updates

app.py

@@ -58,8 +58,8 @@ def app(share=False, use_cluster_feats=False):
     instances, instance_ids = get_instances(cfg['instances_to_explain_path'])
 
     interp      = load_interp_space(cfg)
-    clustered_authors_df = interp['clustered_authors_df'][:200]
-    clustered_authors_df['fullText'] = clustered_authors_df['fullText'].map(lambda l: l[:3]) # Take at most 3 texts per author
+    clustered_authors_df = interp['clustered_authors_df'][:1000]
+    clustered_authors_df['fullText'] = clustered_authors_df['fullText']
 
     with gr.Blocks(title="Author Attribution Explainability Tool") as demo:
         # ── Big Centered Title ──────────────────────────────────────────

config/config.yaml

@@ -1,8 +1,8 @@
 # config.yaml
 instances_to_explain_path: "./datasets/hrs_explanations.json"
 instances_to_explain_url: "https://huggingface.co/datasets/miladalsh/explanation_tool_files/resolve/main/hrs_explanations_luar_clusters_18_balanced.json?/download=true"
-interp_space_path:    "./datasets/sentence_luar_interp_clusters_18.zip/"
-interp_space_url:    "https://huggingface.co/datasets/miladalsh/explanation_tool_files/resolve/main/luar_interp_space_cluster_18.zip?download=true"
+interp_space_path:    "./datasets/sentence_luar_interp_space_2_35/"
+interp_space_url:    "https://huggingface.co/datasets/miladalsh/explanation_tool_files/resolve/main/sentence_luar_interp_space_2_35.zip?download=true"
 gram2vec_feats_path:      "./datasets/gram2vec_feats.csv"
 gram2vec_feats_url:      "https://huggingface.co/datasets/miladalsh/explanation_tool_files/resolve/main/gram2vec_feats.csv?download=true"
 

utils/gram2vec_feat_utils.py

@@ -126,7 +126,7 @@ def highlight_both_spans(text, llm_spans, gram_spans):
 
 
 def show_combined_spans_all(selected_feature_llm, selected_feature_g2v, 
-                            llm_style_feats_analysis, background_authors_embeddings_df, task_authors_embeddings_df, visible_authors, predicted_author=None, ground_truth_author=None, max_num_authors=7):
+                            llm_style_feats_analysis, background_authors_embeddings_df, task_authors_embeddings_df, visible_authors, predicted_author=None, ground_truth_author=None, max_num_authors=4):
     """
     For mystery + 3 candidates:
      1. get llm spans via your existing cache+API

utils/interp_space_utils.py

@@ -79,7 +79,7 @@ def compute_g2v_features(clustered_authors_df: pd.DataFrame, task_authors_df: pd
             clustered_authors_df = pickle.load(f)
     
     else: # Else compute and cache
-        g2v_feats_df = vectorizer.from_documents(author_texts, batch_size=16)
+        g2v_feats_df = vectorizer.from_documents(author_texts, batch_size=8)
 
         print(f"Number of g2v features: {len(g2v_feats_df)}")
         print(f"Number of clustered_authors_df.authorID.tolist(): {len(clustered_authors_df.authorID.tolist())}")
@@ -471,11 +471,11 @@ def identify_style_features(author_texts: list[str], author_names: list[str], ma
             print(f"Cache miss. Computing features for authors: {author_names}")
     
     client = OpenAI(api_key=os.getenv("OPENAI_API_KEY"))
-    prompt = f"""Identify {max_num_feats} writing style features that are commonly found across the following texts. Do not extract spans. Just return the feature names as a list.
+    prompt = f"""Identify {max_num_feats} writing style features that are commonly between the authors texts.
     Author Texts:
-    \"\"\"{chr(10).join(author_texts)}\"\"\"
-    """
 
+    {author_texts}
+    """
     def _make_call():
         response = client.chat.completions.create(
             model="gpt-4o-mini",
@@ -495,7 +495,6 @@ def identify_style_features(author_texts: list[str], author_names: list[str], ma
 
     features = retry_call(_make_call, FeatureIdentificationSchema).features
 
-    print(f"Adding to zoom cache")
     if cache_key and author_names:
         cache[cache_key] = {
             "features": features
@@ -541,10 +540,10 @@ def compute_clusters_style_representation_3(
     background_corpus_df: pd.DataFrame,
     cluster_ids: List[Any],
     cluster_label_clm_name: str = 'authorID',
-    max_num_feats: int = 10,
+    max_num_feats: int = 20,
     max_num_documents_per_author=3,
     max_num_authors=5,
-    max_authors_for_span_extraction=7
+    max_authors_for_span_extraction=4
     ):
 
     print(f"Computing style representation for visible clusters: {len(cluster_ids)}")
@@ -568,20 +567,17 @@ def compute_clusters_style_representation_3(
     
