""" UTILS FILE """ import random import json import numpy as np import matplotlib.pyplot as plt import pandas as pd import pickle import os import mne from surprise import Dataset, Reader, SVD, accuracy, KNNBasic, KNNWithMeans, KNNWithZScore from surprise.model_selection import train_test_split from sklearn.utils import resample from sklearn.metrics import mean_absolute_error from sklearn.metrics import mean_squared_error from scipy import stats import math import altair as alt import matplotlib.pyplot as plt import time from sentence_transformers import SentenceTransformer, util import torch from bertopic import BERTopic ######################################## # PRE-LOADING YOUR_COLOR = '#6CADFD' OTHER_USERS_COLOR = '#ccc' BINS = [0, 0.5, 1.5, 2.5, 3.5, 4] BIN_LABELS = ['0: Not at all toxic', '1: Slightly toxic', '2: Moderately toxic', '3: Very toxic', '4: Extremely toxic'] TOXIC_THRESHOLD = 2.0 alt.renderers.enable('altair_saver', fmts=['vega-lite', 'png']) # Data-loading module_dir = "./" perf_dir = f"data/perf/" # # TEMP reset # with open(os.path.join(module_dir, "./data/all_model_names.pkl"), "wb") as f: # all_model_names = [] # pickle.dump(all_model_names, f) # with open(f"./data/users_to_models.pkl", "wb") as f: # users_to_models = {} # pickle.dump(users_to_models, f) with open(os.path.join(module_dir, "data/ids_to_comments.pkl"), "rb") as f: ids_to_comments = pickle.load(f) with open(os.path.join(module_dir, "data/comments_to_ids.pkl"), "rb") as f: comments_to_ids = pickle.load(f) all_model_names = sorted([name for name in os.listdir(os.path.join(perf_dir)) if os.path.isdir(os.path.join(perf_dir, name))]) comments_grouped_full_topic_cat = pd.read_pickle("data/comments_grouped_full_topic_cat2_persp.pkl") sys_eval_df = pd.read_pickle(os.path.join(module_dir, "data/split_data/sys_eval_df.pkl")) train_df = pd.read_pickle(os.path.join(module_dir, "data/split_data/train_df.pkl")) train_df_ids = train_df["item_id"].unique().tolist() model_eval_df = pd.read_pickle(os.path.join(module_dir, "data/split_data/model_eval_df.pkl")) ratings_df_full = pd.read_pickle(os.path.join(module_dir, "data/ratings_df_full.pkl")) worker_info_df = pd.read_pickle("./data/worker_info_df.pkl") with open(f"./data/users_to_models.pkl", "rb") as f: users_to_models = pickle.load(f) with open("data/perf_1000_topics.pkl", "rb") as f: perf_1000_topics = pickle.load(f) with open("data/perf_1000_tox_cat.pkl", "rb") as f: perf_1000_tox_cat = pickle.load(f) with open("data/perf_1000_tox_severity.pkl", "rb") as f: perf_1000_tox_severity = pickle.load(f) with open("data/user_perf_metrics.pkl", "rb") as f: user_perf_metrics = pickle.load(f) topic_ids = comments_grouped_full_topic_cat.topic_id topics = comments_grouped_full_topic_cat.topic topic_ids_to_topics = {topic_ids[i]: topics[i] for i in range(len(topic_ids))} topics_to_topic_ids = {topics[i]: topic_ids[i] for i in range(len(topic_ids))} unique_topics_ids = sorted(comments_grouped_full_topic_cat.topic_id.unique()) unique_topics = [topic_ids_to_topics[topic_id] for topic_id in range(len(topic_ids_to_topics) - 1)] def get_toxic_threshold(): return TOXIC_THRESHOLD def get_all_model_names(user=None): if (user is None) or (user not in users_to_models): all_model_names = sorted([name for name in os.listdir(os.path.join(perf_dir)) if os.path.isdir(os.path.join(perf_dir, name))]) return all_model_names else: # Fetch the user's models user_models = users_to_models[user] user_models.sort() return user_models def get_unique_topics(): return unique_topics def get_large_clusters(min_n): counts_df = comments_grouped_full_topic_cat.groupby(by=["topic_id"]).size().reset_index(name='counts') counts_df = counts_df[counts_df["counts"] >= min_n] return [topic_ids_to_topics[t_id] for t_id in sorted(counts_df["topic_id"].tolist()[1:])] def get_ids_to_comments(): return ids_to_comments def get_workers_in_group(sel_gender, sel_race, sel_relig, sel_pol, sel_lgbtq): df = worker_info_df.copy() if sel_gender != "null": df = df[df["gender"] == sel_gender] if sel_relig != "null": df = df[df["religion_important"] == sel_relig] if sel_pol != "null": df = df[df["political_affilation"] == sel_pol] if sel_lgbtq != "null": if sel_lgbtq == "LGBTQ+": df = df[(df["lgbtq_status"] == "Homosexual") | (df["lgbtq_status"] == "Bisexual")] else: df = df[df["lgbtq_status"] == "Heterosexual"] if sel_race != "": df = df.dropna(subset=['race']) for r in sel_race: # Filter to rows with the indicated race df = df[df["race"].str.contains(r)] return df, len(df) readable_to_internal = { "Mean Absolute Error (MAE)": "MAE", "Root Mean Squared Error (RMSE)": "RMSE", "Mean Squared Error (MSE)": "MSE", "Average rating difference": "avg_diff", "Topic": "topic", "Toxicity Category": "toxicity_category", "Toxicity Severity": "toxicity_severity", } internal_to_readable = {v: k for k, v in readable_to_internal.items()} # Embeddings for neighbor retrieval model_name = "paraphrase-MiniLM-L6-v2" model = SentenceTransformer(model_name) with open("./data/comments.pkl", "rb") as f: comments = pickle.load(f) embeddings = torch.load("./data/embeddings/21_10_embeddings.pt") # Perspective API recalibration def recalib_v1(s): # convert Perspective score to 0-4 toxicity score # map 0 persp to 0 (not at all toxic); 0.5 persp to 1 (slightly toxic), 1.0 persp to 4 (extremely toxic) if s < 0.5: return (s * 2.) else: return ((s - 0.5) * 6.) + 1 def recalib_v2(s): # convert Perspective score to 0-4 toxicity score # just 4x the perspective score return (s * 4.) comments_grouped_full_topic_cat["rating_avg_orig"] = comments_grouped_full_topic_cat["rating"] comments_grouped_full_topic_cat["rating"] = [recalib_v2(score) for score in comments_grouped_full_topic_cat["persp_score"].tolist()] def get_comments_grouped_full_topic_cat(): return comments_grouped_full_topic_cat ######################################## # General utils def get_metric_ind(metric): if metric == "MAE": ind = 0 elif metric == "MSE": ind = 1 elif metric == "RMSE": ind = 2 elif metric == "avg_diff": ind = 3 return ind def my_bootstrap(vals, n_boot, alpha): bs_samples = [] sample_size = len(vals) for i in range(n_boot): samp = resample(vals, n_samples=sample_size) bs_samples.append(np.median(samp)) p = ((1.0 - alpha) / 2.0) * 100 ci_low = np.percentile(bs_samples, p) p = (alpha + ((1.0 - alpha) / 2.0)) * 100 ci_high = np.percentile(bs_samples, p) return bs_samples, (ci_low, ci_high) ######################################## # GET_AUDIT utils def other_users_perf(perf_metrics, metric, user_metric, alpha=0.95, n_boot=501): ind = get_metric_ind(metric) metric_vals = [metric_vals[ind] for metric_vals in perf_metrics.values()] metric_avg = np.median(metric_vals) # Future: use provided sample to perform bootstrap sampling ci_1 = mne.stats.bootstrap_confidence_interval(np.array(metric_vals), ci=alpha, n_bootstraps=n_boot, stat_fun="median") bs_samples, ci = my_bootstrap(metric_vals, n_boot, alpha) # Get user's percentile percentile = stats.percentileofscore(bs_samples, user_metric) return metric_avg, ci, percentile, metric_vals def plot_metric_histogram(metric, user_metric, other_metric_vals, n_bins=10): hist, bin_edges = np.histogram(other_metric_vals, bins=n_bins, density=False) data = pd.DataFrame({ "bin_min": bin_edges[:-1], "bin_max": bin_edges[1:], "bin_count": hist, "user_metric": [user_metric for i in range(len(hist))] }) base = alt.Chart(data) bar = base.mark_bar(color=OTHER_USERS_COLOR).encode( x=alt.X("bin_min", bin="binned", title=internal_to_readable[metric]), x2='bin_max', y=alt.Y("bin_count", title="Number of