Module divAtScale.src.helpers.semantic_helpers.diversity_measures
Spatial disparity helper functions
Expand source code
"""
Spatial disparity helper functions
"""
import numpy as np
from collections import Counter
from sklearn.metrics.pairwise import cosine_similarity
from scipy.spatial.distance import cosine as cos_dis
from tqdm import tqdm
import random
def comp_w_centroid(s, e, aid2mid):
"""
Compute a sessions activity weighted centroid in embedding space
Args:
s (list) : session (artist ids)
e (numpy.array) : embedding matrix
aid2mid (dict) : matrix index lookup
Returns:
numpy.array : centroid
"""
S_len = len(s)
w = dict(Counter(s))
a_unique = list(set(s))
e_all = np.array([e[int(aid2mid[aid])] for aid in a_unique])
e_w = [e * w[aid] for e, aid in zip(e_all, a_unique)]
w_centroid = np.sum(e_w, axis=0) / S_len
return w_centroid
def gs_score(s, e, aid2mid):
"""
Compute Anderson et al.'s GS-score
Args:
s (list): session
e (numpy.array): embedding matrix
aid2mid (dict): matrix index lookup
Returns:
float : gs_score for a given session
"""
S_len = len(s)
w = dict(Counter(s))
a_unique = list(set(s))
w_vec = np.array([w[aid] for aid in a_unique])
e_all = np.array([e[int(aid2mid[aid])] for aid in a_unique])
e_w = [e * w[aid] for e, aid in zip(e_all, a_unique)]
w_centroid = np.sum(e_w, axis=0) / S_len
sim = cosine_similarity([w_centroid], e_all)[0]
sim = sim * w_vec
gs = sum(sim) / S_len
return gs
def gs_score_inter_sess(S, e, aid2mid):
"""
Compute Anderson et al.'s GS-score between multiple sessions
Args:
S (list): list of sessions (artist ids)
e (numpy.array) : embedding matrix
aid2mid (dict) : matrix index lookup
Returns:
float : gs_score over all sessions
"""
sess_rep = [comp_w_centroid(s, e, aid2mid) for s in S]
centroid = np.mean(sess_rep, axis=0)
sim = cosine_similarity([centroid], sess_rep)[0]
gs = np.mean(sim)
return gs
def sim_inter_aff(S1, S2, e, aid2mid):
"""
Compute cosine similarity between avg. rep of two session centroids
generated from two distinct affordances.
Args:
S (list): list of sessions (artist ids)
e (numpy.array) : embedding matrix
aid2mid (dict): matrix index lookup
Returns:
float : cosine similarity between two affordances
"""
sess_rep1 = [comp_w_centroid(s, e, aid2mid) for s in S1]
centroid1 = np.mean(sess_rep1, axis=0)
sess_rep2 = [comp_w_centroid(s, e, aid2mid) for s in S2]
centroid2 = np.mean(sess_rep2, axis=0)
sim = 1 - cos_dis(centroid1, centroid2)
return sim
def filter_niche_sessions(sess, u_df, subset_a, pop_thresh=0.8):
"""
Remove sessions with too much long tail.
Do not have good structural rep. for these artists.
Args:
sess (list) : collection of sesssions
u_df (pandas.DataFrame) : listening history df
subset_a (list) : pop artists to keep
Returns:
list : filtered sessions
"""
streamed_artists = u_df.artist_id.unique()
pop_a = set(streamed_artists).intersection(set(subset_a))
S_filt = []
for s in sess:
pop_a_in_s = pop_a.intersection(set(s))
p_pop_in_session = len(pop_a_in_s) / len(set(s))
if p_pop_in_session > pop_thresh:
S_filt.append(list(pop_a_in_s))
return S_filt
def comp_gs_intra_or_inter_fs2(lh_df, e, aid2mid, n_sample=-1, scale='intra', subset_a=[], sess_thresh=3):
"""
Top level function to compute gs-scores between (inter)
or wihtin (intra) sessions w.r.t. fs2 filtering.
