Matching Algorithms for Blood Donation (EC20 Poster) Matching Algorithms for Blood Donation Donor blood is a scarce resource: • every 2 seconds, someone in the US needs blood • shortages especially impact developing countries, and in particular, children and women facing complications during childbirth • donation rates are correlated with a country’s wealth; high-income countries have median donation rate of 31.5 donations per 1000 people, while low-income countries have a median of 5.0 donations per 1000 people. 1 * : median donation rates Statistics and quotes from the American Red Cross (ARC) and the World Health Organization (WHO)
Matching Algorithms for Blood Donation (EC20 Poster) Donor Coordination & Recruitment Potential blood donors Web app Donation opportunities (“recipients”) In our setting, blood donors and recipients use a web application to connect with one another. The web app can send notifications to donors, about a particular donation opportunity. These notifications are sent • Browse opportunities Notify donors using a • State a need automatically, by a notification policy. notification policy • receive notifications • Reach out to donors Blood recipients can Blood donors can find Taking the perspective of the web app, we state their need & donation opportunities, study the Facebook Blood Donation Tool , availability and can choose to which connects donors with opportunities receive notifications to donate in several countries around the ● Individuals about opportunities world, with ~70 million registered donors.* ● Hospitals ● Blood drives 2 * : As of June 2020 https://socialgood.fb.com/health/blood-donations/
<latexit sha1_base64="AwIlAF0i9XI5w4ws2M0GIS6Df4w=">ACBXicbVDLSgMxFL3js9ZX1aWbYBEqSJnxgW6Egi6q2Af0I4lk2ba0CQzJlCGbr2F9zq3p249Tvc+iWmj4W2HrhwOdezuUEMWfauO6Xs7C4tLymlnLrm9sbm3ndnarOkoUoRUS8UjVA6wpZ5JWDOc1mNFsQg4rQW965Ff61OlWSQfzCmvsAdyUJGsLHSI71CheQY9Y+aTKbVi7vFt0x0DzxpiQPU5Rbue9mOyKJoNIQjrVueG5s/BQrwinw2wz0TGpIc7tGpxIJqPx1/PUSHVmjMFJ2pEFj9fdFioXWAxHYTYFNV896I/FfLxAzySa89FMm48RQSbBYcKRidCoEtRmihLDB5Zgopj9HZEuVpgYW1zWluLNVjBPqidF7R4fn+WL91N68nAPhxATy4gBLcQhkqQEDBM7zAq/PkvDnvzsdkdcGZ3uzBHzifP4ACl5w=</latexit> <latexit sha1_base64="pZdwSN3gKoiPRqat3S9aM68rAwE=">ACDnicbVDLSsNAFJ3UV62vqLhyM1gEF6UkPtBlwY3uKvQFTSiT6aQdOpmEmRuhP6Dv+BW9+7Erb/g1i9x2mahrQcuHM65h3s5QSK4Bsf5sgorq2vrG8XN0tb2zu6evX/Q0nGqKGvSWMSqExDNBJesCRwE6ySKkSgQrB2Mbqd+5EpzWPZgHC/IgMJA85JWCkn0E2OMSe5lbwV4/Bl3BDW/Ss8tO1ZkBLxM3J2WUo96zv02YphGTQAXRus6CfgZUcCpYJOSl2qWEDoiA9Y1VJKIaT+bvT/Bp0bp4zBWZiTgmfo7kZFI63EUmM2IwFAvelPxXy+IFi5DeONnXCYpMEnh8NUYIjxtBvc54pREGNDCFXc/I7pkChCwTRYMqW4ixUsk9Z51b2oXj1clmv3eT1FdIxO0Bly0TWqoTtUR01EUYae0Qt6tZ6sN+vd+pivFqw8c4j+wPr8Ab6tmvs=</latexit> <latexit sha1_base64="UuoaHY5PqFitslp+IdpMw6nQM1A=">ACnicbVDLSsNAFJ3UV62vVJduBovgQkriA10W3Oiugn1AG8JketMOnUzCzMRSQv/AX3Cre3fi1p9w65c4bPQ1gMXDufcy7mcIOFMacf5sgorq2vrG8XN0tb2zu6eXd5vqjiVFBo05rFsB0QBZwIamkO7UQCiQIOrWB4M/VbjyAVi8WDHifgRaQvWMgo0Uby7fLIz0BPcJcJ3HFOsev5dsWpOjPgZeLmpIJy1H37u9uLaRqB0JQTpTquk2gvI1IzymFS6qYKEkKHpA8dQwWJQHnZ7PUJPjZKD4exNCM0nqm/LzISKTWOArMZET1Qi95U/NcLoVkHV57GRNJqkHQeXCYcqxjPO0F95gEqvnYEIlM79jOiCSUG3aK5lS3MUKlknzrOqeVy/vLyq1u7yeIjpER+gEuegK1dAtqMGomiEntELerWerDfr3fqYrxas/OYA/YH1+QMPQJmI</latexit> <latexit