users"), tooltip=[ alt.Tooltip('bin_min', title=f'{metric} bin min', format=".2f"), alt.Tooltip('bin_max', title=f'{metric} bin max', format=".2f"), alt.Tooltip('bin_count', title=f'Number of OTHER users', format=","), ] ) rule = base.mark_rule(color=YOUR_COLOR).encode( x = "mean(user_metric):Q", size=alt.value(2), tooltip=[ alt.Tooltip('mean(user_metric)', title=f'{metric} with YOUR labels', format=".2f"), ] ) return (bar + rule).interactive() def get_toxicity_severity_bins(perf_metric, user_df, other_dfs, bins=BINS, bin_labels=BIN_LABELS, ci=0.95, n_boot=501): # Note: not using other_dfs anymore y_user = [] y_other = [] used_bins = [] other_ci_low = [] other_ci_high = [] for severity_i in range(len(bin_labels)): metric_others = [metrics[get_metric_ind(perf_metric)] for metrics in perf_1000_tox_severity[severity_i].values() if metrics[get_metric_ind(perf_metric)]] ci_low, ci_high = mne.stats.bootstrap_confidence_interval(np.array(metric_others), ci=ci, n_bootstraps=n_boot, stat_fun='median') metric_other = np.median(metric_others) cur_user_df = user_df[user_df["prediction_bin"] == severity_i] y_true_user = cur_user_df.pred.to_numpy() # user's label y_pred = cur_user_df.rating_avg.to_numpy() # system's label (avg) if len(y_true_user) > 0: used_bins.append(bin_labels[severity_i]) metric_user = calc_metric_user(y_true_user, y_pred, perf_metric) y_user.append(metric_user) y_other.append(metric_other) other_ci_low.append(ci_low) other_ci_high.append(ci_high) return y_user, y_other, used_bins, other_ci_low, other_ci_high def get_topic_bins(perf_metric, user_df, other_dfs, n_topics, ci=0.95, n_boot=501): # Note: not using other_dfs anymore y_user = [] y_other = [] used_bins = [] other_ci_low = [] other_ci_high = [] selected_topics = unique_topics_ids[1:(n_topics + 1)] for topic_id in selected_topics: cur_topic = topic_ids_to_topics[topic_id] metric_others = [metrics[get_metric_ind(perf_metric)] for metrics in perf_1000_topics[topic_id].values() if metrics[get_metric_ind(perf_metric)]] ci_low, ci_high = mne.stats.bootstrap_confidence_interval(np.array(metric_others), ci=ci, n_bootstraps=n_boot, stat_fun='median') metric_other = np.median(metric_others) cur_user_df = user_df[user_df["topic"] == cur_topic] y_true_user = cur_user_df.pred.to_numpy() # user's label y_pred = cur_user_df.rating_avg.to_numpy() # system's label (avg) if len(y_true_user) > 0: used_bins.append(cur_topic) metric_user = calc_metric_user(y_true_user, y_pred, perf_metric) y_user.append(metric_user) y_other.append(metric_other) other_ci_low.append(ci_low) other_ci_high.append(ci_high) return y_user, y_other, used_bins, other_ci_low, other_ci_high def calc_metric_user(y_true_user, y_pred, perf_metric): if perf_metric == "MAE": metric_user = mean_absolute_error(y_true_user, y_pred) elif perf_metric == "MSE": metric_user = mean_squared_error(y_true_user, y_pred) elif perf_metric == "RMSE": metric_user = mean_squared_error(y_true_user, y_pred, squared=False) elif perf_metric == "avg_diff": metric_user = np.mean(y_true_user - y_pred) return metric_user def get_toxicity_category_bins(perf_metric, user_df, other_dfs, threshold=0.5, ci=0.95, n_boot=501): # Note: not using other_dfs anymore; threshold from pre-calculation is 0.5 cat_cols = ["is_profane_frac", "is_threat_frac", "is_identity_attack_frac", "is_insult_frac", "is_sexual_harassment_frac"] cat_labels = ["Profanity", "Threats", "Identity Attacks", "Insults", "Sexual Harassment"] y_user = [] y_other = [] used_bins = [] other_ci_low = [] other_ci_high = [] for i, cur_col_name in enumerate(cat_cols): metric_others = [metrics[get_metric_ind(perf_metric)] for metrics in perf_1000_tox_cat[cur_col_name].values() if metrics[get_metric_ind(perf_metric)]] ci_low, ci_high = mne.stats.bootstrap_confidence_interval(np.array(metric_others), ci=ci, n_bootstraps=n_boot, stat_fun='median') metric_other = np.median(metric_others) # Filter to rows where a comment received an average label >= the provided threshold for the category cur_user_df = user_df[user_df[cur_col_name] >= threshold] y_true_user = cur_user_df.pred.to_numpy() # user's label y_pred = cur_user_df.rating_avg.to_numpy() # system's label (avg) if len(y_true_user) > 0: used_bins.append(cat_labels[i]) metric_user = calc_metric_user(y_true_user, y_pred, perf_metric) y_user.append(metric_user) y_other.append(metric_other) other_ci_low.append(ci_low) other_ci_high.append(ci_high) return y_user, y_other, used_bins, other_ci_low, other_ci_high def plot_class_cond_results(preds_df, breakdown_axis, perf_metric, other_ids, sort_bars, n_topics, worker_id="A"): # Note: preds_df already has binned results # Prepare dfs user_df = preds_df[preds_df.user_id == worker_id].sort_values(by=["item_id"]).reset_index() other_dfs = [preds_df[preds_df.user_id == other_id].sort_values(by=["item_id"]).reset_index() for other_id in other_ids] if breakdown_axis == "toxicity_severity": y_user, y_other, used_bins, other_ci_low, other_ci_high = get_toxicity_severity_bins(perf_metric, user_df, other_dfs) elif breakdown_axis == "topic": y_user, y_other, used_bins, other_ci_low, other_ci_high = get_topic_bins(perf_metric, user_df, other_dfs, n_topics) elif breakdown_axis == "toxicity_category": y_user, y_other, used_bins, other_ci_low, other_ci_high = get_toxicity_category_bins(perf_metric, user_df, other_dfs) diffs = list(np.array(y_user) - np.array(y_other)) # Generate bar chart data = pd.DataFrame({ "metric_val": y_user + y_other, "Labeler": ["You" for _ in range(len(y_user))] + ["Other users" for _ in range(len(y_user))], "used_bins": used_bins + used_bins, "diffs": diffs + diffs, "lower_cis": y_user + other_ci_low, "upper_cis": y_user + other_ci_high, }) color_domain = ['You', 'Other users'] color_range = [YOUR_COLOR, OTHER_USERS_COLOR] base = alt.Chart() chart_title=f"{internal_to_readable[breakdown_axis]} Results" x_axis = alt.X("Labeler:O", sort=("You", "Other users"), title=None, axis=None) y_axis = alt.Y("metric_val:Q", title=internal_to_readable[perf_metric]) if sort_bars: col_content = alt.Column("used_bins:O", sort=alt.EncodingSortField(field="diffs", op="mean", order='descending')) else: col_content = alt.Column("used_bins:O") if n_topics is not None and n_topics > 10: # Change to horizontal bar chart bar = base.mark_bar(lineBreak="_").encode( y=x_axis, x=y_axis, color=alt.Color("Labeler:O", scale=alt.Scale(domain=color_domain, range=color_range)), tooltip=[ alt.Tooltip('Labeler:O', title='Labeler'), alt.Tooltip('metric_val:Q', title=perf_metric, format=".3f"), ] ) error_bars = base.mark_errorbar().encode( y=x_axis, x = alt.X("lower_cis:Q", title=internal_to_readable[perf_metric]), x2 = alt.X2("upper_cis:Q", title=None), tooltip=[ alt.Tooltip('lower_cis:Q', title='Lower CI', format=".3f"), alt.Tooltip('upper_cis:Q', title='Upper CI', format=".3f"), ] ) combined = alt.layer( bar, error_bars, data=data ).facet( row=col_content ).properties( title=chart_title, ).interactive() else: bar = base.mark_bar(lineBreak="_").encode( x=x_axis, y=y_axis, color=alt.Color("Labeler:O", scale=alt.Scale(domain=color_domain, range=color_range)), tooltip=[ alt.Tooltip('Labeler:O', title='Labeler'), alt.Tooltip('metric_val:Q', title=perf_metric, format=".3f"), ] ) error_bars = base.mark_errorbar().encode( x=x_axis, y = alt.Y("lower_cis:Q", title=internal_to_readable[perf_metric]), y2 = alt.Y2("upper_cis:Q", title=None), tooltip=[ alt.Tooltip('lower_cis:Q', title='Lower CI', format=".3f"), alt.Tooltip('upper_cis:Q', title='Upper CI', format=".3f"), ] ) combined = alt.layer( bar, error_bars, data=data ).facet( column=col_content ).properties( title=chart_title, ).interactive() return combined def show_overall_perf(variant, error_type, cur_user, threshold=TOXIC_THRESHOLD, breakdown_axis=None, topic_vis_method="median"): # Your perf (calculate using model and testset) breakdown_axis = readable_to_internal[breakdown_axis] if breakdown_axis is not None: with open(os.path.join(module_dir, f"data/preds_dfs/{variant}.pkl"), "rb") as f: preds_df = pickle.load(f) # Read from file chart_dir = "./data/charts" chart_file = os.path.join(chart_dir, f"{cur_user}_{variant}.pkl") if os.path.isfile(chart_file): with open(chart_file, "r") as f: topic_overview_plot_json = json.load(f) else: preds_df_mod = preds_df.merge(comments_grouped_full_topic_cat, on="item_id", how="left", suffixes=('_', '_avg')) if topic_vis_method == "median": preds_df_mod_grp = preds_df_mod.groupby(["topic_", "user_id"]).median() elif topic_vis_method == "mean": preds_df_mod_grp = preds_df_mod.groupby(["topic_", "user_id"]).mean() topic_overview_plot_json = plot_overall_vis(preds_df=preds_df_mod_grp, n_topics=200, threshold=threshold, error_type=error_type, cur_user=cur_user, cur_model=variant) return { "topic_overview_plot_json": json.loads(topic_overview_plot_json), } ######################################## # GET_CLUSTER_RESULTS utils def get_overall_perf3(preds_df, perf_metric, other_ids, worker_id="A"): # Prepare dataset to calculate performance # Note: true is user and pred is system y_true = preds_df[preds_df["user_id"] == worker_id].pred.to_numpy() y_pred_user = preds_df[preds_df["user_id"] == worker_id].rating_avg.to_numpy() y_true_others = y_pred_others = [preds_df[preds_df["user_id"] == other_id].pred.to_numpy() for other_id in other_ids] y_pred_others = [preds_df[preds_df["user_id"] == other_id].rating_avg.to_numpy() for other_id in other_ids] # Get performance for user's model and for other users if perf_metric == "MAE": user_perf = mean_absolute_error(y_true, y_pred_user) other_perfs = [mean_absolute_error(y_true_others[i], y_pred_others[i]) for i in range(len(y_true_others))] elif perf_metric == "MSE": user_perf = mean_squared_error(y_true, y_pred_user) other_perfs = [mean_squared_error(y_true_others[i], y_pred_others[i]) for i in range(len(y_true_others))] elif perf_metric == "RMSE": user_perf = mean_squared_error(y_true, y_pred_user, squared=False) other_perfs = [mean_squared_error(y_true_others[i], y_pred_others[i], squared=False) for i in range(len(y_true_others))] elif perf_metric == "avg_diff": user_perf = np.mean(y_true - y_pred_user) other_perfs = [np.mean(y_true_others[i] - y_pred_others[i]) for i in range(len(y_true_others))] other_perf = np.mean(other_perfs) # average across all other users return user_perf, other_perf def style_color_difference(row): full_opacity_diff = 3. pred_user_col = "Your predicted rating" pred_other_col = "Other users' predicted rating" pred_system_col = "Status-quo system rating" diff_user = row[pred_user_col] - row[pred_system_col] diff_other = row[pred_other_col] - row[pred_system_col] red = "234, 133, 125" green = "142, 205, 162" bkgd_user = green if diff_user < 0 else red # red if more toxic; green if less toxic opac_user = min(abs(diff_user / full_opacity_diff), 1.) bkgd_other = green if diff_other < 0 else red # red if more toxic; green if less toxic opac_other = min(abs(diff_other / full_opacity_diff), 1.) return ["", f"background-color: rgba({bkgd_user}, {opac_user});", f"background-color: rgba({bkgd_other}, {opac_other});", "", ""] def display_examples_cluster(preds_df, other_ids, num_examples, sort_ascending, worker_id="A"): user_df = preds_df[preds_df.user_id == worker_id].sort_values(by=["item_id"]).reset_index() others_df = preds_df[preds_df.user_id == other_ids[0]] for i in range(1, len(other_ids)): others_df.append(preds_df[preds_df.user_id == other_ids[i]]) others_df.groupby(["item_id"]).mean() others_df = others_df.sort_values(by=["item_id"]).reset_index() df = pd.merge(user_df, others_df, on="item_id", how="left", suffixes=('_user', '_other')) df["Comment"] = df["comment_user"] df["Your predicted rating"] = df["pred_user"] df["Other users' predicted rating"] = df["pred_other"] df["Status-quo system rating"] = df["rating_avg_user"] df["Status-quo system std dev"] = df["rating_stddev_user"] df = df[["Comment", "Your predicted rating", "Other users' predicted rating", "Status-quo system rating", "Status-quo system std dev"]] # Add styling df = df.sort_values(by=['Status-quo system std dev'], ascending=sort_ascending) n_to_sample = np.min([num_examples, len(df)]) df = df.sample(n=n_to_sample).reset_index(drop=True) return df.style.apply(style_color_difference, axis=1).render() def calc_odds_ratio(df, comparison_group, toxic_threshold=1.5, worker_id="A", debug=False, smoothing_factor=1): if comparison_group == "status_quo": other_pred_col = "rating_avg" # Get unique comments, but fetch average labeler rating num_toxic_other = len(df[(df.user_id == "A") & (df[other_pred_col] >= toxic_threshold)]) + smoothing_factor num_nontoxic_other = len(df[(df.user_id == "A") & (df[other_pred_col] < toxic_threshold)]) + smoothing_factor elif comparison_group == "other_users": other_pred_col = "pred" num_toxic_other = len(df[(df.user_id != "A") & (df[other_pred_col] >= toxic_threshold)]) + smoothing_factor num_nontoxic_other = len(df[(df.user_id != "A") & (df[other_pred_col] < toxic_threshold)]) + smoothing_factor num_toxic_user = len(df[(df.user_id == "A") & (df.pred >= toxic_threshold)]) + smoothing_factor num_nontoxic_user = len(df[(df.user_id == "A") & (df.pred < toxic_threshold)]) + smoothing_factor toxic_ratio = num_toxic_user / num_toxic_other nontoxic_ratio = num_nontoxic_user / num_nontoxic_other odds_ratio = toxic_ratio / nontoxic_ratio if debug: print(f"Odds ratio: {odds_ratio}") print(f"num_toxic_user: {num_toxic_user}, num_nontoxic_user: {num_nontoxic_user}") print(f"num_toxic_other: {num_toxic_other}, num_nontoxic_other: {num_nontoxic_other}") contingency_table = [[num_toxic_user, num_nontoxic_user], [num_toxic_other, num_nontoxic_other]] odds_ratio, p_val = stats.fisher_exact(contingency_table, alternative='two-sided') if debug: print(f"Odds ratio: {odds_ratio}, p={p_val}") return odds_ratio # Neighbor search def get_match(comment_inds, K=20, threshold=None, debug=False): match_ids = [] rows = [] for i in comment_inds: if debug: print(f"\nComment: {comments[i]}") query_embedding = model.encode(comments[i], convert_to_tensor=True) hits = util.semantic_search(query_embedding, embeddings, score_function=util.cos_sim, top_k=K) # print(hits[0]) for hit in hits[0]: c_id = hit['corpus_id'] score = np.round(hit['score'], 3) if threshold is None or score > threshold: match_ids.append(c_id) if debug: print(f"\t(ID={c_id}, Score={score}): {comments[c_id]}") rows.append([c_id, score, comments[c_id]]) df = pd.DataFrame(rows, columns=["id", "score", "comment"]) return match_ids def display_examples_auto_cluster(preds_df, cluster, other_ids, perf_metric, sort_ascending=True, worker_id="A", num_examples=10): # Overall performance topic_df = preds_df topic_df = topic_df[topic_df["topic"] == cluster] user_perf, other_perf = get_overall_perf3(topic_df, perf_metric, other_ids) user_direction = "LOWER" if user_perf < 0 else "HIGHER" other_direction = "LOWER" if other_perf < 0 else "HIGHER" print(f"Your ratings are on average {np.round(abs(user_perf), 3)} {user_direction} than the existing system for this cluster") print(f"Others' ratings (based on {len(other_ids)} users) are on average {np.round(abs(other_perf), 3)} {other_direction} than