Args:
lh_df (pandas.DataFrame): listening event df
e (numpy.array) : embedding matrix
aid2mid (dict): matrix index lookup for e
n_sample (int) : if != -1 sample n_sample users at random
scale (str) : 'intra' OR 'inter'
Returns:
list : all users gs scores in Q,P,A,E. Each row corresponds to a user
"""
pdis_res = []
n_failed = 0
P_of_sess_pop = []
uids_keep = []
uids = list(lh_df.anon_user_id.unique())
if n_sample != -1:
uids = random.sample(uids, n_sample)
for uid in tqdm(uids):
spread_os = -1
spread_op = -1
spread_a = -1
spread_e = -1
u_df = lh_df[lh_df.anon_user_id == uid]
sess_os = [list(s) for s in list(u_df[u_df.affordance == 'o_s'].groupby('session_n').artist_id.apply(list))]
sess_op = [list(s) for s in list(u_df[u_df.affordance == 'o_pl'].groupby('session_n').artist_id.apply(list))]
sess_a = [list(s) for s in list(u_df[u_df.affordance == 'a'].groupby('session_n').artist_id.apply(list))]
sess_e = [list(s) for s in list(u_df[u_df.affordance == 'e'].groupby('session_n').artist_id.apply(list))]
if len(subset_a) != 0:
sess_os_filt = filter_niche_sessions(sess_os, u_df, subset_a)
p_pop_os = len(sess_os_filt) / len(sess_os)
sess_os = sess_os_filt
sess_op_filt = filter_niche_sessions(sess_op, u_df, subset_a)
p_pop_op = len(sess_op_filt) / len(sess_op)
sess_op = sess_op_filt
sess_a_filt = filter_niche_sessions(sess_a, u_df, subset_a)
p_pop_a = len(sess_a_filt) / len(sess_a)
sess_a = sess_a_filt
sess_e_filt = filter_niche_sessions(sess_e, u_df, subset_a)
p_pop_e = len(sess_e_filt) / len(sess_e)
sess_e = sess_e_filt
if scale == "intra":
if len(sess_os) >= sess_thresh:
spread_os = np.mean([gs_score(aids, e, aid2mid) for aids in sess_os])
if len(sess_op) >= sess_thresh:
spread_op = np.mean([gs_score(aids, e, aid2mid) for aids in sess_op])
if len(sess_a) >= sess_thresh:
spread_a = np.mean([gs_score(aids, e, aid2mid) for aids in sess_a])
if len(sess_e) >= sess_thresh:
spread_e = np.mean([gs_score(aids, e, aid2mid) for aids in sess_e])
elif scale == "inter":
if len(sess_os) >= sess_thresh:
spread_os = gs_score_inter_sess(sess_os, e, aid2mid)
if len(sess_op) >= sess_thresh:
spread_op = gs_score_inter_sess(sess_op, e, aid2mid)
if len(sess_a) >= sess_thresh:
spread_a = gs_score_inter_sess(sess_a, e, aid2mid)
if len(sess_e) >= sess_thresh:
spread_e = gs_score_inter_sess(sess_e, e, aid2mid)
all_spreads = [spread_os, spread_op, spread_a, spread_e]
if (-1 in all_spreads) == False:
pdis_res.append([spread_os, spread_op, spread_a, spread_e])
P_of_sess_pop.append([p_pop_os, p_pop_op, p_pop_a, p_pop_e])
uids_keep.append(uid)
else:
n_failed += 1
avg_sess_left_post_niche_filt = np.mean(P_of_sess_pop, axis=0)
pdis_res = np.array(pdis_res)
p_failed = n_failed / len(uids)
return pdis_res, p_failed, avg_sess_left_post_niche_filt, uids_keep
def comp_gs_intra_or_inter_fs1(monad_df_filt, e, aid2mid, scale="intra", subset_a=[], sess_thresh=3):
"""
Top level function to compute gs-scores between (inter) or wihtin (intra) sessions w.r.t. fs1 filtering.
Args:
monad_df_filt (pandas.DataFrame): listening history for monadic sessions (i.e. either P|Q|A|E)
e (numpy.array): embedding matrix
aid2mid (dict): matrix index lookup for e
Returns:
list : gs for all users. rows correspond to users
"""
uids = list(monad_df_filt.anon_user_id.unique())
n_failed = 0
GS_all = []
P_of_sess_pop = []
for uid in tqdm(uids):
GS_avg = -1
u_df = monad_df_filt[monad_df_filt.anon_user_id == uid]
sess = [list(s) for s in list(u_df.groupby('session_n').artist_id.apply(list))]
if len(subset_a) != 0:
sess_filt = filter_niche_sessions(sess, u_df, subset_a)
p_sess_post_filt = len(sess_filt) / len(sess)
sess = sess_filt
if scale == "intra":
