Yunchi Yang, Longlong Li, Jianliang Wu, Cunquan Qu
9 min
MISApp is a profile-free framework designed to predict the next mobile app a user will launch, enabling proactive services like resource preallocation and personalized recommendations. Unlike traditional methods that rely on long user histories or simple sequential models, MISApp focuses on short-term session behaviors, making it robust for cold-start scenarios where user profiles are sparse or unavailable. By modeling app transitions as multi-hop graphs and integrating contextual cues like time and location, it captures evolving user intent and higher-order dependencies, outperforming baselines on real-world datasets while staying computationally efficient.
The core innovation is treating app usage sessions—short bursts of app switches—as rich graphs rather than flat sequences. This allows the model to reason about both immediate (1-hop) and distant (2-3 hop) relations, reflecting real user routines like 'open maps, then ride-sharing, then music'.
App usage isn't just a linear chain; it has structure. MISApp builds multi-hop session graphs where nodes are apps in a session, and edges represent transitions at different ranges:
LightGCN (a lightweight GNN) propagates information across these hops, with hop-level attention weighting their importance. This provides interpretable insights—e.g., 2-hop edges might dominate in routine-heavy sessions. Experiments show these graphs better represent behavioral structure than 1-hop alone, boosting prediction accuracy by capturing latent dependencies.
Cold-start is a killer for user-profile-based predictors. MISApp infers intent directly from the current session's recent interactions, modeling intent evolution as shifts influenced by context. No long-term history needed—just the last few apps.
This is key for real-world mobile use: sessions are brief (median ~5 apps), intents flip fast (work → entertainment), and new users lack data. By focusing on session-level dynamics, MISApp achieves strong performance even with zero user history.
Time (hour-of-day) and space (base station ID as location proxy) aren't add-ons—they modulate transitions. E.g., music apps spike evenings; ride-sharing peaks at rush hour.
MISApp embeds these into a unified space via MLPs, then uses cross-modal gated fusion (CMGF)—a sigmoid gate that adaptively blends app-graph features with context. This 'joint reasoning' outperforms separate modeling, as context refines intent (e.g., morning coffee app → news vs. evening → social).
The full pipeline:
It's efficient: LightGCN avoids heavy GNNs, fusion is lightweight. On datasets like AppUsage, MISApp beats RNNs, Transformers, and graph baselines in accuracy (e.g., +5-10% Hit@1) and cold-start, with attention maps aligning to intuitive structures.
Next-app prediction powers smarter phones—prefetch RAM, suggest apps, optimize batteries. MISApp's graph+context approach generalizes to sparse data, inspiring session-based modeling in recommenders, where profiles fail.
Predicting the next mobile app a user will launch is essential for proactive mobile services. Yet accurate prediction remains challenging in real-world settings, where user intent can shift rapidly within short sessions and user-specific historical profiles are often sparse or unavailable, especially under cold-start conditions. Existing approaches mainly model app usage as sequential behavior or local session transitions, limiting their ability to capture higher-order structural dependencies and evolving session intent. To address this issue, we propose MISApp, a profile-free framework for next app prediction based on multi-hop session graph learning. MISApp constructs multi-hop session graphs to capture transition dependencies at different structural ranges, learns session representations through lightweight graph propagation, incorporates temporal and spatial context to characterize session conditions, and captures intent evolution from recent interactions. Experiments on two real-world app usage datasets show that MISApp consistently outperforms competitive baselines under both standard and cold-start settings, while maintaining a favorable balance between predictive accuracy and practical efficiency. Further analyses show that the learned hop-level attention weights align well with structural relevance, offering interpretable evidence for the effectiveness of the proposed multi-hop modeling strategy.
Sam: Yes. They look at the last few apps in the session to capture what the user seems to want lately—like checking recent moves in a game to guess the strategy. That recent pattern acts as a question, weighing how much each map's info fits the moment. This is called hop-level intent attention, dynamically blending the maps based on current needs.
Alex: So the system smartly picks the right range, close-up or zoomed out, depending on the vibe. Time and place must play into that intent too, right?
Sam: They do. For time, they map the hour of the last app use to a pattern—like assigning 'morning rush' or 'late night scroll' to a vector that captures routine feel. For location, they group cell tower spots by nearby landmarks, such as shops or parks, into categories based on similarities, then embed those. These give clues like 'at a stadium' influencing app choices.
Alex: Makes sense—being at a game might mean sports scores next. But blending all that graph stuff with time and place without clashing?
Sam: That's where they use a mixer that lets info flow both ways between the graph patterns and context details, with smart switches to control how much passes—like doors that open based on how well things match. It runs multiple checks across angles for better fit. This cross-modal gated fusion creates a full picture before feeding into the final prediction steps.
