OK, I had some inspiration and updated the scoring model to do as best as it can to incorporate information from tags. The update changes the output, but it’s hard to tell whether it’s an improvement.
Since Victor has (more or less) volunteered to be a guinea pig, here’s a new-and-improved recommendation list (now with more entries):
top 100 recommendations
+----------------------------------------------------------+-------------+
| title | MODEL_SCORE |
+----------------------------------------------------------+-------------+
| Counterfeit Monkey | 143.0130 |
| Treasures of a Slaver's Kingdom | 96.3102 |
| Midnight. Swordfight. | 90.6840 |
| The Wizard Sniffer | 90.3440 |
| Lost Pig | 88.4620 |
| Blue Lacuna | 85.4540 |
| Coloratura | 83.4651 |
| According to Cain | 81.3200 |
| Toby's Nose | 80.6778 |
| Inside the Facility | 80.0299 |
| Cannery Vale | 79.9992 |
| Hadean Lands | 78.2000 |
| Photopia | 76.5000 |
| Violet | 74.8008 |
| With Those We Love Alive | 73.5660 |
| Foo Foo | 73.1510 |
| Shade | 69.9804 |
| Alias 'The Magpie' | 69.8176 |
| 1893: A World's Fair Mystery | 68.0000 |
| Cragne Manor | 66.2662 |
| Bogeyman | 66.2099 |
| Detectiveland | 61.5000 |
| Galatea | 61.3815 |
| Worlds Apart | 61.3536 |
| A Beauty Cold and Austere | 61.1996 |
| Curses | 60.0000 |
| Birdland | 59.9998 |
| The Gostak | 59.9494 |
| How Prince Quisborne the Feckless Shook His Title | 59.5829 |
| robotsexpartymurder | 59.5000 |
| Zozzled | 58.8966 |
| The Wand | 58.6488 |
| Even Some More Tales from Castle Balderstone | 58.5000 |
| Vespers | 57.1116 |
| Trinity | 56.6423 |
| Repeat the Ending | 56.5716 |
| Sub Rosa | 56.1600 |
| Rameses | 56.0000 |
| Taco Fiction | 55.7592 |
| A Mind Forever Voyaging | 55.6569 |
| Trigaea | 54.0000 |
| Bronze | 53.9864 |
| Delightful Wallpaper | 53.5288 |
| A Long Way to the Nearest Star | 53.4000 |
| Fallacy of Dawn | 53.1432 |
| Endless, Nameless | 53.0328 |
| Brain Guzzlers from Beyond! | 52.7228 |
| And Then You Come to a House Not Unlike the Previous One | 52.1148 |
| 4x4 Archipelago | 52.0003 |
| Exhibition | 52.0000 |
| Grimnoir | 51.9996 |
| Varicella | 51.6000 |
| SPY INTRIGUE | 51.5004 |
| HUNTING UNICORN | 51.0000 |
| Worldsmith | 51.0000 |
| Wishbringer | 49.6668 |
| The Impossible Bottle | 49.3658 |
| Skies Above | 48.8568 |
| Andromeda Apocalypse — Extended Edition | 48.6312 |
| Three-Card Trick | 48.5712 |
| Six | 48.5220 |
| Hunger Daemon | 48.3756 |
| Absence of Law | 48.2306 |
| Sorcery! 4 | 48.0000 |
| Chuk and the Arena | 47.3847 |
| Weird City Interloper | 47.1933 |
| Cryptozookeeper | 46.9337 |
| The Edifice | 46.7364 |
| Harmonia | 46.2649 |
| Will Not Let Me Go | 46.2500 |
| Sorcery! 3 | 46.0000 |
| What Heart Heard Of, Ghost Guessed | 45.9415 |
| The Little Match Girl 3: The Escalus Manifold | 45.5560 |
| Sunset Over Savannah | 45.3750 |
| Open Sorcery | 45.0000 |
| Charming | 44.9163 |
| The Hitchhiker's Guide to the Galaxy | 44.3663 |
| Overboard! | 44.2860 |
| A Thousand Thousand Slimy Things | 44.0000 |
| 1 4 the $ | 44.0000 |
| Known Unknowns | 43.8710 |
| The Elysium Enigma | 43.6205 |
| Rogue of the Multiverse | 43.2146 |
| Grooverland | 43.1250 |
| Erstwhile | 42.8130 |
| Tavern Crawler | 41.9050 |
| Illuminismo Iniziato | 41.7650 |
| CYBERQUEEN | 41.5380 |
| Trouble in Sector 471 | 41.5000 |
| Black Knife Dungeon | 40.9090 |
| Scroll Thief | 40.9090 |
| The Master of the Land | 40.7690 |
| What Fuwa Bansaku Found | 40.7140 |
| Suspended | 40.6670 |
| Dr Ludwig and the Devil | 39.8079 |
| Of Their Shadows Deep | 39.6000 |
| Scavenger | 39.4440 |
| Robin & Orchid | 39.4120 |
| Heretic's Hope | 38.9997 |
| A Rope of Chalk | 38.0457 |
+----------------------------------------------------------+-------------+
@VictorGijsbers, since you said that a lot of these would be games you have played but not rated, the hit rate on the above list is actually a good test. How much did you actually like the games on the list that you’ve played? (I would think that Treasures of a Slaver’s Kingdom would be the kind of game you would like, given Kerkerkruip and Turandot, and that comes in at #2. I’m pretty sure that all models based on existing data will insist that everyone will love Counterfeit Monkey.)
Does anyone else want to try?
I would like to try to do something similar by correlating review agreements, but the user field in that data is missing from the backups for some reason.