Сайт рекомендаций похожих игр

Ничего они не доведут, да и нет там никакого алгоритма по сути.

 

Ну как бы... Когда автор темы пишет прямым текстом, что вот это и то не советовать, в теме обязательно кто-нибудь эти названия упомянет. Наталкивает на определенные выводы.

 

Где-то там написано, что информация обновляется рывками время от времени. Что-то мне кажется, что никакой автоматики на сайте вообще нету.

 

На сайте и не нужна автоматика. Судя по описанию, сбор данных и анализ осуществляются независимо от сайта. В принципе аналогичные системы сравнительно неплохо работают там, где есть непосредственные данные о продажах. А тут - поглядим

 

Результаты любого автоматического процесса обработки информации осмысленны ровно настолько, насколько разумно составлен исходный алгоритм. У этого сайта только один критерий - статистика частотности ключевых слов в рецензиях на игры, - и этого явно недостаточно. Возможно, если расширить список критериев и усовершенствовать алгоритмы, результаты будут более впечатляющими. Но в любом случае, это не более чем вспомогательное средство, которое может кому-то помочь, а для кого-то быть абсолютно бесполезным.

 

Hello!

I'm writing in English, because my written Russian is even worse than my English.

First of all thanks for finding out about our site and thanks for some feedback.

We started this project very recently and trying to improve in any possible way.

The main thing is: site is fully automatic, no manual entries (yet) are made to the similar game lists. All similar games are generated from keywords, which we extract from reviews. By now we only depend on keywords and gernes.
The results actually are way better than we expected in the first place. Of course they are far from "very good".

Also the description texts and screenshots are automatically generated from Wikipedia / Google image search. So errors are not a rare thing (like the TERA one, mentioned by drk_patr1ck)

So far the best feature in this project to me personally is keyword search. If I like zombie killing, I can search kill zombies. These results are even quite good, if compared to the similar game lists This is a approach to search by things you want to do in a game or what theme/time/whatever game is about, not so much thinking about gernes or something like that.

Next step is to add some user generated content, allow them to alter the game lists, so we can enliminate some crazy wrong results.

 

Hi! Nice to hear that you're working towards making the search engine more efficient

I haven't explored the keyword search option extensively, but it does look rather promising - in fact, it may indeed be the greatest advantage of your site, so thanks a lot for dropping by and providing the info!

 

Завал.

 

yuppio,
If you're still there, I'd want to give you a clue about how to avoid situations similar to one mentioned by me before. Thing is, in its current state you algorithm often makes decisions heavily based on words like "enemy", "attack", "expirience", "mode", etc. As a result, the algorithm can barely distinguish games between genres and sorts shooter to shooter, rpg to rpg, adventure to adventure and so on, but nothing more.

Currently I see the next solution:
1. Isolating non-representive words like "enemy" and so on from the rest. It seems for me, that it can be even done manually, because the truly problem words are not so plentiful.
2. Giving a weight function to words and assign small weights to problem words.
or
3. Making the decision in several stages: at first you can use all the words to distinguish a genre, and then analyze the derived set of titles without using the problem words to search games with similiar features like, for example, management in shooters.

Either way, you'll get rid at least of some part of non-representative tag words in top.

Sorry for my English. It's not my native language, and I didn't speak it for some time.

 

BTW, it seems that there are currently only single keywords used, whereas obviously key-collocations like "time travel", "science fiction", "space opera" etc., would be as of much use as single words.

 

We have common word table with around 2,5k words which ar ignored in the text and we update it manually from time to time (basically the only thing that we do manually for system to work), to get rid of all these words, that don't mean much for the game. Words you mentioned all already on the list added recently, we need to rerun the keyword generator for all the games for it to take affect (it runs like 10 hours 16k games, 374k keywords and around 190k game connections).

Common words are on the main problem by now, less common words in generated keywords - better results.

The opposite idea of not using "blacklists", but "whitelists" also is good and it would get really good results I think, but the problem is, that it needs a lot of manual work of building this whitelist, which is the thing we are trying to avoid

We can also build all lists manually and get like the best results, but that is not our goal, we are trying to build automated system, that can make good starting lists and after that users can make modifications, less manual work for us, the better

IMHO technically it's far too difficult to develop with simple algorithms , and it might take large amount of time also analyze also word-pairs. And there's is also the same common word problem described above. Because like "common-word other-word" can make good two-word pairs also and so on. But it's worth a try i guess.

 

Hmm, maybe you could simplify automated identification of relevant keywords by statistical analysis of words inside groups of already generally similar games (e.g. similar genre, theme etc.). Or is it exactly how you operate now?

I guess it wouldn't be too troublesome to generate lists for separate groups (genre+theme and/or maybe other parameters), then run an automatic comparison of lists to identify and remove duplicate keywords across the lists, which I suppose could give a rather reliable list of words for each group. If that is possible, you could further run the same procedure for different feature combinations (I suppose MobyGames' tags could serve as a good starting point), getting multiple lists upon which a statistically weighed decision could be derived based upon keywords for each game.

Once again, sorry if this is the method you've already adopted.

Hmm, what about using the same statistical analysis to determine the probability of two words occurring together? I understand that it might take some time and system resources, but to the best of my knowledge, it is doable.