     # Filter out features that are not present in any of the authors
     filtered_spans_by_author = {x[0] : x[1] for x in spans_by_author.items() if x[0] in {'Mystery author', 'Candidate Author 1', 'Candidate Author 2', 'Candidate Author 3'}.intersection(set(cluster_ids))}
-    print('Filtering in features for only the following authors: ', filtered_spans_by_author.keys())
-    filtered_features = []
-    for feature in features:
-        found_in_any_author = False
-        for author_name, author_spans in filtered_spans_by_author.items():
-            if feature in author_spans:
-                found_in_any_author = True
-                break
-        if found_in_any_author:
-            filtered_features.append(feature)
-    features = filtered_features
+    print(filtered_spans_by_author.keys())
+    filtered_spans_by_author = [set([f[0] for f in   x[1].items() if len(f[1]) > 0]) for x in filtered_spans_by_author.items()]
+
+    filtered_set_of_features =  filtered_spans_by_author[0] # all features that appear in all the sets in the filtered_Spans_by_authors list
+    for x in filtered_spans_by_author[1:]:
+        filtered_set_of_features = filtered_set_of_features.intersection(x)
+
+    print('filtered set of features: ', filtered_set_of_features)
 
     return {
-        "features": features,
+        "features": list(filtered_set_of_features),
         "spans": spans_by_author
     }
 
@@ -668,19 +664,17 @@ def compute_clusters_g2v_representation(
         key=lambda x: -x[1]  # Sort by contrastive score
     )
     
-    # Filter out features that are not present in any of the authors
+    # Filter in only features that are present in selected_authors
     selected_authors = {'Mystery author', 'Candidate Author 1', 'Candidate Author 2', 'Candidate Author 3'}.intersection(set(author_ids))
-    print('Filtering in g2v features for only the following authors: ', selected_authors)
-    authors_g2v_feats = background_corpus_df[background_corpus_df['authorID'].isin(selected_authors)][features_clm_name].tolist()
+    # Filter in only features that are present in selected_authors
+    selected_authors_g2v_data = background_corpus_df[background_corpus_df['authorID'].isin(selected_authors)][features_clm_name].tolist()
+    
     filtered_features = []
     for feature, score, z_score in top_g2v_feats:
-        found_in_any_author = False
-        for author_g2v_feats in authors_g2v_feats:
-            if author_g2v_feats[feature] > 0:
-                found_in_any_author = True
-                break
-        if found_in_any_author:
-            filtered_features.append((feature, score, z_score))
+        # Check if the feature has a non-zero value in all of the selected authors
+        if all(author_g2v_feats.get(feature, 0) > 0 for author_g2v_feats in selected_authors_g2v_data):
+            filtered_features.append((feature, score, z_score)) # Only return feature and z-score
+
 
     print('Filtered G2V features: ', [(f[0], f[2]) for f in filtered_features])  # Print feature names and z-scores
     

utils/visualizations.py

@@ -14,7 +14,7 @@ from utils.interp_space_utils import compute_clusters_style_representation_3, co
 from utils.llm_feat_utils import split_features
 from utils.gram2vec_feat_utils import get_shorthand, get_fullform
 from gram2vec.feature_locator import find_feature_spans
-
+import umap
 import plotly.io as pio
 
 def clean_text(text: str) -> str:
@@ -132,8 +132,10 @@ def compute_tsne_with_cache(embeddings: np.ndarray, cache_path: str = 'datasets/
         return cache[hash_key]
     else:
         print("Computing t-SNE")
-        tsne_result = TSNE(n_components=2, learning_rate='auto',
-                           init='random', perplexity=3).fit_transform(embeddings)
+        # tsne_result = TSNE(n_components=2, learning_rate='auto',
+        #                    init='random', perplexity=3).fit_transform(embeddings)
+        tsne_result = umap.UMAP(n_components=2, n_neighbors=15, min_dist=0.0, metric='cosine').fit_transform(embeddings)
+        
         cache[hash_key] = tsne_result
         with open(cache_path, 'wb') as f:
             pkl.dump(cache, f)
@@ -147,7 +149,7 @@ def load_interp_space(cfg):
 
     # Load authors embeddings and their cluster labels
     clustered_authors_df = pd.read_pickle(clustered_authors_path)
-    clustered_authors_df = clustered_authors_df[clustered_authors_df.cluster_label != -1]
+    #clustered_authors_df = clustered_authors_df[clustered_authors_df.cluster_label != -1]
     author_embedding = clustered_authors_df.author_embedding.tolist()
     author_labels    = clustered_authors_df.cluster_label.tolist()
     author_ids      = clustered_authors_df.authorID.tolist()
@@ -276,11 +278,12 @@ def handle_zoom(event_json, bg_proj, bg_lbls, clustered_authors_df, task_authors
     print(f"Task authors: {len(task_authors_df)}, Clustered authors: {len(clustered_authors_df)}")
     merged_authors_df = pd.concat([task_authors_df, clustered_authors_df])
     print(f"Merged authors DataFrame:\n{len(merged_authors_df)}")
-    style_analysis_response = compute_clusters_style_representation_3(
-        background_corpus_df=merged_authors_df,
-        cluster_ids=visible_authors,
-        cluster_label_clm_name='authorID',
-    )
+    style_analysis_response = {'features': [], 'spans': []}
+    # style_analysis_response = compute_clusters_style_representation_3(
+    #     background_corpus_df=merged_authors_df,
+    #     cluster_ids=visible_authors,
+    #     cluster_label_clm_name='authorID',
+    # )
 
     llm_feats = ['None'] + style_analysis_response['features']