How Breeze's recommendations system works
Breeze uses machine learning to decide which profiles we’ll send you. Our overall goal is to maximise the number of dates and thus matches.
We first find the set of profiles we can show you by taking into account the “hard filters” that come from the preferences you set in the app. These filters include, but are not limited to, the following:
- The age range you set;
- The cities or area you would like to date in;
- The genders you would like to be matched with;
- The languages you’d like to date in;
- The date activities you’re interested in.
In addition, you can choose to set a single extra “dealbreaker” of your own: one preference that you turn into a hard filter, so that only profiles matching it are shown to you. Using a dealbreaker is entirely optional, none is set by default, and only one can be active at a time — you decide whether to use it and which preference it applies to. Depending on your choice, this dealbreaker filters the profiles we show you so that you only see people who match it — for example, based on a height range, a maximum distance, showing you only verified profiles, relationship intention, relationship type, whether someone has or wants children, alcohol use, or smoking.
Next, we try to find the profiles with which you have the highest probability of a match, using a recommender system. When doing this for you, the recommender system takes the following information into account:
- Your interaction with profiles you’ve seen before. For example, which profiles you have liked and which profiles you have opened
- Your historic behaviour after getting a match, which includes, but is not limited, to the following:
- Whether you actually went on the date;
- What rating you gave the partner at which the date happened
- Other members with similar tastes and preferences on Breeze
- Information on your profile, which includes, but is not limited to, the following:
- Demographic information (e.g. age, gender, height, the area you live in);
- Tags and factors you have selected (e.g. interests, lifestyle, intention, whether you want children);
- The text you write yourself (e.g. your bio, your answers to the profile questions);
- Other information on your profile (e.g. your work or study, the languages you speak).
- Information on other users’ profiles, which includes, but is not limited to, the following:
- Demographic information (e.g. age, gender, height, the area they live in);
- Tags and factors they have selected (e.g. interests, lifestyle, intention, whether they want children);
- The text they write themselves (e.g. their bio, their answers to the profile questions);
- Other information on their profile (e.g. their work or study, the languages they speak).
- How you and the other user compare, which includes, but is not limited to, the following:
- How far apart you live, and how far each of you would travel to a date location;
- The communities you both belong to, and the interests and lifestyle tags you have in common;
- Whether they fit the preferences you have set but not made a dealbreaker (for example relationship intention, whether someone has or wants children, or smoking).
- How other users interacted with your profile in the past, which includes the following, but is not limited to:
- Whether they liked your profile, and when they did so;
- Interactions with our app, which includes, but is not limited to, the following:
- Whether you have recently opened the app;
For specific profile information, we usually do not directly use the information under points 4 to 6 as it is; more often we derive features from it. Four examples:
- We count how many photos and videos a profile has.
- We use the area you and another user each live in to work out how far apart you are, and how far each of you would travel to a date location.
- We do not read the text on a profile word by word, but convert it, together with the profile’s photos, into a numerical representation of the profile, which the recommender system compares with those of other users and with an average of the profiles you accepted before.
- We do not use what a community is about, such as its name or category, but count how many communities you and another user share.
We handle communities this way for two reasons. We only use the communities you share with someone, which are also the only communities of theirs you can see on their profile. And we leave out what a community is about, such as a particular nationality or ethnicity, because using it could introduce biases we want to prevent. For one community you share, we also use how many members it has, how often its members have matched with each other, how often you have liked its members compared with profiles in general, and how many of them you have decided on; where you share several, we use the one whose members you have liked most. Because some communities relate to a nationality, ethnicity or cultural background, this can still reflect preferences you have shown yourself, in either direction.
Note that as we optimise on dates, the users you like also need to like you back. We already take this probability into account when deciding which profiles to show to you. Therefore, we also process the information above in reverse, to find the probability that the other user will like you. If you’re new to the app, this “reverse information” is the most important. In other words, if you recently joined the app, you’ll mostly be shown profiles which are liked by other users in general. Later, we’ll learn your specific tastes from information in bullets 1-3.
Some suggestions carry a label: we mark a suggestion as worth a closer look when its estimated match probability is high for you, and we mark a profile you declined earlier when we suggest it once more, which we only do for profiles we estimate to be a particularly good match for you. Such a profile returns at most once, and your earlier decisions are only set aside altogether after the period you can set under “Reset history” in your Date Preferences, after which those profiles can be suggested to you as new suggestions again.
Finally, we also take some extra information into account when determining how many suggestions you get per “match round”. We guarantee that if you like a profile, we’ll show your profile to that other user. To avoid those other users getting too many profiles, we send fewer profiles to users who like a higher percentage of their profiles they get. We also show fewer profiles to users for which we have few profiles left. In addition, some users might have high match probabilities with many other users, which means their profiles will be shown to other users relatively often.
We use the same models to decide who sits together at a Group Date table, so that you are seated with people you are likely to get on with. Which Group Dates you can join depends on the city, day and time, and on the type of table, for example a mixed, women-only or themed table. A themed table, for example for Christians or vegans, is open to everyone who can join that Group Date; we actively invite users who selected that religion or diet in their profile.
This explanation was last updated on 18 November 2026.