# asssert we stll have some sessions to work with...
if len(sess) >= sess_thresh:
GS_avg = np.mean([gs_score(s, e, aid2mid) for s in sess])
elif scale == "inter":
if len(sess) >= sess_thresh:
GS_avg = gs_score_inter_sess(sess, e, aid2mid)
if GS_avg != -1:
GS_all.append(GS_avg)
P_of_sess_pop.append(p_sess_post_filt)
else:
n_failed +=1
p_failed = n_failed / len(uids)
return GS_all, p_failed, np.mean(P_of_sess_pop)
def get_sim_between_aff(le_df, e, aid2mid, subset_a=[]):
"""
Top level function to compute semantic (cosine) similarity between all affordances for all users.
Args:
le_df (pandas.DataFrame): listening history df
e (numpy.array): embedding matrix
aid2mid (dict): matrix index lookup for e
Returns:
list :similarity for [Q-P, Q-A, Q-E, P-A, P-E, A-E]. Rows correspond to users
"""
psim_res = []
uids = le_df.anon_user_id.unique()
n_failed = 0
for uid in tqdm(uids):
u_df = le_df[le_df.anon_user_id == uid]
sess_os = [list(s) for s in list(u_df[u_df.affordance == 'o_s'].groupby('session_n').artist_id.apply(list))]
sess_op = [list(s) for s in list(u_df[u_df.affordance == 'o_pl'].groupby('session_n').artist_id.apply(list))]
sess_a = [list(s) for s in list(u_df[u_df.affordance == 'a'].groupby('session_n').artist_id.apply(list))]
sess_e = [list(s) for s in list(u_df[u_df.affordance == 'e'].groupby('session_n').artist_id.apply(list))]
if len(subset_a) != 0:
sess_os = filter_niche_sessions(sess_os, u_df, subset_a)
sess_op = filter_niche_sessions(sess_op, u_df, subset_a)
sess_a = filter_niche_sessions(sess_a, u_df, subset_a)
sess_e = filter_niche_sessions(sess_e, u_df, subset_a)
if (len(sess_op) > 0) & (len(sess_os) > 0) & (len(sess_a) > 0) & (len(sess_e) > 0):
sim_Q_P = sim_inter_aff(sess_os, sess_op, e, aid2mid)
sim_Q_A = sim_inter_aff(sess_os, sess_a, e, aid2mid)
sim_Q_E = sim_inter_aff(sess_os, sess_e, e, aid2mid)
sim_P_A = sim_inter_aff(sess_op, sess_a, e, aid2mid)
sim_P_E = sim_inter_aff(sess_op, sess_e, e, aid2mid)
sim_A_E = sim_inter_aff(sess_a, sess_e, e, aid2mid)
avg_sim = np.mean([sim_Q_P, sim_Q_A, sim_Q_E, sim_P_A, sim_P_E, sim_A_E])
std_sim = np.std([sim_Q_P, sim_Q_A, sim_Q_E, sim_P_A, sim_P_E, sim_A_E])
sim_vec = np.array([sim_Q_P, sim_Q_A, sim_Q_E, sim_P_A, sim_P_E, sim_A_E])
# znorm vec
sim_vec = (sim_vec - avg_sim) / std_sim
psim_res.append(sim_vec)
else:
n_failed += 1
p_failed = n_failed / len(uids)
return psim_res, p_failedFunctions
def comp_gs_intra_or_inter_fs1(monad_df_filt, e, aid2mid, scale='intra', subset_a=[], sess_thresh=3)-
Top level function to compute gs-scores between (inter) or wihtin (intra) sessions w.r.t. fs1 filtering.