the existing system for this cluster") # Display example comments df = display_examples_cluster(preds_df, other_ids, num_examples, sort_ascending) return df # function to get results for a new provided cluster def display_examples_manual_cluster(preds_df, cluster_comments, other_ids, perf_metric, sort_ascending=True, worker_id="A"): # Overall performance cluster_df = preds_df[preds_df["comment"].isin(cluster_comments)] user_perf, other_perf = get_overall_perf3(cluster_df, perf_metric, other_ids) user_direction = "LOWER" if user_perf < 0 else "HIGHER" other_direction = "LOWER" if other_perf < 0 else "HIGHER" print(f"Your ratings are on average {np.round(abs(user_perf), 3)} {user_direction} than the existing system for this cluster") print(f"Others' ratings (based on {len(other_ids)} users) are on average {np.round(abs(other_perf), 3)} {other_direction} than the existing system for this cluster") user_df = preds_df[preds_df.user_id == worker_id].sort_values(by=["item_id"]).reset_index() others_df = preds_df[preds_df.user_id == other_ids[0]] for i in range(1, len(other_ids)): others_df.append(preds_df[preds_df.user_id == other_ids[i]]) others_df.groupby(["item_id"]).mean() others_df = others_df.sort_values(by=["item_id"]).reset_index() # Get cluster_comments user_df = user_df[user_df["comment"].isin(cluster_comments)] others_df = others_df[others_df["comment"].isin(cluster_comments)] df = pd.merge(user_df, others_df, on="item_id", how="left", suffixes=('_user', '_other')) df["pred_system"] = df["rating_avg_user"] df["pred_system_stddev"] = df["rating_stddev_user"] df = df[["item_id", "comment_user", "pred_user", "pred_other", "pred_system", "pred_system_stddev"]] # Add styling df = df.sort_values(by=['pred_system_stddev'], ascending=sort_ascending) df = df.style.apply(style_color_difference, axis=1).render() return df ######################################## # GET_LABELING utils def create_example_sets(comments_df, n_label_per_bin, score_bins, keyword=None, topic=None): # Restrict to the keyword, if provided df = comments_df.copy() if keyword != None: df = df[df["comment"].str.contains(keyword)] if topic != None: df = df[df["topic"] == topic] # Try to choose n values from each provided score bin ex_to_label = [] bin_names = [] bin_label_counts = [] for i, score_bin in enumerate(score_bins): min_score, max_score = score_bin cur_df = df[(df["rating"] >= min_score) & (df["rating"] < max_score) & (df["item_id"].isin(train_df_ids))] # sample rows for label comment_ids = cur_df.item_id.tolist() cur_n_label_per_bin = n_label_per_bin[i] cap = min(len(comment_ids), (cur_n_label_per_bin)) to_label = np.random.choice(comment_ids, cap, replace=False) ex_to_label.extend(to_label) bin_names.append(f"[{min_score}, {max_score})") bin_label_counts.append(len(to_label)) return ex_to_label def get_grp_model_labels(comments_df, n_label_per_bin, score_bins, grp_ids): df = comments_df.copy() train_df_grp = train_df[train_df["user_id"].isin(grp_ids)] train_df_grp_avg = train_df_grp.groupby(by=["item_id"]).median().reset_index() train_df_grp_avg_ids = train_df_grp_avg["item_id"].tolist() ex_to_label = [] # IDs of comments to use for group model training for i, score_bin in enumerate(score_bins): min_score, max_score = score_bin # get eligible comments to sample cur_df = df[(df["rating"] >= min_score) & (df["rating"] < max_score) & (df["item_id"].isin(train_df_grp_avg_ids))] comment_ids = cur_df.item_id.unique().tolist() # sample comments cur_n_label_per_bin = n_label_per_bin[i] cap = min(len(comment_ids), (cur_n_label_per_bin)) to_label = np.random.choice(comment_ids, cap, replace=False) ex_to_label.extend((to_label)) train_df_grp_avg = train_df_grp_avg[train_df_grp_avg["item_id"].isin(ex_to_label)] ratings_grp = {ids_to_comments[int(r["item_id"])]: r["rating"] for _, r in train_df_grp_avg.iterrows()} return ratings_grp ######################################## # GET_PERSONALIZED_MODEL utils def fetch_existing_data(model_name, last_label_i): # Check if we have cached model performance perf_dir = f"./data/perf/{model_name}" label_dir = f"./data/labels/{model_name}" if os.path.isdir(os.path.join(module_dir, perf_dir)): # Fetch cached results last_i = len([name for name in os.listdir(os.path.join(module_dir, perf_dir)) if os.path.isfile(os.path.join(module_dir, perf_dir, name))]) with open(os.path.join(module_dir, perf_dir, f"{last_i}.pkl"), "rb") as f: mae, mse, rmse, avg_diff = pickle.load(f) else: # Fetch results from trained model with open(os.path.join(module_dir, f"./data/trained_models/{model_name}.pkl"), "rb") as f: cur_model = pickle.load(f) mae, mse, rmse, avg_diff = users_perf(cur_model) # Cache results os.mkdir(os.path.join(module_dir, perf_dir)) with open(os.path.join(module_dir, perf_dir, "1.pkl"), "wb") as f: pickle.dump((mae, mse, rmse, avg_diff), f) # Fetch previous user-provided labels ratings_prev = None if last_label_i > 0: with open(os.path.join(module_dir, label_dir, f"{last_i}.pkl"), "rb") as f: ratings_prev = pickle.load(f) return mae, mse, rmse, avg_diff, ratings_prev def train_updated_model(model_name, last_label_i, ratings, user, top_n=20, topic=None): # Check if there is previously-labeled data; if so, combine it with this data perf_dir = f"./data/perf/{model_name}" label_dir = f"./data/labels/{model_name}" labeled_df = format_labeled_data(ratings) # Treat ratings as full batch of all ratings ratings_prev = None # Filter out rows with "unsure" (-1) labeled_df = labeled_df[labeled_df["rating"] != -1] # Filter to top N for user study if topic is None: # labeled_df = labeled_df.head(top_n) labeled_df = labeled_df.tail(top_n) else: # For topic tuning, need to fetch old labels if (last_label_i > 0): # Concatenate previous set of labels with this new batch of labels with open(os.path.join(module_dir, label_dir, f"{last_label_i}.pkl"), "rb") as f: ratings_prev = pickle.load(f) labeled_df_prev = format_labeled_data(ratings_prev) labeled_df_prev = labeled_df_prev[labeled_df_prev["rating"] != -1] ratings.update(ratings_prev) # append old ratings to ratings labeled_df = pd.concat([labeled_df_prev, labeled_df]) print("len ratings for training:", len(labeled_df)) cur_model, perf, _, _ = train_user_model(ratings_df=labeled_df) user_perf_metrics[model_name] = users_perf(cur_model) mae, mse, rmse, avg_diff = user_perf_metrics[model_name] cur_preds_df = get_preds_df(cur_model, ["A"], sys_eval_df=ratings_df_full, topic=topic, model_name=model_name) # Just get results for user # Save this batch of labels with open(os.path.join(module_dir, label_dir, f"{last_label_i + 1}.pkl"), "wb") as f: pickle.dump(ratings, f) # Save model results with open(os.path.join(module_dir, f"./data/preds_dfs/{model_name}.pkl"), "wb") as f: pickle.dump(cur_preds_df, f) if model_name not in all_model_names: all_model_names.append(model_name) with open(os.path.join(module_dir, "./data/all_model_names.pkl"), "wb") as f: pickle.dump(all_model_names, f) # Handle user if user not in users_to_models: users_to_models[user] = [] # New user if model_name not in users_to_models[user]: users_to_models[user].append(model_name) # New model with open(f"./data/users_to_models.pkl", "wb") as f: pickle.dump(users_to_models, f) with open(os.path.join(module_dir, "./data/user_perf_metrics.pkl"), "wb") as f: pickle.dump(user_perf_metrics, f) with open(os.path.join(module_dir, f"./data/trained_models/{model_name}.pkl"), "wb") as f: pickle.dump(cur_model, f) # Cache performance results if not os.path.isdir(os.path.join(module_dir, perf_dir)): os.mkdir(os.path.join(module_dir, perf_dir)) last_perf_i = len([name for name in os.listdir(os.path.join(module_dir, perf_dir)) if os.path.isfile(os.path.join(module_dir, perf_dir, name))]) with open(os.path.join(module_dir, perf_dir, f"{last_perf_i + 1}.pkl"), "wb") as f: pickle.dump((mae, mse, rmse, avg_diff), f) ratings_prev = ratings return mae, mse, rmse, avg_diff, ratings_prev def format_labeled_data(ratings, worker_id="A", debug=False): all_rows = [] for comment, rating in ratings.items(): comment_id = comments_to_ids[comment] row = [worker_id, comment_id, int(rating)] all_rows.append(row) df = pd.DataFrame(all_rows, columns=["user_id", "item_id", "rating"]) return df def users_perf(model, sys_eval_df=sys_eval_df, avg_ratings_df=comments_grouped_full_topic_cat, worker_id="A"): # Load the full empty dataset sys_eval_comment_ids = sys_eval_df.item_id.unique().tolist() empty_ratings_rows = [[worker_id, c_id, 0] for c_id in sys_eval_comment_ids] empty_ratings_df = pd.DataFrame(empty_ratings_rows, columns=["user_id", "item_id", "rating"]) # Compute predictions for full dataset reader = Reader(rating_scale=(0, 4)) eval_set_data = Dataset.load_from_df(empty_ratings_df, reader) _, testset = train_test_split(eval_set_data, test_size=1.) predictions = model.test(testset) df = empty_ratings_df # user_id, item_id, rating user_item_preds = get_predictions_by_user_and_item(predictions) df["pred"] = df.apply(lambda row: user_item_preds[(row.user_id, row.item_id)] if (row.user_id, row.item_id) in user_item_preds else np.nan, axis=1) df = df.merge(avg_ratings_df, on="item_id", how="left", suffixes=('_', '_avg')) df.dropna(subset = ["pred"], inplace=True) df["rating_"] = df.rating_.astype("int32") perf_metrics = get_overall_perf(df, "A") # mae, mse, rmse, avg_diff return perf_metrics def get_overall_perf(preds_df, user_id): # Prepare dataset to calculate performance y_pred = preds_df[preds_df["user_id"] == user_id].rating_avg.to_numpy() # Assume system is just average of true labels y_true = preds_df[preds_df["user_id"] == user_id].pred.to_numpy() # Get performance for user's model mae = mean_absolute_error(y_true, y_pred) mse = mean_squared_error(y_true, y_pred) rmse = mean_squared_error(y_true, y_pred, squared=False) avg_diff = np.mean(y_true - y_pred) return mae, mse, rmse, avg_diff def get_predictions_by_user_and_item(predictions): user_item_preds = {} for uid, iid, true_r, est, _ in predictions: user_item_preds[(uid, iid)] = est return user_item_preds def get_preds_df(model, user_ids, orig_df=ratings_df_full, avg_ratings_df=comments_grouped_full_topic_cat, sys_eval_df=sys_eval_df, bins=BINS, topic=None, model_name=None): # Prep dataframe for all predictions we'd like to request start = time.time() sys_eval_comment_ids = sys_eval_df.item_id.unique().tolist() empty_ratings_rows = [] for user_id in user_ids: empty_ratings_rows.extend([[user_id, c_id, 0] for c_id in sys_eval_comment_ids]) empty_ratings_df = pd.DataFrame(empty_ratings_rows, columns=["user_id", "item_id", "rating"]) print("setup", time.time() - start) # Evaluate model to get predictions start = time.time() reader = Reader(rating_scale=(0, 4)) eval_set_data = Dataset.load_from_df(empty_ratings_df, reader) _, testset = train_test_split(eval_set_data, test_size=1.) predictions = model.test(testset) print("train_test_split", time.time() - start) # Update dataframe with predictions start = time.time() df = empty_ratings_df.copy() # user_id, item_id, rating user_item_preds = get_predictions_by_user_and_item(predictions) df["pred"] = df.apply(lambda row: user_item_preds[(row.user_id, row.item_id)] if (row.user_id, row.item_id) in user_item_preds else np.nan, axis=1) df = df.merge(avg_ratings_df, on="item_id", how="left", suffixes=('_', '_avg')) df.dropna(subset = ["pred"], inplace=True) df["rating_"] = df.rating_.astype("int32") # Get binned predictions (based on user prediction) df["prediction_bin"], out_bins = pd.cut(df["pred"], bins, labels=False, retbins=True) df = df.sort_values(by=["item_id"]) return df def train_user_model(ratings_df, train_df=train_df, model_eval_df=model_eval_df, train_frac=0.75, model_type="SVD", sim_type=None, user_based=True): # Sample from shuffled labeled dataframe and add batch to train set; specified set size to model_eval set labeled = ratings_df.sample(frac=1) batch_size = math.floor(len(labeled) * train_frac) labeled_train = labeled[:batch_size] labeled_model_eval = labeled[batch_size:] train_df_ext = train_df.append(labeled_train) model_eval_df_ext = model_eval_df.append(labeled_model_eval) # Train model and show model eval set results model, perf = train_model(train_df_ext, model_eval_df_ext, model_type=model_type, sim_type=sim_type, user_based=user_based) return model, perf, labeled_train, labeled_model_eval def train_model(train_df, model_eval_df, model_type="SVD", sim_type=None, user_based=True): # Train model reader = Reader(rating_scale=(0, 4)) train_data = Dataset.load_from_df(train_df, reader) model_eval_data = Dataset.load_from_df(model_eval_df, reader) train_set = train_data.build_full_trainset() _, model_eval_set = train_test_split(model_eval_data, test_size=1.) sim_options = { "name": sim_type, "user_based": user_based, # compute similarity between users or items } if model_type == "SVD": algo = SVD() # SVD doesn't have similarity metric elif model_type == "KNNBasic": algo = KNNBasic(sim_options=sim_options) elif model_type == "KNNWithMeans": algo = KNNWithMeans(sim_options=sim_options) elif model_type == "KNNWithZScore": algo = KNNWithZScore(sim_options=sim_options) algo.fit(train_set) predictions = algo.test(model_eval_set) rmse = accuracy.rmse(predictions) fcp = accuracy.fcp(predictions) mae = accuracy.mae(predictions) mse = accuracy.mse(predictions) print(f"MAE: {mae}, MSE: {mse}, RMSE: {rmse}, FCP: {fcp}") perf = [mae, mse, rmse, fcp] return algo, perf def plot_train_perf_results2(model_name): # Open labels label_dir = f"./data/labels/{model_name}" n_label_files = len([name for name in os.listdir(os.path.join(module_dir, label_dir)) if os.path.isfile(os.path.join(module_dir, label_dir, name))]) all_rows = [] with open(os.path.join(module_dir, label_dir, f"{n_label_files}.pkl"), "rb") as f: ratings = pickle.load(f) labeled_df = format_labeled_data(ratings) labeled_df = labeled_df[labeled_df["rating"] != -1] # Iterate through batches of 5 labels n_batches = int(np.ceil(len(labeled_df) / 5.)) for i in range(n_batches): start = time.time() n_to_sample = np.min([5 * (i + 1), len(labeled_df)]) cur_model, _, _, _ = train_user_model(ratings_df=labeled_df.head(n_to_sample)) mae, mse, rmse, avg_diff = users_perf(cur_model) all_rows.append([n_to_sample, mae, "MAE"]) print(f"iter {i}: {time.time() - start}") print("all_rows", all_rows) df = pd.DataFrame(all_rows, columns=["n_to_sample", "perf", "metric"]) chart = alt.Chart(df).mark_line(point=True).encode( x=alt.X("n_to_sample:Q", title="Number of Comments Labeled"), y="perf", color="metric", tooltip=[ alt.Tooltip('n_to_sample:Q', title="Number of Comments Labeled"), alt.Tooltip('metric:N', title="Metric"), alt.Tooltip('perf:Q', title="Metric Value", format=".3f"), ], ).properties( title=f"Performance over number of examples: {model_name}", width=500, ) return chart def plot_train_perf_results(model_name, mae): perf_dir = f"./data/perf/{model_name}" n_perf_files = len([name for name in os.listdir(os.path.join(module_dir, perf_dir)) if os.path.isfile(os.path.join(module_dir, perf_dir, name))]) all_rows = [] for i in range(1, n_perf_files + 1): with open(os.path.join(module_dir, perf_dir, f"{i}.pkl"), "rb") as f: