional_bias_type,
       positional_bias_severity=FLAGS.bias_severity,
+      positional_bias_weights=positional_bias_weights,
       positional_bias_positive_only=FLAGS.bias_positive_only,
       summarize_grads_and_vars=True,
   )

tf_agents/bandits/agents/ranking_agent.py

@@ -38,8 +38,9 @@
 recommendation. The user is responsible for converting the observation to the
 syntax required by the agent.
 """
+
 import enum
-from typing import Optional, Text
+from typing import List, Optional, Text
 
 import tensorflow as tf  # pylint: disable=g-explicit-tensorflow-version-import
 from tf_agents.agents import tf_agent
@@ -127,6 +128,9 @@ class PositionalBiasType(enum.Enum):
   # et al. `Correcting for Selection Bias in Learning-to-rank Systems`
   # (WWW 2020).
   EXPONENT = 2
+  # The bias weight for each slot position is `bias_weights[k]`, where
+  # `bias_weights` is the given bias weight array and `k` is the position.
+  FIXED_BIAS_WEIGHTS = 3
 
 
 class RankingAgent(tf_agent.TFAgent):
@@ -144,6 +148,7 @@ def __init__(
       non_click_score: Optional[float] = None,
       positional_bias_type: PositionalBiasType = PositionalBiasType.UNSET,
       positional_bias_severity: Optional[float] = None,
+      positional_bias_weights: Optional[List[float]] = None,
       positional_bias_positive_only: bool = False,
       logits_temperature: float = 1.0,
       summarize_grads_and_vars: bool = False,
@@ -178,6 +183,8 @@ def __init__(
       positional_bias_type: Type of positional bias to use when training.
       positional_bias_severity: (float) The severity `s`, used for the `BASE`
         positional bias type.
+      positional_bias_weights: (float array) The positional bias weight for each
+        slot position.
       positional_bias_positive_only: Whether to use the above defined bias
         weights only for positives (that is, clicked items). If
         `positional_bias_type` is unset, this parameter has no effect.
@@ -230,6 +237,22 @@ def __init__(
         )
     self._positional_bias_type = positional_bias_type
     self._positional_bias_severity = positional_bias_severity
+    # Validate positional_bias_weights for FIXED_BIAS_WEIGHTS PositionalBiasType
+    if self._positional_bias_type == PositionalBiasType.FIXED_BIAS_WEIGHTS:
+      if positional_bias_weights is None:
+        raise ValueError(
+            'positional_bias_weights is None but should never be for'
+            ' FIXED_BIAS_WEIGHTS PositionalBiasType.'
+        )
+      elif len(positional_bias_weights) != self._num_slots:
+        raise ValueError(
+            'The length of positional_bias_weights should be the same as the'
+            ' number of slots. The length of positional_bias_weights is {} and'
+            ' the number of slots is {}.'.format(
+                len(positional_bias_weights), self._num_slots
+            )
+        )
+    self._positional_bias_weights = positional_bias_weights
     self._positional_bias_positive_only = positional_bias_positive_only
     if policy_type == RankingPolicyType.UNKNOWN:
       policy_type = RankingPolicyType.COSINE_DISTANCE
@@ -409,19 +432,27 @@ def _construct_sample_weights(self, reward, observation, weights):
           chosen_index + 1, self._num_slots, dtype=tf.float32
       )
       weights = multiplier * weights
-    if self._positional_bias_type != PositionalBiasType.UNSET:
-      batched_range = tf.broadcast_to(
-          tf.range(self._num_slots, dtype=tf.float32), tf.shape(weights)
+
+    if self._positional_bias_type == PositionalBiasType.UNSET:
+      return weights
+
+    batched_range = tf.broadcast_to(
+        tf.range(self._num_slots, dtype=tf.float32), tf.shape(weights)
+    )
+    if self._positional_bias_type == PositionalBiasType.BASE:
+      position_bias_multipliers = tf.pow(
+          batched_range + 1, self._positional_bias_severity
       )
-      if self._positional_bias_type == PositionalBiasType.BASE:
-        position_bias_multipliers = tf.pow(
-            batched_range + 1, self._positional_bias_severity
-        )
-      elif self._positional_bias_type == PositionalBiasType.EXPONENT:
-        position_bias_multipliers = tf.pow(
-            self._positional_bias_severity, batched_range
-        )
-      else:
-        raise ValueError('non-existing positional bias type')
-      weights = position_bias_multipliers * weights
+    elif self._positional_bias_type == PositionalBiasType.EXPONENT:
+      position_bias_multipliers = tf.pow(
+          self._positional_bias_severity, batched_range
+      )
+    elif self._positional_bias_type == PositionalBiasType.FIXED_BIAS_WEIGHTS:
+      position_bias_multipliers = tf.tile(
+          tf.expand_dims(self._positional_bias_weights, axis=0),
+          [batch_size, 1],
+      )
+    else:
+      raise ValueError('non-existing positional bias type')
+    weights = position_bias_multipliers * weights
     return weights