Args
monad_df_filt:pandas.DataFrame- listening history for monadic sessions (i.e. either P|Q|A|E)
e:numpy.array- embedding matrix
aid2mid:dict- matrix index lookup for e
Returns
list- gs for all users. rows correspond to users
Expand source code
def comp_gs_intra_or_inter_fs1(monad_df_filt, e, aid2mid, scale="intra", subset_a=[], sess_thresh=3): """ Top level function to compute gs-scores between (inter) or wihtin (intra) sessions w.r.t. fs1 filtering. Args: monad_df_filt (pandas.DataFrame): listening history for monadic sessions (i.e. either P|Q|A|E) e (numpy.array): embedding matrix aid2mid (dict): matrix index lookup for e Returns: list : gs for all users. rows correspond to users """ uids = list(monad_df_filt.anon_user_id.unique()) n_failed = 0 GS_all = [] P_of_sess_pop = [] for uid in tqdm(uids): GS_avg = -1 u_df = monad_df_filt[monad_df_filt.anon_user_id == uid] sess = [list(s) for s in list(u_df.groupby('session_n').artist_id.apply(list))] if len(subset_a) != 0: sess_filt = filter_niche_sessions(sess, u_df, subset_a) p_sess_post_filt = len(sess_filt) / len(sess) sess = sess_filt if scale == "intra": # asssert we stll have some sessions to work with... if len(sess) >= sess_thresh: GS_avg = np.mean([gs_score(s, e, aid2mid) for s in sess]) elif scale == "inter": if len(sess) >= sess_thresh: GS_avg = gs_score_inter_sess(sess, e, aid2mid) if GS_avg != -1: GS_all.append(GS_avg) P_of_sess_pop.append(p_sess_post_filt) else: n_failed +=1 p_failed = n_failed / len(uids) return GS_all, p_failed, np.mean(P_of_sess_pop) def comp_gs_intra_or_inter_fs2(lh_df, e, aid2mid, n_sample=-1, scale='intra', subset_a=[], sess_thresh=3)-
Top level function to compute gs-scores between (inter) or wihtin (intra) sessions w.r.t. fs2 filtering.
Args
lh_df:pandas.DataFrame- listening event df
- e (numpy.array) : embedding matrix
aid2mid:dict- matrix index lookup for e
n_sample (int) : if != -1 sample n_sample users at random scale (str) : 'intra' OR 'inter'
Returns
list- all users gs scores in Q,P,A,E. Each row corresponds to a user
Expand source code
def comp_gs_intra_or_inter_fs2(lh_df, e, aid2mid, n_sample=-1, scale='intra', subset_a=[], sess_thresh=3): """ Top level function to compute gs-scores between (inter) or wihtin (intra) sessions w.r.t. fs2 filtering. Args: lh_df (pandas.DataFrame): listening event df e (numpy.array) : embedding matrix aid2mid (dict): matrix index lookup for e n_sample (int) : if != -1 sample n_sample users at random scale (str) : 'intra' OR 'inter' Returns: list : all users gs scores in Q,P,A,E. Each row corresponds to a user """ pdis_res = [] n_failed = 0 P_of_sess_pop = [] uids_keep = [] uids = list(lh_df.anon_user_id.unique()) if n_sample != -1: uids = random.sample(uids, n_sample) for uid in tqdm(uids): spread_os = -1 spread_op = -1 spread_a = -1 spread_e = -1 u_df = lh_df[lh_df.anon_user_id == uid] sess_os = [list(s) for s in list(u_df[u_df.affordance == 'o_s'].groupby('session_n').artist_id.apply(list))] sess_op = [list(s) for s in list(u_df[u_df.affordance == 'o_pl'].groupby('session_n').artist_id.apply(list))] sess_a = [list(s) for s in list(u_df[u_df.affordance == 'a'].groupby('session_n').artist_id.apply(list))] sess_e = [list(s) for s in list(u_df[u_df.affordance == 'e'].groupby('session_n').artist_id.apply(list))] if len(subset_a) != 0: sess_os_filt = filter_niche_sessions(sess_os, u_df, subset_a) p_pop_os = len(sess_os_filt) / len(sess_os) sess_os = sess_os_filt sess_op_filt = filter_niche_sessions(sess_op, u_df, subset_a) p_pop_op = len(sess_op_filt) / len(sess_op) sess_op = sess_op_filt sess_a_filt = filter_niche_sessions(sess_a, u_df, subset_a) p_pop_a = len(sess_a_filt) / len(sess_a) sess_a = sess_a_filt sess_e_filt = filter_niche_sessions(sess_e, u_df, subset_a) p_pop_e = len(sess_e_filt) / len(sess_e) sess_e = sess_e_filt if scale == "intra": if