Caser
A user's last few interactions, written out as a stack of item embeddings, form a small grid — items down one axis, embedding dimensions across. Once you see it as a grid, the obvious tool is a convolution, and Caser is that idea applied to next-item recommendation.
Horizontal and vertical filters
Caser stacks the embeddings of the L most recent items into an L × d matrix and slides two filter families over it. Horizontal filters span several whole item rows at once, so they respond to short multi-item patterns wherever those sit in the window — tent, then sleeping bag, then torch. Vertical filters run down a single embedding dimension, which amounts to a learned weighted average of the window along each latent factor. The max-pooled outputs of both, concatenated with a user embedding that carries long-term taste, feed a small MLP that scores candidate items.
Point-level and union-level patterns
The motivation is a gap in Markov-chain factorisation such as FPMC, which captures point-level influence — each past item pushing towards the next on its own — but not union-level influence, where a group of items jointly implies what comes next. Horizontal filters capture the union case directly, including patterns with gaps in them.
Where it sits
Caser lands between Markov-chain models and the recurrence of GRU4Rec, and it preceded self-attentive sequence models such as SASRec. Convolution over a fixed window is cheap and fully parallel, with nothing to unroll at serving time, but L hard-caps the history the model can see — which is where the attention-based successors took over. Results are reported with leave-one-out next-item ranking on datasets like MovieLens and Gowalla.
Caser treats a user's recent interactions as an image: it stacks the item embeddings of a fixed-length window into a matrix, then applies horizontal filters, which catch short multi-item patterns anywhere in the window, and vertical filters, which pool each latent dimension. The result joins a user embedding to score the next item. Convolution is cheap and parallel at serving time, but the fixed window caps how far back the model can look.
Picture your last handful of purchases laid out as a short strip of pictures. Instead of reading the strip one item at a time, Caser looks at small patches of it at once — much the way photo software spots an eye or an edge by sliding little windows across an image.
That lets it notice combinations rather than single steps. Someone who bought a tent, a sleeping bag and a torch is probably going camping, whichever order those arrived in. Caser also keeps a separate note of your longer-term taste, so a camping trip does not wipe out everything else it knows about you. Together, those two things let it guess what you will reach for next. Its one real limit is that it only glances at a short stretch of recent history.