Matchmaking
Matchmaking recommends people, teams, mentors, jobs, or peers where the recommended “item” can accept, reject, or be harmed by exposure. Inputs include explicit preferences, behavior, availability, geography, eligibility, safety filters, and interaction history. Targets include reciprocal reply, accepted match, retained relationship, or successful outcome after the match. The output is usually a ranked set with capacity and fairness constraints, not a single prediction.
Framing
This is a two-sided form of matchmaking systems. Unlike ordinary collaborative filtering, scoring only one side’s preference is insufficient; the system must estimate mutual compatibility and manage exposure. Ranking objectives should include quality, reciprocity, diversity, and “do no harm” constraints. Online metrics should include accepted matches, reply rate, negative feedback, repeated exposure, and downstream retention, with online experiments guarded against marketplace imbalance.
GroupLens’ MovieLens 100K is not a matchmaking dataset, but it is a useful recommender baseline artifact: the page reports 100,000 ratings from 1,000 users on 1,700 movies. Reciprocal recommender papers extend this user-item idea to user-user matching.
Executed Artifact
To show why matchmaking must score both sides of a pair, the example below combines each user’s interest with the other side’s attractiveness into one reciprocal score, then greedily selects non-overlapping pairs, exposing how early high-reciprocity choices leave weaker leftovers.
import numpy as np
from sklearn.metrics.pairwise import cosine_similarity
users = np.array([[1, 0, 0.6], [0.9, 0.2, 0], [0, 1, 0.3], [0.1, 0.7, 0.7]])
items = np.array([[1, 0, 0.4], [0.8, 0.1, 0.1], [0, 1, 0.6], [0.2, 0.8, 0]])
interest = cosine_similarity(users, items)
attractiveness = cosine_similarity(items, users).T
score = 0.55 * interest + 0.45 * attractiveness
pairs, used_users, used_items = [], set(), set()
for u, i in sorted(np.ndindex(4, 4), key=lambda x: score[x], reverse=True):
if u not in used_users and i not in used_items:
pairs.append((u, i, round(float(score[u, i]), 3)))
used_users.add(u)
used_items.add(i)
print("greedy_reciprocal_pairs", pairs)
print("min_pair_score", min(p[2] for p in pairs))
print("mean_pair_score", round(sum(p[2] for p in pairs) / len(pairs), 3))Observed output:
greedy_reciprocal_pairs [(1, 1, 0.988), (0, 0, 0.987), (2, 2, 0.969), (3, 3, 0.707)]
min_pair_score 0.707
mean_pair_score 0.913The example is deliberately small: the last accepted pair is much weaker because earlier high-reciprocity choices consumed the best candidates. In a real system, locally high scores can still create bad global outcomes if popular users receive all exposure, which is a feedback loop problem.
Failure Modes
Matchmaking fails when it optimizes clicks instead of mutually wanted outcomes, when it overexposes a small group, or when cold-start users are invisible. Diversity and coverage metrics from diversity, novelty, coverage, and serendipity should be tracked alongside success metrics. Safety filters and appeal processes are part of the product contract, not post-processing decoration.
References
Nav