len(sess_os) >= sess_thresh: spread_os = np.mean([gs_score(aids, e, aid2mid) for aids in sess_os]) if len(sess_op) >= sess_thresh: spread_op = np.mean([gs_score(aids, e, aid2mid) for aids in sess_op]) if len(sess_a) >= sess_thresh: spread_a = np.mean([gs_score(aids, e, aid2mid) for aids in sess_a]) if len(sess_e) >= sess_thresh: spread_e = np.mean([gs_score(aids, e, aid2mid) for aids in sess_e]) elif scale == "inter": if len(sess_os) >= sess_thresh: spread_os = gs_score_inter_sess(sess_os, e, aid2mid) if len(sess_op) >= sess_thresh: spread_op = gs_score_inter_sess(sess_op, e, aid2mid) if len(sess_a) >= sess_thresh: spread_a = gs_score_inter_sess(sess_a, e, aid2mid) if len(sess_e) >= sess_thresh: spread_e = gs_score_inter_sess(sess_e, e, aid2mid) all_spreads = [spread_os, spread_op, spread_a, spread_e] if (-1 in all_spreads) == False: pdis_res.append([spread_os, spread_op, spread_a, spread_e]) P_of_sess_pop.append([p_pop_os, p_pop_op, p_pop_a, p_pop_e]) uids_keep.append(uid) else: n_failed += 1 avg_sess_left_post_niche_filt = np.mean(P_of_sess_pop, axis=0) pdis_res = np.array(pdis_res) p_failed = n_failed / len(uids) return pdis_res, p_failed, avg_sess_left_post_niche_filt, uids_keep def comp_w_centroid(s, e, aid2mid)-
Compute a sessions activity weighted centroid in embedding space
Args
s (list) : session (artist ids) e (numpy.array) : embedding matrix aid2mid (dict) : matrix index lookup
Returns
numpy.array- centroid
Expand source code
def comp_w_centroid(s, e, aid2mid): """ Compute a sessions activity weighted centroid in embedding space Args: s (list) : session (artist ids) e (numpy.array) : embedding matrix aid2mid (dict) : matrix index lookup Returns: numpy.array : centroid """ S_len = len(s) w = dict(Counter(s)) a_unique = list(set(s)) e_all = np.array([e[int(aid2mid[aid])] for aid in a_unique]) e_w = [e * w[aid] for e, aid in zip(e_all, a_unique)] w_centroid = np.sum(e_w, axis=0) / S_len return w_centroid def filter_niche_sessions(sess, u_df, subset_a, pop_thresh=0.8)-
Remove sessions with too much long tail. Do not have good structural rep. for these artists.
Args
sess (list) : collection of sesssions u_df (pandas.DataFrame) : listening history df subset_a (list) : pop artists to keep
Returns
list- filtered sessions
Expand source code
def filter_niche_sessions(sess, u_df, subset_a, pop_thresh=0.8): """ Remove sessions with too much long tail. Do not have good structural rep. for these artists. Args: sess (list) : collection of sesssions u_df (pandas.DataFrame) : listening history df subset_a (list) : pop artists to keep Returns: list : filtered sessions """ streamed_artists = u_df.artist_id.unique() pop_a = set(streamed_artists).intersection(set(subset_a)) S_filt = [] for s in sess: pop_a_in_s = pop_a.intersection(set(s)) p_pop_in_session = len(pop_a_in_s) / len(set(s)) if p_pop_in_session > pop_thresh: S_filt.append(list(pop_a_in_s)) return S_filt def get_sim_between_aff(le_df, e, aid2mid, subset_a=[])-
Top level function to compute semantic (cosine) similarity between all affordances for all users.
Args
le_df:pandas.DataFrame- listening history df
e:numpy.array- embedding matrix
aid2mid:dict- matrix index lookup for e
Returns
list :similarity for [Q-P, Q-A, Q-E, P-A, P-E, A-E]. Rows correspond to users
Expand source code