sha1_base64="MLWDe9oPZvfUOnOHxSOA1JLZMCw=">AC3icbVDLSsNAFL2pr1pfsS7dDBbBhZTEB7osuNFdBfuANoTJdNoOnUzCzKRYQj/BX3Cre3fi1o9w65c4abPQ1gMXDufcy7mcIOZMacf5sgorq2vrG8XN0tb2zu6evV9uqiRhDZIxCPZDrCinAna0Exz2o4lxWHAaSsY3WR+a0ylYpF40JOYeiEeCNZnBGsj+XY59tOxnqIuE6jnCLXK/l2xak6M6Bl4uakAjnqv3d7UkCanQhGOlOq4Tay/FUjPC6bTUTRSNMRnhAe0YKnBIlZfOfp+iY6P0UD+SZoRGM/X3RYpDpSZhYDZDrIdq0cvEf70gXEjW/WsvZSJONBVkHtxPONIRyopBPSYp0XxiCaSmd8RGWKJiTb1ZaW4ixUsk+Z1T2vXt5fVGp3eT1FOIQjOAEXrqAGt1CHBhB4hGd4gVfryXqz3q2P+WrBym8O4A+szx9Z+Zm</latexit> <latexit sha1_base64="3d9oLNhISrRhdgH1oiYT/wRCxc=">AB/nicbVDLSsNAFL2pr1pfVZduBovgqiQ+0GXBje4qmLbQhjKZTtqhM5MwMxFKPgLbnXvTtz6K279EidtFtp64MLhnHs5lxMmnGnjul9OaWV1bX2jvFnZ2t7Z3avuH7R0nCpCfRLzWHVCrClnkvqGU47iaJYhJy2w/FN7rcfqdIslg9mktBA4KFkESPYWKmdoh6TyO9Xa27dnQEtE68gNSjQ7Fe/e4OYpIJKQzjWu5iQkyrAwjnE4rvVTBJMxHtKupRILqoNs9u4UnVhlgKJY2ZEGzdTfFxkWk9EaDcFNiO96OXiv14oFpJNdB1kTCapoZLMg6OUIxOjvAs0YIoSwyeWYKY/R2REVaYGNtYxZbiLVawTFpnde+8fnl/UWvcFfWU4QiO4RQ8uIG3EITfCAwhmd4gVfnyXlz3p2P+WrJKW4O4Q+czx/Su5Wx</latexit> <latexit sha1_base64="hla6shafeQgXxdDl61HBdG6wQc=">AB/nicbVC7SgNBFL0bXzG+opY2g0GwCrs+0DJgo10E84BkCbOT2WTIPJaZ2UBYAv6CrfZ2Yuv2PolTpItNHrgwuGcezmXEyWcGev7n15hZXVtfaO4Wdra3tndK+8fNI1KNaENorjS7QgbypmkDcsp+1EUywiTlvR6Gbmt8ZUG6bkg50kNBR4IFnMCLZOao1Rl0nU7JUrftWfA/0lQU4qkKPeK391+4qkgkpLODamE/iJDTOsLSOcTkvd1NAEkxEe0I6jEgtqwmz+7hSdOKWPYqXdSIvm6s+LDAtjJiJymwLboVn2ZuK/XiSWkm18HWZMJqmlkiyC45Qjq9CsC9RnmhLJ45gopn7HZEh1phY1jJlRIsV/CXNM+qwXn18v6iUrvL6ynCERzDKQRwBTW4hTo0gMAInuAZXrxH79V7894XqwUvzmEX/A+vgHV6ZWz</latexit> Matching Algorithms for Blood Donation (EC20 Poster) Matching Model We formalize an online matching model for this setting, which differs from standard Donors u ∈ U online matching in several key ways • offline , but rate limited : donors can only be matched once every K days Standard Online Blood Donor e = ( u, v ) ∈ E Edges (potential recipients) Same? Matching / Adwords Matching • potential notifications , which may depend on U entirely offline U entirely offline distance or donor/recipient preferences w et ∈ [0 , 1] • Edge weights (likelihood of MA) Fixed time horizon T Fixed time horizon T U removed when U return after K days matched Recipients v ∈ V V have unlimited V have capacity 1 • Both offline recipients (always available), and capacity online recipients (sometimes available) V arrive in batches , V arrive one-by-one p vt ∈ [0 , 1] • Probability that v is available at time t and some are offline | V | >> | U | | V | << | U | t ∈ { 1 , . . . , T } Finite time horizon (days) : 3
Matching Algorithms for Blood Donation (EC20 Poster) Initial Results In simulations we find that a greedy policy (which maximizes edge weight) increases overall matching weight by 5-20%. Total matched Weight (expected # donations) This comes at a cost of ignoring some recipients, which are not well-connected, or have low edge weights. Median recipient weight* In online experiments , using 1.3 million donors, we find that notifications which maximize (estimated) edge weight also Greedy Myopic Policy Random Policy increase overall donor action rate by about 5% ( p <<0.001). LP-based Policy * : normalized by expected weight assigned by the random policy 4
Recommend
More recommend