Alex: Huh. So by keeping ranges separate and fusing carefully, it handles those quick intent shifts without losing the big picture.
Alex: Yeah, but sessions don't happen in isolation—user goals shift over time, right? So how does the system track those changes across multiple bursts of app use?
Sam: It does this by separating stable long-term patterns from quick recent changes. The long-term part comes from the full session info we've discussed, fused with time and place details, then refined through layers that let related pieces talk to each other—like a team reviewing the whole game plan step by step. Researchers use a setup called a Transformer encoder for that stable base. For the short-term, it focuses on just the last few apps, say the most recent three, to catch what's happening right now.
Alex: Okay, so one side holds the big session picture, the other grabs the latest moves. But they need to blend without one overwhelming the other?
Sam: Exactly. The short-term recent apps act like a question, pulling relevant highlights from the long-term base through focused checks—similar to how you'd scan a notebook for notes that match your current thought. This happens in layers too, called a Transformer decoder, mixing self-review with cross-checks to the encoder. The result is a combined view of past habits and fresh intent, ready for the final guess.
Alex: Huh—so it's like evolving the user's goal from steady background to what's popping up now. And that feeds straight into picking the next app?
Sam: Yes. It computes a match score between that blended intent and each possible app, turning it into probabilities—like ranking teammates by how well they fit the play. The paper trains this end-to-end to minimize wrong guesses. In cold-start tests on real datasets like Tsinghua App Usage, it hit about 2 to 5 percent higher accuracy at the top rank than strong baselines, suggesting the intent blending adds a clear edge.
Alex: That's meaningful for new users. Ties the session smarts to real shifts without needing history.
Alex: Okay, so the intent blending seems key. But how do we know those separate hop maps really pull their weight?
Sam: The researchers tested that by building versions without certain parts and measuring the drop-off. Without the multi-hop graphs—sticking to just direct one-step links—the top accuracy fell from 54 percent to 53 percent on the main dataset. That gap shows longer-range connections within a session capture useful patterns, like hidden routines, without muddying the short ones. Dropping time cues caused a smaller dip, but removing location info hit harder on one dataset, as spots like shops hint at app needs.
Alex: Right, so skipping those hurts performance. What about the decoder that evolves intent—does it matter most?
Sam: Yes—the biggest fall came without it, proving tracking shifts from stable to recent matters for dynamic sessions. In tests with no user history, this setup topped charts on two real phone datasets, edging out the next best by 2 to 5 percent at the top pick. Profile-heavy methods tanked without past data, while sequence models struggled with short noisy bursts. The separation of hop levels dodges over-smoothing—where deep links blur nodes into sameness—keeping signals sharp.
Alex: Over-smoothing—like everything blending into gray? Makes sense why separate graphs help. Does it run efficiently, or does all that add overhead?
Sam: It uses fewer parameters than heavy rivals—1.3 million versus 60 million for one top performer—while matching speed and using less memory. That keeps phones responsive, preloading the right app without drain. The paper notes this balance stems from lightweight message-passing and focused fusion.
Alex: Solid evidence each piece fits without excess.
Alex: Those tests confirm the pieces work together without waste. But how stable is it if you tweak the setup, like how much recent history it looks at?
Sam: The researchers checked that by trying different sizes for the recent-apps window—say, looking back one, two, three, or four steps. Smaller to medium sizes worked best, as they grab the most relevant fresh clues without pulling in old noise. They tested session windows from five to nine apps back, peaking at eight for enough history without extra clutter.
Alex: Right, keeps it sharp on the now. And deeper processing layers?
Sam: Two layers gave the best results there too—gains from one, but three or four risked overfitting, where the model memorizes training quirks instead of generalizing.
Alex: Fair—moderation avoids pitfalls. One more thing—how do we know the hop weights aren't just random, but actually match real patterns?
Sam: They checked by comparing the system's hop priorities to a stat on how apps co-occur more than chance—like spotting if browser-to-maps-to-nav truly links stronger. In tough cases where one-hop failed but multi-hop won, the weights lined up well. Messing with key linking apps dropped the right guess sharply, proving it spots real dependencies. A few limits remain, though: capping at three hops might miss very long chains, and performance ties to that eight-app window—stray too far, and noise creeps in.
Alex: That's interpretable too—not a black box. Balanced picture.
Sam: Exactly. The paper suggests this structural session view, fused right, gives a solid profile-free path forward, efficient and explainable.
Alex: A meaningful step for smarter, instant phone aids. Thanks, Sam—that wraps our look at MISApp. Listeners, thanks for joining ResearchPod.