sampling.py

import json
from pathlib import Path

import numpy as np
import pandas as pd
import scipy.stats as sts

import ir_datasets as irds
import ir_measures as irms

from tqdm import tqdm

domains = ['science', 'recreation', 'technology', 'lifestyle', 'writing']

# Define the distribution
domain_distributions = pd.concat({
    "D1": pd.DataFrame({ 
        domains[i]: [sts.distributions.betabinom(a=(i+1), b=(5-i), n=4).pmf(j) for j in range(5) ]
        for i in range(5)
    }),
    "D2": pd.DataFrame({ 
        domains[i]: ([sts.distributions.betabinom(a=1, b=10, n=5).pmf(j) for j in range(5) ]*2)[5-i:10-i]
        for i in range(5)
    }),
    "D3": pd.DataFrame({ 
        domains[i]: ([0]*i + [sts.distributions.betabinom(a=1, b=(5-i)*2, n=4-i).pmf(j) for j in range(5) ])[:5]
        for i in range(5)
    })
}, names=['dist_type', 'session'], axis=0).rename_axis('domains', axis=1)


sampling_cdf = domain_distributions.unstack('dist_type').cumsum(axis=0).pipe(lambda x: x/x.loc[4])


# Sampling for queries
np.random.seed(123)

fps = {
    dt: { i: open(f'./stream_distribution_dss/queries/test_{dt}_{i}.jsonl', 'w') for i in sampling_cdf.index }
    for dt in sampling_cdf.columns.get_level_values('dist_type').unique()
}

for dom in tqdm(domains, desc='queries'):
    cdfs = sampling_cdf[dom]

    for query in tqdm(irds.load(f'lotte/{dom}/test/forum').queries, desc=dom):
        query = { 'query_id': dom+query.query_id, 'text': query.text, 'randval': np.random.random() }
        belong = (query['randval'] < cdfs).agg(lambda x: x.tolist().index(True)).to_dict()
        for dt, session in belong.items():
            fps[dt][session].write(json.dumps(query) + '\n')

for l in fps.values():
    for fp in l.values():
        fp.close()

# Assigning matching qrels for the sampled queries
qrels_pool = [ q._replace(query_id=d+q.query_id, doc_id=d+q.doc_id) for d in domains for q in irds.load(f'lotte/{d}/test/forum').qrels ]
qrels_pool_grouped = {}

for q in qrels_pool:
    if q.query_id not in qrels_pool_grouped:
        qrels_pool_grouped[q.query_id] = []
    qrels_pool_grouped[q.query_id].append(q)

for fn in tqdm(Path("./queries/").glob("*.jsonl"), desc='qrels'):
    _, dt, i = fn.stem.split("_")
    contains_query_ids = [ json.loads(q)['query_id'] for q in fn.open() ]
    with open(f'./qrels/{fn.stem}.qrels', 'w') as fw:
        for query_id in contains_query_ids:
            for qr in qrels_pool_grouped[query_id]:
                fw.write(f"{query_id} {qr.iteration} {qr.doc_id} {qr.relevance}\n")

doc_latest_appear_session = {
    dt: { 
        d.doc_id: i
        for i in range(5) 
        for d in irms.read_trec_qrels(f'./qrels/test_{dt}_{i}.qrels')
    }
    for dt in sampling_cdf.columns.get_level_values('dist_type').unique()
}

def _marginalize_cdf(cdf):
    return cdf/cdf.iloc[-1]
    

# Sampling and assigning documents to sessions
np.random.seed(123)

doc_lists = {
    dt: {i: [] for i in sampling_cdf.index}
    for dt in sampling_cdf.columns.get_level_values('dist_type').unique()
}

for dom in tqdm(domains, desc='docs'):
    cdfs = sampling_cdf[dom]

    for doc in tqdm(irds.load(f'lotte/{dom}/test/forum').docs, desc=dom):
        doc = { 
            'docid': dom+doc.doc_id, # for tevatron
            'text': doc.text, 
            'randval': np.random.random(), 
        }
        
        for dt in cdfs.columns:
            appear_before = doc_latest_appear_session[dt].get(doc['docid'], 99)
            session = (doc['randval'] < cdfs[dt].loc[:appear_before+1].pipe(_marginalize_cdf)).tolist().index(True)
            doc_lists[dt][session].append({**doc, 'before': appear_before})


for dt, ll in doc_lists.items():
    for i, l in ll.items():
        pd.DataFrame(l).to_parquet(f'./docs/test_{dt}_{i}.parquet')