GENAI-2415 Add support for new inferred experiment. Adjust topic restrictions for personalization (#1172)

* Add support for new experiment. Adjust clamping for personalization

* fix name

Author: rolf-moz

merino/curated_recommendations/ml_backends/static_local_model.py

@@ -12,6 +12,7 @@
 from merino.curated_recommendations.protocol import ExperimentName
 
 INFERRED_LOCAL_EXPERIMENT_NAME = ExperimentName.INFERRED_LOCAL_EXPERIMENT.value
+INFERRED_LOCAL_EXPERIMENT_NAME_V2 = ExperimentName.INFERRED_LOCAL_EXPERIMENT_V2.value
 LOCAL_AND_SERVER_V1 = "local-and-server"
 LOCAL_ONLY_V1 = "local-only"
 LOCAL_ONLY_BRANCH_NAME = LOCAL_ONLY_V1
@@ -292,6 +293,8 @@ def get(
             if (
                 experiment_name == INFERRED_LOCAL_EXPERIMENT_NAME
                 or experiment_name == f"optin-{INFERRED_LOCAL_EXPERIMENT_NAME}"
+                or experiment_name == INFERRED_LOCAL_EXPERIMENT_NAME_V2
+                or experiment_name == f"optin-{INFERRED_LOCAL_EXPERIMENT_NAME_V2}"
             ):
                 ## switch on branch name
                 if experiment_branch == LOCAL_AND_SERVER_BRANCH_NAME:

merino/curated_recommendations/protocol.py

@@ -103,6 +103,7 @@ class ExperimentName(str, Enum):
     NEW_TAB_CUSTOM_SECTIONS_EXPERIMENT = "new-tab-custom-sections"
     # Experiment for doing local reranking of popular today via inferred interests
     INFERRED_LOCAL_EXPERIMENT = "new-tab-automated-personalization-local-ranking"
+    INFERRED_LOCAL_EXPERIMENT_V2 = "new-tab-automated-personalization-local-ranking-2"
 
 
 @unique

merino/curated_recommendations/sections.py

@@ -561,6 +561,7 @@ def get_top_story_list(
     extra_count: int = 0,
     extra_source_depth: int = 10,
     rescaler: ExperimentRescaler | None = None,
+    relax_constraints_for_personalization=False,
 ) -> list[CuratedRecommendation]:
     """Build a top story list of top_count items from a full list. Adds some extra items from further down
     in the list of recs with some care to not use the same topic more than once.
@@ -576,6 +577,8 @@ def get_top_story_list(
      rescaler: Optional rescaler associated with the experiment or surface
     Returns: A list of top stories
     """
+    constraint_scale = 2.0 if relax_constraints_for_personalization else 1.0
+
     fresh_story_prob = rescaler.fresh_items_top_stories_max_percentage if rescaler else 0
     total_story_count = top_count + extra_count
 
@@ -588,7 +591,7 @@ def get_top_story_list(
     )
     non_throttled = items[len(items_throttled_fresh) + len(unused_fresh) :]
 
-    balancer: ArticleBalancer = ArticleBalancer(top_count)
+    balancer: ArticleBalancer = ArticleBalancer(round(top_count * constraint_scale))
     topic_limited_stories, remaining_stories = balancer.add_stories(
         items_throttled_fresh, top_count
     )
@@ -597,7 +600,7 @@ def get_top_story_list(
     if len(second_pass_candidates) > extra_source_depth * 2:
         second_pass_candidates = second_pass_candidates[extra_source_depth:]
 
-    balancer.set_limits_for_expected_articles(total_story_count)
+    balancer.set_limits_for_expected_articles(round(total_story_count * constraint_scale))
     topic_limited_stories, remaining_stories = balancer.add_stories(
         second_pass_candidates, total_story_count
     )
@@ -711,6 +714,7 @@ async def get_sections(
         top_stories_count,
         TOP_STORIES_SECTION_EXTRA_COUNT,
         rescaler=rescaler,
+        relax_constraints_for_personalization=personal_interests is not None,
     )
 
     # Get the story ids in top_stories section

tests/unit/curated_recommendations/ml_backends/test_static_local_model.py

@@ -31,6 +31,7 @@
 from merino.curated_recommendations.protocol import ExperimentName
 
 INFERRED_LOCAL_EXPERIMENT_NAME = ExperimentName.INFERRED_LOCAL_EXPERIMENT.value
+INFERRED_LOCAL_EXPERIMENT_NAME_V2 = ExperimentName.INFERRED_LOCAL_EXPERIMENT_V2.value
 
 TEST_SURFACE = "test_surface"
 
@@ -516,7 +517,14 @@ def test_process_passthrough_when_values_missing_even_with_matching_model(
     "experiment,branch,model_id,expect_private_nonempty",
     [
         (INFERRED_LOCAL_EXPERIMENT_NAME, LOCAL_AND_SERVER_BRANCH_NAME, LOCAL_AND_SERVER_V1, True),
+        (
+            INFERRED_LOCAL_EXPERIMENT_NAME_V2,
+            LOCAL_AND_SERVER_BRANCH_NAME,
+            LOCAL_AND_SERVER_V1,
+            True,
+        ),
         (INFERRED_LOCAL_EXPERIMENT_NAME, LOCAL_ONLY_BRANCH_NAME, LOCAL_ONLY_V1, False),
+        (INFERRED_LOCAL_EXPERIMENT_NAME_V2, LOCAL_ONLY_BRANCH_NAME, LOCAL_ONLY_V1, False),
         (
             "optin-" + INFERRED_LOCAL_EXPERIMENT_NAME,
             LOCAL_AND_SERVER_BRANCH_NAME,
@@ -536,7 +544,9 @@ def test_process_passthrough_when_values_missing_even_with_matching_model(
     ],
     ids=[
         "local_and_server_branch",
+        "local_and_server_branch_v2",
         "local_only_branch",
+        "local_only_branch_v2",
         "optin-local_and_server_branch",
         "optin-local_only_branch",
         "local_and_server_branch__no_model",

tests/unit/curated_recommendations/test_sections.py

@@ -865,6 +865,25 @@ def test_basic_topic_limiting(self):
         for ix, item in enumerate(result):
             assert item.receivedRank == ix
 
+    def test_basic_topic_limiting_with_personalization(self):
+        """Extra items should be chosen without repeating topics from top_count items."""
+        items = generate_recommendations(
+            item_ids=["a", "b", "c", "d", "e", "f"],
+            topics=["arts", "arts", "arts", "arts", "food", "government"],
+        )
+        result = get_top_story_list(
+            items,
+            top_count=6,
+            extra_count=0,
+            extra_source_depth=0,
+            relax_constraints_for_personalization=True,
+        )
+        top_ids = [i.corpusItemId for i in result]
+        assert len(result) == 6
+        top_ids[2] == "c"
+        for ix, item in enumerate(result):
+            assert item.receivedRank == ix
+
     def test_includes_extra_items_topic_limiting(self):
         """Extra items should be chosen without repeating topics from top_count items."""
         items = generate_recommendations(