Estimating Dyadic Treatment Effects with Unknown Confounders