tf_agents/bandits/agents/ranking_agent_test.py

@@ -311,6 +311,8 @@ def testTrainAgentScoreFeedback(
           'positional_bias_type': ranking_agent.PositionalBiasType.BASE,
           'positional_bias_severity': 1.2,
           'positional_bias_positive_only': False,
+          'positional_bias_weights': None,
+          'expected_second_weight': 2.2974,  # 2**positional_bias_severity
       },
       {
           'feedback_model': ranking_agent.FeedbackModel.SCORE_VECTOR,
@@ -323,6 +325,8 @@ def testTrainAgentScoreFeedback(
           'positional_bias_type': ranking_agent.PositionalBiasType.EXPONENT,
           'positional_bias_severity': 1.3,
           'positional_bias_positive_only': False,
+          'positional_bias_weights': None,
+          'expected_second_weight': 1.3,  # positional_bias_severity
       },
       {
           'feedback_model': ranking_agent.FeedbackModel.SCORE_VECTOR,
@@ -335,6 +339,36 @@ def testTrainAgentScoreFeedback(
           'positional_bias_type': ranking_agent.PositionalBiasType.BASE,
           'positional_bias_severity': 1.0,
           'positional_bias_positive_only': True,
+          'positional_bias_weights': None,
+          'expected_second_weight': 2.0,  # 2**positional_bias_severity
+      },
+      {
+          'feedback_model': ranking_agent.FeedbackModel.SCORE_VECTOR,
+          'policy_type': ranking_agent.RankingPolicyType.DESCENDING_SCORES,
+          'batch_size': 2,
+          'global_dim': 3,
+          'item_dim': 4,
+          'num_items': 13,
+          'num_slots': 11,
+          'positional_bias_type': (
+              ranking_agent.PositionalBiasType.FIXED_BIAS_WEIGHTS
+          ),
+          'positional_bias_severity': None,
+          'positional_bias_positive_only': True,
+          'positional_bias_weights': [
+              0.1,
+              0.2,
+              0.3,
+              0.4,
+              0.5,
+              0.6,
+              0.7,
+              0.8,
+              0.9,
+              1.0,
+              1.1,
+          ],
+          'expected_second_weight': 0.2,  # positional_bias_weights[1]
       },
   ])
   def testPositionalBiasParams(
@@ -349,6 +383,8 @@ def testPositionalBiasParams(
       positional_bias_type,
       positional_bias_severity,
       positional_bias_positive_only,
+      positional_bias_weights,
+      expected_second_weight,
   ):
     if not tf.executing_eagerly():
       self.skipTest('Only works in eager mode.')
@@ -386,6 +422,7 @@ def testPositionalBiasParams(
         positional_bias_type=positional_bias_type,
         positional_bias_severity=positional_bias_severity,
         positional_bias_positive_only=positional_bias_positive_only,
+        positional_bias_weights=positional_bias_weights,
         optimizer=optimizer,
     )
     global_obs = tf.reshape(
@@ -426,12 +463,7 @@ def testPositionalBiasParams(
     agent.train(experience)
     weights = agent._construct_sample_weights(scores, observations, None)
     self.assertAllEqual(weights.shape, [batch_size, num_slots])
-    expected = (
-        2**positional_bias_severity
-        if positional_bias_type == ranking_agent.PositionalBiasType.BASE
-        else positional_bias_severity
-    )
-    self.assertAllClose(weights[-1, 1], expected)
+    self.assertAllClose(weights[-1, 1], expected_second_weight)
 
 
 if __name__ == '__main__':