mae, mse, rmse, avg_diff = pickle.load(f) all_rows.append([i, mae, "Your MAE"]) df = pd.DataFrame(all_rows, columns=["version", "perf", "metric"]) chart = alt.Chart(df).mark_line(point=True).encode( x="version:O", y="perf", color=alt.Color("metric", title="Performance metric"), tooltip=[ alt.Tooltip('version:O', title='Version'), alt.Tooltip('metric:N', title="Metric"), alt.Tooltip('perf:Q', title="Metric Value", format=".3f"), ], ).properties( title=f"Performance over model versions: {model_name}", width=500, ) PCT_50 = 0.591 PCT_75 = 0.662 PCT_90 = 0.869 plot_dim_width = 500 domain_min = 0.0 domain_max = 1.0 bkgd = alt.Chart(pd.DataFrame({ "start": [PCT_90, PCT_75, domain_min], "stop": [domain_max, PCT_90, PCT_75], "bkgd": ["Needs improvement (< top 90%)", "Okay (top 90%)", "Good (top 75%)"], })).mark_rect(opacity=0.2).encode( y=alt.Y("start:Q", scale=alt.Scale(domain=[0, domain_max])), y2=alt.Y2("stop:Q"), x=alt.value(0), x2=alt.value(plot_dim_width), color=alt.Color("bkgd:O", scale=alt.Scale( domain=["Needs improvement (< top 90%)", "Okay (top 90%)", "Good (top 75%)"], range=["red", "yellow", "green"]), title="How good is your MAE?" ) ) plot = (bkgd + chart).properties(width=plot_dim_width).resolve_scale(color='independent') mae_status = None if mae < PCT_75: mae_status = "Your MAE is in the Good range, which means that it's in the top 75% of scores compared to other users. Your model looks good to go." elif mae < PCT_90: mae_status = "Your MAE is in the Okay range, which means that it's in the top 90% of scores compared to other users. Your model can be used, but you can provide additional labels to improve it." else: mae_status = "Your MAE is in the Needs improvement range, which means that it's in below the top 95% of scores compared to other users. Your model may need additional labels to improve." return plot, mae_status ######################################## # New visualizations # Constants VIS_BINS = np.round(np.arange(0, 4.01, 0.05), 3) VIS_BINS_LABELS = [np.round(np.mean([x, y]), 3) for x, y in zip(VIS_BINS[:-1], VIS_BINS[1:])] def get_key(sys, user, threshold): if sys <= threshold and user <= threshold: return "System agrees: Non-toxic" elif sys > threshold and user > threshold: return "System agrees: Toxic" else: if abs(sys - threshold) > 1.5: return "System differs: Error > 1.5" elif abs(sys - threshold) > 1.0: return "System differs: Error > 1.0" elif abs(sys - threshold) > 0.5: return "System differs: Error > 0.5" else: return "System differs: Error <=0.5" def get_key_no_model(sys, threshold): if sys <= threshold: return "System says: Non-toxic" else: return "System says: Toxic" def get_user_color(user, threshold): if user <= threshold: return "#FFF" # white else: return "#808080" # grey def get_system_color(sys, user, threshold): if sys <= threshold and user <= threshold: return "#FFF" # white elif sys > threshold and user > threshold: return "#808080" # grey else: if abs(sys - threshold) > 1.5: return "#d62728" # red elif abs(sys - threshold) > 1.0: return "#ff7a5c" # med red elif abs(sys - threshold) > 0.5: return "#ffa894" # light red else: return "#ffd1c7" # very light red def get_error_type(sys, user, threshold): if sys <= threshold and user <= threshold: return "No error (agree non-toxic)" elif sys > threshold and user > threshold: return "No error (agree toxic)" elif sys <= threshold and user > threshold: return "System may be under-sensitive" elif sys > threshold and user <= threshold: return "System may be over-sensitive" def get_error_type_radio(sys, user, threshold): if sys <= threshold and user <= threshold: return "Show errors and non-errors" elif sys > threshold and user > threshold: return "Show errors and non-errors" elif sys <= threshold and user > threshold: return "System is under-sensitive" elif sys > threshold and user <= threshold: return "System is over-sensitive" def get_error_magnitude(sys, user, threshold): if sys <= threshold and user <= threshold: return 0 # no classification error elif sys > threshold and user > threshold: return 0 # no classification error elif sys <= threshold and user > threshold: return abs(sys - user) elif sys > threshold and user <= threshold: return abs(sys - user) def get_error_size(sys, user, threshold): if sys <= threshold and user <= threshold: return 0 # no classification error elif sys > threshold and user > threshold: return 0 # no classification error elif sys <= threshold and user > threshold: return sys - user elif sys > threshold and user <= threshold: return sys - user def get_decision(rating, threshold): if rating <= threshold: return "Non-toxic" else: return "Toxic" def get_category(row, threshold=0.3): k_to_category = { "is_profane_frac_": "Profanity", "is_threat_frac_": "Threat", "is_identity_attack_frac_": "Identity Attack", "is_insult_frac_": "Insult", "is_sexual_harassment_frac_": "Sexual Harassment", } categories = [] for k in ["is_profane_frac_", "is_threat_frac_", "is_identity_attack_frac_", "is_insult_frac_", "is_sexual_harassment_frac_"]: if row[k] > threshold: categories.append(k_to_category[k]) if len(categories) > 0: return ", ".join(categories) else: return "" def get_comment_url(row): return f"#{row['item_id']}/#comment" def get_topic_url(row): return f"#{row['topic_']}/#topic" def plot_overall_vis(preds_df, error_type, cur_user, cur_model, n_topics=None, bins=VIS_BINS, threshold=TOXIC_THRESHOLD, bin_step=0.05): df = preds_df.copy().reset_index() if n_topics is not None: df = df[df["topic_id_"] < n_topics] df["vis_pred_bin"], out_bins = pd.cut(df["pred"], bins, labels=VIS_BINS_LABELS, retbins=True) df = df[df["user_id"] == "A"].sort_values(by=["item_id"]).reset_index() df["system_label"] = [("toxic" if r > threshold else "non-toxic") for r in df["rating"].tolist()] df["threshold"] = [threshold for r in df["rating"].tolist()] df["key"] = [get_key(sys, user, threshold) for sys, user in zip(df["rating"].tolist(), df["pred"].tolist())] df["url"] = df.apply(lambda row: get_topic_url(row), axis=1) # Plot sizing domain_min = 0 domain_max = 4 plot_dim_height = 500 plot_dim_width = 750 max_items = np.max(df["vis_pred_bin"].value_counts().tolist()) mark_size = np.round(plot_dim_height / max_items) * 8 if mark_size > 75: mark_size = 75 plot_dim_height = 13 * max_items # Main chart chart = alt.Chart(df).mark_square(opacity=0.8, size=mark_size, stroke="grey", strokeWidth=0.5).transform_window( groupby=['vis_pred_bin'], sort=[{'field': 'rating'}], id='row_number()', ignorePeers=True, ).encode( x=alt.X('vis_pred_bin:Q', title="Our prediction of your rating", scale=alt.Scale(domain=(domain_min, domain_max))), y=alt.Y('id:O', title="Comments (ordered by System toxicity rating)", axis=alt.Axis(values=list(range(0, max_items, 5))), sort='descending'), color = alt.Color("key:O", scale=alt.Scale( domain=["System agrees: Non-toxic", "System agrees: Toxic", "System differs: Error > 1.5", "System differs: Error > 1.0", "System differs: Error > 0.5", "System differs: Error <=0.5"], range=["white", "#cbcbcb", "red", "#ff7a5c", "#ffa894", "#ffd1c7"]), title="System rating (box color)" ), href="url:N", tooltip = [ alt.Tooltip("topic_:N", title="Topic"), alt.Tooltip("system_label:N", title="System label"), alt.Tooltip("rating:Q", title="System rating", format=".2f"), alt.Tooltip("pred:Q", title="Your rating", format=".2f") ] ) # Filter to specified error type if error_type == "System is under-sensitive": # FN: system rates non-toxic, but user rates toxic chart = chart.transform_filter( alt.FieldGTPredicate(field="pred", gt=threshold) ) elif error_type == "System is over-sensitive": # FP: system rates toxic, but user rates non-toxic chart = chart.transform_filter( alt.FieldLTEPredicate(field="pred", lte=threshold) ) # Threshold line rule = alt.Chart(pd.DataFrame({ "threshold": [threshold], "System threshold": [f"Threshold = {threshold}"] })).mark_rule().encode( x=alt.X("mean(threshold):Q", scale=alt.Scale(domain=(domain_min, domain_max)), title=""), color=alt.Color("System threshold:N", scale=alt.Scale(domain=[f"Threshold = {threshold}"], range=["grey"])), size=alt.value(2), ) # Plot region annotations nontoxic_x = (domain_min + threshold) / 2. toxic_x = (domain_max + threshold) / 2. annotation = alt.Chart(pd.DataFrame({ "annotation_text": ["Non-toxic", "Toxic"], "x": [nontoxic_x, toxic_x], "y": [max_items, max_items], })).mark_text( align="center", baseline="middle", fontSize=16, dy=10, color="grey" ).encode( x=alt.X("x", title=""), y=alt.Y("y", title="", axis=None), text="annotation_text" ) # Plot region background colors bkgd = alt.Chart(pd.DataFrame({ "start": [domain_min, threshold], "stop": [threshold, domain_max], "bkgd": ["Non-toxic (L side)", "Toxic (R side)"], })).mark_rect(opacity=1.0, stroke="grey", strokeWidth=0.25).encode( x=alt.X("start:Q", scale=alt.Scale(domain=[domain_min, domain_max])), x2=alt.X2("stop:Q"), y=alt.value(0), y2=alt.value(plot_dim_height), color=alt.Color("bkgd:O", scale=alt.Scale( domain=["Non-toxic (L side)", "Toxic (R side)"], range=["white", "#cbcbcb"]), title="Your rating (background color)" ) ) plot = (bkgd + annotation + chart + rule).properties(height=(plot_dim_height), width=plot_dim_width).resolve_scale(color='independent').to_json() # Save to file chart_dir = "./data/charts" chart_file = os.path.join(chart_dir, f"{cur_user}_{cur_model}.pkl") with open(chart_file, "w") as f: json.dump(plot, f) return plot def get_cluster_overview_plot(preds_df, error_type, threshold=TOXIC_THRESHOLD, use_model=True): preds_df_mod = preds_df.merge(comments_grouped_full_topic_cat, on="item_id", how="left", suffixes=('_', '_avg')) if use_model: return plot_overall_vis_cluster(preds_df_mod, error_type=error_type, n_comments=500, threshold=threshold) else: return plot_overall_vis_cluster2(preds_df_mod, error_type=error_type, n_comments=500, threshold=threshold) def plot_overall_vis_cluster2(preds_df, error_type, n_comments=None, bins=VIS_BINS, threshold=TOXIC_THRESHOLD, bin_step=0.05): df = preds_df.copy().reset_index() df["vis_pred_bin"], out_bins = pd.cut(df["rating"], bins, labels=VIS_BINS_LABELS, retbins=True) df = df[df["user_id"] == "A"].sort_values(by=["rating"]).reset_index() df["system_label"] = [("toxic" if r > threshold else "non-toxic") for r in df["rating"].tolist()] df["key"] = [get_key_no_model(sys, threshold) for sys in df["rating"].tolist()] print("len(df)", len(df)) # always 0 for some reason (from keyword search) df["category"] = df.apply(lambda row: get_category(row), axis=1) df["url"] = df.apply(lambda row: get_comment_url(row), axis=1) if n_comments is not None: n_to_sample = np.min([n_comments, len(df)]) df = df.sample(n=n_to_sample) # Plot sizing domain_min = 0 domain_max = 4 plot_dim_height = 500 plot_dim_width = 750 max_items = np.max(df["vis_pred_bin"].value_counts().tolist()) mark_size = np.round(plot_dim_height / max_items) * 8 if mark_size > 75: mark_size = 75 plot_dim_height = 13 * max_items # Main chart chart = alt.Chart(df).mark_square(opacity=0.8, size=mark_size, stroke="grey", strokeWidth=0.25).transform_window( groupby=['vis_pred_bin'], sort=[{'field': 'rating'}], id='row_number()', ignorePeers=True ).encode( x=alt.X('vis_pred_bin:Q', title="System toxicity rating", scale=alt.Scale(domain=(domain_min, domain_max))), y=alt.Y('id:O', title="Comments (ordered by System toxicity rating)", axis=alt.Axis(values=list(range(0, max_items, 5))), sort='descending'), color = alt.Color("key:O", scale=alt.Scale( domain=["System says: Non-toxic", "System says: Toxic"], range=["white", "#cbcbcb"]), title="System rating", legend=None, ), href="url:N", tooltip = [ alt.Tooltip("comment_:N", title="comment"), alt.Tooltip("rating:Q", title="System rating", format=".2f"), ] ) # Threshold line rule = alt.Chart(pd.DataFrame({ "threshold": [threshold], })).mark_rule(color='grey').encode( x=alt.X("mean(threshold):Q", scale=alt.Scale(domain=[domain_min, domain_max]), title=""), size=alt.value(2), ) # Plot region annotations nontoxic_x = (domain_min + threshold) / 2. toxic_x = (domain_max + threshold) / 2. annotation = alt.Chart(pd.DataFrame({ "annotation_text": ["Non-toxic", "Toxic"], "x": [nontoxic_x, toxic_x], "y": [max_items, max_items], })).mark_text( align="center", baseline="middle", fontSize=16, dy=10, color="grey" ).encode( x=alt.X("x", title=""), y=alt.Y("y", title="", axis=None), text="annotation_text" ) # Plot region background colors bkgd = alt.Chart(pd.DataFrame({ "start": [domain_min, threshold], "stop": [threshold, domain_max], "bkgd": ["Non-toxic", "Toxic"], })).mark_rect(opacity=1.0, stroke="grey", strokeWidth=0.25).encode( x=alt.X("start:Q", scale=alt.Scale(domain=[domain_min, domain_max])), x2=alt.X2("stop:Q"), y=alt.value(0), y2=alt.value(plot_dim_height), color=alt.Color("bkgd:O", scale=alt.Scale( domain=["Non-toxic", "Toxic"], range=["white", "#cbcbcb"]), title="System rating" ) ) final_plot = (bkgd + annotation + chart + rule).properties(height=(plot_dim_height), width=plot_dim_width).resolve_scale(color='independent').to_json() return final_plot, df def plot_overall_vis_cluster(preds_df, error_type, n_comments=None, bins=VIS_BINS, threshold=TOXIC_THRESHOLD, bin_step=0.05): df = preds_df.copy().reset_index(drop=True) # df = df[df["topic_"] == topic] df["vis_pred_bin"], out_bins = pd.cut(df["pred"], bins, labels=VIS_BINS_LABELS, retbins=True) df = df[df["user_id"] == "A"].sort_values(by=["rating"]).reset_index(drop=True) df["system_label"] = [("toxic" if r > threshold else "non-toxic") for r in df["rating"].tolist()] df["key"] = [get_key(sys, user, threshold) for sys, user in zip(df["rating"].tolist(), df["pred"].tolist())] print("len(df)", len(df)) # always 0 for some reason (from keyword search) # print("columns", df.columns) df["category"] = df.apply(lambda row: get_category(row), axis=1) df["url"] = df.apply(lambda row: get_comment_url(row), axis=1) if n_comments is not None: n_to_sample = np.min([n_comments, len(df)]) df = df.sample(n=n_to_sample) # Plot sizing domain_min = 0 domain_max = 4 plot_dim_height = 500 plot_dim_width = 750 max_items = np.max(df["vis_pred_bin"].value_counts().tolist()) mark_size = np.round(plot_dim_height / max_items) * 8 if mark_size > 75: mark_size = 75 plot_dim_height = 13 * max_items # Main chart chart = alt.Chart(df).mark_square(opacity=0.8, size=mark_size, stroke="grey", strokeWidth=0.25).transform_window( groupby=['vis_pred_bin'], sort=[{'field': 'rating'}], id='row_number()', ignorePeers=True ).encode( x=alt.X('vis_pred_bin:Q', title="Our prediction of your rating", scale=alt.Scale(domain=(domain_min, domain_max))), y=alt.Y('id:O', title="Comments (ordered by System toxicity rating)", axis=alt.Axis(values=list(range(0, max_items, 5))), sort='descending'), color = alt.Color("key:O", scale=alt.Scale( domain=["System agrees: Non-toxic", "System agrees: Toxic", "System differs: Error > 1.5", "System differs: Error > 1.0", "System differs: Error > 0.5", "System differs: Error <=0.5"], range=["white", "#cbcbcb", "red", "#ff7a5c", "#ffa894", "#ffd1c7"]), title="System rating (box color)" ), href="url:N", tooltip = [ alt.Tooltip("comment_:N", title="comment"), alt.Tooltip("rating:Q", title="System