def get_sim_between_aff(le_df, e, aid2mid, subset_a=[]): """ Top level function to compute semantic (cosine) similarity between all affordances for all users. Args: le_df (pandas.DataFrame): listening history df e (numpy.array): embedding matrix aid2mid (dict): matrix index lookup for e Returns: list :similarity for [Q-P, Q-A, Q-E, P-A, P-E, A-E]. Rows correspond to users """ psim_res = [] uids = le_df.anon_user_id.unique() n_failed = 0 for uid in tqdm(uids): u_df = le_df[le_df.anon_user_id == uid] sess_os = [list(s) for s in list(u_df[u_df.affordance == 'o_s'].groupby('session_n').artist_id.apply(list))] sess_op = [list(s) for s in list(u_df[u_df.affordance == 'o_pl'].groupby('session_n').artist_id.apply(list))] sess_a = [list(s) for s in list(u_df[u_df.affordance == 'a'].groupby('session_n').artist_id.apply(list))] sess_e = [list(s) for s in list(u_df[u_df.affordance == 'e'].groupby('session_n').artist_id.apply(list))] if len(subset_a) != 0: sess_os = filter_niche_sessions(sess_os, u_df, subset_a) sess_op = filter_niche_sessions(sess_op, u_df, subset_a) sess_a = filter_niche_sessions(sess_a, u_df, subset_a) sess_e = filter_niche_sessions(sess_e, u_df, subset_a) if (len(sess_op) > 0) & (len(sess_os) > 0) & (len(sess_a) > 0) & (len(sess_e) > 0): sim_Q_P = sim_inter_aff(sess_os, sess_op, e, aid2mid) sim_Q_A = sim_inter_aff(sess_os, sess_a, e, aid2mid) sim_Q_E = sim_inter_aff(sess_os, sess_e, e, aid2mid) sim_P_A = sim_inter_aff(sess_op, sess_a, e, aid2mid) sim_P_E = sim_inter_aff(sess_op, sess_e, e, aid2mid) sim_A_E = sim_inter_aff(sess_a, sess_e, e, aid2mid) avg_sim = np.mean([sim_Q_P, sim_Q_A, sim_Q_E, sim_P_A, sim_P_E, sim_A_E]) std_sim = np.std([sim_Q_P, sim_Q_A, sim_Q_E, sim_P_A, sim_P_E, sim_A_E]) sim_vec = np.array([sim_Q_P, sim_Q_A, sim_Q_E, sim_P_A, sim_P_E, sim_A_E]) # znorm vec sim_vec = (sim_vec - avg_sim) / std_sim psim_res.append(sim_vec) else: n_failed += 1 p_failed = n_failed / len(uids) return psim_res, p_failed def gs_score(s, e, aid2mid)-
Compute Anderson et al.'s GS-score
Args
s:list- session
e:numpy.array- embedding matrix
aid2mid:dict- matrix index lookup
Returns
float- gs_score for a given session
Expand source code
def gs_score(s, e, aid2mid): """ Compute Anderson et al.'s GS-score Args: s (list): session e (numpy.array): embedding matrix aid2mid (dict): matrix index lookup Returns: float : gs_score for a given session """ S_len = len(s) w = dict(Counter(s)) a_unique = list(set(s)) w_vec = np.array([w[aid] for aid in a_unique]) e_all = np.array([e[int(aid2mid[aid])] for aid in a_unique]) e_w = [e * w[aid] for e, aid in zip(e_all, a_unique)] w_centroid = np.sum(e_w, axis=0) / S_len sim = cosine_similarity([w_centroid], e_all)[0] sim = sim * w_vec gs = sum(sim) / S_len return gs def gs_score_inter_sess(S, e, aid2mid)-
Compute Anderson et al.'s GS-score between multiple sessions
Args
S:list- list of sessions (artist ids)
e (numpy.array) : embedding matrix aid2mid (dict) : matrix index lookup
Returns
float- gs_score over all sessions
Expand source code
def gs_score_inter_sess(S, e, aid2mid): """ Compute Anderson et al.'s GS-score between multiple sessions Args: S (list): list of sessions (artist ids) e (numpy.array) : embedding matrix aid2mid (dict) : matrix index lookup Returns: float : gs_score over all sessions """ sess_rep = [comp_w_centroid(s, e, aid2mid) for s in S] centroid = np.mean(sess_rep, axis=0) sim = cosine_similarity([centroid], sess_rep)[0] gs = np.mean(sim) return gs def sim_inter_aff(S1, S2, e, aid2mid)-
Compute cosine similarity between avg. rep of two session centroids generated from two distinct affordances.
Args
S:list- list of sessions (artist ids)
- e (numpy.array) : embedding matrix
aid2mid:dict- matrix index lookup
Returns
float- cosine similarity between two affordances
Expand source code
def sim_inter_aff(S1, S2, e, aid2mid): """ Compute cosine similarity between avg. rep of two session centroids generated from two distinct affordances. Args: S (list): list of sessions (artist ids) e (numpy.array) : embedding matrix aid2mid (dict): matrix index lookup Returns: float : cosine similarity between two affordances """ sess_rep1 = [comp_w_centroid(s, e, aid2mid) for s in S1] centroid1 = np.mean(sess_rep1, axis=0) sess_rep2 = [comp_w_centroid(s, e, aid2mid) for s in S2] centroid2 = np.mean(sess_rep2, axis=0) sim = 1 - cos_dis(centroid1, centroid2) return sim