Why Campus Dating Apps Thrive: The Real Power Behind Algorithmic Matching

Algorithmic Matching Model

The SJTU Date platform formulates weekly dating matches as a bipartite‑graph optimal‑matching (Assignment) problem. Let M be the set of male users and F the set of female users. For each possible pair (i,j) a compatibility score c_{ij} is computed. The objective is to choose binary variables x_{ij}\in\{0,1\} that maximize total compatibility while ensuring each user is matched at most once:

max   \sum_{i\in M}\sum_{j\in F} c_{ij} \; x_{ij}
subject to   \sum_{j\in F} x_{ij} \le 1 \;\forall i\in M,
           \sum_{i\in M} x_{ij} \le 1 \;\forall j\in F,
           x_{ij} \in \{0,1\}

This is the classic Assignment Problem and can be solved exactly with the Hungarian algorithm in O(n^3) time, where n is the number of participants. For the observed scale of ~7,000 users the computation is negligible.

Content‑Based Filtering Model

Each user answers a 65‑question psychological questionnaire, which is transformed into a high‑dimensional feature vector u_i. Compatibility between a male‑female pair is estimated by a similarity metric, most commonly cosine similarity: cos(u_i, v_j) = \frac{u_i \cdot v_j}{\|u_i\| \; \|v_j\|} This approach belongs to content‑based filtering: it is transparent, easy to explain, and directly uses the self‑reported preferences.

Stated vs. Revealed Preference

The vectors derived from the questionnaire represent stated preferences ( s_i), whereas actual behavior reflects revealed preferences ( t_i). The bias can be expressed as Δ_i = \|s_i - t_i\|. When Δ_i is large, cosine similarity deviates substantially from the true compatibility, and users may not even be aware of their own latent preferences.

Empirical evidence (Samantha Joel et al., *Psychological Science*, 2017) shows that algorithms can predict aggregate popularity but struggle to forecast attraction between specific individuals, indicating emergent interaction effects that are not captured by independent feature vectors.

Utility Perspective on Social Initiation Cost

In a high‑pressure campus environment the baseline probability of two random students meeting ( p_0) is near zero. The algorithm forces a first contact, raising the meeting probability to p_1≈1. If the conversion from meeting to a lasting relationship ( α) remains unchanged, the expected number of successful pairings rises from N·p_0·α to N·p_1·α, where N is the number of participants. Thus the core contribution is the reduction of the psychological cost of initiating contact, not the precise prediction of long‑term compatibility.

System‑Assigned Matching vs. Parental Selection

Both mechanisms collect information, define a matching standard, and present candidates, thereby lowering the initiator’s social cost. The key differences are:

Ownership of standards: the algorithm uses the user‑filled questionnaire, giving participants a sense of agency; parental selection relies on opaque, experience‑based criteria, perceived as imposed.

Transparency: algorithmic criteria are visible and modifiable by the user, whereas parental criteria are hidden, leading to feelings of misunderstanding beyond mere inaccuracy.

Capabilities and Limits of the Matching Algorithm

The objective function for a romantic relationship is not a static scalar; it evolves over time and is jointly defined by both partners. Consequently, the algorithm solves the “first‑meeting” problem but cannot guarantee sustained compatibility (“stay‑together” problem). In the SJTU Date case, creating 2,400 meetings that would otherwise not occur is valuable, yet the subsequent development of the relationship depends on the individuals.

A cautionary note: as more initial encounters are mediated by a fixed set of algorithmic standards, users’ ability to define “suitable” independently may erode, raising questions about long‑term effects of standardized matching criteria.

Illustrative Figures