The final score S(i,u) for item i and user u is:
[ S(i,u) = \alpha \cdot S_\textcontent(i) + \beta \cdot S_\textcollab(i,u) + \gamma \cdot S_\textfresh(i) ]
Hyper‑parameters (α,β,γ,λ) are tuned on a validation set (grid search).
We used a held‑out log (2 weeks) to compute Normalized Discounted Cumulative Gain (NDCG@10) and Mean Reciprocal Rank (MRR). The hybrid model with bandit personalization outperformed the baseline by: Pthc Top Site
| Metric | Baseline | PTHC Top Site | |--------|----------|---------------| | NDCG@10 | 0.73 | 0.81 | | MRR | 0.42 | 0.53 | | CTR (per session) | 4.1 % | 4.7 % |
Statistical significance was verified with a paired t‑test (p < 0.001).
Report the Site
Block the Site
Educate Others
A Contextual Thompson Sampling bandit adjusts the weight vector w_u per user: The final score S(i,u) for item i and
[ \hatS(i,u) = w_u^\top \cdot \phi(i) ]
where φ(i) is the feature vector concatenating the three sub‑scores above.
After each impression, the bandit receives a binary reward (clicked = 1 or 0) and updates the posterior distribution of w_u. This enables instant adaptation without offline retraining.
A 2023 cross‑sectional survey of 4 800 physiotherapists across 27 countries reported that 78 % of respondents regularly consulted PTHC Top Site for treatment planning. The most cited benefits were: Report the Site