rating", format=".2f"), alt.Tooltip("pred:Q", title="Your rating", format=".2f"), alt.Tooltip("category:N", title="Potential toxicity categories") ] ) # Filter to specified error type if error_type == "System is under-sensitive": # FN: system rates non-toxic, but user rates toxic chart = chart.transform_filter( alt.FieldGTPredicate(field="pred", gt=threshold) ) elif error_type == "System is over-sensitive": # FP: system rates toxic, but user rates non-toxic chart = chart.transform_filter( alt.FieldLTEPredicate(field="pred", lte=threshold) ) # Threshold line rule = alt.Chart(pd.DataFrame({ "threshold": [threshold], })).mark_rule(color='grey').encode( x=alt.X("mean(threshold):Q", scale=alt.Scale(domain=[domain_min, domain_max]), title=""), size=alt.value(2), ) # Plot region annotations nontoxic_x = (domain_min + threshold) / 2. toxic_x = (domain_max + threshold) / 2. annotation = alt.Chart(pd.DataFrame({ "annotation_text": ["Non-toxic", "Toxic"], "x": [nontoxic_x, toxic_x], "y": [max_items, max_items], })).mark_text( align="center", baseline="middle", fontSize=16, dy=10, color="grey" ).encode( x=alt.X("x", title=""), y=alt.Y("y", title="", axis=None), text="annotation_text" ) # Plot region background colors bkgd = alt.Chart(pd.DataFrame({ "start": [domain_min, threshold], "stop": [threshold, domain_max], "bkgd": ["Non-toxic (L side)", "Toxic (R side)"], })).mark_rect(opacity=1.0, stroke="grey", strokeWidth=0.25).encode( x=alt.X("start:Q", scale=alt.Scale(domain=[domain_min, domain_max])), x2=alt.X2("stop:Q"), y=alt.value(0), y2=alt.value(plot_dim_height), color=alt.Color("bkgd:O", scale=alt.Scale( domain=["Non-toxic (L side)", "Toxic (R side)"], range=["white", "#cbcbcb"]), title="Your rating (background color)" ) ) final_plot = (bkgd + annotation + chart + rule).properties(height=(plot_dim_height), width=plot_dim_width).resolve_scale(color='independent').to_json() return final_plot, df def get_cluster_comments(df, error_type, threshold=TOXIC_THRESHOLD, worker_id="A", num_examples=50, use_model=True): df["user_color"] = [get_user_color(user, threshold) for user in df["pred"].tolist()] # get cell colors df["system_color"] = [get_user_color(sys, threshold) for sys in df["rating"].tolist()] # get cell colors df["error_color"] = [get_system_color(sys, user, threshold) for sys, user in zip(df["rating"].tolist(), df["pred"].tolist())] # get cell colors df["error_type"] = [get_error_type(sys, user, threshold) for sys, user in zip(df["rating"].tolist(), df["pred"].tolist())] # get error type in words df["error_amt"] = [abs(sys - threshold) for sys in df["rating"].tolist()] # get raw error df["judgment"] = ["" for _ in range(len(df))] # template for "agree" or "disagree" buttons if use_model: df = df.sort_values(by=["error_amt"], ascending=False) # surface largest errors first else: print("get_cluster_comments; not using model") df = df.sort_values(by=["rating"], ascending=True) df["id"] = df["item_id"] # df["comment"] already exists df["comment"] = df["comment_"] df["toxicity_category"] = df["category"] df["user_rating"] = df["pred"] df["user_decision"] = [get_decision(rating, threshold) for rating in df["pred"].tolist()] df["system_rating"] = df["rating"] df["system_decision"] = [get_decision(rating, threshold) for rating in df["rating"].tolist()] df["error_type"] = df["error_type"] df = df.head(num_examples) df = df.round(decimals=2) # Filter to specified error type if error_type == "System is under-sensitive": # FN: system rates non-toxic, but user rates toxic df = df[df["error_type"] == "System may be under-sensitive"] elif error_type == "System is over-sensitive": # FP: system rates toxic, but user rates non-toxic df = df[df["error_type"] == "System may be over-sensitive" ] elif error_type == "Both": df = df[(df["error_type"] == "System may be under-sensitive") | (df["error_type"] == "System may be over-sensitive")] return df.to_json(orient="records") # PERSONALIZED CLUSTERS utils def get_disagreement_comments(preds_df, mode, n=10_000, threshold=TOXIC_THRESHOLD): # Get difference between user rating and system rating df = preds_df.copy() df["diff"] = [get_error_size(sys, user, threshold) for sys, user in zip(df["rating"].tolist(), df["pred"].tolist())] df["error_type"] = [get_error_type(sys, user, threshold) for sys, user in zip(df["rating"].tolist(), df["pred"].tolist())] # asc = low to high; lowest = sys lower than user (under-sensitive) # desc = high to low; lowest = sys higher than user (over-sensitive) if mode == "under-sensitive": df = df[df["error_type"] == "System may be under-sensitive"] asc = True elif mode == "over-sensitive": df = df[df["error_type"] == "System may be over-sensitive"] asc = False df = df.sort_values(by=["diff"], ascending=asc) df = df.head(n) return df["comment_"].tolist(), df def get_personal_clusters(model, n=3): personal_cluster_file = f"./data/personal_cluster_dfs/{model}.pkl" if (os.path.isfile(personal_cluster_file)): with open(personal_cluster_file, "rb") as f: cluster_df = pickle.load(f) cluster_df = cluster_df.sort_values(by=["topic_id"]) topics_under = cluster_df[cluster_df["error_type"] == "System may be under-sensitive"]["topic"].unique().tolist() topics_under = topics_under[1:(n + 1)] topics_over = cluster_df[cluster_df["error_type"] == "System may be over-sensitive"]["topic"].unique().tolist() topics_over = topics_over[1:(n + 1)] return topics_under, topics_over else: topics_under_top = [] topics_over_top = [] preds_df_file = f"./data/preds_dfs/{model}.pkl" if (os.path.isfile(preds_df_file)): with open(preds_df_file, "rb") as f: preds_df = pickle.load(f) preds_df_mod = preds_df.merge(comments_grouped_full_topic_cat, on="item_id", how="left", suffixes=('_', '_avg')).reset_index() preds_df_mod = preds_df_mod[preds_df_mod["user_id"] == "A"] comments_under, comments_under_df = get_disagreement_comments(preds_df_mod, mode="under-sensitive", n=1000) if len(comments_under) > 0: topics_under = BERTopic(embedding_model="paraphrase-MiniLM-L6-v2").fit(comments_under) topics_under_top = topics_under.get_topic_info().head(n)["Name"].tolist() print("topics_under", topics_under_top) # Get topics per comment topics_assigned, _ = topics_under.transform(comments_under) comments_under_df["topic_id"] = topics_assigned cur_topic_ids = topics_under.get_topic_info().Topic topic_short_names = topics_under.get_topic_info().Name topic_ids_to_names = {cur_topic_ids[i]: topic_short_names[i] for i in range(len(cur_topic_ids))} comments_under_df["topic"] = [topic_ids_to_names[topic_id] for topic_id in comments_under_df["topic_id"].tolist()] comments_over, comments_over_df = get_disagreement_comments(preds_df_mod, mode="over-sensitive", n=1000) if len(comments_over) > 0: topics_over = BERTopic(embedding_model="paraphrase-MiniLM-L6-v2").fit(comments_over) topics_over_top = topics_over.get_topic_info().head(n)["Name"].tolist() print("topics_over", topics_over_top) # Get topics per comment topics_assigned, _ = topics_over.transform(comments_over) comments_over_df["topic_id"] = topics_assigned cur_topic_ids = topics_over.get_topic_info().Topic topic_short_names = topics_over.get_topic_info().Name topic_ids_to_names = {cur_topic_ids[i]: topic_short_names[i] for i in range(len(cur_topic_ids))} comments_over_df["topic"] = [topic_ids_to_names[topic_id] for topic_id in comments_over_df["topic_id"].tolist()] cluster_df = pd.concat([comments_under_df, comments_over_df]) with open(f"./data/personal_cluster_dfs/{model}.pkl", "wb") as f: pickle.dump(cluster_df, f) return topics_under_top, topics_over_top return [], []